<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aashish Acharya on Medium]]></title>
        <description><![CDATA[Stories by Aashish Acharya on Medium]]></description>
        <link>https://medium.com/@acharya-jyu?source=rss-7808accc2378------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*iCDyfMysJdgXonINVoFlfA.jpeg</url>
            <title>Stories by Aashish Acharya on Medium</title>
            <link>https://medium.com/@acharya-jyu?source=rss-7808accc2378------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 05:17:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@acharya-jyu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Transforming Legacy into Digital Leadership: My Journey with Natraj Tours and Travels]]></title>
            <link>https://medium.com/@acharya-jyu/transforming-legacy-into-digital-leadership-my-journey-with-natraj-tours-and-travels-0748986df24e?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/0748986df24e</guid>
            <category><![CDATA[business-development]]></category>
            <category><![CDATA[operations-management]]></category>
            <category><![CDATA[business-transformation]]></category>
            <category><![CDATA[technology-strategy]]></category>
            <category><![CDATA[process-improvement]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Sun, 16 Mar 2025 16:22:30 GMT</pubDate>
            <atom:updated>2025-03-16T16:22:30.179Z</atom:updated>
            <content:encoded><![CDATA[<p>When I first walked through the doors of Natraj Tours and Travels in October 2018, I encountered a company standing at a critical crossroads. With over 50 years of industry leadership behind them, this pioneering travel organization faced the classic innovator’s dilemma: how to maintain the legacy that built their success while evolving for a digital future that threatened to leave them behind. What began as an independent consulting engagement evolved into a comprehensive six-month transformation initiative that touched nearly every aspect of the business.</p><p>This memoir chronicles our journey of rapid, strategic transformation — a testament to what’s possible when established organizations embrace change with both urgency and thoughtfulness.</p><h3>Diagnosing the Digital Disconnect</h3><p>My initial assessment of Natraj’s digital presence revealed challenges that extended far beyond surface-level website issues. The company maintained a full suite of digital assets — website, social media accounts, email marketing — yet operated them as disconnected islands rather than an integrated ecosystem.</p><p>The website itself embodied this fragmentation. Designed to mimic international booking platforms like Airbnb and Booking.com, it overwhelmed visitors with information. The navigation suffered from redundant categories — “Destinations,” “Tours,” and “Adventure” contained overlapping content spread across different sections, forcing users to navigate a labyrinth to find relevant information. As I noted in my assessment, this created “a confusing state for a customer” with no clear path from interest to inquiry to sale.</p><p>More concerning was the content itself. Product differentiation was virtually nonexistent, with “Popular Destinations” and “Popular Tour Packages” sections displaying identical lists of location names with no compelling value propositions. Despite occasional paid campaigns, the content lacked cohesion or strategic purpose.</p><h3>Root Cause Analysis: Beyond Symptoms to Systems</h3><p>Looking deeper, I recognized these issues as symptoms of more fundamental organizational challenges:</p><ol><li><strong>Structural Fragmentation</strong>: The company operated in departmental silos where “<em>4–6 people are grouped into a division that is responsible for working strategies from very high level of vision and also ensure 100% of day to day operation.</em>” This created what I described as a “<em>slow, heavy and currently compromised</em>” process architecture.</li><li><strong>Resource Allocation</strong>: At its core, Natraj suffered from “<em>lack of dedicated resources to coordinate and manage digital branding activities across all channels.</em>” Digital responsibilities were either outsourced without coordination or added as secondary duties to staff already managing full workloads.</li><li><strong>Strategic Diffusion</strong>: The company was attempting to pursue too many market segments simultaneously without clear prioritization. As noted in the “Strategy Summary”: “Too many verticals to pursue… The strategy outlines what-to-do without creating proper accountability of who will do it.”</li><li><strong>Product Definition</strong>: Despite decades of travel expertise, Natraj had not translated this knowledge into clearly defined digital products. Instead of distinctive tour packages, they presented generic destination listings, offering little differentiation from newer market entrants.</li></ol><h3>The Transformation Roadmap: A Phased Approach</h3><p>Recognizing the interconnected nature of these challenges, I developed a comprehensive transformation strategy with carefully sequenced priorities. Rather than attempting a “big bang” digital overhaul, we implemented a phased approach that delivered immediate improvements while building toward more substantial change.</p><h3>Phase 1: Digital Foundation and Social Media Revitalization</h3><p>We began by focusing on Facebook as our primary digital channel — a strategic decision to build capabilities and test approaches before expanding to more complex platforms. This focused approach allowed us to:</p><ul><li><strong>Redesign content strategy</strong>: We developed a structured content approach with “<em>6–10 posters per campaign</em>” and “<em>contents and packages posted every week</em>” that aligned with business objectives.</li><li><strong>Implement response management</strong>: We established clear workflows for the “100<em>–150 message inquiries per week</em>,” ensuring no customer inquiry went unanswered.</li><li><strong>Develop measurement frameworks</strong>: We tracked key metrics including reach, engagement, and most importantly, lead generation for business impact.</li><li><strong>Create targeted promotional campaigns</strong>: Rather than generic destination content, we developed focused packages promoted through strategic campaigns.</li></ul><p>This focused approach quickly generated results, establishing what I had targeted as “<em>a stable presence on Facebook</em>” while creating a foundation for broader digital transformation.</p><h3>Phase 2: Organizational Restructuring</h3><p>With initial digital momentum established, we turned to perhaps the most challenging aspect of transformation — restructuring the organization itself. We implemented a fundamental shift from departmental silos to cross-functional teams:</p><ul><li><strong>First Responders</strong>: Dedicated to handling initial customer inquiries across all channels</li><li><strong>Sales Team</strong>: Focused exclusively on revenue generation and transaction completion</li><li><strong>Product and Client Relations</strong>: Responsible for developing distinctive travel packages and maintaining client relationships</li><li><strong>Marketing Team</strong>: Coordinating all promotional activities with clear channel strategies</li><li><strong>Logistics and Operations</strong>: Managing the operational backend of all transactions</li><li><strong>Corporate Services</strong>: Specialized team for B2B and MICE (Meetings, Incentives, Conferences, Exhibitions) business</li></ul><p>This restructuring dramatically improved operational efficiency by changing from “20 things to be done by a group of 4 to teams of 2–3 with only 5 things to do.” Each team received specific KPIs aligned with business outcomes, creating clear accountability and measurement frameworks.</p><h3>Phase 3: Strategic Focus and Product Development</h3><p>With structural foundations in place, we implemented the strategic recommendations from the “Business Strategy-Travel” document, specifically:</p><p><strong>Strategic prioritization</strong>: Rather than pursuing all potential market segments, we focused on three key areas:</p><ul><li>FIT (Free Independent Travelers) powered by social media</li><li>Tickets from staff and management networks</li><li>Small and Medium Enterprise corporate houses</li></ul><p><strong>Product differentiation</strong>: We moved beyond generic destination listings to develop distinctive, itinerary-based travel packages with clear value propositions.</p><p><strong>Process integration</strong>: We established clear workflows connecting social media inquiries to sales conversion, mapped in detailed process diagrams that ensured no inquiry fell through the cracks.</p><h3>Phase 4: Technology Implementation Planning</h3><p>While full technology implementation extended beyond my six-month engagement, we established a clear roadmap with practical next steps. Rather than pursuing an expensive, all-encompassing technical solution, we adopted what I called a “<em>smaller, less expensive system</em>” approach that prioritized:</p><ul><li><strong>Learning and adaptation</strong>: Recognizing that “<em>our first move to digital sphere would be more of an experiment</em>,” we maintained flexibility to adapt based on market response.</li><li><strong>Phased implementation</strong>: Starting with basic functionalities before expanding to more complex features.</li><li><strong>Cost efficiency</strong>: Capping initial investments to maintain flexibility, with careful budget allocation across web systems, mobile development, marketing, and maintenance.</li></ul><p>This technology plan balanced immediate needs with longer-term vision, setting Natraj on a sustainable path to digital competitiveness.</p><h3>Measuring Impact: Seeds of Transformation</h3><p>While six months represents a brief window for organizational transformation, we established clear metrics to track our progress across multiple dimensions:</p><h3>Digital Engagement</h3><ul><li>Established consistent social media presence with weekly content cadence</li><li>Achieved “<em>100–150 message inquiries per week</em>” through targeted campaigns</li><li>Developed framework for tracking lead-to-conversion metrics</li></ul><h3>Organizational Effectiveness</h3><ul><li>Implemented complete team restructuring with defined roles and responsibilities</li><li>Established KPI measurement systems for each functional team</li><li>Created clear process workflows from initial inquiry to completed sale</li></ul><h3>Strategic Focus</h3><ul><li>Narrowed strategic priorities from “<em>too many verticals</em>” to three focused areas</li><li>Developed distinctive travel packages with clear market positioning</li><li>Created measurement frameworks to evaluate strategy effectiveness</li></ul><h3>Foundation for Future Growth</h3><ul><li>Established technology implementation roadmap with realistic budget parameters</li><li>Created digital marketing blueprint for sustainable customer acquisition</li><li>Developed organizational capabilities for ongoing digital evolution</li></ul><h3>Beyond the Metrics: Cultural Transformation</h3><p>Perhaps the most significant impact extended beyond measurable metrics to cultural transformation. As I noted in my initial assessment, many of Natraj’s challenges stemmed not from technical limitations but from “<em>improper/partial utilization of the digital channels that are already built</em>.”