Development and internal validation of a parsimonious Cox model for FPG-defined incident prediabetes in a Chin
Early identification of individuals at high risk of prediabetes may support timely prevention. Prediction models for large-scale health examination settings should balance predictive accuracy, interpretability, and feasi
Early identification of individuals at high risk of prediabetes may support timely prevention. Prediction models for large-scale health examination settings should balance predictive accuracy, interpretability, and feasibility. The authors aimed to develop and internally validate a Cox proportional hazards-based model for 3-, 4-, and 5-year risk of incident fasting plasma glucose (FPG)-defined prediabetes, and to compare a parsimonious model with an expanded metabolic model. This retrospective cohort study included 100,738 adults without diabetes or FPG-defined prediabetes at baseline from 32 healthcare centers across China. Incident prediabetes was defined by follow-up FPG of 5.6–6.9 mmol/L. Twenty-three routinely collected clinical and biochemical variables, including derived metabolic indices, were evaluated. Cox-LASSO regression was used for variable prioritization. Model discrimination was assessed using time-dependent AUCs, calibration was evaluated using bootstrap internal validation, and sensitivity analyses were performed in five missForest-imputed datasets, with pooled estimates combined using Rubin’s rules. During a median follow-up of 3.1 years, 12,389 participants developed FPG-defined prediabetes. In the parsimonious model, age, BMI, and FPG were independently associated with risk; adjusted hazard ratios per 1-SD increase were approximately 1.32, 1.27, and 2.06, respectively. Apparent AUCs were 0.762, 0.757, and 0.719 at 3, 4, and 5 years, with nearly identical optimism-corrected AUCs. Calibration was acceptable, with an optimism-corrected slope of 0.999 and 5-year O/E ratio of 1.048. The expanded model showed statistically higher AUCs, but absolute gains were modest. Subgroup analyses suggested attenuated 5-year discrimination among older adults. The parsimonious model provided stable, interpretable prediction of FPG-defined incident prediabetes. External validation is required before routine implementation.