Cardiac mortality prediction in patients undergoing PCI based on real and synthetic data

TL;DR

This study tackles predicting 3-year cardiac mortality after PCI in bifurcation lesions by exploiting both real and synthetic tabular data to address severe class imbalance. A broad set of models (including LR, RF, CatBoost, XGBoost, TabPFN, and KAN) were evaluated with extensive oversampling strategies (ARF, TVAE, CTGAN, Gaussian Copula, TabSyn) and edge-case augmentation, coupled with calibration and permutation feature-importance analyses. Results show that without oversampling models achieve high overall accuracy but poor minority-class detection; ARF and GAN-based augmentation markedly improve minority recall and F1, albeit with some AUROC trade-offs, and four clinical features—Age, Ejection Fraction, Peripheral Artery Disease, and Cerebrovascular Disease—consistently drive predictions. External validation on a separate cohort confirms similar patterns, though performance declines due to smaller sample size, reinforcing the importance of calibration and robustness testing. The findings highlight the value of synthetic data to improve imbalanced clinical risk predictions and suggest careful feature selection and stress-testing (edge cases) to ensure clinically meaningful outputs before deployment.

Abstract

Patient status, angiographic and procedural characteristics encode crucial signals for predicting long-term outcomes after percutaneous coronary intervention (PCI). The aim of the study was to develop a predictive model for assessing the risk of cardiac death based on the real and synthetic data of patients undergoing PCI and to identify the factors that have the greatest impact on mortality. We analyzed 2,044 patients, who underwent a PCI for bifurcation lesions. The primary outcome was cardiac death at 3-year follow-up. Several machine learning models were applied to predict three-year mortality after PCI. To address class imbalance and improve the representation of the minority class, an additional 500 synthetic samples were generated and added to the training set. To evaluate the contribution of individual features to model performance, we applied permutation feature importance. An additional experiment was conducted to evaluate how the model's predictions would change after removing non-informative features from the training and test datasets. Without oversampling, all models achieve high overall accuracy (0.92-0.93), yet they almost completely ignore the minority class. Across models, augmentation consistently increases minority-class recall with minimal loss of AUROC, improves probability quality, and yields more clinically reasonable risk estimates on the constructed severe profiles. According to feature importance analysis, four features emerged as the most influential: Age, Ejection Fraction, Peripheral Artery Disease, and Cerebrovascular Disease. These results show that straightforward augmentation with realistic and extreme cases can expose, quantify, and reduce brittleness in imbalanced clinical prediction using only tabular records, and motivate routine reporting of probability quality and stress tests alongside headline metrics.

Figures (6)

• Figure 1: The pipeline of research. a) Data storage and collection b) Prepocessing - data cleaning, embeddings and etc. c) Train test split d) 10 folds datasets for cross-validation e) Hyperparameters tuning and selection of the best f) Models training g) Evaluation on test set
• Figure 2: Three training regimes: a) Standard 10 folds cross validation. b) Synthetic in distribution data used to enrich the dataset. c) Synthetic data and expert validated edge cases.
• Figure 3: Random Forest distribution probabilities on edge cases without oversampling (a) and with ARF (b).
• Figure 4: Risk coverage analysis on the test set across different models. (a) Random Forest (b) XGBoost
• Figure 5: Radar plots of classification metrics on test (Without classification, ARF, Edge cases, ARF + Edge cases).