<?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 Ludivine Siau on Medium]]></title>
        <description><![CDATA[Stories by Ludivine Siau on Medium]]></description>
        <link>https://medium.com/@lu-syo?source=rss-6d0de68df3d6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PD3Fpi7dvOgh3aUTkAn2yg.jpeg</url>
            <title>Stories by Ludivine Siau on Medium</title>
            <link>https://medium.com/@lu-syo?source=rss-6d0de68df3d6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 12:28:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lu-syo/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[Week 31: Denied access and pilot results]]></title>
            <link>https://medium.com/retraining-epoch/week-31-denied-access-and-pilot-results-67b7c79919c3?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/67b7c79919c3</guid>
            <category><![CDATA[research]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[social-justice]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Thu, 23 Apr 2026 19:42:04 GMT</pubDate>
            <atom:updated>2026-04-23T19:42:04.297Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Good news:</strong> This week I finalised my end-to-end pipeline, and it’s running well on the pilot dataset.</p><p><strong>Bad news:</strong> <a href="https://www.bbc.co.uk/news/articles/cpvxgl3n138o">Data extracted from the UK Biobank (UKB) has just been discovered on sale in China,</a> and the access to the platform is now suspended until they can successfully restrict data extraction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*HPz8AzyVOLM9c6za4tV8hw.png" /></figure><p>Tough luck. Just when I was ready to port my pipeline to the full UKB cohort. The platform is expected to reopen in May. Fingers crossed, that means only a couple of weeks to wait.</p><p>In the meantime, I’ll use this forced intermission to break down my pilot results and share some progress here.</p><h3>Refined Structural Causal Model (SCM)</h3><p>Since my last update, I’ve made several key architectural decisions to make my Causal Effect Variational Autoencoder for Health Equity (CEVAE-HE) more robust and effective in isolating unfair bias.</p><p><strong>Reminder</strong>: In my causal model, EHR features are partitioned into sociologically-influenced (<em>X_desc</em>), biologically-correlated (<em>X_corr</em>), and independent (<em>X_ind</em>) pathways.</p><p><strong>What’s new in this refined SCM?</strong></p><ul><li>Only the latent space has a direct influence on the outcome. The observed features are all confounders.</li><li>The latents Ucorr and <em>U_desc</em> are facets of a general Health Latent <em>H</em></li><li>I’ve made the <em>S_bio</em>/<em>S_soc</em> decomposition of the sensitive attribute explicit, and implemented it in the CEVAE-HE. By feeding <em>S_soc</em> into the “desc” part of the network and <em>S_bio</em> into the “corr” part, I can flip only Ssoc to simulate a sociological counterfactual, i.e. a world where the participant would be treated without systemic bias.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/812/1*fnS_7JZeKaOnJaIFYzwUEQ.png" /></figure><h3>Fixing the discriminator</h3><p>I spent a long time fighting with the TC Loss of my VAE, which is meant to minimise statistical dependency between the latent Udesc and the sensitive attribute.</p><p>🤦 The (stupid) reason it wasn’t working? I was rewarding the network if the discriminator was always wrong (null accuracy) instead of making it as good as a random guess (0.5 accuracy). The discriminator was oscillating between being perfectly right and perfectly wrong, leading the adversary objective nowhere.</p><p>With that fixed, I also introduced:</p><ul><li><strong>Latent Counterfactual Invariance loss:</strong> I previously had a Fair loss that was forcing the CEVAE-HE to act as a fair predictor. I realised that the objective of the CEVAE-HE was not to be fair, but to produce causally accurate latents and counterfactuals under our defined causal constraints. These constraints are that the latent <em>U_desc</em> should be 1/ independent of the sensitive attribute and 2/ should remain constant by sociological counterfactual. The former is enforced by the TC loss. And to enforce the latter, the Latent Counterfactual Sensitivity loss penalises any divergence of the latent <em>U_desc</em> in the sociological counterfactual world.</li><li><strong>Latent Redundancy loss, </strong>which forces the network to minimise the predictive information shared by <em>U_corr </em>and <em>U_desc</em>, to make them as independently useful as possible.</li></ul><h3>Stochastic Pathway Analysis</h3><p>Assigning features to descriptive the <em>X_desc</em>, <em>X_corr </em>and <em>X_ind </em>pathways solely via clinical intuition risks overlooking selection biases or “sneaky” data collection artifacts. To formalise this assignment, I’ve created a Stochastic Pathway Sensitivity (SPS) analysis process.</p><p>Instead of making assumptions about the features, I use a bootstrap loop (N=50) to empirically validate assignments. In each iteration, features are randomly partitioned between <em>X_desc </em>and <em>X_corr </em>and the CEVAE-HE is trained to compute four metrics:</p><ul><li><strong>Feature Counterfactual Sensitivity:</strong> How much does the reconstruction of the feature change in the counterfactual world? This indicates how influenced a feature is by the sensitive attribute.</li><li><strong>Predictive Utility:</strong> AUROC and AUPRC of the CEVAE-HE’s native predictor</li><li><strong>Mean Absolute Counterfactual Effect (MACE):</strong> Quantifying counterfactual fairness, it measures how many predictions change when I flip <em>S_soc</em>.</li><li><strong>Total Effect (TE) Error:</strong> The delta between the observed effect and the network’s estimated effect of the sensitive attribute on the target outcome, serving as a proxy for global causal accuracy.</li></ul><p>Post-bootstrap, feature assignments are made using the averaged metrics. Features showing near-zero counterfactual sensitivity are assigned to <em>X_ind</em>. For the remaining features, configurations are mapped onto a Pareto Frontier (representing the optimal solutions for the trade-off between MACE and TE Error). The final configuration is selected by identifying the solution on this frontier that optimises causal fidelity while maintaining utility above a pre-defined clinical threshold (e.g. the CHARGE-AF score for AFib prediction).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*7ZZWHaYhQAPdYrEjpHQRxg.png" /><figcaption>Feature Counterfactual Sensitivity measured across 50 random pathway configurations, respectively 3e-4 and 8e-3 for fbs and ecg.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sGgb7oU5eEXjPfgDgen0vA.png" /></figure><p>On the pilot dataset (UCI Heart Disease), we find <em>fbs </em>(a binary indicator for Fasting Blood Sugar) and <em>ecg </em>(categorical feature for the normal/abnormal result of the ECG) to be sex-independent. And we have two configurations on the Pareto Frontier:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/1*LlsvIzQ-8Uhl-FQ0orCNuQ.png" /></figure><p>We choose the configuration that maximises fairness (lowest max(Female MACE, Male MACE)) and utility, despite a higher causal error. We need to keep in mind that features are on a spectrum between sociologically and biologically influenced and that their assignment to a pathway is an optimisation problem rather than a search for an absolute truthful representation of the real world’s causal relationships.</p><h3>Closing the loop</h3><p>After training and finetuning the CEVAE-HE on the Training dataset, I can finally run it on the Test set to generate the latent variables and the counterfactuals for each participant.</p><p>I use the Test dataset and its corresponding latents and counterfactuals to cross-validate baseline and fair classifiers in a 100-iteration bootstrap. I train four different “fair” models:</p><ul><li><strong>Fair A</strong>, the “pure” model: trained only on the latents <em>[U_desc, U_corr] </em>and the independent features <em>X_ind</em>.</li><li><strong>Fair B</strong>, the hybrid: trained on the <em>U_desc</em> latent (i.e. the clean signal from sociologically-influenced features) and all features free of sociological influence from the sensitive attribute (<em>X_corr</em>, <em>X_ind</em>)</li><li><strong>Fair C &amp; D</strong>, ablation experiments: trained respectively on <em>[X_desc, U_corr, X_ind]</em> and <em>[U_desc, U_corr]</em>, to isolate the bias even further</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sua92jKmfruQbi7jsh-Epw.png" /><figcaption>End-to-end pipeline</figcaption></figure><h3><strong>Pilot results</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/989/1*ELGHt8ajPSJujsCfxYmugQ.png" /><figcaption>Classification performance on the UCI Heart Disease dataset</figcaption></figure><p>The results show a clear Global Utility vs Group Fairness trade-off:</p><ul><li>The baseline model shows a significant gap in FNR between female and male patients (8.18% average FNR disparity)</li><li>Fair A overcorrects that disparity and reverses the bias, now putting men at a disadvantage, but with a Male FNR (17.01%) that is lower than the baseline Female FNR (20.45%)</li><li>Fair B nearly closes the gap between Male and Female patients (1.72% average FNR disparity) with the highest group FNR of 16.21% for Female patients.</li></ul><p>It’s tempting to look at the higher average error rates, or the reversal in disparity, and feel like we’ve failed. Do we want a world that maximizes utility for the majority (Baseline), or a world that raises the minimum for the most vulnerable (Fair A and B)? It might be a context-dependent question that should be answered on a case-by-case basis. In healthcare, the latter is probably preferable as long as solutions provide sufficient minimum utility.</p><h3>What’s next?</h3><ol><li>Dig further in the results and write them up properly for my dissertation</li><li>As soon as the UKB reopens for business I’m porting my pipeline to the Atrial Fibrillation cohort that I have already pre-processed in the platform. Biggest technical (and analysis) change: handling a categorical sensitive attribute (ethnicity) instead of a binary one (sex).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67b7c79919c3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-31-denied-access-and-pilot-results-67b7c79919c3">Week 31: Denied access and pilot results</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 23: Building a Causal Effect VAE for Health Equity]]></title>
