A Multi-omics Framework Based on Machine Learning as a Predictor of Cognitive Impairment Progression in Early Parkinson's Disease.

INTRODUCTION

Cognitive impairment (CI) is a common non-motor symptom of Parkinson's disease (PD). However, the diagnosis and prediction of CI progression in PD remain challenging. We aimed to explore a multi-omics framework based on machine learning integrating comprehensive radiomics, cerebrospinal fluid biomarkers, and genetics information to identify CI progression in early PD.

METHODS

Patients were first diagnosed with PD without CI at baseline. According to whether CI progressed within 5 years, patients were divided into two groups: PD without CI and PD with CI. Radiomics signatures were extracted from patients' T1-weighted MRI. We used machine learning methods to construct radiomics, hybrid, and multi-omics models in the training set and validated the models in the testing set.

RESULT: In the two groups, we found 7, 23, and 25 radiomics signatures with significant differences in the parietal, temporal, and frontal lobes, respectively. The radiomics model using the 25 signatures of the frontal lobe had an accuracy of 0.833 and an AUC (area under the curve) of 0.879 to predict CI progression. In addition, the hybrid model fused with the cerebrospinal fluid Aβ level had an accuracy of 0.867 and an AUC of 0.916. In our study, the multi-omics model showed the best predictive performance. The accuracy of the multi-omics model was 0.900, and the average AUC value after five-fold cross-validation was 0.928.

CONCLUSION

Radiomics signatures have a recognition effect in the CI progression in early PD. Multi-omics frameworks combining radiomics, cerebrospinal fluid biomarkers, and genetic information may be a potential predictor of CI progression in PD.