Routine health records can help identify effective treatment strategies for individual patients over time. However, changes to patients’ health and previous treatments affect future outcomes. Treatment decisions depend on patients’ clinical histories, and standard causal methods can become unstable when certain treatments are rarely observed for patients with particular histories. Methods such as inverse probability weighting (IPW) can produce extreme weights, while g-computation and longitudinal targeted maximum likelihood estimation may yield unreliable predictions when data provide little support for alternative treatments. On the other hand, deep learning methods may recommend treatment strategies that are poorly supported by observed data, making their estimated benefits unreliable. We therefore propose a framework in which deep learning captures complex patient histories while causal inference ensures that only well-supported strategies are recommended and evaluated, to estimate dynamic treatment regimes (DTRs) from routine HIV programme data. The main methodological contribution is an adaptive overlap-constrained approach to learning treatment regimes. Most existing approaches impose the same minimum support threshold at every decision stage. Our method adjusts the threshold based on the available data at each stage and for each patient history. As a result, restrictions are stricter when data are limited and more relaxed when data are sufficient. Deep learning models capture changes in patients’ health over time, while doubly robust methods account for time-varying confounding. Together, these components support the learning of treatment regimes while restricting poorly supported treatment choices. We use simulations to assess how these restrictions affect policy value and estimation reliability under treatment–confounder feedback, limited treatment overlap, model misspecification, and missing viral load measurements. In addition. we compare the proposed method with g-computation, longitudinal targeted maximum likelihood estimation (LTMLE), sequential double machine learning, IPW and deep Q-learning. We assess performance using bias, root mean squared error, confidence interval coverage, policy regret, and the frequency of unsupported treatment recommendations. The framework will be applied to Zimbabwean electronic patient monitoring records to compare personalised antiretroviral dispensing and viral load follow-up strategies for sustained viral suppression. Sensitivity analyses will examine the effects of missing viral load measurements and unmeasured confounding. The study aims to establish when adaptive overlap-constrained learning can produce more reliable and interpretable treatment recommendations for HIV care using routine programme data.
Keywords: Dynamic treatment regimes; causal inference; deep learning; positivity; overlap constraints; HIV care.