            <link>https://medium.com/retraining-epoch/week-23-building-a-causal-effect-vae-for-health-equity-3c75d87aaf42?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c75d87aaf42</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[causal-inference]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[social-justice]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Sun, 01 Mar 2026 20:58:59 GMT</pubDate>
            <atom:updated>2026-03-02T08:12:16.778Z</atom:updated>
            <content:encoded><![CDATA[<p>*Cracks hands* Let’s get this Causal Effect VAE for Health Equity (CEVAE-HE) working!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/331/1*k5v6d6o6horfEWB9ys4GRg.png" /><figcaption>Figure A. Causal Model of the CEVAE-HE</figcaption></figure><p><strong>Quick reminder about our goal:</strong> we are aiming to determine the latent variables <em>Udesc </em>and <em>Ucorr </em>that are respectively the direct hidden confounders of <em>Y</em> and <em>Xcorr </em>(features correlated with the protected attribute <em>A</em> but not caused by it), and of <em>Y</em> and <em>Xdesc </em>(features directly influenced by <em>A</em>). As these variables cannot be possibly collected in the real world (they might not even represent any real single concept) or exactly calculated from the observed features, we are trying to approximate them using a neural network, before using them to generate counterfactuals for our patients, i.e. image patients with the exact same underlying health and their protected attribute changed. This is all the job of our CEVAE-HE.</p><p>Important things to note in our causal model:</p><ul><li><em>Ucorr</em> and <em>Udesc </em>are independent by definition</li><li><em>Udesc </em>is independent from <em>A</em>, as all information about <em>A</em> is inherited by <em>Xdesc </em>from <em>A</em> itself</li><li><em>Ucorr </em>will carry information about <em>A</em>, as <em>Xcorr </em>is correlated with <em>A</em> but not directly caused by it</li></ul><p>Here’s what our architecture for the CEVAE-HE includes:</p><ul><li>Two primary encoder networks that map the observed features to the latent variables <em>Ucorr </em>and <em>Udesc</em></li><li>Tasked with reconstructing the original <em>Xdesc </em>and <em>Xcorr</em>, and predicting the outcome, from the latent variables, the protected attribute <em>A</em> and the independent features <em>Xind</em></li><li>An adversarial discriminator network D that penalises any mutual information between <em>Udesc </em>and <em>(Ucorr, A)</em></li><li>An auxiliary encoders network for patients whose outcome is unknown (i.e. patients for whom we need to predict <em>Y</em>), learning its weight from the other encoders network (abduction network) through distillation</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vcWnwV1Mw90GSMR46BtNtA.jpeg" /><figcaption>Figure B. CEVAE-HE neural network in training</figcaption></figure><p>We give the model a multi-part loss structure to achieve the following objectives:</p><ul><li>Accurate data reconstruction of <em>Xdesc </em>and <em>Xcorr </em>(reconstruction losses)</li><li>Accurate prediction of <em>Y</em> (prediction loss)</li><li>Statistical independence of <em>Udesc </em>and both <em>(Ucorr, A)</em> (Total Correlation loss or TC loss)</li><li>Fair prediction of <em>Y</em> (Fairness loss)</li></ul><p>Meanwhile, the discriminator <em>D</em> is trained to maximize its ability to distinguish samples from the joint distribution<em> P(A, Udesc, Ucorr) </em>from samples drawn from the marginal distributions <em>P(A, Ucorr)*P(Udesc)</em> (simulated with shuffled <em>(Udesc, [A, Ucorr])</em> pairs)</p><p>After implementing this architecture, training the model on the UCI heart disease dataset for 100 epochs with a 80/20/10 train/val/test split, this is where I am:</p><ul><li><strong>Test performance: </strong>0.9200 accuracy, 0.9373 ROC AUC, 0.0855 Brier score, 0.0976 FNR, 0.0588 FPR</li><li><strong>Final training accuracy:</strong> 0.98</li><li>All the loss components plateau after 10ish epochs</li><li>The<strong> canonical correlation analysis</strong> between <em>Ucorr </em>and <em>Udesc </em>reveals a score of 0.689 (we want CCA ~ 0)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/695/1*BdloO1xxMkSVgCVo-PXBeQ.png" /><figcaption>Figure C. Total VAE loss in training and validation</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*zcdcq-KwcVobevY55kgBqg.png" /><figcaption>Figure D. All components of the VAE loss, and the Discriminator loss, in training</figcaption></figure><p>The model is overfitting on the training set, and it is not learning to disentangle <em>Udesc </em>and <em>Ucorr</em>. It is likely over-prioritising the prediction of <em>Y</em> over the independence of the latent variables and finding a shortcut for feature reconstruction. Potential issues off the top of my head:</p><ul><li>A vanishing gradient in the discriminator, which can’t challenge the VAE enough to separate <em>Udesc </em>from <em>(Ucorr, A)</em></li><li>The prediction loss is overpowering the other components of the VAE loss, especially the TC loss</li><li>The dimensions of <em>Udesc </em>and <em>Ucorr </em>allow the model to encode too much info from the features, essentially passing them directly to the decoders to reconstruct them and the prediction</li></ul><h3>Clustering analysis of the latent spaces</h3><p>First, let’s look at what the latent variables look like in training by projecting them in a 2D space (using T-distributed Stochastic Neighbour Embedding) and checking for clustering by protected attribute and by target outcome. We compare the latent variables coming out of the encoders on the abduction path (knowing <em>Y</em>) and the inference path (<strong>not </strong>knowing <em>Y</em>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cf18SWXEW1wHFTHNlsCSdA.png" /><figcaption>Figure E. Clustering analysis of the latent variables in training</figcaption></figure><p>Observations:</p><ul><li>Abducted <em>Udesc </em>is clustered by outcome and mixed by protected attribute,</li><li>Abducted <em>Ucorr </em>is mixed by both dimensions. It might indicate that the model is entirely relying on <em>Udesc </em>on the abduction path to pass on information about <em>Y</em> and making an easy job of its prediction.</li><li>Inferred <em>Udesc </em>looks completely different to the abducted <em>Udesc</em></li></ul><p>Let’s reduce the dimension of <em>Udesc </em>from 4 to 2 and see what happens.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lBMJ6GhG8CQcPaS7ntSztg.png" /><figcaption>Figure F. Clustering analysis of the latent variables after reducing the dimension of Udesc to 2</figcaption></figure><p>More information about the outcome is coming through <em>Ucorr </em>now that <em>Udesc </em>has fewer dimensions, but the two latent variables remain highly correlated (CCA: 0.704). It looks like instead of creating two orthogonal sources of info about the exogenous uncertainty, we’ve created redundant ones. We do want <em>Udesc </em>and <em>Ucorr </em>to carry predictive info about <em>Y</em>, but <strong>different </strong>information.</p><p>The discriminator and the TC loss are clearly not doing their job to enforce the separation.</p><p>Let’s increase the weight of the TC loss (5) and decrease the weight of the predictive loss (1) in the VAE total loss.</p><p>⇒ We get a slightly more realistic accuracy of 0.8533 for our test set (ROC AUC: 0.9323), but our latent variables are still highly entangled with a CCA of 0.691.</p><p>What about a different culprit: vanishing gradients in the discriminator? Let’s check the average gradient norm in the discriminator and our encoders:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*e094uo5olJeMVrm0rYTb0Q.png" /><figcaption>Figure G. Evolution of the average gradient norm across epochs in the discriminator and the encoders</figcaption></figure><p>The gradients in the discriminator and the <em>Udesc </em>encoder are stabilising between 0 and 0.5, but not quite vanishing.</p><p>Hmm. I’m going to leave this aside for now.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*v29LRBvQv32v-LABnJRiPg.png" /><figcaption>Figure H. All components of the VAE loss, and the Discriminator loss, in training (Udesc dim=2, Ucorr dim=4)</figcaption></figure><p>There are a few things bothering me in the profile of the loss components, starting with the prediction loss being very low from the start. But that’s enough investigation for a Sunday. I’ll get back to this tomorrow!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c75d87aaf42" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-23-building-a-causal-effect-vae-for-health-equity-3c75d87aaf42">Week 23: Building a Causal Effect VAE for Health Equity</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 23: Previously on “Lu’s causal discovery adventures”…]]></title>
            <link>https://medium.com/retraining-epoch/week-23-previously-on-lus-causal-discovery-adventures-1e6a0420dcca?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e6a0420dcca</guid>
            <category><![CDATA[social-justice]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Sat, 28 Feb 2026 17:33:31 GMT</pubDate>
