Machine Learning-Based Plasma Protein Risk Score Improves Atrial Fibrillation Prediction Over Clinical and Genomic...

Area of Science:

• Cardiovascular Medicine
• Proteomics
• Machine Learning

Background:

• Clinical factors and polygenic risk scores offer moderate accuracy in predicting incident atrial fibrillation (AF).
• The potential of large-scale proteomic profiling to enhance AF risk estimation remains largely unexplored.

Purpose of the Study:

• To develop and validate a machine learning model (Pro-AF) utilizing serum protein levels for predicting incident AF risk.
• To compare the performance of the Pro-AF model against established clinical (CHARGE-AF) and genetic (polygenic risk score) risk prediction tools.

Main Methods:

• A machine learning model was trained on 32,631 UK Biobank participants using 121 serum protein levels.
• Model performance was evaluated using time-dependent area under the receiver operating characteristic curve (AUC) and net reclassification improvement in internal and hold-out test sets.
• Pro-AF was also assessed in a simplified 5-protein version.

Main Results:

• Pro-AF demonstrated superior discrimination for 5-year incident AF compared to CHARGE-AF and the polygenic risk score (AUC internal: 0.761 vs 0.719 vs 0.686; hold-out: 0.763 vs 0.702 vs 0.682).
• The model showed good calibration and provided substantial net reclassification improvement over CHARGE-AF.
• A simplified 5-protein Pro-AF model retained high discriminative value (AUC internal: 0.750; hold-out: 0.759).

Conclusions:

• A machine learning-based protein score (Pro-AF) effectively discriminates 5-year incident AF risk, outperforming clinical and genetic factors.
• Large-scale proteomic analysis holds promise for identifying individuals at high risk for AF, facilitating targeted screening and preventive interventions.