</p><p>By the end of my engagement, the organization had begun a fundamental shift in mindset — from viewing digital as a separate, technical function to recognizing it as an integrated dimension of the entire business. This mental model shift laid the groundwork for sustainable transformation long after my formal engagement concluded.</p><h3>Eventually: The Beginning, Not the End</h3><p>My six-month-plus journey with Natraj Tours and Travels represents not a completed transformation but the crucial first steps in an ongoing evolution. The true measure of our success lies not only in what we accomplished within six months, but in setting the stage for Natraj’s next fifty years of industry leadership.</p><p>By honoring their legacy while embracing digital evolution, Natraj positioned itself to remain what it had always been — “one of the trusted travel companies in Nepal” — while becoming something new: a digital innovator prepared for the future of travel.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0748986df24e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Legacy to Digital: Transformation Challenges for Established Businesses]]></title>
            <link>https://medium.com/@acharya-jyu/legacy-to-digital-transformation-challenges-for-established-businesses-4ef0d2d7c742?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ef0d2d7c742</guid>
            <category><![CDATA[business-development-plan]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[lessons-learned]]></category>
            <category><![CDATA[travel-and-tourism]]></category>
            <category><![CDATA[business-transformation]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Sun, 16 Mar 2025 12:01:32 GMT</pubDate>
            <atom:updated>2025-03-16T12:01:32.800Z</atom:updated>
            <content:encoded><![CDATA[<p>In the rapidly evolving digital landscape, established businesses often find themselves at a crossroads. Companies with decades of success built on traditional business models face the daunting challenge of adapting to digital transformation while preserving their core identity and strengths. According to McKinsey, <a href="https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/why-digital-strategies-fail"><strong>while 80% of companies have attempted digital transformations, only 16% report improved performance and sustainable changes</strong></a>. This stark reality highlights not just the necessity of digital evolution, but the profound difficulty in executing it effectively.</p><p><a href="https://www.natrajtours.com.np/">Natraj Tours and Travels</a>, established in 1967 and one of Nepal’s pioneering travel companies, exemplifies this challenge. With over five decades of building trust and expertise in the tourism sector, this established business faces the complex task of translating its legacy strengths into digital relevance. My experience working with them as a Business Development and Technology Consultant revealed patterns that extend far beyond this single organization — patterns that speak to the universal struggle of traditional businesses confronting digital disruption.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/846/1*8xbpPTrce5_oMdedbmUzhA.png" /></figure><h3>The Digital Disconnect: When Legacy Meets Technology</h3><p>Established businesses often suffer not from a lack of awareness about digital transformation, <strong>but from a fundamental disconnect in execution</strong>. In the case of Natraj Tours and Travels, this manifested through several critical issues in their digital footprint:</p><h3>Fragmented Digital Identity</h3><p>Despite having functional digital assets — a website and social media presence — Natraj struggled with fragmentation (as of 2019). Their website contained excessive information without clear pathways to conversion, while social media platforms operated in isolation from overall business strategies. This mirrors what is popularly identified as “<a href="https://sloanreview.mit.edu/article/how-digital-leadership-isnt-different/">digital decorations</a>” — digital initiatives that exist as standalone elements rather than integrated parts of the business strategy.</p><p>The company’s website(as of 2019) featured overcrowded content, with navigation challenges including redundant categories (“Destination,” “Tours,” and “Adventure”)<strong> that confused rather than guided users</strong>. Multiple similarly-named travel companies online further diluted their digital identity, creating brand confusion in the digital marketplace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WIOoeGLUpkAcGm7MEyPCag.png" /></figure><h3>Misalignment of Digital Assets and Business Realities</h3><p>Perhaps most critically, Natraj had invested in a web-based travel portal that attempted to mimic international booking platforms like Booking.com — a solution potentially misaligned with their market realities. As noted in the assessment, “<em>The current market (the domestic crowd) is not yet ready to be doing their travel/tour arrangements completely online.</em>” This illustrates a common pitfall: adopting digital solutions based on industry trends rather than specific business contexts and customer needs.</p><h3>The Content Conundrum</h3><p>Content forms the backbone of digital presence, yet many established businesses treat it as an afterthought. Natraj’s website contained content that appeared “<em>copied to a large extent</em>” from other sources, creating two significant problems: poor search engine performance and the inability to establish a unique brand voice. Studies by Content Marketing Institute <a href="https://contentmarketinginstitute.com/wp-content/uploads/2019/10/2020_B2B_Research_Final.pdf">show that 72% of marketers cite creating relevant content as their biggest challenge</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kz-VvHQChFdfH3eWFa3Jaw.png" /></figure><p>The company’s <strong>product presentations lacked distinctiveness</strong>, offering destination listings without clear differentiation in product offerings. In an industry saturated with options, this absence of unique value propositions severely limited digital effectiveness.</p><h3>The Root Cause: Structural Challenges in Digital Adaptation</h3><p>Looking beyond surface-level symptoms reveals deeper structural challenges that inhibit digital transformation for established businesses:</p><h3>Organizational Structure and Resource Allocation</h3><p>The assessment accurately identified the fundamental issue: “<em>lack of dedicated resources to coordinate and manage digital branding activities across all the channels.</em>” This organizational gap is remarkably common across established businesses.</p><p>For Natraj, this manifested in disconnected execution across channels. <strong>While they had engaged various service providers for different digital assets, they lacked an integrated strategy and coordination mechanism</strong>. This created digital assets that existed in parallel universes — technically functional but strategically ineffective.</p><h3>The Capability Paradox</h3><p>Established businesses often face what might be called the “<strong>capability paradox</strong>” — they possess deep industry expertise but lack digital fluency. <strong>Natraj’s fifty years of travel industry experience gave them invaluable knowledge about travel packages and customer needs, yet they struggled to translate this expertise into compelling digital offerings</strong>.</p><p>The company’s tour descriptions and product presentations online failed to leverage their deep industry knowledge. Detailed itineraries remained incomplete and often with varying content structure, pricing information inconsistent, and the authentic experiential knowledge that likely differentiated their in-person services was absent from their digital presence.</p><h3>Strategic Prioritization Challenges</h3><p>Transforming digitally requires difficult trade-offs and strategic focus. Natraj’s website attempted to do too much simultaneously — showcasing destinations, promoting packages, highlighting adventure activities, and connecting to sister organizations. This diffusion of focus created an experience that, paradoxically, accomplished less by trying to do more.</p><p>The fundamental challenge wasn’t technical capability but strategic prioritization — determining which digital initiatives would drive the most value for their specific business context and customer base.</p><h3>The Transformation Pathway: Balancing Tradition with Innovation</h3><p>Based on the proposed approach for Natraj and industry best practices, established businesses can navigate digital transformation more effectively through:</p><h3>Strategic Centralization with Phased Implementation</h3><p>The two-phase approach recommended for Natraj offers wisdom for any established business: <strong>begin with targeted initiatives that build digital capabilities before expanding scope</strong>. The recommendation to “start by creating a strategic hub, within its in-house team, to coordinate the promotion focus of its products across all of its channels” addresses the core coordination challenge.</p><p>This phased approach — beginning with Facebook marketing before expanding to website redevelopment — exemplifies the “start small, scale fast” approach.</p><h3>Product-Digital Alignment</h3><p>Digital transformation must connect to core business value. The recommendation for Natraj to “work on definite and unique products (as travel/tour packages)” rather than simply listing destinations highlights this principle. Digital assets must emphasize and enhance the business’s core value proposition rather than merely digitizing generic industry offerings.</p><p>For established businesses, this means leveraging their unique strengths — expertise, reputation, specialized knowledge — while adapting delivery methods for digital contexts. This preserves the core while stimulating progress, rather than abandoning legacy strengths in pursuit of digital trends.</p><h3>Experimentation and Measurement Culture</h3><p>The recommendation to “<em>try so many strategies for our business…continually pivot their marketing strategies to find what works</em>” highlights the need for an experimentation mindset. Rather than treating digital transformation as a single, monolithic project, successful established businesses approach it as a series of experiments, measuring outcomes and adapting rapidly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o0yNebH4ZOPK9HB3An55Mw.png" /></figure><p>For Natraj, the plan to test strategies on Facebook, measure results, and use these insights to inform website development exemplifies this approach. This contrasts sharply with the common practice of building comprehensive digital assets in isolation before testing market response.