<?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 Dania Obaid Sayed on Medium]]></title>
        <description><![CDATA[Stories by Dania Obaid Sayed on Medium]]></description>
        <link>https://medium.com/@dos21012?source=rss-39914fb6dfdc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*9uG9Im1oAGb_nYTK</url>
            <title>Stories by Dania Obaid Sayed on Medium</title>
            <link>https://medium.com/@dos21012?source=rss-39914fb6dfdc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:31:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dos21012/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[From Understanding to Experimenting: My Replicate Journey]]></title>
            <link>https://medium.com/@dos21012/from-understanding-to-experimenting-my-replicate-journey-32729b509e3f?source=rss-39914fb6dfdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/32729b509e3f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tuberculosis]]></category>
            <category><![CDATA[ai-in-healthcare]]></category>
            <dc:creator><![CDATA[Dania Obaid Sayed]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 09:30:23 GMT</pubDate>
            <atom:updated>2026-11-05T11:35:07.092Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><em>Continuing my Learn Article: Can We Find Tuberculosis Before It Finds More People?</em></blockquote><h3>From researching an idea to actually building it</h3><p>After completing my <a href="https://medium.com/@dos21012/tuberculosis-can-we-find-it-before-it-finds-more-people-54d6080530de">Learn Article</a> about how artificial intelligence could contribute to earlier tuberculosis (TB) detection, I wanted to take my research further. Throughout the article, I explored technologies that already exist, including Google’s Health Acoustic Representations (HeAR), AI-powered cough analysis, computer-aided chest X-ray interpretation and other emerging approaches to identifying TB.</p><p>However, I started asking myself a different question: What if I actually tried to recreate part of this technology myself? I wanted to move beyond researching how these systems worked and experiment with the technology myself.</p><blockquote>This became the foundation of my Replicate project.</blockquote><p>My research introduced me to Google HeAR, which focuses on learning representations from health-related sounds. I became interested in the idea that an AI model could learn patterns from audio recordings that might not be immediately noticeable to a human listener.</p><p>I also looked at the difference between traditional audio processing and machine learning. Audio processing allows us to measure characteristics such as frequency and energy, while machine learning can identify patterns and classify sounds based on what a model has learned.</p><p>My original objective was not simply to produce a graph of a cough. I wanted to investigate how audio could be transformed into numerical data and how a model could interpret that information.</p><h3>My first actual prototype: What Does a Cough Look Like?</h3><p>I decided to begin with something achievable: creating a Google Colab notebook called <em>What Does a Cough Look Like?</em></p><p>I chose Google Colab because it allowed me to write and execute Python code without needing to set up a complete programming environment on my computer.</p><p>I started with two main libraries:</p><ul><li><strong>Librosa:</strong> For loading, processing and analysing audio recordings.</li><li><strong>Matplotlib:</strong> For creating visual representations of the audio.</li></ul><p>My first task was to upload cough recordings and make Python read them correctly. I started with three audio recordings and explored their characteristics.</p><h3>Turning audio into visual information</h3><p>I created two main types of visualizations.</p><blockquote><strong>Waveforms</strong></blockquote><p>A waveform displays how the amplitude of an audio signal changes over time. This allowed me to see the structure of a cough recording and how its intensity varied throughout the recording.</p><blockquote><strong>Spectrograms</strong></blockquote><p>A spectrogram displays how the frequency content of a recording changes over time. It helped me investigate the different frequencies present in the audio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9gksq2pHLHioXchi9cUKGw.png" /><figcaption><em>My first successful waveform showing how the amplitude of a cough recording changes over time, and my first spectrogram visualizing how different frequencies appear throughout the cough recording.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GsGV45sc7riusXCnn4tdkA.png" /><figcaption><em>Comparing the visual characteristics of my three original cough recordings using waveforms and spectrograms.</em></figcaption></figure><h3>Extracting numerical features from cough recordings</h3><p>After successfully creating the visualizations, I realized that graphs alone wouldn’t provide all the information I needed. I wanted Python to calculate actual numerical measurements from each recording.</p><p>Using Librosa, I started extracting different audio features.</p><p>These included:</p><ul><li><strong>Duration: </strong>How long the recording lasts.</li><li><strong>Sample rate:</strong> How many audio samples are recorded per second.</li><li><strong>RMS energy:</strong> A measurement related to the signal’s average energy.