Hasso-Plattner-Institut
Prof. Dr.-Ing. Bert Arnrich
  
 

Active Learning in Personalized OCD Recognition

Master's Thesis

Fabian Stolp, Supervisor: Kristina Kirsten

Obsessive–compulsive disorder (OCD) is a mental disorder, which can lead to a high degree of suffering. At the same time, it is quite frequent and diagnosed in at least two percent of the world population [1]. Recently, more research is being conducted to leverage digital technology for diagnosis, treatment, and follow-up of this medical condition [2]. In this context, recognizing compulsions using sensor data from wearable devices could give valuable insights to the patient as well as the therapist. For example, it could help to quantitatively evaluate how well a therapy works. Another use case could be to inform a patient when during the follow-up of a therapy, compulsions, which seemed to be already cured, reappear. Though, compulsions that occur in OCD are manifold [3]. Therefore, creating a generic recognition system that works well for every individual patient is very difficult. Thus, the personalization of such a system for the single patient is one of the major challenges in this field.

For our use case, we want to be able to break down the diversity of compulsions by confining the activities, which are to be recognized as compulsions, to a small set that is personalized to the single patient. Therefore, we need to supervise the learning algorithms and need to be able to determine how the activities look like, which they are supposed to detect.

When supervised machine learning methods are used, there is a need for labeled data. Getting these labels can require a lot of time and effort if they were not recorded directly during the data collection. Fortunately, whether a compulsion occurred or even if it is going to occur is something an OCD patient can tell, as one of OCDs key properties is that a patient is aware of their own compulsions [3]. Therefore, a recognition system could learn from the patient how their compulsions look in the sensor data by asking them to state at specific points in time whether they experienced one or not. With that, the system could get the information to label the data right away.

In this master thesis, we want to study how current research in online, active learning for HAR can be efficiently and unobtrusively applied to the area of OCD recognition. The key idea of such an online active learning system is depicted in the following figure.

References

[1] Wayne K. Goodman et al. Obsessive-compulsive disorder. 2014. DOI: 10.1016/j.psc.2014. 06.004.

[2] Florian Ferreri et al. “How new technologies can improve prediction, assessment, and in- tervention in obsessive-compulsive disorder (e-ocd): Review”. In: Journal of Medical Internet Research 21.12 (2019), pp. 1–15. ISSN: 14388871. DOI: 10.2196/11643.

[3] Jonathan S. Abramowitz, Steven Taylor, and Dean McKay. “Obsessive-compulsive disor- der”. In: The Lancet 374.9688 (2009), pp. 491–499. ISSN: 01406736. DOI: 10.1016/S0140- 6736(09)60240-3. URL: dx.doi.org/10.1016/S0140-6736(09)60240-3.