ABSTRACT

Symptoms overlap among different types of psychiatric disorder, which makes diagnosis challenging. There has been evidence that exploring new biotypes may help solve the problem. Using neuroimaging data, clustering technique has been applied to identify biotypes. however deep learning has not been well studied for this goal. Here, we propose a new semi-supervised deep clustering method, which takes full advantage of slight amount of labeled samples to adaptively extract useful information from unlabeled samples with high confidence level to facilitate the semi-supervised clustering. We validate our method using the Fashion MNIST data and then apply it to 137 patients with autism spectrum disorder (ASD) and 137 patients with schizophrenia (SZ) to explore psychiatric biotypes by using functional magnetic resonance imaging (fMRI) data. Compared with traditional methods, our method obtains better clustering performance for the Fashion MNISTdata and identifies interesting transdiagnostic biotypes in ASD and SZ.In summary. our proposed semi-supervised deep clustering method is promising in extracting biotypes of psychiatric disorders.