A number of studies have been investigating the use of mobile phone sensing to predict mood in unipolar (depression) and bipolar disorder. However, most of these studies included a small number of people making it difficult to understand the feasibility of this method in practice. This paper reports on mood prediction from a large (N = 129) sample of bipolar disorder patients. We achieved prediction accuracies of 89% in personalized machine learning models that were trained for each individual patient.
These result were published at the ACM International Joint Conference on Ubiquitous and Pervasive Computing in Singapore.
This paper reports data from the MONARCA II Randomized Clinical Tria (RCT) involving 129 Bipolar Disorder patients. Using smartphone sensor data coupled with self-reported daily mood collected via the Monsenso system, we developed and evaluated predictive machine learning models tuned to discriminate a) euthymic from depressive and manic states, and b) euthymic from depressive states. We applied two types of models, one that rely on prior knowledge about the patient (i.e. requiring a user to collect data for several days before calibrating the model) and one that does not require any inputs from the patient. We called these two models personalized and generic, respectively. We found that personalized models outperformed the generic models, which performed close to a baseline.
Detailed prediction scores are listed in the table below and details can be found in the paper.
Reference