酰胺质子转移成像核团特征对帕金森病的诊断价值:基于可解释性机器学习OA

Objective A three-classification model based on APT imaging radiomics was constructed to accurately distinguish normal controls(NM),early-stage and middle-to-late-stage Parkinson's disease(PD).Through explainability analysis,the diagnostic value of key brain regions and features was clarified.Methods A total of 99 subjects from the Second Affiliated Hospital of Xinjiang Medical University from January 2022 to January 2024 were retrospectively enrolled,including 37 healthy controls,36 patients with early-stage PD,and 26 patients with advanced-stage PD.All subjects underwent brain APT sequence scanning.Six brain nuclei,caudate nucleus(CN),putamen(PUT),globus pallidus(GP),red nucleus(RN),substantia nigra(SN),and nucleus accumbens(NAc),were manually segmented to extract 107 radiomic features.Key features were selected to construct diagnostic models using six machine learning algorithms.Model performance was evaluated using area under the receiver operating characteristic curve(AUC)and accuracy.Shapley additive explanations(SHAP)analysis was employed to decipher model decision logic and quantify feature contributions across disease stages.Results After three-step feature screening,15 key radiomics features were identified.The combined LR model demonstrated optimal performance:training set macro-AUC=0.889(95%CI:0.827-0.943),micro-AUC=0.895(95%CI:0.837-0.946);The test set macro-AUC was 0.859(95%CI:0.707-0.975)and micro-AUC was 0.854(95%CI:0.704-0.967),significantly outperforming other models.SHAP analysis revealed key feature contribution patterns across PD stages:SN and RN features were critical for early-stage and normal classification,while GLCM autocorrelation coefficients of the PUT nucleus and RN features were core contributors for mid-to-late stage classification.Conclusion The LR combined model based on APT radiomics effectively achieves PD three-category diagnosis and staging.SN,RN,and PUT nuclei serve as core imaging biomarkers for PD pathological progression.SHAP analysis clearly elucidates the model's decision-making mechanism,providing an imaging tool for PD precision diagnosis and treatment that combines performance with interpretability.