Mental state in clinical research is assessed subjectively through discussion and rating scales. Initial research confirms that objectively collected data from sensors can be a gamechanger in detection of episodes of bipolar disorder (BD). However, the key barriers to use sensors in BD monitoring remain open: (i) lack of easily adaptable computational methods for BD episodes prediction; (ii) lack of reliable benchmark datasets for training of the algorithms. Furthermore, in the majority of the state of the art, the episode prediction problem is stated as a supervised learning task and collecting numerous labels is almost infeasible in the BD monitoring context.
BIPOLAR has access to two large digital anonymised data sets already collected from sensors of BD patients which will guide the research and experimental development.
BIPOLAR aims at the development of highly novel computational intelligence methods for sensor-based, semi-supervised and uncertainty-aware prediction of BD episodes.
The objectives of the BIPOLAR are: (1) to develop and evaluate methods for sensor data aggregation and feature retrieval (2) to model uncertainty of depressive and manic symptoms; (3) to develop semi-supervised software prototype for prediction of shifts of patients’ mental state (4) to demonstrate the solution in two real-life use cases. BIPOLAR will use an agile approach in its research-based development including evaluation in specific psychiatric scenarios characterized with low sensor-data labeling percentage, high uncertainty of psychiatric labelling and high-variability between individuals.