ADAPTIVE MONITORING OF DEPRESSION PATIENT POPULATION: A SELECTIVE SENSING APPROACH
30 million Americans use antidepressant medication. Inadequate follow-up monitoring has been identified as a main challenge in managing the depression patient population. This study aims to create patient-specific adaptive monitoring schedules and dynamically allocate limited sensing resources to detect high risk individuals of severe depression.
The proposed method integrates depression trajectory modeling, prognosis, and selective sensing into a unified framework. We characterized individuals’ depression progressions as Markov chains and predicted the risk to severe depression using similarity-based collaborative modeling (SCM) by exploiting the latent trajectory-pattern structure embedded in the population and incorporating the similarity information between individuals. We formulated the selective sensing as an optimization problem to optimally allocate the sensing resources to detect a subgroup of high risk individuals in each monitoring period. Sensing results were further incorporated into the SCM to update the prognosis for all individuals.
We applied the proposed method to monitor 610 subjects with ongoing depression treatment from the Mental Health Research Network dataset. Depression severity is assessed by Patient Health Questionnaire (PHQ)-9 scores over time. We ran the algorithm to adaptively monitor the 610 subjects over 15 time periods (each consist of two weeks) under a sensing capability of 100 patients in each period. The adaptive method was compared with two alternative monitoring strategies: individual and cohort-based monitoring. Individual monitoring predicted depression progression independently while cohort-based monitoring assumed subjects in the same trajectory-pattern group have the same progression model. Both individual and cohort-level monitoring selected the top-ranked 100 individuals based on their predicted risks.
The comparison of prediction performance, which is measured by the correlation between predicted risks of severe depression and ground truth risks, is shown in Fig.1 (a). The comparisons of detection performance, which are measured by the percentage of severe individuals being detected and the average true risk among detected individuals, are shown in (b) and (c), respectively. The proposed method had better performance on both risk prediction and high-risk patient detection over the entire study duration.
Fig.1 Performance Comparison (first 5 time units were used as training data to substantiate the SCM).