Interpretable Machine Learning Prediction of Drug-Induced QT Prolongation: Electronic Health Record Analysis

Area of Science:

• Cardiology
• Pharmacology
• Medical Informatics

Background:

• Drug-induced long-QT syndrome (diLQTS) poses a significant risk to hospitalized patients.
• Machine learning models can predict diLQTS risk, with deep learning showing superior accuracy but lacking interpretability.

Purpose of the Study:

• To investigate the trade-off between interpretability and predictive accuracy in diLQTS risk models.
• To compare a deep learning algorithm with a more interpretable cluster analysis-based algorithm for diLQTS risk prediction.

Main Methods:

• A cohort of 35,639 inpatients treated with 39 high-risk medications was analyzed.
• Predictors included over 22,000 diagnoses and medications; diLQTS was defined as corrected QT interval >500 ms.
• Two models were developed: a 6-layer deep learning network and a 4-cluster interpretable model.

Main Results:

• Class III antiarrhythmics increased diLQTS risk across all clusters.
• Propofol increased risk in non-critically ill patients without cardiovascular disease; ondansetron decreased risk.
• The interpretable model had lower accuracy (AUC 0.65) than deep learning (AUC 0.78) but comparable calibration.

Conclusions:

• An interpretable modeling approach for diLQTS prediction is less accurate but more clinically applicable than deep learning.
• Future research should consider this interpretability-accuracy trade-off when developing clinical prediction models.