Comparing Machine Learning Models and Statistical Models for Predicting Heart Failure Events: A Systematic Review and...

Area of Science:

• Cardiology
• Medical Informatics
• Biostatistics

Background:

• Machine learning (ML) models are increasingly proposed for medical applications, yet their comparative performance against statistical models for predicting heart failure (HF) events remains underexplored.
• This study systematically reviews and compares the predictive performance, clinical feasibility, and reliability of statistical and ML models for HF event prediction.

Approach:

• A systematic literature search identified studies published between 2011 and 2021 that developed or validated statistical or ML models for predicting all-cause mortality or readmission in HF patients.
• The Prediction Model Risk of Bias Assessment Tool assessed study quality, and random-effects meta-analysis pooled model performance using c-statistics.

Key Points:

• Included 202 statistical and 78 ML model studies. Pooled c-indices for ML models in predicting mortality (0.777) were not consistently superior to statistical models (0.733).
• For readmission prediction, statistical models (0.678) outperformed ML models (0.660). Head-to-head comparisons yielded similar results, indicating no significant ML advantage.
• ML models faced limitations due to excessive predictor use, hindering clinical feasibility. Risk of bias analysis revealed more technical pitfalls in ML models compared to statistical models.

Conclusions:

• ML models do not demonstrate a significant advantage over statistical models for predicting HF events.
• The clinical feasibility and reliability of ML models in HF prediction are currently inferior to statistical models.
• Further research is needed to clarify the efficacy of ML models across different HF subgroups and address technical limitations.