PrevCardioOncAI: Machine Learning Algorithms for Predicting Cardiovascular Disease in Cancer Survivors

Comprehensive Summary

In this study, researchers evaluated the efficacy of AI algorithms to predict cardiovascular disease (CVD) and other cardiovascular events in cancer patients. ECG and laboratory data from the over 4,000 patients was separated into a training cohort (3,835 patients) and an additional validation cohort (329 patients). Along with predicting five cardiac outcomes CVD, heart failure, atrial fibrillation, and coronary artery disease (CAD)) regardless of when the cancer was diagnosed, researchers determined AI’s capacity to predict CVD post-diagnoses (de novo). Researchers choose to utilize regularized logistic regression, decision tree, and random forest models, for their abilities in overfitting reduction, interpretability, and transferability, respectively. In the general predictive analysis, the logistic regression model produced the highest mean AUROCs for CVD (0.806), atrial fibrillation (0.783), and CAD (0.792), with it only being outperformed by the random forest model for heart failure prediction (0.845 vs 0.851). The additional validation set produced similar results. Logistic regression also produced the highest AUROC (0.826) in de novo CVD predictions, with the random forest (0.802) and decision tree (0.735) models following. Additionally, the predictive analysis of four other advanced algorithms (BART, XGBoost, LightGBM, and CatBoost) was used to evaluate the robustness of the other models. In pairwise comparisons of AUROC between these models and the logistic regression model, XGBoost significantly outperformed logistic regression for heart failure and CAD prediction.

Outcomes and Implications

Cardiologists today work to identify cancer survivors at-risk for developing cardiovascular disease. AI models have the ability to identify patterns in clinical data and make predictions based on cardiovascular risk, thus helping clinicians personalize cancer survivor’s care. The numerous AI models today all provide their own strengths and weaknesses. Like a decision tree model, some algorithms can predict outcomes while also identifying the variables that most influenced their decision, giving further insight to clinicians about potential risk factors. Additionally, this study detailed the demographics and epidemiology of each patient, which may provide valuable insight into outcome differences among different demographics and carves the path to develop further specialized risk recognition and treatment.