Risk prediction of patients with severe mental disorder using machine learning

Abstract

An Electronic Health Record is a digital database comprising of all the information related to a patient’s medical history and treatment journey. It is a digital record of all the clinical activities, medications and lab tests that the patient has undergone. All the details regarding treatment costs are also available in an Electronic Health Record. They have immense advantage in the field of healthcare offering detailed insights into a patient’s medical background enabling the healthcare advisors and physicians to deliver quality care to the patient. However, healthcare costs are rising worldwide, including treatment costs , hospital and ICU stays, and medication expenses. In order to address this challenge it is essential to develop predictive models that can forecast the probable treatment path and also the future potential risk associated with a patient’s disorder. This can also help to predict the future treatment expenses In this thesis, we propose a machine learning approach to predict the risk of patients suffering from severe Mental Disorder with limited medical history available in Electronic Health Record. Here, we aim to predict the future risk and possible clinical activities of newly admitted Mental Disorder Patients, using the records of existing Mental Disorder patients available in Electronic Health Record. We have made use of Hierarchical DBSCAN ( Hierarchical Density Based Spatial Clustering of Applications with Noise) and Edit Distance to group the existing Mental Disorder Patients according to the similarity of their clinical activities. Markov Model have been used to test the interpretability of our approach and using Maximum Likelihood, we have assigned the new patient to the group with highest likelihood. Then, we have predicted future medications, diagnostics and lab tests for a certain number of days in future using Markov Model. We have compared our risk-prediction model with other state-of-the-art models. This comparative analysis suggests that our proposed algorithm has outperformed other architectures.