</li><li><strong>Spectral centroid:</strong> The approximate centre of the sound’s frequency distribution.</li><li><strong>Spectral bandwidth:</strong> How spread out the frequencies are.</li><li><strong>Dominant frequency:</strong> The frequency component with the greatest magnitude.</li><li><strong>Zero-crossing rate:</strong> How frequently the audio signal crosses zero.</li></ul><p>These measurements helped me understand the acoustic characteristics of the recording. However, I also learned that these values alone could not tell me whether someone had TB.</p><h3>A major change: Why I moved away from directly using HeAR</h3><p>As I continued developing the project, I encountered an important limitation with my original approach.</p><p>I wanted to replicate Google HeAR, but I realized that simply using the same model would not demonstrate that I had independently recreated its functionality.</p><p>I also encountered difficulties accessing and working with the model itself.</p><p>I needed an alternative that would allow me to experiment with machine learning using tools I could actually implement within my available time.</p><p>This led me to explore <strong><em>YAMNet</em></strong>.</p><h3>Discovering YAMNet</h3><p>YAMNet is a pre-trained machine-learning model that recognizes different categories of sounds.</p><p>Unlike HeAR, which focuses on health-related audio representations, YAMNet is designed for general audio-event classification.</p><p>I decided to explore it because it could recognize sound categories such as coughing, speech and other everyday sounds.</p><p>This gave me a way to experiment with machine learning without relying directly on the model I originally wanted to replicate.</p><h3>My first successful machine-learning output</h3><p>One of my successful outputs showed:</p><p>Embedding shape: (6, 1024)</p><p>An embedding is a numerical representation that captures information about an input. In my case, YAMNet transformed the audio into vectors containing numerical values that represented learned characteristics of the sound.</p><p>The output shape of (6, 1024) meant that I had six embedding vectors, each containing 1,024 numerical values.</p><p>I initially found this concept confusing because the model wasn’t giving me a simple explanation of what each number represented. However, I began understanding that these values work together as a representation that a machine-learning system can use for further analysis.</p><p>I also had to be careful when interpreting the classification scores. A cough score of 0.4043 did not mean there was a 40.43% chance of tuberculosis. It was a score associated with the model’s cough sound category, not a medical prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/926/1*Eo-2I7fjYdo85NilqA8r_A.png" /><figcaption><em>My first successful YAMNet classification results, showing the different sound categories recognized by the model and their corresponding scores.</em></figcaption></figure><h3>The biggest challenge: Finding a suitable TB dataset</h3><p>I felt that my project was still missing an important part.</p><p>My original research question was connected to tuberculosis, not simply cough recognition.</p><p>I wanted to investigate whether the acoustic characteristics of cough recordings from people with TB differed from those of people without TB.</p><p>To do this, I needed a dataset containing properly labelled recordings.</p><p>This turned out to be one of the most time-consuming parts of my entire project.</p><p>I had already tried working with two other datasets. I downloaded and investigated both, hoping they would contain the TB-related recordings I needed. However, when I examined the available recordings and information, I couldn’t identify any confirmed TB recordings in either dataset. This meant that, even though I had spent time downloading and exploring them, I still didn’t have the data I needed for my experiment.</p><p>This was one of the points where I genuinely had to accept defeat. I had already spent a considerable amount of time trying to find suitable recordings, and it was frustrating to keep reaching dead ends.</p><p>However, I decided not to abandon the project. Instead, I continued searching for another dataset that would allow me to investigate the connection between cough sounds and tuberculosis.</p><h3>My first dataset attempts</h3><p>I investigated an alternative cough dataset and attempted to process a larger collection of recordings.</p><p>One of my attempts took approximately an hour to process, which made it difficult to continue experimenting efficiently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/830/1*sgUjgPseCxUGt_3AvHDEOw.png" /><figcaption><em>One of my earlier dataset-processing attempts, which took an hour with zero results and no progress.</em></figcaption></figure><h3>Discovering the TB-screen dataset</h3><p>After struggling with the earlier dataset attempts, I explored another option: the TB-screen dataset associated with research by Sharma and colleagues.</p><p>This looked promising because it contained recordings collected from participants with pulmonary tuberculosis and control participants.</p><p>The research described approximately 149 pulmonary-TB participants and 46 controls, with a large collection of passive and forced cough recordings.</p><p>I found the Zenodo record and downloaded the TB-screen code ZIP, which was approximately 284.2 MB.