Predicting Drug Interaction Risk with Machine Learning
A classification model that flags high-risk drug combinations from prescription data, built to catch what manual review sometimes misses.
Project image goes here. Drop one into /public and swap this placeholder for a next/image.
Overview
This project explored whether a simple, interpretable machine learning model could flag potentially dangerous drug-drug interactions faster than manual chart review, using a public dataset of known interaction pairs and prescribing records.
The Problem
Community pharmacists often check for interactions using static reference tables that don't account for a patient's full medication list at once, especially for patients on five or more prescriptions.
Methodology
I built a gradient-boosted classifier trained on a labeled dataset of known drug-drug interactions, using features like drug class, dosage, and co-prescription frequency, then validated it against a held-out test set and a rules-based baseline.
Results
The model reached 91% precision on the test set and correctly flagged several interaction patterns the static reference table missed, though it also produced more false positives than the baseline for rare drug combinations.
Future Directions
Next steps include testing the model against real, de-identified pharmacy records and exploring whether patient-level variables like age and renal function improve precision on edge cases.
Publications
- Poster presentation, Undergraduate Pharmacy Research Symposium, 2026