            <atom:updated>2026-02-28T17:33:31.683Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dmY7CsXPm9hgIVErqgE--Q.jpeg" /><figcaption>Algonquin park last year, when I still had no clue that causal inference or counterfactuals were even a thing…</figcaption></figure><p>Last time I shared my progress in my project, I was trying to use causal reasoning and inference to mitigate unfair sex-biased pathways in a cardiovascular disease prediction model.</p><p><a href="https://medium.com/retraining-epoch/week-14-mitigating-bias-in-heart-disease-predictive-models-using-causal-modelling-2f29b3f876f7">My first experiment</a> treated the model as a black box and aimed to correct the direct bias between protected attribute and the predicted outcome. Although it led to some reduction in False Negative Rate disparity between sex groups, it still left significant disparities in performance, likely due to the influence of the protected attribute leaking through other features.</p><p>The challenge in targeting bias from an attribute that has relationships with other features as well as the target outcome in the medical domain is to isolate unfair bias (sociological aspect of the protected attribute that influences a patient’s healthcare) from fair bias (biological aspect that is relevant to clinical outcomes), and remove the former while preserving the latter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/335/1*J5iBMwmJ8_ulBmoqaAQuwA.png" /><figcaption>Figure A. <em>Sex, as a protected attribute, carries clinically relevant and irrelevant information. This causal model assumes its direct causal relationship with clinical features X and with the target outcome Y</em></figcaption></figure><p>The basis of my second experiment was therefore to postulate the existence of a latent variable U independent of the protected attribute and predictive of a patient’s innate health. If we could infer the distribution of this variable from the observed dataset, we could then use it to train a fair predictive model. The pipeline involved a Markov Chain Monte Carlo (MCMC) sampler built on the causal model in Figure B. to infer the posterior of U from the observed features, the construction of an augmented fair dataset using age (independent of sex) and samples from U, and the training of a fair model predicting Y from that fair dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/665/1*bRFUrimYnfBa5ZbE0hJtSA.png" /><figcaption>Figure B. Causal model for the MCMC. Training a “fair” on the infered latent variable U and the independent feature age effectively disables all pathways from the protected attribute to the target outcome.</figcaption></figure><p>The results in terms of disparity reduction were good: the FNR disparity between Male and Female patients decreased from -11.79% in the baseline model trained on all features to 2.47% in the fair model (p-value &lt; 0.001). The predictive performance of the model was only moderately degraded (fairness is always a trade-off with accuracy), from 0.863 to 0.788 accuracy and from 0.927 to 0.860 ROC AUC.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/612/1*FQpUitKmQe_YoHCkFAa6Rg.png" /><figcaption>Figure C. Visualisation of the reduction in FNR disparity between sex groups</figcaption></figure><p>Not too bad, but not enough.</p><p>First of all, this pipeline imposes a <strong>big limitation on my target protocol</strong>. My chosen metric to measure fairness at the individual level is the <strong>Individual Equalised Counterfactual Odds</strong> (IECO)¹. Counterfactual Fairness asserts the equality of a model’s prediction for an individual and their counterfactual image, i.e. the exact same individual in all aspects but their protected attribute and the variables causally influenced by the protected attribute should get the exact same prediction. IECO augments that definition of counterfactual fairness by conditioning it on the counterfactual outcome, i.e. an individual and their counterfactual image whose outcome has not changed in the counterfactual world should receive the same prediction. This definition of counterfactual fairness takes into account the fair bias that might exist in the clinical domain from biological signals of a protected attribute.</p><p>Still with me? Okay. So, to be able to measure IECO, I need to generate counterfactuals for individuals, including what their real outcome would be in a counterfactual world. However, the MCMC in my pipeline had to be parameterised only on features and not the target outcome Y, to make sure that no knowledge of Y leaked into U and to enable the abduction of U for patients whose outcome is unknown. So I can’t use it for my counterfactual generation. I could maybe create a second MCMC, modelled on the features and U to infer Y, but there’s a better solution: replacing the MCMC with a <strong>Variational Autoencoder</strong> (VAE)! Instead of using Bayesian inference, we train a neural network to encode features into the latent variable and predict Y which can then be used to generate full counterfactuals.</p><p>Second problem with my second experiment: we have theoretically removed <strong>ALL </strong>influence of the protected attribute on the prediction, including the clinically relevant signal that might be needed to accurately predict the outcome. Uh oh.</p><p>Here comes my third experiment, building on the <strong>Disentangled Causal Effect Variational Encoder</strong> (DCEVAE) proposed by Kim et al. (2021)² and adapted to health equity problems. This new approach proposes that features belong in three buckets:</p><ul><li><strong><em>Xcorr</em>:</strong> Features which might be correlated with the protected attribute but precede its sociological component, such as biological sex markers or baseline clinical measurements existing prior to the perception of the sociological attribute.</li><li><strong><em>Xdesc</em>: </strong>Descendants of the sociological component of the protected attribute that are susceptible to systemic bias, such as clinician interpretations of results or symptoms reported by patients.</li><li><strong><em>Xind</em>:</strong> Features fully independent and uncorrelated with the protected attribute (e.g. age with sex)</li></ul><p>We then postulate the existence of <strong>two </strong>latent variables representing the unobserved confounding factors in the dataset: <em>Udesc </em>and <em>Ucorr</em>, direct parents of their respective feature bucket. This causal model (Figure D) allows us to differentiate between legitimate biological differences and unfair sociological bias.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/417/1*eZELFbElYxdYWl4MqSgRGA.png" /><figcaption>Figure D. Causal model for CCEVEA-HE and the disentanglement of a protected attribute’s influence in a dataset</figcaption></figure><p>Sounds really good to me. Now my job is to build a VAE with the job of inferring <em>Udesc </em>and <em>Ucorr</em> from the features, before accurately reconstructing the features and the outcome from the latent variables.</p><p>Before I dive into the architecture of the network, I need exercise. I’m off to the gym! See you a bit later.</p><p>Stay tuned for the <strong>design of the CEVAE-HE</strong> (Causal Effect VAE for Health Equity) as well as my <strong>first struggles with its implementation</strong>… 🙄</p><ol><li>Pfohl, S.R., Duan, T., Ding, D.Y., Shah, N.H., 2019. Counterfactual Reasoning for Fair Clinical Risk Prediction, in: Proceedings of the 4th Machine Learning for Healthcare Conference. Presented at the Machine Learning for Healthcare Conference, PMLR, pp. 325–358. <a href="https://doi.org/10.1609/aaai.v35i9.16990">https://doi.org/10.1609/aaai.v35i9.16990</a></li><li>Kim, H., Shin, S., Jang, J., Song, K., Joo, W., Kang, W., Moon, I.-C., 2021. Counterfactual Fairness with Disentangled Causal Effect Variational Autoencoder. Proceedings of the AAAI Conference on Artificial Intelligence 35, 8128–8136. <a href="https://doi.org/10.1609/aaai.v35i9.16990">https://doi.org/10.1609/aaai.v35i9.16990</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e6a0420dcca" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-23-previously-on-lus-causal-discovery-adventures-1e6a0420dcca">Week 23: Previously on “Lu’s causal discovery adventures”…</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Found them! They’re right behind Dave, Rohan and Chen]]></title>
            <link>https://medium.com/retraining-epoch/found-them-theyre-right-behind-dave-rohan-and-chen-5ed15fe168c4?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ed15fe168c4</guid>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[diversity]]></category>
            <category><![CDATA[women-in-tech]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Mon, 23 Feb 2026 20:54:20 GMT</pubDate>
            <atom:updated>2026-02-23T20:54:20.177Z</atom:updated>
            <content:encoded><![CDATA[<p>I made the mistake to open LinkedIn after dinner…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*kULbEejsbq8gR5nUwPx4zA.png" /></figure><p>… and this post made me skip my dessert and open my laptop again to write down my thoughts 😤</p><p>Women have heard this self-righteous lament in Tech for twenty years, and now it’s AI’s turn. To the event organisers and recruiters asking <em>“Where are the women?”</em>: <strong>We’re right here</strong>, buddies. We’ve always been here.</p><p>Aside from the handful of researchers who have actually been living by the light of PyTorch and sliding down the gradient descent for a decade, the tech industry’s attention is currently dominated by guys who are simply very good at shouting about their “AI expertise”.</p><p>But the reality of the AI boom is that <strong>everyone is new to this.</strong> The self-proclaimed “experts” bragging about agentic workflows or AI-driven business transformation are learning as they go. There’s nothing inherently wrong with that, but let’s be honest: what makes them visible is their ability to “prompt-engineer” their way into the spotlight rather than superior technical depth. I would bet that men, in aggregate, are much more skilled at “faking it until they make it” <em>(my LinkedIn feed certainly supports this hypothesis)</em>, so if you’re only looking for the loudest voices, you’re just hiring for confidence, not competence or potential.</p><p>If you want diversity in AI, don’t look for self-proclaimed gurus. Look for mission-oriented contributors and leaders who have built careers on resilience, adaptability and high-stakes problem-solving. Look for people with diverse domains and backgrounds who are deeply curious about the evolution of AI and are already thinking three steps ahead about implications, real value and safety.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ujZc31F_Sb3J4UGKB9NZrw.jpeg" /><figcaption>Look! She’s right there!</figcaption></figure><p>Let’s talk about me for five minutes 💁‍♀️ After all, that’s what seems to work for my male counterparts:</p><p>I’ve been sitting on an application for a fellowship at Anthropic for days. It would be my dream role. At the frontier of safety research, alongside amazing mentors and teammates, the opportunity to work on alignment and interpretability, in a company whose approach to AI safety stands out in the right way. It perfectly aligns with my career goals, my values, and I could leverage my diverse background: my technical skills, my current MRes research, my experience as a product/delivery manager, etc…</p><p>And yet, I managed to convince myself that my 13 years of experience meant nothing because my pivot into AI research was “too new”. I told myself: <em>I’m not technical enough. I’m not enough of a nerd. I don’t speak the same language as these guys. I don’t think like them.