</p><p>At first, I thought I had downloaded the actual recordings I needed. However, I discovered that the code package and the complete audio collection were separate.</p><p>The actual audio collection was reported to be approximately 395.4 GB!!</p><p>This meant that downloading the code ZIP did not automatically give me access to all the original recordings.</p><p>I had to investigate the files that were already available and understand how the dataset was organized.</p><h3>Investigating thousands of files</h3><p>Once I had the TB-screen material available in Google Colab, I began inspecting its contents.</p><p>I discovered that the dataset directory contained:</p><ul><li>4,259 files in total.</li><li>1,267 audio recordings.</li><li>125 CSV files.</li></ul><p>This was much more information than I had initially expected!</p><p>However, another problem appeared.</p><p>When I first searched through the filenames, I couldn’t find clear TB-positive or TB-negative labels.</p><p>My initial scan returned:</p><ul><li>0 filenames containing TB-positive labels.</li><li>0 filenames containing TB-negative labels.</li></ul><p>This did not mean that the dataset had no labels. It meant that the labels weren’t necessarily stored in the filenames.</p><p>I realized that I needed to investigate the CSV files and other metadata instead.</p><h3>Understanding the dataset’s metadata</h3><p>I then started investigating the available metadata files, including:</p><ul><li>Metadata.csv</li><li>Dataset_key.csv</li><li>Passive_coughs.csv</li><li>Forced_coughs.csv</li></ul><p>During this process, I discovered that some of the dataset’s CSV files contained much more useful information than the filenames themselves.</p><p>For example, I inspected a file called folds/T1_0.csv.</p><p>It contained 5,031 rows and included columns such as label, class, and path.</p><p>Importantly, I found examples where the label was “TB”and the corresponding class was 1.</p><p>This was a major step forward!!</p><p>I also inspected another file:</p><p>result_baseline/T1_vggish_1.csv</p><p>This contained <em>4,673</em> rows including columns such as:</p><ul><li>Actual</li><li>Predicted</li><li>score</li><li>path</li></ul><p>I discovered that this file contained existing baseline-model results, including predictions and scores.</p><p>However, I still needed to understand how these results were generated and whether they could be used to compare my own original recordings. Existing model results were not automatically equivalent to having the original audio and a compatible model for my own comparison.</p><h3>Expanding my own analysis to 18 recordings</h3><p>While investigating the larger TB-screen collection, I also worked on organizing and analyzing a separate working collection of recordings.</p><p>Instead of focusing only on my original three cough samples, I managed to create a structured output containing 18 recordings.</p><p>For each recording, I extracted several acoustic characteristics, including:</p><ul><li>Participant identifier.</li><li>Filename.</li><li>Group.</li><li>Duration.</li><li>RMS energy.</li><li>Spectral centroid.</li><li>Spectral bandwidth.</li><li>Dominant frequency.</li><li>Zero-crossing rate.</li></ul><p>This allowed me to see multiple recordings together. However, I needed to keep the different collections separate and verify their labels before treating them as TB-positive or non-TB data.</p><p>The 18-recording analysis was a useful technical milestone, but it did not automatically establish a validated TB comparison.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uLxqIr77FXptKjy3TKlDOw.png" /><figcaption><em>My expanded audio-analysis table containing participant information, recording groups and extracted acoustic features.</em></figcaption></figure><h3>My next intended experiment: Comparing the recordings</h3><p>The next step I did was that I created two additional tables to organize the recordings and their acoustic measurements in a way that would make comparisons between the labelled groups easier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aVNFimeev9S9fhVH-ZYlyA.png" /><figcaption>My first additional table, organizing the recording information and measurements to help me examine the different participant groups.</figcaption></figure><p>I then created a second table to continue organizing the information needed for the comparison.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L1fr2_2RihpX5PHqz4-Dag.png" /><figcaption>My second additional table, providing another organized view of the recordings and their acoustic characteristics.</figcaption></figure><p>After organizing the tables, I created five graphs comparing the TB-labelled and non-TB-labelled groups using RMS Energy, Spectral Centroid, Spectral Bandwidth, Dominant Frequency and Zero-Crossing Rate. The graphs showed differences as well as overlapping values, reminding me that these features alone could not reliably distinguish between the groups, especially with a limited number of participants.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/1*jIoZBVdLg6ycdJMU6e9zOA.png" /></figure><p>After analyzing the graphs, I wanted to return to the actual audio recordings. Numerical values were useful, but I also wanted to see what the sounds themselves looked like. I generated waveforms for the TB-labelled participants. These waveforms showed how the amplitude of each recording changed over time, allowing me to examine the sound patterns of the different participants.