Artificial Intelligence in Drug Discovery and Health Care Systems: Machine Learning Applications in Modern Pharmacy and Patient Care

Abstract

Artificial intelligence (AI) is transforming the value chain in pharmaceuticals from molecular design to delivery of care but most available literature is based on narratives, focused on discovery and care separately, and gives a single accuracy score without including calibration, utility, and fairness evidence as required by the contemporary standards and regulators. This paper introduces AI-PHARM, a consistent machine-learning framework for evaluation with five common layers of data, representation, learning, evaluation and governance applied consistently to both ends of the value chain with the same methodology using publicly available recent datasets. In the drug-discovery track, six model families were compared on all six in vitro Absorption-Distribution-Metabolism-Excretion (ADME) endpoints of the publicly available Biogen dataset (3,521 compounds). The heterogeneous stacked ensemble model was found to be the most accurate on all endpoints (average R² = 0.533), significantly outperforming the best individual model on three endpoints, and applicability domain analysis showed increasing prediction error for new-to-chemistry compounds. For the pharmacists’ screening of dysglycemia based on patientcare, five different classifiers were tested using 229,781 unique adult patients’ records in the CDC Diabetes Health Indicators Dataset; a well-calibrated gradient boosting model yields an AUC ROC of 0.814 and near-perfect calibration (expected calibration error = 0.003) and positive net clinical benefit according to decision curve analysis, and a subgroup audit identifies differences in error rates due to population differences in terms of age and income. An example of a negative result—a finding of no value in applying the ensemble in the classification problem at hand—is given by the framework’s fundamental idea: Increased model complexity needs to be supported empirically, and discrimination, calibration, clinical usefulness, and fairness need to be considered together. Publicly available datasets and an explicitly defined experiment design have been used for all experiments