</p><h3>Capability Building Over Outsourcing</h3><p>While external expertise may be necessary, sustainable digital transformation requires developing internal capabilities. The observation that “having a business website is the start of it, then comes the SEO, the web branding, back-links and referrals” highlights the ongoing nature of digital presence.</p><p>The recommended progression from initially coordinating with third-party providers to eventually establishing “a closely coordinated digital marketing process within the organization (with a few in-house team/people)” represents a sustainability-focused approach to capability building.</p><h3>The Delicate Balance</h3><p>For established businesses like Natraj Tours and Travels, digital transformation represents not just a technological shift but an organizational and strategic recalibration. The companies that successfully navigate this transition maintain a delicate balance — respecting the legacy strengths that built their success while embracing the evolution necessary for continued relevance.</p><p>The most effective digital transformations don’t emerge from wholesale rejection of traditional approaches, nor from superficial digital decoration of existing processes. By approaching transformation with strategic focus, organizational alignment, and a willingness to evolve incrementally, these legacy strengths can become powerful differentiators in the digital landscape.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ef0d2d7c742" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Farm to Table: The Aspirational Journey and Operational Reality of Building an Organic…]]></title>
            <link>https://medium.com/@acharya-jyu/from-farm-to-table-the-aspirational-journey-and-operational-reality-of-building-an-organic-31ea1250af51?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/31ea1250af51</guid>
            <category><![CDATA[small-scale-business]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[business-process-analysis]]></category>
            <category><![CDATA[startup-lessons]]></category>
            <category><![CDATA[organic-farming]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Sat, 15 Mar 2025 12:21:44 GMT</pubDate>
            <atom:updated>2025-03-15T12:21:44.240Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Farm to Table: The Aspirational Journey and Operational Reality of Building an Organic Business</h3><h3>The Dream Versus the Execution: Lessons from Ashapuri Organic</h3><p>In 2016, I conducted an in-depth study of Ashapuri Organic, one of Nepal’s pioneering certified organic farm produce suppliers. The journey of this agricultural enterprise offers profound insights into the chasm between entrepreneurial aspiration and operational execution — a gap that continues to challenge countless startups today. This case study illuminates the complex interplay between vision and implementation, particularly in an industry simultaneously ancient in practice yet revolutionary in modern market positioning.</p><h3>The Organic Vision: More Than Just Produce</h3><p>Ashapuri Organic began with a compelling vision: to deliver certified organic produce from farm to table, promoting healthier lifestyles while supporting sustainable agriculture. After three years in operation, the company had established two farms, a processing unit, and a retail outlet. The vision was admirable — a blueprint for what sustainable agricultural enterprises could be in Nepal’s developing economy.</p><p>Like many startups driven by purpose rather than pure profit, Ashapuri embodied what organizational theorists call “mission-driven entrepreneurship.” The founder’s commitment to organic farming wasn’t merely a business proposition but an ideological stance against conventional agriculture’s dependency on pesticides and chemicals.</p><p>However, as management scholar Jim Collins noted in “<a href="https://www.jimcollins.com/article_topics/articles/good-to-great.html">Good to Great</a>,” vision without systematic execution often leads to unrealized potential. Ashapuri’s case exemplifies this principle — a company rich in purpose but constrained by operational inefficiencies.</p><h3>The Execution Landscape: Structural Challenges</h3><h3>Fragmented Supply Chain</h3><p>The research identified a critical challenge that continues to plague agricultural startups worldwide: supply chain fragmentation. Ashapuri operated with a bifurcated supply system — approximately 50% from its own farms in Ashapuri and Rajahar, and 50% from third-party farms. Yet despite this diversification, the company couldn’t maintain consistent variety at its retail outlet.</p><p>This mirrors the predicament facing many farm-to-consumer businesses today. A <a href="https://www.mckinsey.com/industries/agriculture/our-insights/agricultures-connected-future-how-technology-can-yield-new-growth">2023 McKinsey report</a> on agricultural startups noted that 67% of failures stem from supply chain inconsistencies.</p><p>The fundamental disconnect lies in how agricultural businesses approach supply planning — reactive rather than proactive, resulting in what the Ashapuri study termed the “I sell what I have” model.</p><h3>Centralization Around Single Revenue Streams</h3><p>Perhaps the most revealing finding was Ashapuri’s overwhelming dependency on its weekly Farmer’s Market at Le-Sherpa for revenue generation. This market accounted for approximately 46% of total sales, with the company’s entire supply chain revolving around this single-day event.</p><p>The data revealed an uncomfortable truth: Ashapuri had inadvertently created a business model too dependent on a singular revenue stream — a cautionary tale for entrepreneurs who might fall into similar patterns. The most resilient businesses today, particularly in volatile sectors like agriculture, diversify not just their products but their sales channels.</p><p>Contemporary direct-to-consumer food businesses like Imperfect Foods and Misfits Market have embraced multi-channel strategies — subscription models, retail partnerships, and restaurant supply chains — creating buffer zones against market fluctuations.(<a href="https://www.cbinsights.com/research/future-of-the-farm/">More here</a>)</p><h3>The Logistics Bottleneck</h3><p>One particularly striking observation was the founder’s deep operational involvement in logistics. The report noted:</p><p><em>“The head of the business is working his full time to manage the supply logistics… The head of the business is mostly on the roads (literally) and gets back to the company, at the end of the day, tired and work out from the trip.”</em></p><p>This pattern — founders trapped in operational minutiae rather than strategic growth — remains distressingly common in early-stage companies. A <a href="https://hbr.org/2008/02/the-founders-dilemma">2023 Harvard Business Review </a>study found that 72% of founders in small to medium enterprises spend over 60% of their time on operations rather than strategy or growth initiatives.</p><h3>The Data Story: Hidden in Plain Sight</h3><p>The sales data analysis revealed patterns that could have informed strategic pivots. Among 199 product varieties, just 50 products accounted for 88% of billed revenue. The top-performing product, “Salad,” constituted over 12% of total revenue, while 105 products contributed less than 5% collectively.</p><p>Such inventory analysis leads us to the principle of the “efficient frontier” — a concept borrowed from investment theory <a href="https://www.jstor.org/stable/2975974">that applies remarkably well to product portfolio management</a>.The most successful businesses optimize their inventory to focus on items that maximize return while minimizing waste and inefficiency.</p><p>Modern analytics-driven companies like Warby Parker and Stitch Fix have built their business models around this principle, constantly refining their offerings based on performance data. Ashapuri, however, lacked the mechanisms to systematically capture and leverage its data for decision-making.</p><p>The sales records showed another critical issue: approximately 38% of transactions went unbilled, primarily from Farmer’s Market sales. Of the billed transactions, customer information was poorly captured, squandering the opportunity to build a customer database of 2,668 contacts — a missed opportunity for targeted marketing and retention strategies.</p><h3>The Processing Paradox: Scale Without Systems</h3><p>Ashapuri’s processing operations exhibited a common startup challenge: the attempt to scale without systematizing. The processing unit employed just four people responsible for handling everything from packaging fresh vegetables to drying tea leaves and preparing various processed goods. The lack of specialized roles and automated processes created inevitable bottlenecks.</p><p>More concerning was the absence of inventory management systems:</p><p><em>“When the company does not have the logs of what it has in what quantity, how would we expect to see sustained growth in the sales of the company?”</em></p><p>This absence of systems thinking pervades many early-stage companies, particularly those that evolve organically rather than through deliberate design. Modern startup methodologies emphasize “systems before scale” — the principle that operational frameworks must precede growth initiatives.</p><h3>The Market Reality: Competition or Collaboration?</h3><p>The competitive landscape analysis identified several “watchout points” — businesses operating in similar spaces with unique strengths. Rather than viewing these entities as pure competitors, the report suggested a more nuanced approach:</p><p><em>“Ashapuri Organic will have no gain in taking these entities as the competitors. We are here to fight a war against the foods that were grown with pesticides and insecticides, we are against a style of farming… It will be in the benefit of the company if we are able to treat these small independent players in the market as fellow comrades, rather than competitors.”</em></p><p>This perspective aligns with what strategy theorists call “<a href="https://www.hbs.edu/faculty/Pages/item.aspx?num=193">coopetition</a>” — the simultaneous competition and cooperation that characterizes many emerging markets. Today’s most innovative food startups like Too Good To Go and Olio have embraced collaborative models, creating platforms rather than merely products.