During a hospital stay, a patient’s medication regimen often changes as clinicians start, stop, or continue therapies. This work studies whether models can predict which ATC‑level‑3 drug classes are added or removed between 24 hours after admission and discharge, using only data that are available within the first 24 hours and prior completed admissions. Conventional evaluation that compares the full discharge regimen rewards models for copying unchanged drugs, masking their ability to detect real changes. To avoid this leakage, the authors build a benchmark that predicts net ATC‑3 additions and removals. Candidates for addition are classes not active at the 24‑hour mark, while removal candidates are those active at that time.
On this benchmark they propose R‑GEAN (Regimen‑Guided Edit Action Network), an asymmetric candidate‑scoring network with separate addition and removal predictors. Across 240,480 admissions from 82,286 patients, R‑GEAN achieves an edit‑composite score of 0.464, outperforming the strongest baseline (0.435). Re‑implemented RETAIN, GAMENet and MICRON baselines obtain 0.428, 0.420 and 0.288 respectively. R‑GEAN’s advantage lies mainly in correctly identifying drug classes that become inactive at discharge; rare additions and admissions with multiple changes remain challenging. Rankings based on micro‑F1 over the reconstructed discharge regimen correlate weakly with the edit composite (Spearman r = 0.20), indicating that the two evaluation perspectives capture different aspects. A “continuation” baseline, which predicts no changes, attains the highest complete‑regimen score, highlighting the divergence between full‑regimen and edit‑level metrics. The benchmark therefore measures observed prescribing changes rather than treatment appropriateness.
Review: R‑GEAN shows promise in detecting drug discontinuations under limited information, offering a finer‑grained evaluation for clinical decision support, yet its ability to predict new prescriptions needs further improvement.