FewShotMetabolic: Parameter-Efficient Transfer Learning for Rapid Metabolic Risk Prediction in Data-Scarce Obesity...

Area of Science:

• Metabolic disease research
• Precision medicine
• Computational biology

Background:

• Developing metabolic risk prediction models for obesity is hindered by the need for large datasets, especially for rare obesity phenotypes.
• Current technologies struggle to account for inter-phenotype variability, limiting clinical research translation for underrepresented groups.

Purpose of the Study:

• To introduce the FewShotMetabolic (FSM) Framework, a parameter-efficient approach for creating individualized metabolic risk models.
• To enable accurate risk prediction using limited data (10 points) for diverse obesity phenotypes.
• To facilitate knowledge sharing between obesity phenotypes while preserving unique metabolic signatures.

Main Methods:

• Developed the FewShotMetabolic (FSM) Framework, a parameter-efficient model.
• Utilized selective pathway fine-tuning to maintain distinct metabolic signatures for each obesity phenotype.
• Enabled knowledge transfer across different obesity phenotypes.

Main Results:

• Achieved 87.3% accuracy (AUC = 0.923) in classifying metabolic syndrome risk across six obesity subtypes.
• Demonstrated a significant reduction in learnable parameters (17.2% of fine-tuned models).
• Estimated 2-hour post-meal glycemic response with an RMSE of 14.2 mg/dL (Pearson r = 0.876).
• Validated performance in external cohorts with AUC > 0.89.

Conclusions:

• The FSM Framework provides a practical solution for precision obesity medicine.
• Enables the development of predictive models for historically underrepresented populations.
• Advances the application of machine learning in metabolic disease research.