A Leakage-Safe, Facility-Specific Framework for Multi-Label Drug Recommendation: A Comparative Study of Deep and Classical Models on Small Hospital Data

Physicians treating chronic ischemic heart disease (ICD-۱۰: I۲۵.۱) routinely combine several medications at once, and the choice depends heavily on a patient's clinical trajectory and on what a given hospital actually keeps in stock. Most drug-recommendation models sidestep this reality: they train on large public datasets and lean on external pharmacological knowledge — drug names, chemical structure, or language-model embeddings — to represent medications, which risks recommending drugs a specific facility never prescribes. We took a different route. Using ۷۰۶ clinical records from ۹۵ patients across two hospitals (Shahid Beheshti and Kamkar), we represent every drug by nothing more than a numerical ID, so the model's output space is restricted to drugs already observed in that hospital's own prescribing history — and can be retrained cheaply whenever a formulary changes. The starting point was a BiLSTM network augmented with Multi-Head Attention, evaluated with patient-level splitting to avoid leakage, Focal Loss and rare-drug filtering (۳۱۸ → ۱۴۶ drugs) to handle severe class imbalance, and a precision-first threshold chosen only on the validation set. Averaged over four random seeds, this model reached a macro-F۱ of ۰.۱۹±۰.۰۳ and a micro-averaged AUC of ۰.۸۶±۰.۰۳. We then asked a harder question: does the added complexity of attention actually earn its keep on a dataset this small? Running the identical pipeline on eight alternatives — a frequency baseline, Random Forest, XGBoost, MLP, a plain LSTM, GRU, BiLSTM without attention, and a pure Transformer — showed that it does not: XGBoost matched or beat the attention model on macro-F۱ (۰.۲۳۵±۰.۰۳۳ vs. ۰.۱۹۴±۰.۰۳۲) while training in ۱۶ seconds instead of over five minutes, and a plain BiLSTM without attention did much the same. We read this as a finding in its own right rather than a setback: at hospital-scale data, simpler and far cheaper models are at least as accurate, which strengthens rather than undermines the case for lightweight, facility-specific retraining as a practical path for resource-constrained clinical settings.

کلیدواژه ها:

drug recommendation ، multi-label learning ، ischemic heart disease ، BiLSTM ، attention ، class imbalance