Abstract

Psychiatric health care relies heavily on self assessments and -monitoring. The treatment’s success of patients with bipolar disorder depends on an early detection of depressive and manic states. Objective behavioral measurements, representing relevant aspects of the mental illness, can be collected by smartphone applications.

This thesis investigates the application of changepoint detection algorithms to smartphone usage data, in order to recognize bipolar state changes. We pursue the goal to autonomously learn the user behavior, and detect changes therein, without requiring any manual configurations, user inputs or learning targets. This enables the application of the approach to new, unseen data.

We introduce a change detection process consisting of data collection and -processing, feature selection and -extraction, changepoint detection and evaluation. Expert knowledge is used to select disease relevant measurements, while unsupervised feature selection methods are used to further narrow the subset down to user relevant features. The changepoint detection algorithms ChangeFinder and Bayesian Online Changepoint Detection are implemented and evaluated according to the objectives of this thesis. Results are presented on the basis of three datasets. Hypothetical data is used to evaluate the algorithmic performance according to different use cases, such as outliers and changepoints of various causes and amplitudes. We show the detection of incisive events in smartphone data of ordinary users, among them a ligament rupture and a changing relationship status. Usage data of bipolar patients, which originates from an ongoing clinical study, is used to demonstrate the detection of bipolar state changes. We conclude, that the Bayesian Online Changepoint Detection algorithm is better suited for our objectives. It incorporates prior knowledge about the domain. Thus it gives us the possibility to further improve the results and configure the algorithm to find certain types of changes, such as changes between bipolar states. No manual parameter selection is required and the performance on our data is promising. However, further evaluation on bipolar usage data is required.