Machine-learning-based cardiovascular mortality prediction using a cumulative PM_2.5 exposure metric in high risk...

Classical risk factors for cardiovascular disease (CVD) are well established. Although the association between air pollution (AP) and CVD is also well recognized, its integration into cardiovascular mortality risk prediction models remains limited, particularly in the context of machine learning approaches. We conducted a study within the prospective EP-PARTICLES cohort to quantify the improvement in CVD mortality risk prediction achieved by integrating environmental factors into prediction models using the 'ePM-years Index' which reflects the proportion of cumulative long-term PM_2.5 exposure exceeding a reference threshold relative to total exposure and advanced machine learning (ML) techniques. The study included 6935 patients (mean age 64.5 ± 10.1 years; 46% women) with a median observation period of 3140 days (IQR 2377-4169 days). We analysed 21 classical risk factors with CV death as the primary endpoint. The 'ePM-years Index' was calculated for each patient. The χ² test and Monte Carlo strategy selected the most discriminating predictors. ML techniques were applied to improve the accuracy of mortality risk predictions and to identify key risk factors. Among the patients, 2646 (38.2%) had prevalent CVD, and 927 (13.4%) died from CVD. Feature selection identified 14/21 (66.7%) most discriminative predictors. The best ML model without the ePM-years Index achieved the Matthews correlation coefficient (MCC) of 0.281 ± 0.021, correctly classifying 72.59% of patients. Adding the 'ePM-years Index' improved the ML model's performance to an MCC of 0.657 ± 0.027, correctly classifying 92.75% of patients. The Area Under the Curve for the ML model exploiting the 'ePM-years Index' ranged from 0.88 to 0.91. Long-term exposure to PM_2.5 was associated with higher CVD mortality risk in high risk groups. ML incorporating PM_2.5 demonstrated greater predictive abilities and clinical utility. Including environmental factors into CVD mortality risk scales improves risk stratification, and ML tools can help build simpler, more clinically useful models.Trial registration The study was registered at ClinicalTrials.gov (NCT05198492).