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A18Ao4QuRfO38disAwVQUQ.png" /><figcaption>Waveforms of the TB-labelled participants, showing how the amplitude of each audio recording changed over time.</figcaption></figure><p>I then generated spectrograms for the same TB-labelled recordings. While waveforms show changes in amplitude, a spectrogram shows how frequencies change over time. The horizontal axis represents time, the vertical axis represents frequency, and the colours show sound intensity. This gave me another way to examine the audio and explore its frequency patterns. I wanted to use both visualizations because they reveal different characteristics of the same sound.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7pW5chXwKS_yK5A03_7cYQ.png" /><figcaption>Spectrograms of the TB-labelled participants, showing the frequency patterns present throughout their recordings.</figcaption></figure><p>By this stage, I had analyzed the recordings using numerical tables, comparison graphs, waveforms and spectrograms. However, I still wanted to connect the experiment to my original idea. I decided to compare my original recordings, including the fake-laugh recording, with the TB-labelled and non-TB-labelled reference groups. This was an interesting part of the project because the fake laugh was not a genuine TB cough, yet I wanted to see how its acoustic measurements compared with the reference data.</p><p>For the final comparison, I used the extracted acoustic features and compared the original recording with the average measurements of the two labelled groups. The notebook calculated how close the recording’s measurements were to each group’s reference average. The smaller the distance, the more acoustically similar the recording was to that group’s average based on the features included in the calculation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ytcZteqhVV5J1d59_5RDNg.png" /><figcaption>My final table between the original recording and the TB-labelled and non-TB-labelled reference groups.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xovJvIAeqBv-0oVHx3e89w.png" /><figcaption>Comparing the waveforms and spectrograms between the original recording and the TB-labelled and non-TB-labelled reference groups.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GoPMN353-XBengfG7O9epw.png" /><figcaption>My final comparison between them, showing which group’s average acoustic measurements were closer to the recording.</figcaption></figure><p>The final output showed that my original fake-laugh recording was acoustically closer to the TB-labelled group than to the non-TB-labelled group. Its distance from the TB-labelled average was 1.022, compared with 1.295 from the non-TB-labelled average. This was a surprising result because the recording was a fake laugh rather than a genuine cough from someone with tuberculosis. The fake laugh was closer to the TB-labelled group because some of its numerical audio features may have been similar. This shows that numerical similarity does not necessarily mean two recordings contain the same type of sound.</p><p>This helped me understand acoustic similarity does not prove that someone has tuberculosis. A fake laugh being closer to the TB-labelled group does not mean that the person has TB, and it certainly does not mean that the system can diagnose the disease. It only means that, based on the selected measurements, the recording was closer to that group’s average. Other sounds and respiratory conditions can share similar acoustic characteristics.</p><h3>Feedback, limitations and changes in direction</h3><p>I had managed to move from analyzing a few original recordings to working with labelled groups, creating visual comparisons and calculating acoustic similarities. However, I also realized that producing a result is not the same as proving that the result is medically accurate.</p><p>Throughout the project, I had to reconsider what I was actually trying to demonstrate. My original goal required much more than finding patterns in data. It also requires reliable datasets, carefully tested methods and proper medical validation. There were several important distinctions I had to understand.</p><p>First, YAMNet is a general audio-classification model, not a TB-specific model.</p><p>Second, its cough-classification score does not represent the likelihood of someone having tuberculosis.</p><p>Third, acoustic features such as frequency and energy can overlap between different recordings and respiratory conditions.</p><p>Finally, a dataset containing TB labels does not automatically mean that a model trained on it will be accurate or medically reliable.</p><p>This changed how I described my prototype.</p><p>Instead of presenting it as an AI system that diagnoses TB, I focused on demonstrating how audio processing and machine learning can be used to investigate respiratory sounds.</p><p>I also had to recognize that my original goal of replicating HeAR was more ambitious than what I could realistically complete within my available time and technical experience.</p><h3>What my prototype currently does</h3><p>After all these experiments, errors, dataset investigations and changes in direction, I have developed a working audio-analysis prototype with two main components.</p><blockquote>Component 1: Traditional audio processing</blockquote><p>Using Librosa and Matplotlib, my notebook can:</p><ul><li>Load cough recordings.