Metabolomics-Based Machine Learning for Predicting Mortality: Unveiling Multisystem Impacts on Health

Area of Science:

• Metabolomics
• Biomarkers
• Mortality Prediction

Background:

• Reliable predictors for long-term all-cause mortality in middle-aged and older populations are crucial.
• Previous metabolomics studies were limited by small sample sizes, limited metabolite measurement, and conventional statistical methods.

Purpose of the Study:

• To identify novel metabolites associated with all-cause mortality using advanced techniques.
• To assess the predictive value of identified metabolites for mortality risk.

Main Methods:

• Utilized liquid chromatography-tandem mass spectrometry to measure over 1000 metabolites in 10,197 men from the METSIM study.
• Applied three machine learning methods (logistic regression, XGBoost, Welch's t-test) alongside conventional statistics.
• Performed Cox regression analyses to evaluate hazard ratios and confidence intervals.

Main Results:

• Identified 32 impactful metabolites associated with all-cause mortality (25 increasing, 7 decreasing risk), including 20 novel ones across various pathways.
• These 25 metabolites improved all-cause mortality prediction beyond clinical and laboratory risk factors (HR 1.89 vs 1.76).
• Found 13 metabolites associated with increased cardiovascular disease mortality risk, but none with cancer mortality.

Conclusions:

• Novel metabolites identified significantly enhance the prediction of all-cause mortality in men.
• Metabolomics, particularly with machine learning, offers a powerful approach to discover mortality biomarkers.
• The findings have implications for cardiovascular disease risk stratification and understanding mortality pathways.