A rule-based prognostic model for type 1 diabetes by identifying and synthesizing baseline profile patterns

Abstract

Objective: To identify the risk-predictive baseline profile patterns of demographic, genetic, immunologic, and metabolic markers and synthesize these patterns for risk prediction. Research Design and Methods: RuleFit is used to identify the risk-predictive baseline profile patterns of demographic, immunologic, and metabolic markers, using 356 subjects who were randomized into the control arm of the prospective Diabetes Prevention Trial-Type 1 (DPT-1) study. A novel latent trait model is developed to synthesize these baseline profile patterns for disease risk prediction. The primary outcome was Type 1 Diabetes (T1D) onset. Results: We identified ten baseline profile patterns that were significantly predictive to the disease onset. Using these ten baseline profile patterns, a risk prediction model was built based on the latent trait model, which produced superior prediction performance over existing risk score models for T1D. Conclusion: Our results demonstrated that the underlying disease progression process of T1D can be detected through some risk-predictive patterns of demographic, immunologic, and metabolic markers. A synthesis of these patterns provided accurate prediction of disease onset, leading to more cost-effective design of prevention trials of T1D in the future.