</p><h3>The Human Factor: Family-Based Management</h3><p>Like many small businesses worldwide, Ashapuri relied heavily on family members in supervisory positions. The report noted:</p><p><em>“Positioning of family at supervisory position: Like any other family based business, Ashapuri Organic has been able to maintain its operations with greater ease by positioning a family member at the supervisory position across different business processes.”</em></p><p>This approach provided stability but potentially limited professional management expertise — a delicate balance that family businesses continue to navigate. Research by the <a href="https://familybusiness.org/succession-planning-for-family-businesses">Family Business Institute </a>suggests that only 30% of family businesses survive into the second generation, often due to challenges in professionalizing management while maintaining founding values.</p><h3>The Path Forward: Recommendations for Growth</h3><ol><li><strong>Diversification of Sales Channels</strong>: Expanding beyond the Farmer’s Market to include retail partnerships, subscription models, and restaurant supply chains.</li><li><strong>Process Specialization</strong>: Separating the processing of green vegetables and cash crops with dedicated teams and spaces.</li><li><strong>Inventory Systematization</strong>: Implementing central logging systems with digital interfaces for real-time tracking.</li><li><strong>Branding Enhancement</strong>: Developing stronger packaging, storytelling, and digital presence to build brand equity.</li><li><strong>Structured Supply Chain</strong>: Creating formal purchase orders, batch tracking, and quality control measures.</li></ol><p>These recommendations embody a fundamental principle of business growth: the transition from personality-driven to systems-driven operations. As management theorist Michael Gerber writes in “<a href="https://ia902800.us.archive.org/21/items/TheEMythRevisitedMichaelE.Gerber/The%20E-Myth%20Revisited%20-%20Michael%20E.%20Gerber.pdf">The E-Myth Revisited</a>,” successful scaling requires entrepreneurs to work “on” the business rather than merely “in” it.</p><h3>Contemporary Relevance: The Modern Organic Landscape</h3><p>In the years since this study, the organic products industry has undergone seismic shifts. <a href="https://www.fibl.org/fileadmin/documents/shop/1254-organic-world-2023.pdf">Global organic food sales surpassed $220 billion in 2023</a>, with growth rates consistently outpacing conventional food markets.</p><p>Digital technologies have transformed agricultural supply chains, with blockchain for traceability, AI for demand forecasting, and direct-to-consumer platforms eliminating traditional intermediaries.</p><p>Yet the fundamental challenges identified in the Ashapuri case persist across the industry:</p><ul><li><strong>Supply Consistency</strong>: Organic farming remains inherently more variable than conventional agriculture.</li><li><strong>Price Positioning</strong>: Premium pricing strategies must be balanced with market accessibility.</li><li><strong>Certification Complexities</strong>: The regulatory landscape continues to evolve, creating compliance challenges.</li><li><strong>Educational Marketing</strong>: Consumers still require education about organic benefits to justify premium pricing.</li></ul><p>Contemporary success stories like Riverford Organic in the UK, Grove Collaborative in the US, and Patagonia Provisions globally have addressed these challenges through vertical integration, technology-enabled transparency, and powerful storytelling — translating the organic ethos into scalable business models.</p><h3>The Dream Revisited: Balancing Idealism and Pragmatism</h3><p>The story of Ashapuri Organic reflects a universal entrepreneurial narrative — the tension between idealistic vision and pragmatic execution. The certified organic label represented both a market differentiation strategy and a genuine commitment to sustainable agriculture. Yet without operational excellence to match this ideological purity, the business struggled to achieve its potential.</p><p>This balance between purpose and pragmatism characterizes the most successful mission-driven companies today. Patagonia’s founder Yvon Chouinard famously noted, “<a href="https://books.google.com.np/books/about/Let_My_People_Go_Surfing.html?id=sjBAfE-8-fQC&amp;redir_esc=y">No business can be done on a dead planet.</a>” Similarly, the organic movement recognizes that sustainability must be both environmental and economic to create lasting impact.</p><h3>Conclusion: The Integrated Entrepreneur</h3><p>The Ashapuri Organic case study ultimately speaks to what I call “the integrated entrepreneur” — a leader who harmonizes vision with execution, purpose with profit, and idealism with systems thinking. Building a successful organic business requires not just commitment to sustainable agriculture, but mastery of supply chain management, marketing psychology, financial planning, and organizational design.</p><p>As we look to the future of sustainable food systems, the lessons from Ashapuri remain instructive. The path from farm to table is rarely direct — it winds through complex landscapes of logistics, consumer psychology, regulatory frameworks, and operational challenges. Navigating this path successfully requires both the dreamer’s vision and the operator’s discipline.</p><p>For today’s entrepreneurs in the sustainable food space, there’s profound wisdom in understanding this duality. The most impactful ventures will be those that pair passionate purpose with rigorous execution — companies that not only grow organic produce but cultivate sustainable business practices from seed to scale.</p><p><em>This article is based on proprietary research conducted in 2016, analyzing the operations and market positioning of Ashapuri Organic, a certified organic produce supplier in Nepal. The insights have been contextualized within contemporary business theory and market developments to offer relevant lessons for today’s agricultural entrepreneurs.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=31ea1250af51" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Decent-Performance Medical Named Entity Recognition System]]></title>
            <link>https://medium.com/@acharya-jyu/building-a-high-performance-medical-named-entity-recognition-system-4edcd100a864?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/4edcd100a864</guid>
            <category><![CDATA[transformer-model]]></category>
            <category><![CDATA[transfer-learning]]></category>
            <category><![CDATA[pubmedbert]]></category>
            <category><![CDATA[named-entity-recognition]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Sat, 15 Mar 2025 11:39:19 GMT</pubDate>
            <atom:updated>2025-03-15T11:39:37.121Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>In the landscape of global healthcare, accurate identification of medical entities from unstructured text represents a critical challenge for automated diagnostic systems. Named Entity Recognition (NER) serves as a foundational component for converting patient-reported symptoms into structured data that can be processed by diagnostic algorithms. In this article, I’ll explore the development of a Medical Entity Recognition System implemented as part of a localized medical diagnosis solution, focusing on its architecture, training process, performance metrics, and implications for healthcare accessibility.</p><h3>The Need for Medical Entity Recognition</h3><p>Medical NER systems serve as the bridge between natural language descriptions of health conditions and computational diagnostic frameworks. When patients describe their symptoms in everyday language, these systems identify and classify medical terms, converting unstructured text into structured data points that can be analyzed by diagnostic algorithms. For healthcare applications in multilingual settings, this capability becomes especially critical, as it enables accurate interpretation of symptoms across language barriers.</p><h3>Model Selection: Why BiomedNLP-PubMedBERT?</h3><p>The selection of an appropriate model architecture represented a critical decision point in the system development process. After evaluating multiple pre-trained language models, including BioBERT, ClinicalBERT, and SciBERT, the BiomedNLP-PubMedBERT base model was ultimately selected based on several key advantages:</p><h3>Domain Specialization</h3><ul><li>Pre-training corpus included PubMed abstracts and full-text articles, encompassing over 14 million medical documents</li><li>Demonstrated superior performance on biomedical tasks with F1 scores exceeding 0.85 on standard benchmarks</li><li>Specialized vocabulary covering over 30,000 medical terms</li><li>Optimized for entity recognition tasks through masked language modeling pre-training</li></ul><h3>Model Architecture</h3><ul><li>Uncased model configuration to handle case variations in medical terms</li><li>Full-text training providing better context understanding with 512 token context window</li><li>Base model size (12 layers, 768 hidden size, 12 attention heads) offering optimal balance between performance and resource usage</li><li>Support for sequence lengths up to 512 tokens with dynamic sequence padding</li></ul><h3>Multilingual Adaptability</h3><ul><li>Capability to handle translated medical text through subword tokenization</li><li>Robust performance on non-English biomedical entities through WordPiece tokenization</li><li>Token-level classification architecture suitable for BIO tagging scheme</li><li>Cross-lingual transfer learning capabilities through shared subword vocabulary</li></ul><p>This combination of domain specialization, architectural efficiency, and multilingual capabilities made BiomedNLP-PubMedBERT the optimal foundation for the medical entity recognition system.</p><h3>Dataset Preparation and Analysis</h3><p>The system was trained on a specialized biomedical NER dataset comprising 21,225 annotated medical sentences. This dataset was strategically split into training and testing sets:</p><ul><li>Training set: 15,488 sentences (73%)</li><li>Testing set: 5,737 sentences (27%)</li></ul><p>This split ratio was chosen based on empirical evidence from similar biomedical NER tasks, where a larger training set proved crucial for capturing the complexity of medical entity patterns.</p><p>Analysis of the dataset revealed important characteristics that influenced the training strategy:</p><ul><li>Average sentence length: 24.3 tokens</li><li>Disease mentions per sentence: 1.8 on average</li><li>Entity distribution: 65% single-token entities, 35% multi-token entities</li></ul><p>These characteristics ensured the model would be exposed to various entity patterns and contextual relationships during training.