Machine Learning for Smartphone-based Monitoring and Treatment of Unipolar and Bipolar Disorders

Abstract

Bipolar disorder is a common mental illness characterized by unusual changes in mood and energy and is regarded as one of the most important causes of disability worldwide. Smartphones provide a unique platform for unobtrusive disease monitoring and management and are almost constant companions to their users. By replacing traditional paper-based self-assessments with a smartphone-based system, users can unobtrusively collect and monitor their own data. Modern smartphones additionally enable pervasive collection of detailed objective data that can track a wide range of human behaviors relevant to monitoring mental illness. Digital data collection has the additional advantage of making data available for immediate analysis by humans and computers, which can support disease monitoring and treatment tasks in and between outpatient consultations at their treatment center. Automated analysis of smartphone-based data can potentially detect early warning signs and predict disease outcomes, which can facilitate early intervention and thus potentially mitigate severity of affective episodes and prevent costly hospitalizations. The overall objective of the PhD has been to establish methods and algorithms for analysis of behavioral smartphone data from patients with bipolar disorder aiming at pattern recognition and prediction of recurring depressive and manic symptoms by applying data mining and machine learning techniques. This report presents a summary of research conducted during the author’s PhD studies and includes three research manuscripts. We show that by applying hierarchical Bayesian regression models we are able to forecast subjective mood up to seven days ahead based on short self-assessment histories. Using the same hierarchical modelling approach, we can produce daily estimates of clinical severity ratings of depression and mania from self-assessments. We also show how to utilize uncertainty in the estimated severity ratings to compute individual scores indicating risk of relapse. Finally, we show how simple features of objective smartphone data can discriminate between patients with bipolar disorder and healthy control individuals during different affective states. Based on the current research, we are confident that predictive analysis based on data collected with smartphones has the potential to improve disease monitoring and treatment in bipolar disorder. To accomplish this goal, predictions must be accurate, interpretable and actionable. Modern machine learning techniques have proved unparalleled in revealing complex patterns and achieving high predictive accuracy in a wide range of domains. To unlock the potential of advanced methods and drive research forward, detailed datasets from large groups of patients are needed.