PREDICTIVE PERFORMANCE OF MACHINE LEARNING MODELS FOR DETECTION OF INCIDENT HEART FAILURE USING MULTI-CENTRE DATA.

OBJECTIVE: To mitigate the health and economic burden associated with heart failure (HF), primary prevention is of utmost importance. To effectively address the diverse range of individual health profiles, prediction models must be trained and validated on a wide range of population-based cohorts. We evaluated the predictive performance of Survival Gradient Boosting (SGB) and Coxnet (CN) for the detection of incident HF. In addition, we compared the models' performance trained only on a single-centre cohort with those trained on diverse multi-centre cohorts. DESIGN AND METHOD: We utilized 6 cohorts from the Heart 'omics' in AGEing (HOMAGE) project to evaluate two computational approaches. The dataset consists of 31,382 individuals (mean age 66 ± 10 years old), free of HF at the baseline. During a median follow-up of 5.36 years, 1,130 non-fatal HF events occurred. SGB is a non-linear machine learning method based on training regression trees with the objective of optimizing Cox partial likelihood. Coxnet is a standard linear Cox proportional hazard model, regularized by both L1 and L2 norms. To estimate the predictive accuracy of a single-centre model on a specific test cohort, we averaged the performance of single models trained on remaining cohorts one at a time. To develop models in multi-centre settings, we used 5 cohorts in every iteration as a training set and the remaining 1 cohort as a test set. We used 33 features, including clinical, ECG, and biochemical data. RESULTS: Overall, multi-centre models achieved a higher c-index than those trained on single cohorts (Figure). For example, in the FLEMENGHO cohort, multi-centre SGB achieved 0.819, while only 0.664 in a single-centre setting. Additionally, non-linear SGB achieved a considerably higher c-index than linear CN in all but one cohort, especially in multi-centre settings, for example in the HVC cohort (SGB 0.689, CN 0.614). CONCLUSIONS: With a greater number and variety of training cohorts, the model learns a wider range of specific individual health characteristics. Flexible machine learning algorithms such as gradient boosting can be used to capture these diverse distributions and produce more precise, personalized models for the prediction of adverse events.

Publication year:2022

Accessibility:Closed