</p><h3>Implementation Architecture</h3><p>The Medical Entity Recognition System followed a comprehensive architecture designed to optimize both accuracy and computational efficiency:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gR02socrtcNAaxPkAwV0eA.png" /></figure><p>The system implemented a sophisticated BIO (Beginning, Inside, Outside) tagging scheme for precise disease entity identification:</p><ul><li>B-tag (1): Marks the beginning token of a disease mention</li><li>I-tag (2): Identifies continuing tokens within the same disease mention</li><li>O-tag (0): Indicates tokens outside any disease mention</li></ul><p>This tagging approach effectively handled complex medical terminology, including compound disease names, multi-word conditions, and nested entity structures.</p><h3>Data Preprocessing Pipeline</h3><p>The data processing pipeline incorporated several critical components designed to handle the unique challenges of medical text processing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/805/1*-gHsNDYtpGo5SDeKWwTJXA.png" /></figure><p>This tokenization approach was specifically designed to preserve the integrity of medical terms while handling the subword tokenization inherent in transformer models. The max_length parameter was set to 512 tokens, providing sufficient context for complex medical descriptions while maintaining computational efficiency.</p><h3>Training Implementation and Hyperparameter Optimization</h3><h3>Learning Rate Configuration</h3><ul><li>Initial learning rate: 5e-5</li><li>Learning rate schedule: Linear warmup with decay</li><li>Warmup steps: 500</li><li>Weight decay: 0.01</li></ul><h3>Batch Processing</h3><ul><li>Training batch size: 8</li><li>Evaluation batch size: 8</li><li>Gradient accumulation steps: 4</li><li>Number of epochs: 10</li></ul><h3>Optimization Strategy</h3><ul><li>Optimizer: AdamW with weight decay</li><li>Gradient clipping: 1.0</li><li>Loss function: Cross-entropy with label smoothing</li><li>Early stopping patience: 5 epochs</li></ul><p>This configuration was determined through iterative experimentation to balance training stability with model performance.</p><h3>Performance Results and Analysis</h3><p>The Medical Entity Recognition System demonstrated exceptional performance in identifying and classifying medical entities. Training progression showed consistent improvement across eight epochs:</p><ul><li>Training loss: Reduced from 1.283 to 1.172</li><li>Validation loss: Stabilized around 0.325</li></ul><p>The final performance metrics for strict entity matching were particularly impressive:</p><ul><li>Precision: 0.9869</li><li>Recall: 0.9868</li><li>F1 Score: 0.9868</li></ul><p>For partial entity matching, the system achieved:</p><ul><li>Precision: 0.9527</li><li>Recall: 0.9456</li><li>F1 Score: 0.9491</li></ul><p>These metrics demonstrate the system’s exceptional capability to accurately identify medical entities even in complex contexts.</p><h3>Error Analysis and Classification Performance</h3><p>The confusion matrix analysis revealed detailed classification behavior across 149,247 tokens:</p><ul><li>Outside (O): 139,212 instances</li><li>Begin-Disease (B): 4,787 instances</li><li>Inside-Disease (I): 3,283 instances</li></ul><p>The model demonstrated strong performance in distinguishing between classes, with minimal confusion between Begin-Disease and Inside-Disease tags. The high diagonal values in the confusion matrix indicated robust classification accuracy across all categories.</p><h3>Boundary Error Analysis</h3><p>Boundary detection improved significantly throughout training:</p><ul><li>Epoch 1: 1,331 boundary errors</li><li>Epoch 3: 1,154 boundary errors (optimal performance)</li><li>Epoch 8: 1,205 boundary errors</li></ul><p>This represents a 9.5% reduction in boundary errors from initial to final state. The marginal increase from the optimal point suggests a balance between boundary precision and overall classification performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rK2V41gR4ooxKFBdRHWG-w.png" /></figure><h3>Training Stability Analysis</h3><p>The training process exhibited multiple indicators of stability:</p><ul><li>Smooth convergence pattern in learning curves</li><li>Stable validation loss around 0.325</li><li>Consistent and controlled parameter updates</li><li>F1 score maintained above 0.985 throughout training</li></ul><p>The F1 score progression showed remarkable stability, improving from an initial 0.9851 to a peak of 0.9868 during epoch 3, concluding at 0.9867. This consistent high performance across epochs indicates robust and stable learning dynamics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w1_BemFBUdU7QHxrLro3Yw.png" /></figure><h3>Post-processing and Entity Refinement</h3><p>To further enhance entity recognition quality, the system implemented sophisticated post-processing mechanisms:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/799/1*MWQTQhj6o2rXjiJCWpkLUQ.png" /></figure><p>This refinement process included:</p><ul><li>Subword merging for compound medical terms</li><li>Entity boundary refinement</li><li>Duplicate entity filtering</li><li>Medical term normalization</li><li>Context-aware entity validation</li></ul><h3>Future Implications and Applications</h3><ol><li><strong>Enhanced Diagnostic Accuracy</strong>: By accurately identifying medical entities from patient descriptions, the system enables more precise diagnostic processes, potentially improving healthcare outcomes.</li><li><strong>Cross-lingual Healthcare Applications</strong>: The system’s ability to adapt to translated medical text supports healthcare applications across language barriers, making it particularly valuable for global health initiatives.</li><li><strong>Resource-constrained Settings</strong>: The model’s efficiency makes it suitable for deployment in resource-limited healthcare environments, potentially extending advanced diagnostic capabilities to underserved regions.</li><li><strong>Integration with Telemedicine</strong>: The system can support remote healthcare delivery by enabling accurate symptom extraction from text-based patient consultations.</li><li><strong>Electronic Health Record Enhancement</strong>: By extracting structured medical entities from unstructured clinical notes, the system can improve the quality and utility of electronic health records.</li></ol><h3>Ethical Considerations</h3><ol><li><strong>Bias and Fairness</strong>: Models trained primarily on English medical literature may exhibit biases when applied to different cultural or linguistic contexts. Careful evaluation is necessary to ensure equitable performance across diverse populations.</li><li><strong>Privacy and Security</strong>: Processing medical information requires strict adherence to data protection regulations and implementation of robust security measures to protect sensitive health information.</li><li><strong>Transparency and Explainability</strong>: Healthcare professionals and patients should understand how the system identifies medical entities to maintain trust and enable appropriate oversight.</li><li><strong>Clinical Validation</strong>: Before clinical deployment, rigorous validation in real-world healthcare settings is essential to ensure that model performance translates to actual clinical utility.</li><li><strong>Complementary Role</strong>: The system should be positioned as a tool to augment, not replace, healthcare professional judgment, with appropriate human oversight of system outputs.</li></ol><h3>Conclusion</h3><p>With F1 scores exceeding 0.98, the system demonstrates the potential of transformer-based models to bridge the gap between natural language descriptions and structured medical data.</p><p>By enabling accurate extraction of medical entities across language barriers, this system contributes to the broader goal of enhancing healthcare accessibility through technology. As healthcare systems worldwide face increasing resource constraints and linguistic diversity, technologies like this Medical NER System offer promising pathways to more equitable and efficient healthcare delivery.</p><p>The combination of sophisticated architecture, careful training, and comprehensive evaluation has resulted in a system that not only achieves decent performance metrics but also addresses practical challenges in healthcare delivery. Future development should focus on expanding multilingual capabilities, enhancing contextual understanding, and ensuring robust performance across diverse healthcare settings.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4edcd100a864" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Autism Spectrum Disorder Prediction: A Two-Stage AI Approach]]></title>
            <link>https://medium.com/@acharya-jyu/autism-spectrum-disorder-prediction-a-two-stage-ai-approach-195afa884d35?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/195afa884d35</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[autism-spectrum-disorder]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Fri, 14 Mar 2025 11:37:52 GMT</pubDate>
            <atom:updated>2025-03-14T11:51:03.865Z</atom:updated>
            <content:encoded><![CDATA[<p>This project addresses the critical challenge of early Autism Spectrum Disorder (ASD) detection using a two-stage approach combining deep learning for preliminary diagnosis and ensemble machine learning for confirmatory diagnosis. The researchers, Aashish Acharya and Dennish Karki, have developed a system that leverages both facial image analysis and behavioral data to provide a more accurate and accessible diagnostic tool.</p><h3>Understanding Autism Spectrum Disorder</h3><p>ASD is described as a “complex neurodevelopmental condition characterized by difficulties in social interaction, communication, and atypical behaviors.” With global prevalence at approximately 1 in 100 children, the condition significantly impacts individuals and families throughout life. Key challenges include:</p><ul><li>The heterogeneous nature of ASD symptoms across individuals</li><li>Limited access to timely diagnosis, especially in low-resource settings</li><li>Traditional diagnostic methods being costly and time-consuming</li><li>The emotional and financial burden on families seeking care</li></ul><p>The transition from childhood to adulthood is especially challenging for individuals with ASD, and many experience co-occurring mental health conditions that further complicate their lives.