</li><li>Extract duration and sample rate.</li><li>Calculate RMS energy.</li><li>Calculate spectral centroid and bandwidth.</li><li>Identify dominant frequency.</li><li>Calculate zero-crossing rate.</li><li>Generate waveforms.</li><li>Generate spectrograms.</li></ul><blockquote>Component 2: Machine-learning audio classification</blockquote><p>Using YAMNet, my notebook can:</p><ul><li>Process audio recordings through a pre-trained model.</li><li>Recognize general sound categories.</li><li>Generate sound-classification scores.</li><li>Produce numerical audio embeddings.</li></ul><p>These components allow me to explore cough recordings from both a signal-processing and machine-learning perspective.</p><p>The TB-screen investigation also allowed me to explore the labelled groups and carry out an acoustic comparison with my original recording. However, this was an experiment based on numerical similarity, not a medically validated TB detection system.</p><p>I did not recreate HeAR itself, but I replicated parts of the general audio-analysis process using Python, Librosa and YAMNet which created my own TB-related comparisons and experiments.</p><h3>What I learned from the entire experience</h3><p>Looking back, this project was much more complicated than I initially expected.</p><p>I had to understand how audio is processed, learn how machine-learning models interpret sound, investigate numerical embeddings, work through programming errors and spend a considerable amount of time searching for suitable research data.</p><p>The dataset investigation was particularly important because it showed me that finding data is not the same as finding useful data. I had to understand where the recordings came from, how they were labelled and whether the information could actually support my research question.</p><p>Most importantly, I learned that changing direction doesn’t necessarily mean that a project has failed. Moving from HeAR to YAMNet, investigating different datasets and reconsidering my original objective were all part of developing a more realistic and technically achievable project.</p><p>I haven’t recreated Google HeAR, and I haven’t developed a medically validated TB detection system. What I have developed is an early cough-audio analysis prototype that combines signal processing and machine learning, while investigating how these approaches could contribute to future respiratory-health research.</p><p>I could test this further by more labelled cough recordings and test the approach on sounds it had never analysed before. I could then measure how accurately it distinguishes between the groups and how often it makes mistakes.</p><p>For me, the most valuable part of this experience wasn’t simply getting my code to work. It was learning how to approach a complicated problem, deal with uncertainty, question my results and find another way forward when my original approach didn’t work. In the end, I figured it out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=32729b509e3f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tuberculosis: Can We Find It Before It Finds More People?]]></title>
            <link>https://medium.com/@dos21012/tuberculosis-can-we-find-it-before-it-finds-more-people-54d6080530de?source=rss-39914fb6dfdc------2</link>
            <guid isPermaLink="false">https://medium.com/p/54d6080530de</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tuberculosis]]></category>
            <dc:creator><![CDATA[Dania Obaid Sayed]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 12:47:49 GMT</pubDate>
            <atom:updated>2026-10-18T09:45:18.586Z</atom:updated>
            <content:encoded><![CDATA[<p>What if a curable disease is still killing more than a <em>million</em> people a year?</p><p>That is the reality of tuberculosis, or TB. The disease is preventable and treatable, but millions of people still become sick, and many cases are not found quickly enough.</p><p>That made me wonder: <strong>Could emerging technology help us find TB earlier, especially in places where getting to a hospital or specialist is difficult?</strong></p><h3>What Exactly Is TB?</h3><p>Tuberculosis is an infectious disease caused by the bacterium <em>Mycobacterium tuberculosis</em>. It most commonly affects the lungs, although it can affect other parts of the body. TB can spread through the air when a person with infectious pulmonary TB coughs, speaks, or sneezes.</p><p>According to the World Health Organization (WHO), about one-quarter of the world’s population is estimated to have been infected with TB bacteria. Most people who are infected do not develop TB disease, but some eventually become sick.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/483/1*it4PSBjqoBPoUDlavB071w.jpeg" /><figcaption>TB bacteria microscopic</figcaption></figure><h3><strong>What Is the Actual Problem?</strong></h3><p>The biggest problem is not simply that TB exists. One of the biggest problems is that people can have TB without being diagnosed quickly.</p><p>In 2024, an estimated <strong>10.7 million people developed TB</strong>, and about <strong>1.23 million people died from it</strong>. Around <strong>8.3 million people were officially diagnosed and notified</strong>, leaving a large gap between the estimated number of people who became sick and the number who were identified by health systems.</p><p>That detection gap matters because someone who is not diagnosed cannot easily start the right treatment or take steps to reduce transmission.