Synopsis

Keywords: Machine Learning/Artificial Intelligence, Parkinson's Disease

With the ageing of the population, Parkinson's disease (PD) has presented a severe challenge to public health. Here, a deep-learning framework named the AMDGM model was proposed to predict PD patients at an early stage. Firstly, multi-modal image-based models were respectively generated using the AMDGM model. Then, a weighted ensemble network was created as the final model. The proposed method achieved the best AUC performance of 0.872 in the testing cohort, better than others. And the proposed method can predict PD patients early to help clinical radiologists formulate more targeted treatments in the future.

Introduction

Parkinson's disease (PD) is the world's second most common neuropsychiatric disease¹. In practice, traditional PD diagnosis methods may be subjective since they depend on analyzing motions that are occasionally faint to human eyes and hence challenging to be categorized, potentially leading to misdiagnosis². Therefore, there is an urgent demand in predicting PD patients at the early stage in the clinic. Currently, Graph Convolutional Networks (GCN) in deep learning are widely concerned as they can effectively integrate multi-modal features and model the correlation between samples. Previous studies have shown the ability of GCN to classify PD and Alzheimer's disease (AD) on MRI with high accuracy^3-4. And our group proposed a deep-learning method to differentiate PD patients and healthy controls (HC), an ensemble-learning framework using a graph convolution neural network of differentiable graph module for multi-modal MRI data. The experiments showed that the proposed method could successfully achieve early diagnosis o PD patients based on multi-modal MRI data.

Methods

The network we use to present our approach is auto-metric DGM(AMDGM), whose structure is based on the DGM model⁵. The proposed model improves upon previously developed DGM designs, using the few-shot approach to achieve inductive learning of independent tests and imbalanced positive and negative data samples. An ensemble technique was employed to integrate T1, DTI, and fMRI modalities data. In this way, it received prediction scores from the image-based models as inputs and formed outputs by the weights of the prediction scores.
The DGM model (Figure 1A) proposes a generic method for learning the graph based on the characteristics of each layer's output. It takes the feature matrix X as input and yields a graph G as output, and the initial adjacency matrix A is optional as it can be defined following the domain knowledge or directly following the network using the Gumbel-Top-k trick⁶. To address the issue of imbalanced data samples, inspired by AMGNN model⁷, we proposed a new model AMDGM (Figure 1B). In this approach, a tiny graph created by randomly chosen samples serves as the input, and the labels of unknown nodes in the graph structure are obtained through multiple AMDGM layers, and then the training and testing strategy was applied (Figure 1C). Samples are initially chosen randomly from the dataset to serve as the model's input X. It contains q known samples per category and one unknown sample whose label is defined as Y:
$$X=\left\{ \left\{ \left( x_1,c_1 \right) ,...,\left( x_{\left( N-1 \right)},c_{\left( N-1 \right)} \right) \right\} ,\left\{ \overline{x} \right\} ;c_i∈\left\{ 1,C \right\} \right\} $$Where x is the sample, $$$\overline{x}$$$ is the unknown sample, c is the label, C is the number of categories, and $$$N=C∙q+1$$$.
Then the matrix X was put into DGM model as the initial input and the label of the unknown node is obtained. The parameters of the network can be updated through supervised learning. The updated parameters serve as as the initial parameters for training the subsequent new iteration. Finally, the AMDGM architecture was used for prediction challenges in a single input modality (Figure 2A) and directly integrated the three input modalities using an ensemble-learning approach for information fusion (Figure 2B).
The Parkinson’s Progression Markers Initiative (PPMI) database includes 123 PD patients and 41 HC, with T1, DTI, and fMRI modalities data⁸ (Table 1), and the protocols were approved by our Institutional Reviews Board (IRB). Data were randomly split into a training dataset (n = 114) and an independent test dataset (n = 50). All T1, DTI, and fMRI modalities data are individually preprocessed by the software of SPM, PANDA, and DAPBI^9-11. Then the preprocessed images are segmented into AAL-116 template¹² and used the REF algorithm for feature selection¹³ due to the high dimensionality of features.
The Adam optimizer and the binary cross-entropy loss function were implemented to optimize the network, with the learning rate set to 0.001 and batch size of 4 for 10 epochs and iterations were 100. The number of GCN layers used was 2, and the number of q was 10. The weights were $$$W_1=0.2,W_2=0.3,W_3=0.5$$$. The following metrics were calculated to evaluate performance using a five-fold cross-validation procedure: AUC, accuracy, sensitivity, and specificity.

Results

Performance of the single input modality of AMDGM and proposed AMDGM on the dependent test set described in Table 2. The best classification results were obtained for the proposed AMDGM model (T1+DTI+fMRI modalities) with an AUC of 0.872 in the testing cohort. Figure 3 shows the most high-frequency 20 features extracted by the REF algorithm. Figure 4 shows the ROC curve for the single input modality of AMDGM models’ performance on the test set. Figure 5 shows the ROC curve for our proposed AMDGM models’ performance on the test set.

Discussion and Conclusion

In this study, the proposed AMDGM model achieved the best AUC performance of 0.872, indicating that integrating multi-modal data on MRI scanners is better than a single MRI modality. And this model can rectify the imbalanced data sample issue. The current test results indicate that the proposed method could be an efficient, reliable way to predict PD patients at an early stage. More significant sample sizes from more than one site are preferred in future investigations.

Acknowledgements

This work was partially supported by the National Natural Science Foundation of China (61871373, 62271474, 81830056, U1805261, 81729003, 81901736, 12026603, 12026603 and 81971611), the Strategic Priority Research Program of Chinese Academy of Sciences (XDB25000000 and XDC07040000), the High-level Talent Program in Pearl River Talent Plan of Guangdong Province (2019QN01Y986), the Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province (2020B1212060051), the Science and Technology Plan Program of Guangzhou (202007030002), the Key Field R&D Program of Guangdong Province (2018B030335001), the Shenzhen Science and Technology Program, Grant Award (JCYJ20210324115810030), and the Shenzhen Science and Technology Program (Grant No. KQTD20180413181834876, and KCXF20211020163408012).