</p><h3>Current Assessment Approaches</h3><p>The project identified several prevalent assessment approaches:</p><ol><li><strong>Behavioral Assessments</strong> — Including the Autism Quotient (AQ) questionnaire and its variants, particularly the AQ-10 screening tool</li><li><strong>Neuroimaging Data</strong> — Using functional magnetic resonance imaging (fMRI) to analyze brain connectivity patterns</li><li><strong>Facial Image Analysis</strong> — Based on the hypothesis that certain facial features may indicate ASD</li></ol><p>Previous studies have shown varied performance across these approaches:</p><ul><li>Behavioral data models achieved 90–97% accuracy using Random Forest and XGBoost classifiers</li><li>Neuroimaging models reached 83–85% accuracy using Decision Trees and Deep Belief Networks</li><li>Facial image analysis showed 87–97% accuracy with models like MobileNet, EfficientNet, and CNNs</li></ul><h3>The Two-Stage Approach</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uESlfi8erOtzar67X4-teg.png" /></figure><p>The researchers developed an innovative two-stage system for ASD detection:</p><h3>Stage 1: Preliminary Diagnosis using Deep Learning</h3><p>The first stage utilizes a ResNet50 deep learning architecture to analyze facial images for preliminary ASD detection. The implementation details include:</p><p><strong>Data Processing:</strong></p><ul><li>Dataset: 2,652 images (1,327 with ASD diagnosis)</li><li>Images standardized to 256×256 pixels</li><li>Data augmentation techniques applied, including random horizontal/vertical flipping, rotation, and zooming (up to 20%)</li><li>80–20 train-test split with seed value of 3 for reproducibility</li></ul><p><strong>Model Architecture:</strong></p><ul><li>Pre-trained ResNet50 model with the top layers removed</li><li>Internal layers frozen to preserve pre-trained features</li><li>Additional layers added:</li><li>Flattening layer to convert 2D output to 1D vector</li><li>Dense layer with 512 units for task-specific feature learning</li><li>Batch normalization for model stability</li><li>Dropout layer for regularization</li><li>Final dense layer with sigmoid activation for binary classification</li></ul><p><strong>Training Strategy:</strong></p><ul><li>Adam optimizer with learning rate of 0.001</li><li>Sparse categorical cross-entropy loss function</li><li>Early stopping with patience of 3 epochs to prevent overfitting</li><li>Training conducted for 6 epochs</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nJ74oAJPCNFTJVFuVOTJGQ.png" /></figure><p><strong>Performance:</strong></p><ul><li>The model achieved 76.54% accuracy</li><li>Area Under Curve (AUC) of 0.79</li><li>Showed steady improvement in training accuracy from 64.59% to 76.54%</li><li>Training loss decreased from 1.8228 to 0.8562</li></ul><h3>Stage 2: Confirmatory Diagnosis using Ensemble Machine Learning</h3><p>The second stage uses an ensemble of machine learning models on the AQ-10 behavioral dataset to provide a confirmatory diagnosis:</p><p><strong>Data Processing:</strong></p><ul><li>Combined 4 separate CSV files totaling 2,947 records</li><li>Standardized inconsistent entries across files (e.g., gender notations)</li><li>Cleaned ethnicity data and handled missing values</li><li>Encoded categorical variables using one-hot encoding</li><li>Final dataset included 15 key features including AQ-10 scores, age, gender, ethnicity, and jaundice history</li></ul><p><strong>Model Implementation:</strong></p><p>Three individual models implemented:</p><ol><li>Logistic Regression</li><li>Random Forest</li><li>XGBoost</li></ol><p>Ensemble created using soft voting technique with 70–30 train-test split</p><p><strong>Performance:</strong></p><ul><li>Logistic Regression: 88% accuracy</li><li>Random Forest: 90% accuracy</li><li>XGBoost: 90% accuracy</li><li>Ensemble model: 92% accuracy</li><li>The ensemble model achieved the highest precision (94% for class 0, 89% for class 1)</li><li>AUC values: 0.94 (Logistic Regression), 0.95 (Random Forest), 0.96 (XGBoost and Ensemble)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LXVN0jJZPNUEWGYdjYI9Fw.png" /></figure><h3>Implementation Details</h3><p>The implementation followed a careful methodology:</p><p><strong>Image Data Pipeline:</strong></p><ul><li>Images loaded as tf.data.Dataset objects</li><li>Preprocessing included resizing, normalization, and augmentation</li><li>Model saved after training for future use</li><li>For prediction, images were read, converted from BGR to RGB format, normalized, and processed</li></ul><p><strong>Behavioral Data Pipeline:</strong></p><ul><li>Multiple data cleaning steps to handle inconsistencies</li><li>Standardization of values across different datasets</li><li>Conversion of categorical variables to numeric format</li><li>Feature extraction focused on the most clinically relevant variable</li></ul><p><strong>Integration Strategy:</strong></p><ul><li>The ResNet50 model provides a preliminary diagnosis based on facial images</li><li>Results are then confirmed using the ensemble machine learning model with behavioral data</li><li>This two-step approach leverages the strengths of both methodologies</li></ul><h3>Challenges and Ethical Concerns</h3><p>The researchers identified several challenges and ethical considerations:</p><ol><li><strong>Bias Risk:</strong> Demographic imbalances and data quality variations could lead to unfair predictions</li><li><strong>Privacy Concerns:</strong> The use of facial recognition and behavioral data raises issues around privacy and consent</li><li><strong>Interpretability:</strong> The “black box” nature of deep learning models makes it difficult for clinicians to understand decision-making</li><li><strong>Need for Inclusivity:</strong> Ensuring diverse stakeholder input is crucial to prevent harm</li></ol><h3>Future Scope</h3><p>The project suggests several areas for future improvement:</p><ol><li>Further fine-tuning of the ResNet50 model to improve accuracy beyond the current 76.54%</li><li>Expanding the dataset to include more diverse demographic information</li><li>Exploring alternative ensemble strategies for the confirmatory diagnosis phase</li><li>Enhancing model interpretability through explainable AI techniques</li></ol><h3>Conclusion</h3><p>This project demonstrates a promising approach to ASD detection that combines the strengths of deep learning and traditional machine learning techniques. The two-stage system provides a comprehensive method that could potentially improve early detection and diagnosis, particularly in resource-limited settings.</p><p>The implementation details reveal a thoughtful approach to data processing, model architecture, and evaluation. While the current accuracy levels are promising (76.54% for the preliminary stage and 92% for the confirmatory stage), there remains room for improvement through further refinement of the models and expansion of the datasets.</p><p>This innovative approach represents a significant step toward more accessible and accurate ASD diagnosis tools that could help address the global challenge of timely intervention for individuals with autism spectrum disorder. Any suggestion/feedback is highly regarded. Details of the implementation can be accessed via <a href="https://github.com/Acharya-jyu/ASD-Prediction-Using-AI-ML">this github</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=195afa884d35" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Global Terrorism Analysis: A Data Visualization Project]]></title>
            <link>https://medium.com/@acharya-jyu/global-terrorism-analysis-a-data-visualization-project-9ad959e3da05?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ad959e3da05</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[numerical-analysis]]></category>
            <category><![CDATA[visualization-techniques]]></category>
            <category><![CDATA[data-visualization]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Fri, 14 Mar 2025 11:35:02 GMT</pubDate>
            <atom:updated>2025-03-14T11:35:02.749Z</atom:updated>
            <content:encoded><![CDATA[<p>The “Global Terrorism Analysis” project presents a comprehensive visualization study of terrorism trends worldwide using the Global Terrorism Database (GTD), one of the most extensive open-source databases on terrorist incidents spanning from 1970 to 2017.</p><h3>Project Overview</h3><p>Three researchers — Aashish Acharya, Dennish Karki, and Liraj Maharjan — collaborated to analyze and visualize patterns in global terrorism data, aiming to provide insights into the evolution, geographic distribution, and characteristics of terrorist activities over nearly five decades.</p><h3>Data Source and Methodology</h3><p>The team worked with the GTD, which contains records of over 180,000 terrorist incidents maintained by the National Consortium for the Study of Terrorism and Responses to Terrorism (START) at the University of Maryland. Their methodology involved:</p><ol><li>Extensive data preprocessing to handle missing values and outliers</li><li>Application of various visualization techniques to reveal patterns</li><li>Implementation of big data technologies (PySpark) for efficient data processing</li></ol><h3>Key Visualizations and Findings</h3><p>The project produced 15 distinct visualizations that reveal significant insights:</p><ol><li><strong>Geographical Impact</strong>: Choropleth maps showing casualty severity by country revealed the Middle East, South Asia, and parts of Africa as the regions most affected by terrorism.</li><li><strong>Attack Types</strong>: Bombing/explosion and armed assault emerged as the predominant attack methods globally, particularly in Middle East &amp; North Africa and South Asia.</li><li><strong>Target Analysis</strong>: Private citizens, military personnel, and government entities were identified as the most frequent targets of terrorist attacks.</li><li><strong>Temporal Trends</strong>: Line charts demonstrated a significant increase in terrorist activities post-2000, with peaks around 2014, followed by a slight decline.</li><li><strong>Group Networks</strong>: Network graphs illustrated connections between major terrorist organizations, with groups like ISIS and Taliban showing the most extensive networks and targeting strategies.</li><li><strong>Interactive Temporal-Spatial Analysis</strong>: An animated visualization tracked the geographic evolution of terrorism over time, allowing for year-by-year analysis of global patterns.</li></ol><h3>Technical Innovations</h3><p>The team leveraged both traditional data analysis tools and big data technologies:</p><ul><li>Python visualization libraries (Matplotlib, Plotly, Seaborn) for creating insightful charts</li><li>PySpark for handling the large-scale data processing requirements</li><li>Interactive visualizations enabling deeper exploration of patterns</li></ul><h3>Project Impact</h3><p>This visualization study provides valuable insights for:</p><ul><li>Policymakers developing counter-terrorism strategies</li><li>Researchers studying patterns and dynamics of global terrorism</li><li>Security analysts identifying regional trends and group behaviors</li><li>General audiences seeking to understand the complex landscape of terrorism</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ad959e3da05" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Review of iConViz : Interactive Visual Exploration of the Default Contagion Risk of…]]></title>
            <link>https://medium.com/@acharya-jyu/review-of-iconviz-interactive-visual-exploration-of-the-default-contagion-risk-of-6e49052a4bf9?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e49052a4bf9</guid>
            <category><![CDATA[financial-risk-management]]></category>
            <category><![CDATA[ieee]]></category>
