EM-Optimized Parameter Learning Framework for Interpretable Modeling of Therapeutic Activity and Adverse Side Effects in Personalized Hypertension Treatment

Accurate prediction of pharmacological activity and associated therapeutic side effects is essential for advancing personalized medicine, particularly in hypertension therapy where patients’ responses vary depending on their health profiles and comorbidities. However, existing machine learning approaches face challenges in minimizing uncertainty specifically when targeting exploiting explainable models for ethical and personalized hypertension treatments. Moreover, pharmacological-hypertension datasets often contain missing and heterogeneous data due to incomplete side-effect reports which consequently reduce likelihood estimation and posterior calibration of the learned models. To address these challenges, a novel uncertainty-aware probabilistic learning framework based on Bayesian Networks (BNs) optimized via the Expectation–Maximization (EM) algorithm is proposed to reduce uncertainties through higher likelihood estimation for hypertension medications. This implementation extends previous EM-based probabilistic models by integrating uncertainty evaluation using influence strength and sensitivity analysis to enhance interpretability and reliability. The BN–EM framework learns dependencies between drug-related attributes such as drug classes, pregnancy categories, and brand names to predict side effects and activity levels. Three probabilistic experiments were conducted to assess the proposed parameter learning approach. The BN–EM model outperformed Tree-Augmented Naïve Bayes (TAN) and Naïve Bayes (NB), achieving the highest micro-averaged accuracy of 97.03%, compared to 95.48% and 89.55%, respectively. These findings highlight the potential of EM-optimized probabilistic modeling to support transparent and clinically interpretable personalized hypertension therapy.