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

Abstract

Prediction of Torsades de Pointes (TdP) risk using hiPSC-CM assays remains challenging, as many models fail to capture nonlinear patterns and exhibit unstable performance across different drugs. We examined whether a stacking ensemble can improve robustness when only two simple <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>Δ</mi> <mi>F</mi> <mi>P</mi> <mi>D</mi> <mi>c</mi></mrow> </math> -derived predictors are available. Two electrophysiological predictors were derived from MEA-based FPDc measurements: the maximum change ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>Δ</mi> <mi>F</mi> <mi>P</mi> <mi>D</mi> <mi>c</mi></mrow> </math> ) and the interpolated <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>Δ</mi> <mi>F</mi> <mi>P</mi> <mi>D</mi> <mi>c</mi></mrow> </math> at Cmax. These features were used to train a stacking model comprising random forest (RF), XGB, and a shallow artificial neural network (ANN). Performance was assessed on 16 unseen CiPA reference compounds using AUC, likelihood ratios, pairwise accuracy, and classification error. Class weighting addressed class imbalance, and likelihood-based metrics assessed diagnostic consistency. The stacking ensemble consistently outperformed all single classifiers. The XGB-based meta-classifier achieved perfect discrimination for both AUC1 and AUC2 (1.000), with pairwise accuracy reaching 1.000 and classification error remaining below 0.125 across repeated evaluations. Combining a simple MEA-derived feature set with a stacking architecture provides a more reliable framework for early TdP risk assessment than single-model approaches. External validation using additional MEA datasets and the integration of interpretable modeling strategies will be important for future translational use.