Machine learning of connectome-wide functional connectivity in spinal muscular atrophy

While traditional view posed SMA as a pure motor neuron disorder, accumulating evidence pointed to the broader involvement of central nervous system. This study aimed to combined connectome-wide resting-state functional connectivity with machine learning algorithms to investigate the underlying brain mechanisms in SMA and capture the potential neuroimaging biomarker. Resting-state functional magnetic resonance imaging were acquired from 43 SMA (26 SMA type 2 and 17 type 3) and 45 healthy controls (HCs). Connectome-wide functional connectivity, defined based on the Automated Anatomical Labeling 116 atlas (AAL116), combined with machine learning algorithms including logistic regression (LR), support vector machine (SVM) and random forest (RF) were conducted for SMA vs HCs and SMA type 2 vs 3 classification. SHapley Additive exPlanations (SHAP) algorithm was used to quantify the contribution of individual connections. Connectome-wide functional connectivity allowed individual classification of SMA and HCs with stable performance across machine learning models. The majority of the most discriminative functional connectivity involved the high-level associative regions and cerebellum, subserving sensory and motor processing. Furthermore, the functional connectivity between right superior parietal gyrus and right inferior temporal gyrus was positively correlated with motor function. In conclusion, our study provided connectome-wide evidence that functional connectivity patterns involving high-level associative cortices and the cerebellum could distinguish SMA from HCs using machine learning. The identified connectivity features offer new insight into the distributed brain pathology of SMA beyond the traditional classical motor neuron view, and may serve as promising neuroimaging marker for this disease.