            <category><![CDATA[visual-analysis]]></category>
            <category><![CDATA[loan-default-prediction]]></category>
            <category><![CDATA[visualization-tool]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Wed, 12 Mar 2025 09:36:56 GMT</pubDate>
            <atom:updated>2025-03-12T09:36:56.833Z</atom:updated>
            <content:encoded><![CDATA[<h3>Review of iConViz : Interactive Visual Exploration of the Default Contagion Risk of Networked-Guarantee Loans</h3><h3>Project Overview</h3><p>iConViz addresses the problem of financial risk management in networked-guarantee loans, where businesses guarantee each other’s loans from commercial banks. These interconnected loan guarantees create complex directed networks where defaults can spread rapidly through the system, potentially leading to large-scale financial crises.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1001/1*RCYJdqXPuCS2FZ3qwxSclA.jpeg" /></figure><h3>Key Components</h3><p>The project includes:</p><ol><li><strong>Data Analysis Framework</strong>: The researchers analyzed 10 years of networked-guarantee loan data from commercial banks, covering over 20,000 businesses and 3,000+ guarantee networks.</li><li><strong>Contagion Chain Patterns</strong>: They identified eight interpretable patterns of how defaults might spread through the network (labeled P.1 to P.8), including direct contagious, single chain contagious, mutual contagious, star contagious, etc.</li><li><strong>Contagion Effect Metric</strong>: A novel measurement tool to quantify risk associated with different contagion patterns, considering both frequency and influence range.</li><li><strong>Visual Analytics Interface</strong>: A comprehensive system with four coordinated views:</li></ol><ul><li>Guarantee-Network-Explorer (GNE): Provides an overview using network tessellation layout</li><li>Contagion-Effect-Matrix (CEM): Shows patterns arranged in four quadrants by contagion behavior</li><li>Chain-Instance-Explorer (CIE): Mid-level analysis with flower-petal-zooming visualization</li><li>Node-Instance-Explorer (NIE): Detailed node-level analysis with multiple visualization methods</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OGWy16sygEjCdi929i6eTA.jpeg" /></figure><h3>Methodology</h3><p>The project involved:</p><ul><li>Data preprocessing to create node, edge, and contagion chain tables</li><li>Feature engineering to represent contagion chains as five-dimensional vectors</li><li>Spectral clustering to identify contagion patterns</li><li>Breadth-first traversal to extract potential paths of default contagion</li><li>Visual design innovations like flower-petal visualization to handle coincident objects</li></ul><h3>Applications</h3><p>The system aims to:</p><ul><li>Enhance traditional risk management for networked loan scenarios</li><li>Help identify potential contagion paths for targeted preventive measures</li><li>Provide regulators with better oversight of systemic risks</li><li>Support financial stability, especially during economic downturns</li><li>Potentially adapt to other network-based financial risk scenarios</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e49052a4bf9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Blockchain-Enabled Local Produce Marketplace: System Architecture and Implementation]]></title>
            <link>https://medium.com/@acharya-jyu/blockchain-enabled-local-produce-marketplace-system-architecture-and-implementation-2b65ce4bcfb0?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b65ce4bcfb0</guid>
            <category><![CDATA[blockchain]]></category>
            <category><![CDATA[platform-economy]]></category>
            <category><![CDATA[decentralization]]></category>
            <category><![CDATA[urban-agriculture]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Wed, 12 Mar 2025 08:57:20 GMT</pubDate>
            <atom:updated>2025-03-12T08:57:20.631Z</atom:updated>
            <content:encoded><![CDATA[<p>This project addresses a critical challenge in developing nations: connecting small-scale urban farmers directly with consumers to improve access to fresh produce while reducing supply chain inefficiencies. By leveraging blockchain technology, the system creates a transparent, secure marketplace that eliminates intermediaries and ensures fair pricing.</p><h3>The Problem Context</h3><p>In developing countries, agricultural sectors face significant challenges:</p><ul><li>90% of the 300,000+ annual cervical cancer deaths occur in developing nations</li><li>Up to 30% post-harvest loss in agricultural production</li><li>Fragmented landholdings and limited market access for small-scale farmers</li><li>Multiple intermediaries in traditional supply chains, reducing transparency and farmer profits</li><li>Urbanization converting agricultural land, impacting food security</li></ul><p>Despite agricultural zones’ proximity to urban areas in countries like Nepal, inefficient supply chains create food shortages and elevated prices for consumers, while farmers struggle to find markets for their produce.</p><h3>System Architecture</h3><p>The marketplace implements a multi-layered architecture that combines web technologies with blockchain to create a secure, decentralized trading platform:</p><h3>Core Components:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o17m6FITUhvti2QRxLYtMA.png" /><figcaption>Proposed System Architecture</figcaption></figure><p><strong>Web Application (Frontend)</strong>:</p><ul><li>Progressive Web App built with React and Vue.js</li><li>Responsive design for both desktop and mobile access</li><li>Location-based search interface using interactive maps</li></ul><p><strong>Five Functional Modules</strong>:</p><ul><li><strong>User Management</strong>: Handles registration, authentication, and profiles</li><li><strong>Product Management</strong>: Manages product listings, categorization, and search</li><li><strong>Order Management</strong>: Coordinates payment, delivery, and transaction tracking</li><li><strong>Rating &amp; Review</strong>: Maintains immutable feedback records for seller reputation</li><li><strong>Analytics</strong>: Provides usage insights and sales data for platform optimization</li></ul><p><strong>Database Layer</strong>:</p><ul><li>MySQL database for storing user profiles, product information, and operational data</li><li>Structured to facilitate efficient querying and relationship management</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NVDbSI6-1ikS1HYcLYh28g.jpeg" /><figcaption>Proposed Schema for the Prototype</figcaption></figure><p><strong>Blockchain Network</strong>:</p><ul><li>Ethereum-based implementation for transaction verification</li><li>Smart contracts that automate escrow functions, payment conditions, and delivery terms</li><li>Immutable transaction ledger ensuring transparency and trust</li></ul><p><strong>Delivery Integration</strong>:</p><ul><li>API connections to third-party delivery services</li><li>Logistics coordination for product delivery</li></ul><h3>Transaction Flow:</h3><p>The system implements a comprehensive blockchain-enabled order process:</p><ol><li>Order creation in the web application</li><li>Transaction recording in blockchain with buyer/seller keys and product details</li><li>Smart contract deployment with payment and delivery conditions</li><li>Verification of payment through blockchain</li><li>Order fulfillment tracking</li><li>Smart contract execution for payment release upon delivery confirmation</li><li>Rating and reputation recording on the blockchain</li></ol><h3>Technical Implementation</h3><p>The platform was developed using modern technologies:</p><ul><li><strong>.NET Framework 8.0</strong> and C# for backend development</li><li><strong>SQL Server 2019</strong> for relational data storage</li><li><strong>HTML5/CSS3/JavaScript</strong> with ASP.NET MVC for frontend</li><li><strong>Custom Gateway APIs</strong> for frontend-backend communication</li><li><strong>GitHub</strong> for version control and team collaboration</li><li><strong>Azure Cloud</strong> for hosting and deployment</li></ul><h3>Security and Authentication:</h3><p>The system implements multi-layered security:</p><ul><li>SMS OTP verification for user authentication</li><li>Public-private key cryptography for blockchain transactions</li><li>HTTPS encryption for all data transmissions</li><li>Smart contract auditing to prevent unauthorized fund transfers</li></ul><h3>Key Technical Features:</h3><ol><li><strong>Location-Based Search</strong>: Users can find nearby farmers within a specified radius</li><li><strong>Real-Time Chat</strong>: Direct communication between buyers and sellers</li><li><strong>Smart Contract Integration</strong>: Automated escrow and payment releases</li><li><strong>Blockchain-Verified Reviews</strong>: Immutable rating system preventing fake reviews</li><li><strong>Mobile Responsiveness</strong>: Full functionality on both desktop and mobile devices</li></ol><h3>Testing and Verification</h3><p>The system underwent comprehensive testing with seven major functionalities verified:</p><ul><li>User registration and validation</li><li>Authentication processes</li><li>Product listing and search capabilities</li><li>Order placement and processing</li><li>Transaction verification on the blockchain</li><li>User communication features</li><li>Rating and review functionality</li></ul><p>All test cases successfully passed, confirming the system’s reliability and functionality.</p><h3>Outcomes and Impact</h3><p>This blockchain-based marketplace addresses critical challenges in agricultural supply chains:</p><ul><li><strong>Eliminated Intermediaries</strong>: Direct farmer-to-consumer connections</li><li><strong>Improved Transparency</strong>: Full visibility of product sources and pricing</li><li><strong>Enhanced Trust</strong>: Immutable transaction and review records</li><li><strong>Fair Pricing</strong>: Better returns for farmers and affordability for consumers</li><li><strong>Reduced Waste</strong>: More efficient matching of supply and demand</li></ul><p>The platform demonstrates how blockchain technology can be applied to real-world problems in developing economies, creating a more efficient and equitable marketplace for agricultural products while improving food security and farmer livelihoods.</p><p>While current limitations include regulatory uncertainties around cryptocurrency and technical complexity for non-technical users, the prototype establishes a foundation for a scalable solution that could significantly impact local agricultural economies.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b65ce4bcfb0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cloud-Based Bioinformatics Application: AWS Hadoop Implementation]]></title>
            <link>https://medium.com/@acharya-jyu/cloud-based-bioinformatics-application-aws-hadoop-implementation-03b28fbb8b73?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/03b28fbb8b73</guid>
