Improving predictive performance in incident heart failure using machine learning and multi-center data.

Area of Science:

• Cardiology
• Biostatistics
• Computational Health

Background:

• Primary prevention of heart failure (HF) is crucial for mitigating its societal burden.
• Early risk stratification for HF necessitates advanced computational methods, including machine learning (ML), to analyze complex individual patterns in large datasets.

Purpose of the Study:

• To compare the predictive performance of incident HF risk models.
• To evaluate flexible ML models against linear models.
• To assess models trained on single-center versus multi-center data.

Main Methods:

• Utilized meta-data from 30,354 individuals across 6 cohorts with a median follow-up of 5.40 years.
• Evaluated survival gradient boosting (SGB), CoxNet, PCP-HF risk score, and a stacking method.
• Employed iterative predictions with one cohort as the external test set and remaining cohorts as the training set (single- or multi-center).

Main Results:

• Multi-center models consistently outperformed single-center models.
• SGB achieved a higher c-index (0.735) than CoxNet (0.694) in the pooled population.
• A stacking method integrating multiple models achieved the highest PR/AUC (0.804) for 10-year HF risk prediction, significantly outperforming PCP-HF (0.551).

Conclusions:

• Training models on diverse, multi-center data enhances their ability to learn a wider range of individual health characteristics.
• Flexible ML algorithms are effective in capturing diverse data distributions for more precise HF prediction models.