Data-driven Diagnosis in Psychiatry

Abstract

Diagnosis of mental disorders is a symptom-based approach in current society, causing high prevalence and co-morbidity in the field of mental healthcare. Improving treatment has called for adopting machine learning. One of these machine learning techniques is cluster analysis that aid in stratifying patients to discover sub-types or increased symptom severity. In mental healthcare, research with cluster analysis is limited to single algorithmic approaches, and machine learning is often overlooked in terms of applying it adequately. Thus, researchers often face the dilemma of choosing an appropriate clustering algorithm, meanwhile, machine learning steps are neglected. In this study, a meta-algorithmic model (MAM) is developed that is based on the cluster ensemble. It guides researchers in performing machine learning correctly and alleviates the problem of choosing a clustering algorithm. Moreover, the cluster ensemble does not force researchers to optimize algorithms, allowing standard algorithms to be used, meanwhile, it performs equal to or better than an optimized single clustering algorithm. We evaluated the MAM on several synthetic and real datasets. Afterward, our cluster ensemble is utilized on actual data from the Psychiatry Department of the UMCU, consisting of DSM-IV diagnosed patients with schizophrenia or psychosis and a HoNOS assessment. We identified three distinct clusters that separate patients on the severity of symptoms. With one cluster showing increased severity for patients in their social environment due to correlation between positive and negative symptoms.