</p><p>TB symptoms can also develop gradually. Common symptoms include a persistent cough, chest pain, fever, night sweats, tiredness, and weight loss.</p><h3>My Take</h3><p>The number that stands out to me is not just the number of deaths. It is the gap between people who are estimated to have TB and people who are actually diagnosed.</p><p>That makes detection feel like one of the most important places for innovation.</p><h3>How Is TB Diagnosed Today, and Why Can It Be Difficult?</h3><p>Diagnosing tuberculosis involves more than looking at someone’s symptoms. Healthcare workers may begin by asking about symptoms such as a persistent cough, fever, and night sweats. However, these symptoms can also occur with other illnesses, so they cannot confirm TB on their own.</p><p>One common tool is a chest X-ray. The X-ray creates an image of the lungs, allowing healthcare workers to look for unusual areas that could be associated with TB. However, other lung conditions can cause similar changes, and some people with TB may not have clear signs on their X-rays. This means an X-ray can help identify people who need further testing, but it cannot always confirm the disease.</p><p>Laboratory tests provide more specific evidence. For example, a sputum sample, which contains mucus coughed up from the lungs, can be tested for TB bacteria or their genetic material. Molecular tests look for genetic sequences associated with the bacteria, and some can also identify signs of drug resistance. These tests are important because knowing whether TB is present — and whether certain medicines may work — is essential for choosing the right treatment.</p><p>However, testing can be harder in communities with limited laboratory facilities, trained staff, or medical equipment. People may also have to travel long distances to reach a clinic. These challenges help explain why someone can have TB but remain undiagnosed.</p><p>This is where I think emerging technology could make a difference. Instead of <em>replacing</em> existing tests, AI and portable diagnostic devices could help healthcare workers identify people who need testing and bring some testing services closer to them.</p><h3>So Where Can AI Actually Help?</h3><p>AI does not replace a doctor or magically diagnose TB from one piece of information. Instead, it can help healthcare workers screen more people, prioritize who needs more testing, and bring screening tools into places that have fewer resources.</p><h3>How Does AI Read a Chest X-ray?</h3><p>AI systems used for TB screening are trained using large collections of chest X-rays that have been assessed or labelled for relevant findings. During training, the model learns patterns in the images that are associated with TB, such as certain distributions of unusual shadows or changes in lung tissue.</p><p>When a new X-ray is uploaded, the model processes the image and compares its patterns with what it learned during training. Depending on the system, it may produce an abnormality score or a score indicating how strongly the image resembles patterns associated with TB. Healthcare workers can use this result to help decide which people may need further investigation.</p><p>For example, a higher score may mean that an X-ray should receive closer attention. However, the score is not a diagnosis or a direct measurement of how much TB a person has. Its meaning depends on how the system was designed and validated, and the final decision still requires appropriate medical assessment and testing.</p><p>Tools such as CAD4TB from Delft Imaging and qXR from Qure.ai use AI to analyse chest X-rays. Their value is that they can support screening, especially in places where access to specialist image interpretation is limited.</p><p>However, AI can make mistakes. Other lung diseases may resemble TB, and the quality of the image or differences between patient populations can affect performance. A promising score must therefore lead to appropriate follow-up rather than being treated as proof of infection.</p><p>What I find interesting is that AI can help narrow down which images need attention, but the score still needs to be interpreted carefully. It seems useful as a way to support healthcare workers, not replace them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jFqKPD7rev15Nem8aT7GHQ.jpeg" /><figcaption>AI-assisted chest X-ray</figcaption></figure><h3>How Can AI Analyze a Cough?</h3><p>Cough-based AI works differently because it analyses sound instead of an image. A recording of a person’s cough is converted into digital audio data. A model can process the recording into a representation of its sound patterns and use patterns learned from previously collected examples to estimate whether the cough resembles those associated with a particular respiratory condition.</p><p>For example, a model may examine patterns in the sound over time, including changes in frequency and intensity. It then produces a prediction based on what it learned during training. The precise features and methods vary between systems.</p><p><strong>Swaasa</strong>, developed by Salcit Technologies, is one example of technology being investigated for cough-based respiratory screening. <strong>Google’s HeAR</strong> model is another example of research into learning from health-related sounds, including coughs and breathing.