Target variable formalization in improving the performance of models for predicting adverse cardiac events.

<jats:p>Introduction. Efficacy of prognostics models depends also on the clinically strict definition of the target variable, not only on the algorithms and predictors. Combining acute and chronic forms of ischemic heart disease may increase labeling noise and reduce predictive accuracy in the cardiology. Aim. To substantiate clinically and logically grounded approaches to defining the target variable for predicting adverse cardiac events while keeping the machine learning algorithm, predictor set, and internal validation framework fixed. Materials and methods. A retrospective study was conducted using depersonalized data from a regional healthcare information system. The analysis included 38,905 clinical episodes in 20,114 patients with ischemic heart disease. CatBoost was used for modeling, with a fixed set of 40 clinical, demographic, and laboratory features. Seventy-four target variable configurations were compared, including binary and multiclass schemes, with temporal isolation of sequential episodes taken into account. Model performance was assessed using 5-fold stratified cross-validation based on ROC-AUC, recall, and specificity. Results. The ROC-AUC was 0.76 in the baseline setting. Optimization of the labeling scheme increased the ROC-AUC to 0.84. The best-performing configuration had a ROC-AUC of 0.84, recall of 0.54, and specificity of 0.92. Several alternative schemes achieved higher sensitivity but reduced specificity, making them perspective for screening applications. Conclusion. Clinically and logically grounded optimization of the target variable is an independent stage in the prognostic models’ development and can improve internal validation performance and adapt the model to the intended clinical case.</jats:p>