Marvin Mirtschin, Juan Carlos Niño Rodriguez, Anne Radunski, Julius Rudolph, Fabian Stolp
Obsessive-compulsive disorder (OCD) is an often underestimated mental disorder in severity as well as in incidence. It is characterized by obsessions, which are intrusive thoughts that lead to the execution of compulsive actions. Digital approaches could support diagnosis, treatment as well as follow-up of the disease. Though, at the moment, they are mostly limited to digital questionnaires.
In this master project, sensor data from wearables and indoor positioning data were used to predict the occurrence of compulsions with machine learning techniques. The motivation behind the approach is to help patients as well as therapists by making the severity of the disorder quantifiable with an objective count of the number of occurrences of compulsions and an objective measurement of the time spent executing them. Other use cases are imaginable, such as online feedback to the patient when the system recognizes a compulsion's execution.
The master students created a protocol for the recording of simulated OCD activities. They collected 6 hours of data and video material from 11 healthy participants who followed the protocol. Afterward, the students annotated the collected data. The annotation was made possible through the video material, which showed the executed actions and was synchronized with the remaining collected data. Subsequently, the data was used to build machine learning models for classification. In the end, the performance of the models was evaluated. The data included acceleration and gyroscope data from smartphone sensors and indoor localization data in the form of Bluetooth signal strengths.
The analysis of the results showed that the overall concept is viable. High-quality predictions could be achieved when focusing on small sets of compulsions. This insight and an interview with psychologists from the university clinic for psychotherapy of the Humboldt University of Berlin showed the need for personalized classification models. There is a great diversity in which compulsions can occur in the overall population, while individual patients only execute a small set of different compulsions. Therefore, personalized machine learning models could be much more precise. At the same time, it could be shown that the quality of predictions strongly improved when considering not only acceleration and gyroscope but also indoor localization data.