</p><p>The important limitation is that a cough is not unique to TB. Several illnesses can cause similar sounds, and background noise, recording devices, and differences between people may affect a model’s predictions. Cough AI should therefore be treated as a possible screening aid, not a stand-alone TB test.</p><p>I think this technology is interesting because recording a cough could be simpler than traveling to a hospital for an initial assessment. However, its real value depends on whether it can work reliably in different communities and connect people to proper diagnostic testing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A9rOvanxa3riAawinA_4OA.png" /></figure><h3>Portable Technology</h3><p>Technology can also help by changing <strong>where</strong> testing happens.</p><p>Portable digital X-ray systems can bring chest imaging closer to remote communities instead of requiring every person to travel to a large hospital.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jy0h9s9kZIusrDe6To91NA.jpeg" /><figcaption>Portable X-ray in remote community</figcaption></figure><p>Another example is <strong>Truenat</strong>, a portable molecular testing platform developed by Molbio Diagnostics. It is designed to work in peripheral and lower-resource settings and can be used for molecular testing for TB.</p><p>This matters because molecular tests can help confirm infection and provide information that a screening tool alone cannot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*4VnMBBDtMhovBDsoSeS3cw.jpeg" /><figcaption>Portable molecular testing</figcaption></figure><h3>Who Is Already Working on This?</h3><p>There is already a large ecosystem of organizations working on TB detection and diagnosis.</p><p>The <strong>World Health Organization (WHO)</strong> develops recommendations and performance standards.</p><p><strong>FIND</strong> works on diagnostic innovation and evaluation. <a href="https://www.finddx.org/">finddx.org</a></p><p><strong>Delft Imaging</strong> developed CAD4TB, a system that supports automated analysis of chest X-rays for TB screening. <a href="https://www.delft.care/">Delft Imaging</a></p><p><strong>Qure.ai</strong> developed qXR, an AI-based chest X-ray analysis platform. <a href="https://www.qure.ai/">Qure.ai</a></p><p><strong>Salcit Technologies</strong> is working on cough-based respiratory screening through Swaasa. <a href="https://www.salcit.com/">Salcit</a></p><p><strong>Molbio Diagnostics</strong> developed Truenat for portable molecular testing. <a href="https://www.molbiodiagnostics.com/">Molbio Diagnostics</a></p><h3>How CAD4TB works, step by step</h3><p>CAD4TB is made by <strong>Delft Imaging</strong> with Radboud University Medical Center in the Netherlands. Version 6, released in 2018, was the first version to use deep learning (<a href="https://doi.org/10.1038/s41598-020-62148-y">Murphy et al., 2020</a>). Here is what happens from X-ray to decision.</p><p><strong>What goes in.</strong> A digital chest X-ray taken from the front (a posterior-anterior view).</p><p><strong>What happens inside.</strong></p><ol><li><em>Normalization.</em> X-ray machines produce images that look different from one another, so the software first adjusts the image to reduce those differences.</li><li><em>Quality check.</em> It confirms the image is actually a valid front-view chest X-ray.</li><li><em>Lung segmentation.</em> A deep neural network outlines the lung fields so the analysis focuses on the right region.</li><li><em>Texture analysis.</em> A deep convolutional neural network studies the texture of the lung tissue. It was trained on annotated X-rays to recognize patterns that tend to go with TB. In plain terms, it learned from examples which patterns of light and dark lung texture are more often seen in TB. The developers don’t publish the training data, citing commercial sensitivity, which is a transparency limit worth knowing about (<a href="https://doi.org/10.1038/s41598-020-62148-y">Murphy et al., 2020</a>).</li></ol><p><strong>What comes out.</strong> In under 15 seconds, the software produces a score from 0 to 100 (higher means more likely TB) and a colored heatmap marking regions it considers suspicious, from blue (normal) to red (abnormal) (<a href="https://doi.org/10.1038/s41598-020-62148-y">Murphy et al., 2020</a>). In the paper’s example, an X-ray scored 91.7 and the patient’s Xpert test was positive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*liePn3sV1EutoMetz5tlQQ.jpeg" /><figcaption>How CAD4TB transforms chest X-ray features into a numerical score diagram</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AbEEioWbvtbqUWFUj_uLDg.jpeg" /><figcaption><em>Original X-ray beside the heatmap version</em></figcaption></figure><p><strong>How the score becomes a decision.</strong> The score is not a diagnosis. A program must pick a <strong>threshold</strong>: people scoring above it get a sputum test, and people below it are sent home. For example, one MSF product sheet describes a setup where a score above 40 triggers a GeneXpert test (<a href="https://www.unicat.msf.org/cat/product/92522">MSF CAD4TB product sheet</a>). Lowering the threshold catches more TB but sends more healthy people for costly tests. Raising it saves tests but misses more cases. Choosing that setting well is one of the hardest parts of using the tool.</p><h3><strong>What Still Isn’t Solved?</strong></h3><p>Even though these technologies are promising, there are still major challenges.</p><ul><li><strong>AI can make mistakes.