Minimax Optimization in Deep Learning for Drug Recommendation: A Structured Review of Safety Objectives and Evaluation Practice

Abstract

Medication recommendation is now one of the most active applications of deep learning to electronic health records, and the published numbers suggest steady progress: Jaccard scores on MIMIC-III have climbed from 0.4509 for GAMENet in 2019 to roughly 0.53 for the strongest recent models. Read side by side rather than one at a time, that progress is harder to credit. This review organises the literature into two groups, one predicting a medication set from longitudinal records and one drawing on other signals entirely, and compares them on methodology, evaluation data, safety objective, and what their reported gains rest on. The convergence on a Jaccard of about 0.53 turns out to be largely an artefact of evaluation practice. Baseline rows are copied between papers rather than re-run; the same model appears with a 31 percent spread in its own headline safety metric across sources; and the label space is a therapeutic class rather than a prescribable drug, so that scoring the same data at finer granularity drops Jaccard to between 0.16 and 0.26. The field's only safety measure derives from a single spontaneous-reporting resource with no severity grading, and no reviewed model has been validated prospectively or at an external institution. We set out the gap under six headings and propose testable next steps: a controlled re-evaluation of the canon under one pipeline, prescribable-granularity labels, severity-weighted and deviation-based safety metrics, and hybrid designs pairing a structured sequence model with retrieval-grounded reasoning while reporting medication burden alongside accuracy.