NEW IHCs AVAILABLE: ADD HER2 OR FOLR1(FRɑ) TO xT OR xR /// LEARN MORE NEW IHCs AVAILABLE: ADD HER2 OR FOLR1(FRɑ) TO xT OR xR ///
11/08/2022

EHR-Based Machine Learning Model Predicts Drug-Induced QT Prolongation With Superior Performance Compared to Clinical Risk Predictors

AHA Scientific Sessions 2022 PRESENTATION
Authors Linyuan Jing, Thomas Morland, Christopher Kelsey, Sushravya Raghunath, John Pfeifer, Jeffrey Ruhl, Brandon Fornwalt and Christopher M Haggerty

Introduction: Many medications are associated with QTc prolongation (LQT). Current LQT predictors, including Tisdale and RISQ-PATH, have limited utility due to lack of generalizability to key use cases and modest model performance. A highly accurate LQT risk predictor utilizing data often available at the time of drug starts could broadly support LQT risk management.

Hypothesis: An electronic health record (EHR)-based machine learning (ML) model is superior in predicting drug-induced LQT in a clinical population vs. Tisdale and RISQ-PATH.

Methods: We identified all patients who began taking QT-prolonging medications (‘QTdrug’, per CredibleMeds) at Geisinger and had a follow-up 12-lead ECG within 1 year of medication start date while still on the drug. In case of multiple QTdrugs with overlapping dates, the last start date was used. Using 5-fold cross-validation, we trained an XGBoost ML model to predict LQT (QTc >500 ms) within 1 year of QTdrug start date using EHR data as input, including 131 features across age, sex, smoking history, vitals, lab tests, comorbidities, baseline ECG metrics (within 3 years before drug start date) and QTdrug usage (ever on QTdrug, number of current QTdrugs).

Results: QTdrug records were identified for 362,086 patients, and 6% had LQT within 1 year. The ML model demonstrated superior performance (Figure A) in predicting LQT (area under the curve (AUROC): 0.86, average precision score (AUPRC): 0.41) compared to RISQ-PATH (AUROC: 0.70, AUPRC: 0.19) and Tisdale (AUROC: 0.76, AUPRC: 0.21, calculated in a hospitalized subset N=113544). At the same specificity as RISQ-PATH (98%), the ML model had higher sensitivity (30% vs 10%) and positive predictive value (55% vs 29%). Similar comparisons were observed for Tisdale (Figure B).

Conclusions: An EHR-based ML model outperforms commonly used risk calculators for predicting drug-induced LQT. This model can be used to stratify patients by risk of developing drug-induced LQT in a clinical setting.

VIEW THE PUBLICATION

VIEW THE SLIDE DECK