</em></p><p>It took a classmate’s encouragement to realise I was gaslighting myself. I’m sure many women fall into the same gaslighting trap, from themselves and from the industry.</p><p>I should know my own worth.</p><p>I’m a high-speed learner. Two months ago I knew nothing about causal theory; Last week I was playing with tensors for the first time; And this week I’m training a Causal Effect Variational Encoder to conduct surgery on hidden confounders in clinical datasets. If anything, my “newness” is proof of high-velocity upskillability, independent thought and technical curiosity. In a field where the state-of-the-art changes every Monday, this should give me a definite advantage.</p><p>I’m also a bridge-builder. Between people and disciplines, between ways of working, between ideas and concepts. I’m the “right arm” you want when you have tough and messy problems to solve and you need to navigate uncertainty at speed.</p><p>I know I can bring enormous value to AI safety and governance, even on the technical side. But I fear I won’t get the chance to get through the door because I don’t look like the “AI Guy” on your panel.</p><p>Brag over. The ROI on my potential is massive. <strong>So is the potential of many, many women in AI and adjacent industries.</strong> Instead of asking “Where are the women?”, look at the proxies you’re using in your definition of “expertise” or “excellence”: who do they accidentally favour and who do they exclude?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ed15fe168c4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/found-them-theyre-right-behind-dave-rohan-and-chen-5ed15fe168c4">Found them! They’re right behind Dave, Rohan and Chen</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 18: Riding the causal wave… where to?]]></title>
            <link>https://medium.com/retraining-epoch/week-18-riding-the-causal-wave-where-to-2cd2c9c912bc?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/2cd2c9c912bc</guid>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[social-justice]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Fri, 23 Jan 2026 09:58:44 GMT</pubDate>
            <atom:updated>2026-01-23T09:58:44.460Z</atom:updated>
            <content:encoded><![CDATA[<p>Exams are a week behind me, and I am deep (deep, deep, deep) in my project.</p><p>I hit a wall of doubt in early January after meeting my supervisors. Their feedback always throws me into a spin: <em>Am I focused on the right question? Is this clinically valuable? Can I even solve this? Am I doing too much? Or not enough?</em> AAAAAAH.</p><p>So, I’ve refined my focus. Again. I know, I keep doing that. It probably won’t be the last time, especially with that little voice in my head whispering, “This won’t work, and it’s not even useful anyway.” For now, I’m pushing through to see how these ideas actually take shape.</p><p>For the past week, I’ve been in a flow state so intense I’ve had to remind myself to perform basic human functions like eating, walking and sleeping at a reasonable hour. I have a lot of other plates spinning: rewriting my stand-up set (yes, I’m a new-born comedian when I’m not training ML models), creating Insta reels (for my comedian alter ego), writing for The AI Library, and trying to kickstart a side hustle to alleviate my liquidity issues (although, as long as I don’t start it, I can hope that it’ll work). But I couldn’t pull myself away from my code for a second. I had to slap myself this morning to stay out of Google Colab and write this post instead.</p><p>Maybe I should just ride the wave while it’s strong. Who cares if the newsletter is late? But I can’t help feeling guilty about dropping balls, even the ones that don’t really matter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sv7jQFzMd6ogJkKE9DwzSg.png" /><figcaption>My Flow Toolkit</figcaption></figure><h3>So, what is my new experiment?</h3><p>I’m currently getting to grips with <strong>counterfactual fairness</strong>.</p><p>You could try to make AI models fair by simply making them “unaware” of a person’s sex or race, but the bias still “leaks” through other variables that act as proxy of the protected attribute.</p><p>My experiment uses causal modelling to remove unfair sex-biased pathways from an ML model predicting cardiovascular disease. In the clinical world, “Sex” is complicated: it’s both biological (affecting biomarkers in a clinically-relevant way) and sociological (affecting how a doctor interprets your pain or how you report symptoms). My goal is to protect the model from <strong>sociological bias</strong> while keeping the legitimate clinical signals.</p><p>I’m using the observed clinical features to infer a “fair” unobserved variable that you could call the <strong>Age- and Sex-independent Physiological Integrity</strong>. Essentially, I’m extracting a version of the patient’s health that exists independently of their sex and age and directly impacts their clinical outcome. I then train the model using this “clean” variable instead of the potentially contaminated raw features.</p><p>To verify that I’ve created a fair model, I then ask: <em>“Would the heart disease prediction for this specific individual stay the same if we lived in a world where their sex was different?”</em> If the answer is yes, we’ve effectively removed unfair sex bias from the model and created individual fairness.</p><h3>Flowing… where?</h3><p>So, I’m still having the time of my life studying. But where is this wave taking me?</p><p>Since Christmas, I’ve been feeling three distinct pressures:</p><ol><li>Making the most of this MRes experience</li><li>Turning this degree into a meaningful career pivot</li><li>Returning to a decent salary as soon as possible!</li></ol><p>I know where I want to go: <strong>AI safety and governance</strong>. I also know where I <em>don’t</em> want to go: back to “tech delivery” and all that jazz. But the imposter syndrome is real. Even with 13 years of experience, I struggle to believe I can find a role in this industry that doesn’t require a PhD.</p><p>I’ve looked at doctoral options, but I can’t survive in the UK for three more years on a £30k stipend (at best). If I stay here, it has to be for financial stability. However, I’d consider being a “cash-poor” student in Canada. For some reason, a studio flat with shared toilets feels more like an adventure and less like a step backward if it’s in a new country that I love.</p><p>In the meantime, I’ve applied to several summer fellowship programmes, as well as AI governance training to widen my network and get some “real world” safety experience. I’d even consider pausing the MRes for the summer if the right internship came along.</p><p>Gosh, this all makes me feel like I’m twenty again. I still can’t tell if I like it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2cd2c9c912bc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-18-riding-the-causal-wave-where-to-2cd2c9c912bc">Week 18: Riding the causal wave… where to?</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 14: Mitigating bias in heart disease predictive models using causal modelling]]></title>
            <link>https://medium.com/retraining-epoch/week-14-mitigating-bias-in-heart-disease-predictive-models-using-causal-modelling-2f29b3f876f7?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/2f29b3f876f7</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[social-justice]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Tue, 23 Dec 2025 15:48:26 GMT</pubDate>
            <atom:updated>2025-12-23T15:48:26.491Z</atom:updated>
            <content:encoded><![CDATA[<p>What is more enjoyable in the days before Christmas than concluding an experiment on causal detection and mitigation of sex bias in a heart disease prediction model?</p><p>Before I conduct a larger-scale analysis of racial bias mitigation in Atrial Fibrillation (AF) prediction models within the UK Biobank, I need to run a few experiments to de-risk my methodology.</p><p>I ran the pilot experiment below on a smaller open-access heart disease dataset to see if causal modelling can be used to identify the “unfair pathways” driving biased model predictions.</p><h3>Method to validate</h3><p>Hui and Lau¹ introduced an intuitive method to detect and mitigate bias on a sensitive attribute — such as gender, age or race — in a black-box predictive model. By modelling and inferring the causal relationship between the sensitive attribute, the predicted outcome and the ground-truth outcome, they quantify the strength of the bias introduced by the model and correct its predictions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MB25Nlt5u2DWUzgMq8sbow.jpeg" /><figcaption><strong>Fig A. </strong>Conceptual model of the causal modelling approach to detect unfair pathways</figcaption></figure><p>The conceptual causal model above can be described with the following structural equations, where <em>y </em>is the actual outcome, <em>ŷ</em> the predicted outcome, a the protected attribute, ε error terms with a zero mean, and β and γ causal parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/257/1*5CNLuKcEh2af5I-qCEb9DA.png" /></figure><p>Hui and Lau use the <em>lavaan </em>R library to train a causal model on the protected attribute, the ground-truth outcome and predictions of the black-box model, resulting in an approximation of the causal parameters of the model, and specifically β₂.</p><p>In an unbiased model, all effects of the protected attribute are mediated through the actual outcome. This is equivalent to setting β₂ to zero in the above equations to get the corrected prediction <em>ỹ</em>. Therefore they can deduce the corrected prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/352/1*PmdoCzphb2VpA8xzjbTqpA.png" /></figure><p>Using this method on a Logistic Regression model trained on the UCI Adult dataset, Hui and Lau demonstrate its ability to detect a statistically significant causal relationship between the sensitive attribute (sex) and the predicted outcome (whether a person’s earnings exceed 50K). By applying the correction to the model’s prediction, they show a reduced difference in True Positive Rate between sex groups and an increase to the overall accuracy of the model.