Multi-task learning based hyper-network construction for MCI classification in Parkinson’s disease

The recognition of mild cognitive impairment(MCI) is crucial for the prevention and treatment of dementia in Parkinson’s disease. Most of the current methods for identifying MCI are based on neuropsychological assessments, which have a certain degree of subjectivity. The use of graph theory methods can quantitatively analyze the variability of patients' brains as well as identify patients with MCI. However, the brain involves multiple brain regions working together during activity, which traditional graph theory cannot represent. This paper proposes a sparse representation method based on multi-task learning to construct a fused brain hypernetwork using processed functional near-infrared spectroscopy(fNIRS) time series signals from both congruent and incongruent tasks of healthy controls and Parkinson’s disease patients with mild cognitive impairment (PD-MCI). In the hypernetwork, more than two nodes which interact with each other will be connected in a hyperedge. High-order functional connections between brain regions were extracted from the fused brain hypernetwork as input features to build a disease classification model. The results indicate that there are functional differences in the brain between PD-MCI patients and healthy controls(HC), particularly in the prefrontal cortex. Furthermore, our proposed method outperformed single-task based hypernetwork construction methods in terms of classification performance between PD-MCI patients and healthy controls.