طراحي سيستم توصيه‌كننده‌ي دارو براي بيماران كرونايي با استفاده از روش‌هاي آماري و يادگيري ماشين

Designing a drug recommendation system for COVID-19 patients using statistical methods an‎d machine learning

مهندسي پزشكي

The COVID-19 pan‎demic created unprecedented challenges for healthcare systems, one of the most important being the selec‎tion of an effective drug therapy for each patient given the heterogeneity of clinical responses. Therefore, relying solely on uniform protocols for all patients is not efficient an‎d highlights the need for personalized medicine. In this context, using data from 2,285 patients hospitalized at Bellvitge Hospital (the COVID-SEMI study), this research developed an AI- an‎d causal-inference–based clinical decision-support framework that simultaneously addresses two key objectives: outcome prediction an‎d causal estimation of treatment effects. For modeling, in addition to classical classification models such as XGBoost an‎d Ran‎dom Forest, deep neural networks were also used. Specifically, a Long Short-Term Memory (LSTM) network was employed to leverage sequential structure, an‎d a Multi-Layer Perceptron (MLP) was used as a simpler alternative (with a fully feed-forward architecture). Because machine learning models learn only correlational patterns an‎d do not by themselves guarantee causal inference, the causal layer of the study was designed by combining machine learning/deep learning with causal inference methods. Accordingly, to overcome the challenge of time-varying confounders, target trial emulation an‎d the parametric g-formula were used, an‎d heterogeneous treatment effects such as the Individual Treatment Effect (ITE) an‎d the Conditional Average Treatment Effect (CATE) were estimated for five main drugs—steroids, hydroxychloroquine, remdesivir, Kaletra, an‎d tocilizumab—based on which a personalized treatment system was designed. A key point is that the causal inference framework is not dependent on a specific network architecture; in other words, even when replacing the LSTM with an MLP, goodness-of-fit an‎d especially calibration remained appropriate (ECE was very small an‎d below 0.01 for both architectures). This indicates that the g-formula–based causal component can be implemented on top of different predictive models, provided that the outcome/treatment models are trained to be sufficiently well-calibrated an‎d stable. In this section, the reported values correspond to the LSTM model because overall it performed slightly better than the MLP. The model showed good performance for next-day outcomes: at the person-day level, the AUC-ROC for next-day discharge was 0.803±0.015 an‎d for next-day death 0.795±0.010; in contrast, for next-day ICU admission performance was 0.717±0.023, an‎d for death by the end of hospitalization at the patient level within a hazard-function modeling framework, the reported value was 0.749±0.033. Finally, from an interpretability perspective, SHAP analysis showed that age was the most important variable in the model’s decisions, followed by markers of inflammation an‎d organ function, including C-reactive protein level, creatinine, an‎d lymphocyte count. This study demonstrated that combining outcome prediction with causal inference can produce decision-oriented indicators such as preventable mortality. Based on counterfactual estimates, the share of preventable deaths among all deaths ranged from 39.3% to 49.4% (the highest value corresponding to corticosteroids: 49.4%), suggesting that the system could potentially reduce a substantial portion of adverse outcomes. Therefore, a data-driven personalized treatment system can significantly enhance the efficiency an‎d effectiveness of clinical decision-making.

5

مارتين وكويتز

157596