</p><h3>The experiment</h3><p>Using the study of statistical gender bias in cardiology algorithms conducted by Straw et al.² as a reference, we apply the causal modelling bias mitigation method to an ML model trained on the Heart Disease (CAD) dataset³ , and we compare the bias quantified through statistical fairness metrics to the results of the causal analysis.</p><h3>Hypotheses</h3><ol><li>Causal modelling (as per Hui and Lau’s method) detects a direct causal path from the protected attribute (sex) to the prediction outcome</li><li>The mitigation of this causal path reduces the observed sex-based performance disparity (False Negative Rate difference) in the cardiac disease prediction model.</li></ol><h3>Dataset</h3><p>The Heart Disease dataset combines five curated heart datasets into a total of 1,190 records including demographics (age, sex), clinical measurements (e.g. resting blood pressure, fasting blood sugar), and the binary target class capturing a Coronary Artery Disease diagnosis.</p><h3>Protocol</h3><h4>Data pre-processing</h4><p>We remove duplicates and missing values to obtain a cleaned dataset of 746 records.</p><h4>Baseline model training</h4><p>We use a bootstrapping approach with 100 runs and a random 70% training / 30% test dataset split stratified by the target outcome.</p><p>For each test run, we:</p><ul><li>Train and finetune a Random Forest classifier.</li><li>Measure the model’s overall accuracy, AUC ROC, and error rates, i.e. False Negative Rate (FNR) and False Positive Rate (FPR), on the test data, as well as its performance stratified by sex along the same metrics.</li><li>Measure the statistical significance of the difference in error rates between sex groups, using a Barnard’s test</li><li>Train a causal model using the <em>semopy </em>package, and extract the causal parameter β₂ between sex and prediction, and its P-value.</li><li>Calculate the corrected prediction using β₂, as per Hui and Lau, if the causal relationship is statistically significant (P-value ≤ 0.05)</li><li>Measure the corrected performance of the model, overall and stratified by sex</li></ul><h3>Analysis &amp; results</h3><h4>Disparity in the baseline model</h4><p>Analysis of the baseline model’s performance and error rates across all runs show similar disparities as the ones observed by Straw et al.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/633/1*cBchM3TQphX88T6HGrICuA.png" /><figcaption><strong>Table 1. </strong>Compared global performance of the baseline model in both experiments</figcaption></figure><p>Both experiments reveal statistically significant differences in error rates between sex groups (measured as Female rate — Male rate), which is also reflected in a disparity in ROC AUC. We measure a False Negative Rate (FNR) disparity of -12.18 % (P-value &lt;0.05), indicating that females are more greatly affected by the model’s failure to predict positive cases. A significant False Positive Rate (FPR) disparity is also confirmed (+5.15%, P-value &lt;0.05), with males being disproportionately over-flagged by the model. Despite these differences, the model’s accuracy remains consistent across groups (overall accuracy of 0.859, 95% CI: 0.855–0.863).</p><h4>Causal analysis</h4><p>The causal model detects a statistically significant causal relationship between the sex attribute and the model’s prediction in 89% of the runs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*amCbDBgFOG8JdQY9iPBzmg.jpeg" /><figcaption><strong>Fig B.</strong> At simulation level, how often does the causal model flag an unfair pathway exactly when the error rates show an inequitable predictive outcome?</figcaption></figure><p>We analyse the association in each training run between the detection of a sex-prediction causal path and the statistical significance of the error rate differences between groups. We find that while the disparity in error rates is significant in aggregate across all runs, it rarely is at run level, as shown in Fig B.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/544/1*raRiv0QF-2Hv1o8LCqhfrA.png" /><figcaption><strong>Fig C. </strong>Correlation heatmap between disparity metrics and causal parameter across all 100 simulations</figcaption></figure><p>To better understand the association between the causal detection of bias and its statistical measurement across runs, we look at the correlation between disparity metrics, showing no strong relationship between the causal parameter and statistical metrics.</p><h4>Bias mitigation</h4><p>We use a paired t-test to analyse the difference in False Negative Rate before and after correcting the model’s prediction with the causal model mitigation. The correction results in an average 53.65% relative disparity reduction (or absolute 2.12% reduction in error rate difference) over the 100 runs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/521/1*H0gvtLM_7vuktN8zzxgbUg.png" /><figcaption><strong>Table 2.</strong> Disparity in False Negative Rate before and after causal correction</figcaption></figure><p>The overall accuracy of the model remains unaffected by the correction. The FPR is slightly lower as a result, while the FNR is higher than the baseline, demonstrating that fairness improvements at group level can result in compromises in aggregate performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/388/1*VbR53tVC6uPkIPAUIlX-yQ.png" /><figcaption><strong>Table 3.</strong> Model performance before and after causal correction</figcaption></figure><h3>Conclusion</h3><p>This experiment successfully validated that causal modelling can detect and partially mitigate bias on a single protected attribute in a cardiovascular prediction model.</p><p>While the global performance audit confirmed the presence of significant disparities, specifically a 12.18% higher False Negative Rate for females, simulation-level frequentist tests (Barnard’s) often failed to flag these disparities, likely due to limited statistical power in small test data subgroups. In contrast, the causal model identified a statistically significant unfair pathway between sex feature and prediction in 89% of simulations.</p><p>The observed lack of correlation between the detection of an unfair causal pathway and measured disparities in error rates suggests that causal and statistical methods are sensitive to different layers of algorithmic bias. While the causal model measures the internal unfair pathway created by the model between the protected attribute and the prediction, FNR and FPR measure the consequent statistical patterns. Moreover, the error rates are measured from binary classifications, and therefore small biased shifts in continuous prediction estimates only manifest as statistical disparities if the classification threshold is crossed.</p><p>Although the causal mitigation strategy yielded a statistically significant 2.12% absolute reduction in FNR disparity, the persistence of a 10.07% gap suggests that post-hoc causal corrections, such as the studied causal modelling approach, should be part of a broader “fairness-by-design” framework rather than a sufficient standalone solution.</p><p>We did not reproduce other mitigation methods from Straw et al. to compare their impact or explore the potential combination with causal modelling. This remains a valuable area for further experiments and the construction of a bias mitigation pipeline for Atrial Fibrillation prediction.</p><h3>References</h3><p>The source code for this experiment can be found on GitHub: <a href="https://github.com/LuSyo/Health-ML-Bias-Experiments/blob/master/notebooks/01_CAD_bias_causal_modelling.ipynb">https://github.com/LuSyo/Health-ML-Bias-Experiments/blob/master/notebooks/01_CAD_bias_causal_modelling.ipynb</a></p><ol><li>W. Hui and W. K. Lau, “Detecting and Mitigating Algorithmic Bias in Binary Classification using Causal Modeling,” <em>2024 4th International Conference on Computer Communication and Information Systems (CCCIS)</em>, Phuket, Thailand, 2024, pp. 47–51, doi: <a href="https://doi.org/10.1109/CCCIS63483.2024.00016">10.1109/CCCIS63483.2024.00016</a></li><li>I. Straw et al., “Sex-Based Performance Disparities in Machine Learning Algorithms for Cardiac Disease Prediction: Exploratory Study”, <em>J Med Internet Res 2024</em>, doi: <a href="https://doi.org/10.2196/46936">10.2196/46936</a></li><li>M. Siddhartha, <em>“Heart Disease Dataset (Comprehensive)”</em>, <em>IEEE Dataport</em>, November 5, 2020, doi:<a href="https://dx.doi.org/10.21227/dz4t-cm36">10.21227/dz4t-cm36</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2f29b3f876f7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-14-mitigating-bias-in-heart-disease-predictive-models-using-causal-modelling-2f29b3f876f7">Week 14: Mitigating bias in heart disease predictive models using causal modelling</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 13: my first experiment with ML bias mitigation]]></title>
            <link>https://medium.com/retraining-epoch/week-13-my-first-experiment-with-ml-bias-mitigation-e7a2ad79f0f9?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/e7a2ad79f0f9</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[social-justice]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Wed, 17 Dec 2025 10:37:05 GMT</pubDate>