</strong> A screening tool can miss disease or flag something that turns out not to be TB.</li><li><strong>Performance can vary.</strong> Results can depend on populations, healthcare settings, X-ray machines, recording devices, and the data used to train the models.</li><li><strong>Access is still a problem.</strong> A technology can be excellent but have little impact if communities cannot afford it, operate it, or connect it to healthcare workers.</li><li><strong>Cough-based AI is still developing.</strong> A cough can be caused by many different conditions, so audio alone should not be treated as a definitive TB diagnosis.</li><li><strong>Screening is only useful if people can receive follow-up testing and treatment.</strong></li></ul><h3>My Hot Take</h3><p><strong>My hot take is that the future of TB detection shouldn’t be one giant machine sitting in a hospital.</strong></p><p>I think it should be a system that brings smaller pieces of healthcare to the people who need them.</p><p>A phone could help with initial screening, portable imaging could bring X-rays into communities, and portable molecular testing could help confirm cases closer to where people live.</p><p>I wouldn’t want AI to replace doctors. I’d want it to help healthcare workers find the people who might otherwise be missed.</p><h3>What Happens After Detection?</h3><p>Finding TB early only matters if detection leads to care.</p><p>Once TB is confirmed, the person needs access to appropriate treatment and follow-up. Technology can also help after diagnosis by improving communication, tracking appointments, supporting healthcare workers, and helping people stay connected to care.</p><h3>Where Is the Field Heading?</h3><p>The future of TB detection is likely to focus on making diagnosis <strong>faster, more accessible, and more accurate</strong>.</p><p>Researchers are working toward better ways to detect drug-resistant TB, improve screening, expand access to portable tests, and use AI responsibly.</p><p>Another important goal is finding people who may not have obvious symptoms before they pass the disease to others.</p><p>I think the biggest shift will be from healthcare that waits for people to reach a hospital toward healthcare that can <strong>reach people where they are</strong>.</p><p>AI, portable imaging, molecular testing, and better data systems could all be pieces of that change.</p><p>TB is an old disease, but the tools we can use to fight it are becoming increasingly modern.</p><p>The main challenge I kept coming back to is <strong>detection</strong>: finding people sooner, especially when distance, cost, or limited healthcare resources make diagnosis difficult.</p><p>AI-assisted chest X-rays, cough-based AI, portable X-rays, and molecular tests such as Truenat are not magic solutions. They each have limitations, and none should replace medical professionals.</p><p>But together, they show how emerging technology could help close the gap between people who have TB and people who are actually diagnosed.</p><p>To me, the most exciting possibility is not simply making technology smarter.</p><p><strong>It is making healthcare easier to reach.</strong></p><h3>About Me</h3><p>Hi, I’m Dania!</p><p>I became interested in tuberculosis because I wanted to understand how a disease that is preventable and curable can still affect millions of people every year.</p><p>What really caught my attention was the <strong>detection gap</strong>. I started looking into whether emerging technology could help find people who might otherwise be missed.</p><p>I found it really interesting how AI, cough analysis, portable X-rays, and other technologies could work together to make detection easier, especially for people who live far from hospitals.</p><p>TB made me especially interested in the intersection between <strong>AI, healthcare, and accessibility</strong>, and I want to explore how technology could make early detection possible for more people around the world.</p><h3>Sources</h3><ul><li><strong>World Health Organization — Tuberculosis Fact Sheet<br></strong><a href="https://www.who.int/news-room/fact-sheets/detail/tuberculosis">Tuberculosis</a></li><li><strong>World Health Organization — Global Tuberculosis Programme<br></strong><a href="https://www.who.int/teams/global-tuberculosis-programme">https://www.who.int/teams/global-tuberculosis-programme</a></li><li><strong>World Health Organization — Global Tuberculosis Reports<br></strong><a href="https://www.who.int/teams/global-tuberculosis-programme/tb-reports">Global tuberculosis reports, 1997 to 2025</a></li><li><strong>World Health Organization — Systematic Screening for Tuberculosis Disease<br></strong><a href="https://www.who.int/publications/i/item/9789240022676">WHO consolidated guidelines on tuberculosis: module 2: screening: systematic screening for tuberculosis disease</a></li><li><strong>FIND — Diagnostics for All<br></strong><a href="https://www.finddx.org/">finddx.org</a></li><li><strong>Delft Imaging<br></strong><a href="https://www.delft.care/">Delft Imaging</a></li><li><strong>Qure.ai<br></strong><a href="https://www.qure.ai/">Qure.ai</a></li><li><strong>Salcit Technologies<br></strong><a href="https://www.salcit.com/">Salcit</a></li><li><strong>Molbio Diagnostics / Truenat<br></strong><a href="https://www.molbiodiagnostics.com/">Molbio Diagnostics</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54d6080530de" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>