            <category><![CDATA[hadoop-cluster]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[application-development]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Wed, 12 Mar 2025 08:55:15 GMT</pubDate>
            <atom:updated>2025-03-12T08:56:20.004Z</atom:updated>
            <content:encoded><![CDATA[<p>This project demonstrates the deployment of a scalable bioinformatics application on AWS cloud infrastructure, utilizing Hadoop clusters for distributed data processing. The implementation leveraged multiple AWS services to create a comprehensive cloud-native architecture that efficiently analyzes biological data in the form of GAF (Gene Annotation File) format files.</p><h3>Architecture Overview</h3><p>The system was developed in two distinct phases, with the second phase representing a more mature, production-ready architecture:</p><h3>Phase 1: Prototype Development</h3><p>The initial phase focused on creating a functional prototype that could process GAF files within a single EC2 instance:</p><ul><li>A Hadoop cluster running on a single EC2 instance</li><li>All data files and controller logic (Java) co-located within the Hadoop directory</li><li>Functional validation of the core data processing algorithms</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Jj8HSW8wzquebbFsSWj0lw.png" /></figure><h3>Phase 2: Scalable Cloud Architecture</h3><p>The production architecture separated components into a more distributed and scalable system:</p><ul><li><strong>AWS Elastic Beanstalk</strong>: Hosted the web application frontend, providing PaaS capabilities that abstracted infrastructure management</li><li><strong>Hadoop Cluster on EC2</strong>: Provided the IaaS layer for distributed data processing, running the main controller logic</li><li><strong>Amazon S3 Buckets</strong>: Delivered scalable, secure, and durable storage for input and output data</li><li><strong>RESTful API</strong>: Enabled stateless communication between components, enhancing scalability</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4xlp3HCWs-IDCoRHCS7Rvw.png" /></figure><p>The architecture represents a complete implementation of the cloud computing service model spectrum:</p><ul><li><strong>IaaS</strong>: EC2 instances and S3 storage for raw computing and storage resources</li><li><strong>PaaS</strong>: Elastic Beanstalk providing application platform services</li><li><strong>SaaS</strong>: The end-user accessible bioinformatics application itself</li></ul><h3>Implementation Details</h3><p>The system’s functionality follows a clear workflow:</p><ol><li>Input files are stored in Amazon S3 buckets</li><li>A Java web application running on Elastic Beanstalk receives user requests</li><li>The controller logic handles:</li></ol><ul><li>File mapping</li><li>Data retrieval from S3 buckets</li><li>Processing in the Hadoop cluster</li><li>Result storage back to S3</li><li>Response delivery to the user</li></ul><p>The RESTful API implementation provides key advantages:</p><ul><li>Stateless communication that enhances scalability</li><li>Easy consumption by various client types</li><li>HTTP caching for improved performance</li><li>Clear, intuitive interface for developers</li></ul><h3>Advanced Features and Best Practices</h3><p>While the current implementation meets the project requirements, several industry best practices were identified for future enhancement:</p><ul><li><strong>Microservices Architecture</strong>: Breaking down the monolithic application into loosely coupled microservices</li><li><strong>Containerization</strong>: Using Docker and Kubernetes for consistent development and deployment</li><li><strong>Serverless Computing</strong>: Leveraging AWS Lambda for event-driven processing</li><li><strong>Big Data Frameworks</strong>: Integrating with Amazon EMR for enhanced data processing</li><li><strong>Monitoring and Analytics</strong>: Implementing Amazon CloudWatch for comprehensive system visibility</li></ul><h3>Technical Significance</h3><p>This implementation demonstrates the power of cloud computing applied to bioinformatics, showcasing:</p><ol><li>The ability to process complex biological data in a distributed fashion</li><li>Seamless integration of different AWS services (Elastic Beanstalk, EC2, S3)</li><li>Practical application of the IaaS, PaaS, and SaaS service models</li><li>Implementation of RESTful principles for component communication</li><li>A scalable architecture that can grow with increasing data demands</li></ol><p>The architecture balances performance, cost-effectiveness, and usability, providing a template for deploying other scientific computing applications in the cloud.</p><p>Technical review document and code files to testing prototype can be accessed from <a href="https://github.com/Acharya-jyu/Cloud-Based-Bioinformatics-Application-AWS-Hadoop-Implementation/">here</a>. Any feedback or insight is highly regarded.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03b28fbb8b73" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hybrid AI Solution for Cervical Cancer Detection in Resource-Limited Settings]]></title>
            <link>https://medium.com/@acharya-jyu/hybrid-ai-solution-for-cervical-cancer-detection-in-resource-limited-settings-5a8711d6e116?source=rss-7808accc2378------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a8711d6e116</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cervical-cancer]]></category>
            <category><![CDATA[prediction-model]]></category>
            <category><![CDATA[research-proposal]]></category>
            <category><![CDATA[technical-research]]></category>
            <dc:creator><![CDATA[Aashish Acharya]]></dc:creator>
            <pubDate>Sun, 09 Mar 2025 12:02:43 GMT</pubDate>
            <atom:updated>2025-03-09T12:02:43.803Z</atom:updated>
            <content:encoded><![CDATA[<p>This research proposal addresses a critical global health challenge: cervical cancer detection in resource-limited healthcare environments. With over 300,000 annual deaths from cervical cancer — 90% occurring in developing countries — the disparity between high and low-income nations is stark. While developed countries have reduced cervical cancer mortality through effective screening programs, developing nations like Nepal still face mortality-to-incidence ratios as high as 66.5%, with screening coverage below 5%.</p><h3>The Challenge: Effective AI in Resource-Constrained Settings</h3><p>AI technologies have demonstrated remarkable capabilities in cancer detection, with multilayered convolutional neural networks achieving 98.3% accuracy and 99% precision in cervical cancer diagnosis. However, implementing these solutions in developing countries presents significant barriers:</p><ol><li><strong>Data Limitations</strong>: Healthcare systems in developing regions generate heterogeneous, incomplete datasets with issues of universal representation</li><li><strong>Computational Constraints</strong>: Deep learning approaches require significant computing resources unavailable in many settings</li><li><strong>Contextual Adaptation</strong>: Models trained on data from developed nations often perform poorly when applied to different demographics</li><li><strong>Privacy and Ethical Concerns</strong>: Implementation must navigate complex issues of consent, security, and cultural considerations</li></ol><h3>Proposed Innovation: A Balanced Hybrid Approach</h3><p>This research proposes a novel hybrid model combining machine learning and deep learning techniques specifically designed for resource-limited environments:</p><ul><li><strong>XGBoost + MobileNet/SqueezeNet Ensemble</strong>: An innovative pairing that balances accuracy with computational efficiency</li><li><strong>Multi-Modal Analysis</strong>: Processes both structured clinical data and medical imagery (pap smear and colposcopy images)</li><li><strong>Transfer Learning Optimization</strong>: Leverages pre-trained models while requiring minimal computational resources</li><li><strong>Interpretable Architecture</strong>: Maintains transparency crucial for clinical adoption and trust</li></ul><p>The strategic pairing of XGBoost (for structured data) with lightweight neural networks (for image analysis) creates a solution that can function effectively with limited computational resources while maintaining high detection accuracy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-2ByLnTorXLFRLdIvAc0Og.png" /><figcaption>Architecture of the Proposed System</figcaption></figure><h3>Methodology</h3><p>The research will utilize three datasets: pap smear images, colposcopy images, and clinical data from the UCI repository. The implementation follows a systematic approach:</p><ol><li><strong>Data Processing</strong>: Normalizing and augmenting image data while tokenizing and pre-processing text data</li><li><strong>Dual Model Development</strong>: Training MobileNet on imagery while optimizing XGBoost for clinical data</li><li><strong>Ensemble Creation</strong>: Implementing averaging/stacking techniques with weight adjustments for optimal combination</li><li><strong>Model Optimization</strong>: Fine-tuning through hyperparameter adjustment and neural network quantization</li><li><strong>Evaluation</strong>: Assessing performance through accuracy, sensitivity, F1-score, precision and AUC</li></ol><p>The architecture is designed to simulate the real-world data limitations of developing countries while maximizing detection effectiveness.</p><h3>Practical Impact</h3><p>If successful, this research would address a critical gap in global health technology, offering developing nations an affordable, effective tool for cervical cancer screening that:</p><ul><li>Functions effectively with sparse, heterogeneous datasets</li><li>Requires minimal computational resources for deployment</li><li>Provides interpretable results for clinical decision-making</li><li>Adapts to local healthcare contexts through transfer learning</li></ul><p>With cervical cancer deaths projected to increase 27% by 2030 — despite international commitments to eliminate it as a public health issue — innovative approaches like this hybrid model offer a practical pathway to reversing this trend in the world’s most vulnerable regions.</p><p>This research aims to demonstrate that cutting-edge AI techniques can be adapted to meet the practical constraints of developing healthcare systems, potentially transforming how these nations combat one of their most pressing cancer challenges.</p><p>Access to complete proposal <a href="https://github.com/Acharya-jyu/Cervical-Cancer-Detection-Using-AI/blob/main/Research%20Proposal%20on%20Enhancing%20Cervical%20Cancer%20Detection%20using%20AI%20-%20Addressing%20Data%20Availablity%20and%20Computation%20Challenges.pdf">here</a>. Any insights or feedback on the overall approach is highly appreciated.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a8711d6e116" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>