Development and Validation of a Predictive Model for Glycemic Response to Dorzagliatin in Type 2 Diabetes: A Retrospective Real-World Study

Abstract

Patient response to dorzagliatin, a novel glucokinase activator for type 2 diabetes mellitus (T2DM), exhibits significant heterogeneity in clinical practice, posing a challenge for precision medicine. This study aimed to develop and validate a predictive model for identifying optimal responders prior to treatment initiation. We conducted a retrospective cohort study involving 300 T2DM patients hospitalized between April 2024 and April 2025. Patients were categorized into a responder group (n=146, fasting/postprandial glucose reduction >0.5 mmol/L) and a non-responder group (n=154, reduction <0.3 mmol/L). Multivariable logistic regression and random forest machine learning algorithms were employed. Key predictors identified included diabetes duration, HOMA-β index, and use of GLP-1 receptor agonists and metformin. The logistic regression model demonstrated high predictive accuracy (87.0%), while the random forest model achieved an AUC of 0.99 in the test set. A comprehensive model integrating the most significant variables yielded a superior AUC of 0.900. This clinically applicable prediction model facilitates the pre-treatment identification of patients most likely to benefit from dorzagliatin, thereby advancing personalized T2DM management.