Machine learning for detecting Wilson's disease by amplitude of low-frequency fluctuation

Wilson's disease (WD) is a genetic disorder with the A7P7B gene mutations. It is difficult to diagnose in clinic. The purpose of this study was to confirm whether amplitude of low-frequency fluctuations (ALFF) is one of the potential biomarkers for the diagnosis of WD. The study enrolled 30 healthy controls (HCs) and 37 WD patients (WDs) to obtain their resting-state functional magnetic resonance imaging (rs-fMRI) data. ALFF was obtained through preprocessing of the rs-fMRI data. To distinguish between patients with WDs and HCs, four clusters with abnormal ALFF-z values were identified through between-group comparisons. Based on these clusters, three machine learning models were developed, including Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR). Abnormal ALFF z-values were also combined with volume information, clinical variables, and imaging features to develop machine learning models. There were 4 clusters where the ALFF z-values of the WDs were significantly higher than that of the HCs. Cluster1 was in the cerebellar region, Cluster2 was in the left caudate nucleus, Cluster3 was in the bilateral thalamus, and Cluster4 was in the right caudate nucleus. In the training set and test set, the models trained with Cluster2, Cluster3, and Cluster4 achieved area of curve (AUC) greater than 0.80. In the Delong test, only the AUC values of models trained with Cluster4 exhibited statistical significance. The AUC values of the Logit model (P = 0.04) and RF model (P = 0.04) were significantly higher than those of the SVM model. In the test set, the LR model and RF model trained with Cluster3 had high specificity, sensitivity, and accuracy. By conducting the Delong test, we discovered that there was no statistically significant inter-group difference in AUC values between the model that integrates multi-modal information and the model before fusion. The LR models trained with multimodal information and Cluster 4, as well as the LR and RF models trained with multimodal information and Cluster 3, have demonstrated high accuracy, specificity, and sensitivity. Overall, these findings suggest that using ALFF based on the thalamus or caudate nucleus as markers can effectively differentiate between WDs and HCs. The fusion of multimodal information did not significantly improve the classification performance of the models before fusion.