            <atom:updated>2025-12-23T15:52:58.900Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*po_r5Q9WdvrKvOJdLJkWmw.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@bradyn?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Bradyn Trollip</a> on <a href="https://unsplash.com/photos/person-holding-white-and-blue-plastic-blocks-pxVOztBa6mY?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>First term over! I can fully focus on revisions for the exams in Jan, and on my dissertation project (ok, and a little bit of Christmas…)</p><p>I’ve made a lot of progress on my project since kick off in November. I’ve completed the first part of my literature review to map state-of-the-art models trained on EHR data for Atrial Fibrillation prediction. I screened 76 research papers, excluded 45 that were out of scope for my focus (mainly because the models were built on ECG data) and reviewed 31 (including 10 systematic reviews). Out of these, I have identified at least three models that I could train on the UK Biobank dataset and use as uncorrected baselines to validate a racial bias assessment and mitigation framework.</p><p>While a few of them recognised the mono-ethnicity or racial imbalance of their dataset, none of the papers I reviewed included a stratified bias audit of their models’ predictive performance. This confirmed the research gap that I aim to address.</p><p>After completing this, I dove into the second part of my literature review: fairness assessment and bias mitigation in machine learning models. So far, I’ve screened 57 studies and reviewed the full text of 16 of them. What am I looking for?</p><ul><li>Insights on algorithmic fairness metrics, their respective purpose, limitations, and implications when optimised for in healthcare</li><li>Evidence of the applicability, effectiveness and caveats of different bias assessment and mitigation methods on ML models, in healthcare or not; And more specifically, methods leveraging causal modelling.</li></ul><p>I’ve already set aside a few studies introducing techniques that I would like to test, including a bias detection and mitigation method demonstrated by Hui and Lau (2024) that relies on causal modelling¹ and a framework leveraging counterfactuals².</p><p><em>“Eh… what’s a counterfactual?”</em> you ask.</p><p>It’s one of the many concepts I have been getting <em>very </em>intimate with this past month. I have been eating and sleeping statistical disparities and causal fairness! I’m still wrapping my head around most of the theory, but it is slowly making more and more sense.</p><p>Here’s a <strong>causal model</strong> describing why I might miss my train in the morning:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tHUJNLGN_UEvQ1IcHc5VgQ.jpeg" /></figure><p>The arrows are directed <strong>causal paths or relationships</strong> between the variables in the model. My wake up time affects my energy and stress level, which in turn affect my walking pace and whether the train is missed.</p><p>A <strong>counterfactual </strong>describes what would have happened if a specific past circumstance (the fact) had been different (the counter-fact). Consider the situation where I wasted time on the way and missed my train. A counterfactual is the question: <strong><em>“Would I have missed my train had I not wasted time on the way, everything else being equal?”</em></strong></p><p>Reading studies where these concepts are put in action really helps me grasp their value. And my own experiments will undoubtedly unlock some <em>“oh my god I get it now”</em> moments. Speaking of which…</p><h3>First experiments</h3><p>Time to get our hands dirty!</p><p>Before I start messing with the UK Biobank dataset, I want to test methods on open datasets and put together a rough pipeline to train baseline models, identify and quantify bias on protected attributes, and mitigate that bias.</p><p>My first experiment tests the causal modelling method introduced by Hui and Lau to detect and correct bias on a protected attribute.</p><p>Straw et al³ previously assessed the gender bias in two models trained on cardiac disease datasets. I used one of the models (trained on the <a href="https://ieee-dataport.org/open-access/heart-disease-dataset-comprehensive">Heart Disease (CAD) dataset</a>) as a baseline for my experiment, replicating its training and measuring its statistical bias on gender.</p><p>This reference allows me to test my hypotheses:</p><ol><li>Causal modelling (as per Hui and Lau’s method) will detect a statistically significant direct causal path from the protected attribute (sex) to the prediction outcome, in alignment with the statistical bias detected by Straw et al.</li><li>Mitigating this direct causal path will reduce the observed sex-based performance disparity (and specifically disparity in False Negative Rate between gender groups) in the cardiac disease prediction model</li></ol><p>I am working with Python 3, using <em>scikit-learn</em> to train prediction models and <em>semopy </em>for causal modelling.</p><p><strong>If you’re curious, have a look at the </strong><a href="https://github.com/LuSyo/Health-ML-Bias-Experiments/blob/master/notebooks/01_CAD_bias_causal_modelling.ipynb"><strong>experiment Jupyter notebook</strong></a><strong>.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/821/1*9uBLDfzYtoyArVFuNQ2mHA.png" /></figure><h4>Results</h4><p>The experiment involved 100 bootstrapped runs of a Random Forest classifier trained on the Heart Disease dataset, with the causal mitigation applied conditionally when a significant causal path was detected between the protected attribute (sex) and the model’s prediction.</p><p>I ran my procedure in Google Colab, on CPU only. It took about an hour to complete (~35–45s per training run).</p><p>The baseline model was successfully reproduced. As in the original study, the model exhibits significant disparities in error rates, confirming the need for an intervention:</p><ul><li><strong>Equal Opportunity failure:</strong> the False Negative Rate (FNR) disparity was <strong>-11.97 % </strong>(FNR_male &lt; FNR_female, P-value &lt;0.05), indicating that females were more greatly affected by the model’s failure to predict positive cases.</li><li><strong>Predictive Parity failure:</strong> a significant False Positive Rate (FPR) disparity was also confirmed (+4.9%, P-value &lt;0.05), indicating males were disproportionately over-flagged.</li></ul><p>All good so far.</p><p>But then, the Hui and Lau mitigation protocol was found to be inconsistently reliable and overall ineffective in mitigation bias for the baseline model:</p><ul><li><strong>Low reliability:</strong> the causal model detected bias in only 44% of the runs</li><li><strong>No correlation: </strong>the detection of the <em>sex → prediction</em> causal path was found to be statistically independent from the observed error rates disparities</li><li><strong>Mitigation failed:</strong> a paired t-test revealed that the causal modelling mitigation method failed to reduce the FNR disparity (P ~ 1.0)</li></ul><h4>Next steps</h4><p>I still need to draw conclusions from these results, by first re-checking what assumptions made by Hui and Lau might not be verified in this experiment and would render the method ineffective.</p><p>Although I was hoping to validate my hypotheses, these findings will help me choose the appropriate methods for my bias detection and mitigation pipeline.</p><p>Next I plan to test the <em>FlipTest </em>method introduced by Black and Yeom⁴, perhaps on the same baseline model if it’s appropriate.</p><h4>Errata, 23 Dec 2025</h4><p>I discovered important flaws in my protocol since I wrote this post that led to negative results for the causal bias detection and mitigation. I have since corrected and re-run the experiment. To read more about the results, head over to my <a href="https://medium.com/retraining-epoch/week-14-mitigating-bias-in-heart-disease-predictive-models-using-causal-modelling-2f29b3f876f7">Week 14 post</a>.</p><p>¹ <em>Detecting and Mitigating Algorithmic Bias in Binary Classification using Causal Modeling</em>, Hui and Lau, 2024, <a href="https://doi.org/10.1109/CCCIS63483.2024.00016">https://doi.org/10.1109/CCCIS63483.2024.00016</a></p><p>² <em>Metrics and Algorithms for Identifying and Mitigating Bias in AI Design: A Counterfactual Fairness Approach</em>, Moon and Ahn, 2025, <a href="https://doi.org/10.1109/ACCESS.2025.3556082">https://doi.org/10.1109/ACCESS.2025.3556082</a></p><p>³ Sex-Based Performance Disparities in Machine Learning Algorithms for Cardiac Disease Prediction: Exploratory Study, Straw et al., 2024, <a href="https://doi.org/10.2196/46936">https://doi.org/10.2196/46936</a></p><p>⁴ <em>FlipTest: Fairness Testing via Optimal Transport</em>, Black and Yeom, 2020, <a href="https://doi.org/10.1145/3351095.3372845">https://doi.org/10.1145/3351095.3372845</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e7a2ad79f0f9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/week-13-my-first-experiment-with-ml-bias-mitigation-e7a2ad79f0f9">Week 13: my first experiment with ML bias mitigation</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 9: Project scope nailed down, brain activated]]></title>
            <link>https://medium.com/retraining-epoch/scope-refinement-0c39c7d715e1?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c39c7d715e1</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 21:16:42 GMT</pubDate>
            <atom:updated>2025-11-19T07:01:32.026Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WyNUDkT_BIhhdq3UP3FJLw.jpeg" /><figcaption>Photo by Possessed Photography on Unsplash</figcaption></figure><p>In my first post, I was on the starting line, about to kick-off my MRes dissertation project on an exciting topic: “Developing equitable patient-level prediction models in chronic cardiovascular disease management”.</p><p>Since then, I’ve met with my supervisor and co-supervising team. I shared with them my early understanding of the project and my plan.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tXGCqcFQUP-TWZ4kgSzh-g.jpeg" /><figcaption>My original project scope</figcaption></figure><p>The feedback was unanimous and incredibly useful: <strong>Scope it down!</strong></p><p>They guided me to see that the real value would come from going deep, not wide. Pick <strong>one </strong>disease, choose a specific set of protected characteristics, and focus on the bias assessment and mitigation. That way I’ll have better chances to produce actionable insights.</p><p>They also confirmed that focusing on <strong>causal analysis</strong> to understand why a model is showing disparities would be a novel approach in this particular domain in AI-enabled healthcare.</p><h3>Finding the right focus</h3><p>I spent last week reframing everything. And I have to admit I felt a bit lost at first.</p><p>After a quick exploratory search of the literature, I settled on <strong>Atrial Fibrillation (AF)</strong>, a common cardiovascular condition that is a critical risk factor for other heart diseases but is often missed because it can be intermittent. This makes it a perfect subject for predictive ML models to find at-risk patients before they present with complications.</p><p>I also decided to focus specifically on <strong>ethnic bias</strong>. It’s a critical area, as models developed on non-diverse datasets and without bias assessment and mitigation risk amplifying health disparities. And from my initial digging, this seems to be a very unexplored area for existing AF risk ML models.