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

Creating risk prediction models based on obesity type requires large datasets, which are often not available for rare obesity phenotypes. Available technologies do not allow for consideration of the variability between different phenotypes, making it difficult to translate clinical research to underrepresented populations. We propose the FewShotMetabolic (FSM) Framework, a parameter-efficient framework to enable the creation of individualized metabolic risk models using only 10 data points per obesity phenotype. With the FSM Framework, we can connect information about the different obesity phenotypes so that knowledge can be shared between them, while keeping the unique metabolic signature associated with each obesity phenotype through selective pathway fine-tuning. The FSM model demonstrated an accuracy of 87.3% (AUC = 0.923) for metabolic syndrome risk classification among the six different obesity subtypes and achieved learnable parameters equal to 17.2% of fine-tuned models. The FSM model also produced an RMSE of 14.2 mg/dL for estimating the 2-hour post-meal glycemic response (Pearson r = 0.876). Results from external cohort testing (N=2 cohorts) indicated an AUC of $>$ 0.89. Thus, the FSM Framework offers a practical way to apply precision obesity medicine and creates an opportunity to develop predictive models for populations that have historically been underrepresented.