In silico tools hold potential to improve prediction of drug-induced Torsade de Pointes (TdP), a rare but lethal adverse drug reaction, responsible for the withdrawal of several drugs from the market. However, computational cellular models do not usually consider inter-individual variability, which may be crucial when predicting rare events such as TdP. Computational tools are also used for risk stratification, and binary classifiers have been previously proposed. The validation process of these classifiers requires external data sets to assure reliability. In the present study we propose a ternary TdP-risk classifier validated with an external set of drugs and compare the effect of incorporating inter-individual variability in a computational model or using a single model. First, we built a population of models by randomly modifying all ionic conductances of a modified version of the ventricular action potential model published by O’Hara et al. The population of models was calibrated according to experimental data and 848 physiologically plausible models were selected. Then, effects of the 12 training CiPA drugs were simulated on the single baseline model and on the calibrated population. Ternary classifiers, based on support vector machines and logistic regression, were built using biomarkers extracted from simulation results. Classifiers were validated using the 16 validation CiPA drugs as an external data set. The classification accuracy increased to 80.1% when using the population of models, with respect to an accuracy of 62.4% obtained using the baseline model. Simulations with population of models allowed the identification of the ionic features of individuals more prone to develop TdP, such as lower conductances of IKr, IKs or INaK and higher conductances of ICaL, INaL or INCX. The methodology presented provides new opportunities to assess drug-induced TdP, taking into account inter-individual variability and may be helpful to improve current cardiac safety screening methods.