</p><p>Finally, to make the causal analysis “simpler”, I’m narrowing the data modality. I’ll be leaving models trained on electrocardiogram data aside for now and focusing entirely on models trained with tabular <strong>Electronic Health Record (EHR) data</strong>: lab results, baseline measurements, biomarkers, lifestyle factors, etc…</p><p>So that’s it, my updated project title is a bit of a mouthful, but I’m feeling good about it:</p><blockquote>“Causal analysis of ethnic bias in EHR-based machine learning models for Atrial Fibrillation prediction”</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8tFNsc7efdkNA_9UPv_jQg.jpeg" /><figcaption>My updated high-level plan</figcaption></figure><p>If you’re curious about the details of my plan, here’s my <a href="https://drive.google.com/file/d/1h9J8RUqvgTgr_NEmLDg2zh0SJjZ40vUO/view?usp=sharing"><strong>preliminary project brief</strong></a><strong>.</strong></p><h3>Hit the ground running</h3><p>I’m already deep in the first objective: the literature review. I’m screening papers and mapping existing ML models built to predict AF risk. I’ve already found a dozen models, including a couple trained on the UK Biobank, which could be perfect for replicating a baseline.</p><p>I am <strong>loving</strong> the process. Reading, analysing and synthesising research. This is the exact opposite of the numbness I felt in my tech job. My brain is working, it’s fully activated. It’s so satisfying to understand new complex concepts (even when some of it still goes over my head), to feel curious and critical.</p><h3>Good news</h3><p>Yesterday, I received my “approval on principle” for access to the UK Biobank! This is <em>months</em> earlier than I had factored into my timeline.</p><p>Because I was only expecting to get access in February, based on the average 15 weeks processing on their website, I had added a prototyping phase in my plan to de-risk the timeline. I was planning to build and test my bias assessment and mitigation pipeline on an open-access dataset first, while waiting for UKB access.</p><p>But that phase is still valid, even with early access to the UKB. On top of the flat access fee to the dataset, the UK Biobank Research Analysis Platform where you have to perform your analysis of the data has a cost: you pay for compute, storing the data you create, and egress. So I’m not going to go and play there straight away like an idiot. I don’t want to waste resources on “stab in the dark” experiments. I’ll get a good grasp on my approach first.</p><h3>Next steps</h3><p>In parallel to screening papers for ML models predicting AF risk, I’m starting to look into fairness metrics this week.</p><p>And I’m aiming to start playing with some open datasets in December. I can’t wait!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c39c7d715e1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/scope-refinement-0c39c7d715e1">Week 9: Project scope nailed down, brain activated</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Human disconnection: How AI fills the void and risks deepening it]]></title>
            <link>https://medium.com/retraining-epoch/human-disconnection-how-ai-fills-the-void-and-risks-deepening-it-0c045b45125a?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c045b45125a</guid>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[loneliness]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Mon, 10 Nov 2025 08:01:02 GMT</pubDate>
            <atom:updated>2025-11-10T08:01:02.508Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dIg9R4lzMcMvQxh2HBwDtA.jpeg" /><figcaption>Photo by Sergey Zolkin on Unsplash</figcaption></figure><h3>In short</h3><p>The rise in loneliness in the western world, particularly among young adults, has created a vulnerability ripe for AI companions to exploit. Many argue that AI offers a solution to isolation, and research does show that it provides immediate relief. But it also suggests that this easily accessible “fix” is temporary and could actually undermine a person’s ability to engage in the more challenging, but ultimately more rewarding and precious, human relationships necessary for their long-term well-being and a healthier society.</p><h3>Quick breakdown</h3><p>The rate of loneliness among young adults in the US has increased every year between 1976 and 2019. In the UK in 2023/24, 9–10% of adults aged 16 to 34 report that they felt lonely often or always ¹. Meanwhile, the decline in real-world social engagement is stark: the US Surgeon General reported in 2023 that the amount of time people spent engaging with friends socially in-person decreased from 60 minutes per day in 2003 to 20 minutes per day in 2020 ². That’s a decrease of 20 hours per month spent engaging with friends.</p><p>We’re at a crossroads. AI is rising at the perfect time to either help us get out of this funk, or plunge us deeper into it.</p><h4>The problem with the AI quick fix</h4><p>We don’t need to look further in the future. AI companions and chatbots are already advanced enough that they offer a seductive and low-friction alternative to the complexity of human connection. Always agreeable, empathetic, eager to please, praise and compliment, and most importantly, always there. AI will never leave you or be mad at you.</p><p>Some studies suggest that interacting with AI companions can alleviate feelings of loneliness. Their effectiveness is largely explained by the psychological construct of “feeling heard,” the perception that the communication is received with empathy and respect ³.</p><p>But is this genuine connection or just the paracetamol of loneliness?</p><p>One study found that university students who reported greater social support from AI simultaneously felt less supported by other people ⁴. And individuals who already lack human companionship are considered more vulnerable to developing an AI relationship in place of human ones ⁴. The long-term reliance on chatbots could create dependencies and degrade social skills, eventually exacerbating isolation ³. Even replacing human staff with AI in tasks like customer service, for the sake of efficiency and lower friction, could remove a form of socialising that counters loneliness and allows people to practice social skills.</p><h4>The AI Labs’ response to the crisis of connection</h4><p>The major AI labs are clearly aware of the emotional terrain they’re operating on, although they like to pretend they’re not. And still, their response is more akin to damage control after the fact to barely disguise their embrace of the lucrative companion space.</p><p>For instance, after a lawsuit was filed against OpenAI alleging that ChatGPT encouraged a teenager to take his own life, the company swiftly announced new safety updates, including mental health guardrails and age controls ⁵⁻⁶.</p><p>But Sam Altman then posted a <a href="https://x.com/sama/status/1978129344598827128">mind-blowing statement on X</a>:</p><blockquote>“Now that we have been able to mitigate the serious mental health issues and have new tools, we are going to be able to safely relax the restrictions in most cases.”</blockquote><p><em>Oh really?</em> You have now fully fixed the issue and you’re ready to move on and loosen the leash on your models?</p><p>The company’s work on mental health guardrails itself highlights how signs of distress are difficult to define and teach to a model because they are rare and show up in various forms ⁷. They claim to have reduced unwanted responses to challenging mental health situations in GPT-5 by 65%. Yet, they hammer on the same old Tech product trick we’ve all served to our stakeholders to shut them up: “Don’t worry, we’ll take an iterative approach to solve these nasty problems”. It excuses the potential for catastrophic failure in the name of speed and progress.</p><p>In the same X post, Altman announces that ChatGPT will soon allow erotica, to align with their principle to <em>“treat adult users like adults”</em>. This is a move that could potentially open up new low-friction forms of intimate connection, which is worrying in an era where forming healthy romantic relationships is already a challenge. A new road to avoid confronting the opposite sex is exactly the “help” some vulnerable people don’t need.</p><p>OpenAI is only the latest lab to tap into the companion gold chest while fighting off legal troubles:</p><ul><li><strong>Meta </strong>received scrutiny after reports of its AI engaging in “sensual” chats with children ⁸. The company is also betting on personal AI at consumer scale, with a ‘Superintelligence’ team to achieve this.</li><li><strong>xAI (Grok)</strong> launched AI companions “of very bad taste” with no adult-use restrictions, available to engage in explicit conversations.</li><li><strong>Character.AI</strong> is another popular alternative, particularly among teens. It’s being investigated by the Texas Attorney General for potentially misleading children with deceptive AI-generated mental health services ¹². To their (little and late) credit, it’s also the first company to announce they would ban minors from using their chatbots ⁹, although they haven’t yet shared a plan for how they’ll transition their loyal user base of teens out of the platform without causing incidents.</li></ul><p>The move to lower friction and raise engagement is a core business model in tech, and it’s no different with Gen AI chatbots. It seems to me like labs are content to exploit user dependence, including with minors. And only the legal baton seems to stir them into a frenzy of safety measures.</p><h3>What we should care about</h3><p>This issue hit me hard during a recent workshop on the ideal future in the age of AGI. When asked to imagine the world we want, my first thought was a future with more human connections, less isolation and stronger communities.</p><p>One person immediately asked: <em>“How will AGI help us do that?”</em></p><p>I replied that while AGI might give us some tools to help us, I hope that we would tackle that problem ourselves, AI or not.</p><p>The next person said: <em>“That’s a lovely vision, but I don’t see how that’s possible… with the current state of the world, and all…”</em></p><p><em>Sigh.