Arrhythmias & electrophysiology
Machine learning–enabled assessment of risk of drug-induced QT prolongation at the time of prescribing
Authors' conclusion
Development and retrospective validation of 3 machine learning-based models for predicting drug-induced long QTc at the time of new QTdrug starts demonstrated superior performance compared with current clinical risk calculators
Study context
- Population
- Patients starting QTc-prolonging medications with follow-up 12-lead electrocardiograms (ECGs) within 1 year182,448 with baseline ECGs
- Participants
- 345,371 patients
- Exposure
- Machine learning models vs Tisdale and RISQ-PATH risk scores
- Outcome
- Prediction of QTc ≥500 ms (area under the receiver operating characteristic (AUROC))
- Follow-up
- Within 1 year
Methodological detailsprediction model development and validation, 5-fold cross-validation
- Prediction model development and validation
- 5-fold cross-validation
Main finding
As reported in the abstract
QTc ≥500 ms (AUROC)
| Group | AUC (baseline ECG subset) | AUC (overall, vs RISQ-PATH) | AUC (predominantly inpatients, vs Tisdale) | AUC (overall) | AUC (predominantly inpatients) | PPV (vs RISQ-PATH operating point) | PPV (vs Tisdale operating point) | PPV |
|---|---|---|---|---|---|---|---|---|
| XGBoost (XGB) | 0.869 | 0.859 | 0.855 | Not reported in the abstract | Not reported in the abstract | 61.5% | 54.0% | Not reported in the abstract |
| Deep neural network | 0.864 | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract |
| Combined | 0.874 | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract |
| RISQ-PATH | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | 0.701 | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | 28.35% |
| Tisdale | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | Not reported in the abstract | 0.770 | Not reported in the abstract | Not reported in the abstract | 28.3% |
AUC (baseline ECG subset)
XGBoost (XGB)0.869
Deep neural network0.864
Combined0.874
RISQ-PATHNot reported in the abstract
TisdaleNot reported in the abstract
AUC (overall, vs RISQ-PATH)
XGBoost (XGB)0.859
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATHNot reported in the abstract
TisdaleNot reported in the abstract
AUC (predominantly inpatients, vs Tisdale)
XGBoost (XGB)0.855
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATHNot reported in the abstract
TisdaleNot reported in the abstract
AUC (overall)
XGBoost (XGB)Not reported in the abstract
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATH0.701
TisdaleNot reported in the abstract
AUC (predominantly inpatients)
XGBoost (XGB)Not reported in the abstract
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATHNot reported in the abstract
Tisdale0.770
PPV (vs RISQ-PATH operating point)
XGBoost (XGB)61.5%
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATHNot reported in the abstract
TisdaleNot reported in the abstract
PPV (vs Tisdale operating point)
XGBoost (XGB)54.0%
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATHNot reported in the abstract
TisdaleNot reported in the abstract
PPV
XGBoost (XGB)Not reported in the abstract
Deep neural networkNot reported in the abstract
CombinedNot reported in the abstract
RISQ-PATH28.35%
Tisdale28.3%
Observational data: associations, not effects.
Further findings
Long QTc incidence · within 1 year
- Long QTc within 1 year, all QTdrug patients
- 5.7%
Original publication
Machine learning–enabled assessment of risk of drug-induced QT prolongation at the time of prescribing. Heart Rhythm. Published 24 October 2025. doi:10.
Summarised from the published abstract; every number is checked against it