A Text-Mining Algorithm for Identifying Drug Pairs Potentially Involved in Prescribing Cascades: Development and Internal Validation Within the Prescribing Inappropriateness Assessment (PINA) Digital Tool.
TL;DR
An algorithm designed to identify drug pairs potentially compatible with PCs and evaluated its classification performance may represent a scalable and reproducible strategy for supporting structured medication review and facilitating the recognition and clinical evaluation of potential PCs.
Abstract
Background
Prescribing cascades (PCs) occur when an adverse drug reaction (ADR) is intentionally or unintentionally treated with an additional medication, potentially perpetuating inappropriate polypharmacy. Digital tools that systematically support the identification of potential PCs during medication review are lacking. Here, we describe the development of an algorithm designed to identify drug pairs potentially compatible with PCs and evaluate its classification performance.
Methods
The algorithm was developed as a component of the Prescribing INappropriateness Assessment (PINA) digital tool, a software platform designed to support clinicians in conducting medication reviews. The algorithm logic identifies potential PCs by detecting intersections between ADRs and therapeutic indications extracted from Italian Summaries of Product Characteristics (SmPCs) and encoded as Medical Dictionary for Regulatory Activities (MedDRA) terms using a natural language processing pipeline. Four algorithm configurations were tested, differing in the MedDRA semantic levels used for matching: (1) Preferred Terms (PTs) only; (2) PTs with semantic extension to High Level Terms (HLTs); (3) PTs with extension to Standardised MedDRA Queries (SMQs); and (4) PTs with combined HLT/SMQ extension. Performance was assessed against a reference standard comprising 69 literature-supported positive drug pairs and 138 randomly generated negative drug pairs in a 1:2 ratio. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, F1 score, and Youden's index were calculated.
Results
All models demonstrated excellent specificity (≥ 99%). Sensitivity increased with semantic expansion: 50.7% for PT-only matching, 73.9% for PT+HLT, 68.1% for PT+SMQ, and 85.5% for PT+HLT/SMQ. The combined HLT/SMQ extension model achieved the highest overall performance, with an accuracy of 94.7%, a PPV of 98.3%, an NPV of 93.2%, an F1 score of 91.4, and a Youden's index of 84.8. False-positive rates remained minimal across all models.
Conclusions
The newly developed algorithm demonstrated high classification performance, particularly when PT-level matching was extended using both HLTs and SMQs. This MedDRA-driven approach based on regulatory data may represent a scalable and reproducible strategy for supporting structured medication review and facilitating the recognition and clinical evaluation of potential PCs.
