UCSF

The Diagnostic and Statistical Manual (DSM) has served as the gold standard for psychiatric diagnosis for the past several decades in the USA, and DSM diagnoses mirror mental health and substance abuse diagnoses in ICD-9 and ICD-10. However, DSM diagnoses have severe limitations when used as phenotypes for studies of the pathophysiology underlying mental disorders, as well as for clinical treatment and research. In this paper, we use a novel approach of deconstructing DSM diagnostic criteria, and using expert knowledge to inform feature selection for unsupervised machine learning. We are able to identify clusters of symptoms that stratify subjects with the same DSM disorders into cohorts with increased clinical and biological homogeneity. These findings suggest that itemized self-report symptom data should inform a new taxonomy for psychiatry, and will enhance the bi-directional translation of knowledge from the bench to the clinic through a common terminology.