| Objective: To explore the interrelationship between brain functional networks and features in functional magnetic resonance imaging (fMRI) of patients with Alzheimer''s disease (AD), and to construct mixed-function networks (MFN), and apply them in machine learning classification models to improve the accuracy of AD classification. Methods: 102 AD patients and 227 healthy subjects in the Alzheimer''s Neuroimaging Initiative (ADNI) dataset were retrospectively analyzed. The partial correlation brain network of the blood oxygen level dependent (BOLD) signal was calculated and fused with low-frequency wave amplitude (ALFF), fractional low-frequency wave amplitude (fALFF) and local consistency (ReHo) features to construct MFN. Network topology parameters were extracted, and a variety of machine learning classification models were constructed based on MFN topological parameters, accuracy, precision, recall and area under the curve (AUC) were used to evaluate the predictive efficiency of the models. Results: By constructed MFN and calculated intra group to inter group ratio (IIGR), 35 features could be obtained from ALFF, fALFF and ReHo feature topological parameter analysis, after rank sum test and FDR correction, there were statistical differences among 28 features (P<0.05). The classification results show that, all the five classifiers have high classification performance on the test data set. The accuracy, precision and recall rates of random forest (RF), adaptive lifting algorithm (AdaBoost), guided aggregation algorithm (Bagging) and support vector machine (SVM) were all 99.7%, and the AUC values were up to 100%, 99.5%, 99.1% and 99.5%, respectively. The accuracy (98.5%), precision (98.5%), recall (98.5%), and AUC (99.1%) of the multi-layer perceptron (MLP) were slightly lower than other models, but remained excellent. It was worth noting that RF has the highest AUC value of all models at 100.0%, while Bagging has the lowest AUC value (99.1%) in the integrated approach. The results of performance comparison show that, MFN classification model can significantly improve the recognition and classification of AD disease, and greatly improve the performance of various indicators of the classifier. The results showed that, MFN classification model was superior to intelligent classification based fusion, DBN-based multitask learning, PVT-TSVM, unsupervised learning and clustering, SVM and SVM of degree 3 polynomial kernel function in key indicators such as accuracy (99.13%), AUC (99.42%), recall rate (99.46%) and specificity (99.42%) With plasma proteins, machine learning algorithms. It was further proved that MFN classification model has good generalization ability and robustness in AD disease classification. Conclusion: The AD classification model constructed based on brain mixed function network topology parameters and machine learning can improve the accuracy of AD classification. |
