ABSTRACT
Objective:To explore the value of machine learning models based on multiple structural MRI features for diagnosis of Parkinson disease (PD).Methods:The clinical and imaging data of 60 PD patients (PD group) diagnosed in the Neurology Department of the Second Affiliated Hospital of Soochow University from November 2017 to August 2019 and 56 normal elderly people (NC group) recruited from the community were retrospectively analyzed. All subjects underwent brain MR imaging. Multiple structural MRI features were extracted from cerebellum, deep nuclei and of brain cortex based on different partition templates. The Mann-Whitney U test, as well as least absolute shrinkage and selection operator regression were used to select the most discriminating features. Finally, logistic regression (LR) and linear discriminant analysis (LDA) classifier combined with the 5-fold cross-validation scheme were used to construct the models based on structural features of cerebellum, deep nuclei and cortex, and a combined model based on all features. The receiver operating characteristic curves were drawn, and the diagnostic performance and clinical net benefit of each model were evaluated by the area under curve (AUC) and the decision curve analysis (DCA). Results:In total, four cerebellum (asymmetry index of Lobule Ⅵ volume, asymmetry index of Lobule ⅦB cortical thickness, asymmetry index of total gray matter volume and absolute value of right Lobule Ⅵ gray matter volume), 3 deep nuclei (absolute value of right nucleus accumbens volume, absolute and relative value of total nucleus accumbens volume) and 3 cortex features (local gyration index of left PFm, local fractal dimension of right superior frontal gyrus and sulcal depth of left superior occipital gyrus) were selected as the most discriminating features, and the related models were constructed. In validation set, the AUC of cerebellum, deep nuclei, cortex and combined models for diagnosis of PD based on LR classifier were 0.692, 0.641, 0.747 and 0.816; the AUC of cerebellum, deep nuclei, cortex and combined models for diagnosis of PD based on LDA classifier were 0.726, 0.610, 0.752 and 0.818. The diagnostic efficiency of the combined models based on LR and LDA classifiers were significantly better than those of other models ( P<0.05). The DCA curve demonstrated that the combined models based on LR and LDA classifiers showed the highest clinical net benefit. Conclusion:The combined models with all structural features of cerebellum, deep nuclei and cortex included based on LR and LDA classifiers showed favorable performance and clinical net benefit for diagnosis of PD, which have the potential application value in clinical diagnosis.