XGBoost-SHAP-based interpretable diagnostic framework for early cognitive impairment in type 2 diabetes mellitus

Abstract

Abstract Objective To develop and validate a radiomic-clinical model to assess early cognitive impairment in type 2 diabetes mellitus (T2DM) using the XGBoost algorithm. Methods We retrospectively enrolled 193 patients with T2DM from two medical centers. According the Montreal Cognitive Assessment (MoCA), patients were categorized into normal control (NC) and mild cognitive impairment (MCI) groups. We used ComBat to normalize and gather the data distributions of two centers. The Elastic Net Regression were used to filter redundant and irrelevant features. Based on the eXtreme Gradient Boosting Machine algorithm (XGBoost), clinical factors and radiomic features was used to construct the combined model. The SHAP method explained the model by prioritizing the importance of features, in terms of assessment contribution. Results The radscore, along with two clinical factors (education level and drinking), were used to build the combined model. The AUCs for predicting MCI in the training set, testing set, and validation set were 0.802, 0.817, and 0.852, respectively. The radscore was the most important feature for discriminating MCI/NC classification, and higher SHAP values of radscore were associated with a higher risks of MCI onset. Conversely, higher SHAP values of education level and drinking were associated with a lower risks of MCI onset. However, the contribution of drinking to the model was minimal. Conclusion The radiomic-clinical model, utilizing the XGBoost algorithm, can be an auxiliary tool for predicting early cognitive impairment in T2DM.