Stacking Ensemble Machine Learning for Cardiac Safety Assessment Using hiPSC-CM MEA Data.

Area of Science:

• Cardiovascular pharmacology
• Computational toxicology
• Stem cell biology

Background:

• Predicting drug-induced Torsades de Pointes (TdP) risk with induced pluripotent stem cell-derived cardiomyocyte (hiPSC-CM) assays is challenging.
• Existing models often struggle with nonlinear patterns and drug-specific performance instability.

Purpose of the Study:

• To evaluate if a stacking ensemble model can enhance the robustness of TdP risk prediction.
• To assess the effectiveness of using two simple $\Delta FPDc$ -derived predictors.

Main Methods:

• Derived two electrophysiological predictors ($\Delta FPDc$ maximum change and interpolated $\Delta FPDc$ at Cmax) from MEA-based FPDc measurements.
• Trained a stacking model using random forest (RF), XGB, and an artificial neural network (ANN).
• Assessed performance on 16 unseen CiPA reference compounds using AUC, likelihood ratios, pairwise accuracy, and classification error, addressing class imbalance with class weighting.

Main Results:

• The stacking ensemble consistently outperformed individual classifiers.
• The XGB-based meta-classifier achieved perfect discrimination (AUC = 1.000) and pairwise accuracy (1.000).
• Classification error was maintained below 0.125 across evaluations.

Conclusions:

• A stacking architecture combined with simple MEA-derived features offers a more reliable framework for early TdP risk assessment.
• Future work should include external validation and interpretable modeling for translational application.