All Times EDT
Keywords: Logistic regression, random forest model, machine learning, Tdp
Introduction: Torsades de pointes (TdP) is a rare but potentially fatal ventricular arrhythmia that can be induced as a side effect by drugs. Several pre-clinical paradigms have been proposed to assess the drug-induced TdP risk in vitro. However, a comprehensive statistical modeling that provides accurate prediction on drug-induced TdP risks from experimental observations is still lacking.
Methods:
We utilized two datasets (hiPSC-CMs and rabbit ventricular wedge assay) to build statistical models that predict low, medium, and high TdP risks. Both datasets include the same 28 drugs with known levels of TdP risks categorized by domain experts. Logistic regression and random forest models were trained on the previous datasets with the utilization of ordinal information in drug risks. Leave-one-drug-out cross validation was used to evaluate the model prediction performance.
Results:
The three-risk prediction accuracies of logistic regression and random forest are significantly higher than no information rate. The two models also outperformed the baseline multinomial logistic regression without ordinal risk information. Among the model and dataset combinations, random forest trained on wedge dataset achieved the best prediction performance, producing 85.71% accuracy by observations and 92.86% accuracy by drugs. The stratified bootstrapping shows that random forest is more stable and conservative in identifying low risk drugs.
Conclusions/Implications:
Our statistical models exhibited high prediction accuracy in the identification of drug-induced TdP risks. The model prediction result can be used as one evidence in the process of drug-risk assessment. The model performance also shows that the wedge dataset is more informative in the modeling of drug-risk prediction.