A metabolomics-driven machine learning model for osteoporosis risk prediction.

Area of Science:

• Metabolomics
• Machine Learning
• Bone Health Research

Background:

• Osteoporosis is a growing public health concern in aging populations, characterized by reduced bone mineral density (BMD) and increased fracture risk.
• Current machine learning (ML) models for osteoporosis prediction using clinical data have limitations in accuracy and generalizability.
• Metabolomics, crucial for bone metabolism, has been underexplored in osteoporosis risk prediction.

Purpose of the Study:

• To identify novel metabolites associated with osteoporosis risk.
• To assess the predictive utility of these metabolites, particularly when integrated with clinical data.
• To develop an enhanced ML model for osteoporosis risk prediction using metabolomic and clinical data.

Main Methods:

• Utilized hip BMD measurements, clinical data, and metabolomic profiles from 2041 participants (aged ≥40).
• Developed and evaluated ML models, comparing a clinical-only model with one integrating metabolomic data.
• Employed receiver operating curve (AUC) analysis for model performance evaluation and feature importance analysis for predictor identification.

Main Results:

• Identified 44 metabolites significantly associated with osteoporosis risk (p < .05), with 25 previously linked to bone metabolism.
• The integrated model (clinical + metabolomics) showed improved predictive performance (AUC = 0.763) compared to the clinical-only model (AUC = 0.741, p = .017).
• Key predictors included clinical factors (weight, grip strength, height, age, sex) and metabolites (e.g., N-acetylcarnosine, hypotaurine, homoarginine).

Conclusions:

• A distinct metabolomic signature for osteoporosis risk was identified.
• Integrating metabolomics with clinical data significantly enhances ML-based prediction accuracy for osteoporosis.
• These findings support the use of metabolomics for earlier osteoporosis detection and personalized management in aging populations.