</em> We’ve become so focused on technology as the solution that we’ve lost faith in our ability to solve our own human problems. Many of us are simply giving up, surrendering to disconnected life. We’re scared or discouraged to reach out to our peers because we’re losing our social skills and choosing the path of least resistance. To be fair, the modern world isn’t making it easy for us to connect in real life, but I still believe we could climb over these barriers together. What’s the worst that can happen?</p><p>Human connection is not a luxury. It’s a basic human need. It’s been proven to be a determinant of health and longevity ²⁻¹⁰. Indeed, poor social connections are linked to a wide range of chronic diseases and conditions. For example:</p><ul><li>Strong evidence links social disconnection to heart disease and stroke outcomes, with poor social relationships associated with a 29% increase in the risk of heart disease and a 32% increase in the risk of stroke ²⁻¹⁰⁻¹¹.</li><li>Chronic loneliness is associated with accelerated cognitive decline and an approximately 50% increased risk of dementia in older adults ².</li></ul><p>More evidently, loneliness is strongly associated with depression and anxiety ²⁻¹⁰. It is the most reliable predictor for suicide ideation and attempt ².</p><h4>So what?</h4><p>What if connectedness was part of the top quality of life and societal health we optimise for? It’s as precious as our natural resources, and we should be nourishing it as a core part of what it means to build a resilient and sustainable society.</p><p>How about making it part of a standard business charter, for instance? Like the Good Business Charter?</p><p>If AI truly frees up our time and drives down the cost of essential services, the “good” use of that surplus surely isn’t endless productivity. It’s investing in the real-world connections and communities we crave. Why don’t we make this our goal?</p><h3>Sources</h3><ol><li><a href="https://www.gov.uk/government/statistics/community-life-survey-202324-annual-publication/community-life-survey-202324-background-and-headline-findings">Community Life Survey 2023/24: Background and headline findings</a> (4 Dec 2024, UK Gov Official Statistics, Department for Culture, media &amp; Sport)</li><li><a href="https://www.hhs.gov/sites/default/files/surgeon-general-social-connection-advisory.pdf">Office of the Surgeon General. Our epidemic of loneliness and isolation: the U.S. Surgeon General’s advisory on the healing effects of social connection and community.</a> (2023, US Department of Health and Human Services, Office of the Surgeon General)</li><li><a href="https://doi.org/10.1093/jcr/ucaf040">AI Companions Reduce Loneliness</a> (25 Jun 2025, Journal of Consumer Research, Julian De Freitas, Harvard Business School)</li><li><a href="https://www.tandfonline.com/doi/full/10.1080/03075079.2024.2326956">When artificial intelligence substitutes humans in higher education: the cost of loneliness, student success, and retention</a> (13 Mar 2024, Studies in Higher Education, Joseph Crawford, University of Tasmania)</li><li><a href="https://www.axios.com/2025/11/07/openai-chatgpt-lawsuits-safety">OpenAI faces seven more suits over safety, mental health</a> (7 Nov 2025, Axios)</li><li><a href="https://openai.com/index/building-more-helpful-chatgpt-experiences-for-everyone/">Building more helpful ChatGPT experiences for everyone</a> (2 Sep 2025, OpenAI)</li><li><a href="https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations">Strengthening ChatGPT’s responses in sensitive conversations</a> (8 Nov 2025, OpenAI)</li><li><a href="https://www.bbc.co.uk/news/articles/c3dpmlvx1k2o">Meta investigated over AI having ‘sensual’ chats with children</a> (18 Aug 2025, BBC News)</li><li><a href="http://character.ai/">Character.AI</a> <a href="https://apnews.com/article/characterai-kids-minors-18-ban-chatbot-5d203e9f22c62c153936ccc776a0ed09">is banning minors from interacting with its chatbots</a> (29 Oct 2025, AP)</li><li><a href="https://assets.publishing.service.gov.uk/media/5fb66cf98fa8f54aafb3c333/6.4882_DCMS_Loneliness_Strategy_web_Update_V2.pdf">A connected society: A strategy for tackling loneliness</a> (Oct 2018, UK Gov, Department for Digital, Culture, Media and Sport)</li><li><a href="https://www.nature.com/articles/s41562-023-01617-6">A systematic review and meta-analysis of 90 cohort studies of social isolation, loneliness and mortality</a> (19 Jun 2023, Nature, Wang et al., Harbin Medical University)</li><li><a href="https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-investigates-meta-and-characterai-misleading-children-deceptive-ai">Attor­ney Gen­er­al Ken Pax­ton Inves­ti­gates Meta and Char​ac​ter​.AI for Mis­lead­ing Chil­dren with Decep­tive AI-Gen­er­at­ed Men­tal Health Services</a> (18 Aug 2025, Texas Attorney General, Ken Paxton)</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c045b45125a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/human-disconnection-how-ai-fills-the-void-and-risks-deepening-it-0c045b45125a">Human disconnection: How AI fills the void and risks deepening it</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 7: From Tech burnout to studying AI]]></title>
            <link>https://medium.com/retraining-epoch/from-tech-burnout-to-studying-ai-62a204f58f40?source=rss-6d0de68df3d6------2</link>
            <guid isPermaLink="false">https://medium.com/p/62a204f58f40</guid>
            <category><![CDATA[students]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ludivine Siau]]></dc:creator>
            <pubDate>Wed, 05 Nov 2025 21:42:29 GMT</pubDate>
            <atom:updated>2025-11-19T07:02:50.645Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QkdJmpJlT_lYliQ0WS5cdw.jpeg" /><figcaption>UCL Cruciform Building, Feb 2018 — Photo by <a href="https://unsplash.com/@illeqi?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Luke Wang</a> on <a href="https://unsplash.com/photos/brown-concrete-building-during-daytime--b57E2FSvnI?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>At the start of 2025, after 13 years working in Tech, I made a major decision: I applied for a Masters of Research (MRes) focused on AI and Healthcare.</p><h4>My job sucked, my career was stuck</h4><p>End of 2024. As a senior delivery/product manager, I had lost all purpose in my current job and, more widely, in my career. The project I was delivering was going nowhere, I was working with difficult people, and the downstream results of my work were non-existent. The job market was brutal, and the entire industry felt deeply unappealing anyway: full of products no one asked for, made-up value propositions, and no real problems solved.</p><p>I tried exploring sustainability as potential area of interest, but I quickly realised that others were far more competent and passionate than I was.</p><p>Then, the AI hype went full steam. concerns about its impact on white-collar jobs, like middle management in Tech *cough cough*, started to rise.</p><p>I couldn’t see any future for my career. And this perspective was especially dark as I was recovering from depression.</p><h4>Serendipity to the rescue</h4><p>An old friend and classmate posted on LinkedIn about starting an MSc after a job burnout. I reached out, we caught up on Zoom, and his enthusiasm about his new direction was infectious.</p><p>Wait a minute… Could <em>I</em> go back to school?</p><p>After some deep reflection and Googling for masters programmes, three things became clear:</p><ol><li>After 13 years in Tech, only having myself to care for and being financially conservative, I realised the financial and logistical worries about a second degree were <strong>solvable</strong>.</li><li>I <em>needed </em>to study <strong>on campus</strong>. i was tired of working remotely. I wanted to meet people, enjoy the buzz of a university, and get out of my local routine.</li><li>I quickly found the topic I had to study. It’s potent, almost urgent, it’s currently being exploited for useless corporate profit, and it might be coming for my job. I decided to <strong>ride the beast</strong> in a fabulous way instead of letting it trample me: I would study AI applied to healthcare, a field that genuinely matters.</li></ol><h4>Hungry for learning</h4><p>My application to UCL for an MRes (I chose the research-focused path, it felt more practical than a ‘normal’ taught MSc) was accepted, and I started studying at the end of September. Five weeks in, I couldn’t be happier.</p><p>Coming back to school after a decade in the professional world, I am far more appreciative of the learning experience than I was in my early twenties in engineering school. Maybe because this is an intentional life choice, and a costly one. I am more curious, attentive, and <em>hungry</em>. Compared to 20 year-old Lu, I have the experience, reflection and critical thinking skills now to truly enjoy the opportunity to fill the gaps in my knowledge. And I’m also determined to get my money’s worth, I’ll admit.</p><h4>Let’s get our hands dirty, shall we?</h4><p>Now, I’m ready for the real stuff I signed up for: my research project and dissertation.</p><p>I received confirmation of my topic last week, and it was my top choice (woop woop!)</p><p>My focus will be on “Developing equitable patient-level prediction models in chronic disease management”.</p><p>I will be assessing bias in machine learning models specifically used for predicting chronic cardiovascular diseases, studying their causes, and exploring mitigation strategies. I chose this subject because my primary interest in AI lies in <strong>safety, ethics and governance</strong>. I wanted to get my hands dirty, build models, dig into safety measures (like bias mitigation), and take a wider perspective on the systemic challenges of training and using AI in healthcare.</p><p>We’re kicking off with my supervisors next week. In the meantime I’ve started digging into the literature and planning the project (there will be Trello and Miro boards, I can’t help…) But I’ll share more about the research process in the next posts.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=62a204f58f40" width="1" height="1" alt=""><hr><p><a href="https://medium.com/retraining-epoch/from-tech-burnout-to-studying-ai-62a204f58f40">Week 7: From Tech burnout to studying AI</a> was originally published in <a href="https://medium.com/retraining-epoch">Retraining Epoch</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>