DEPRESSION MANAGEMENT USING ELECTRONIC HEALTH RECORD: PERSONALIZED PROGRESSION PREDICTION AND MONITORING

Purpose: Depression is a common, complex, and dynamic mental disorder. According to the CDC, about 9% of Americans suffer from depression at least occasionally. Therefore, mitigating depression has become a national health priority. Prediction of disease progression is of significant importance in primary care and mental health clinic settings. This study aims to build an easily implementable prediction model for individual patient's depression trajectory and design a personalized monitoring scheduling system from electronic health record (EHR).

Method: Our dataset contains longitudinal Patient Health Questionnaire (PHQ)-9 scores over 4 years for 3,159 patients that have at least six PHQ-9 observations while on treatment. A nature history model was proposed to predict a patient's depression progression. To build a database of patients' depression natural history, firstly, the estimated PHQ-9 score trajectory of each patient was obtained by fitting b-spline on the observed scores. The model sampled the spline-derived values at bi-week intervals, and selected each set of three sequential values and converted them into a time triplet. The model then stratified the time triplets by the characteristics of the patient including age, sex, treatment status, and Charlson comorbidity score. To predict the PHQ-9 score of a new patient at the next time interval, the model searched the entire database to find the most similar patient by using a multivariate nearness measure approach. By adopting the prediction model, a simulation-based monitoring system was further established, where a visit schedule table can be designed for each patient according to a given scheduling criteria, for instance, visit in 2 months if score<10 (mild), 1 month if 10<score<15 (moderate), 2 weeks if score>15 (major).

Result: Extensive experiments were conducted by using five-fold cross-validation under distinct scenarios including long-term predictions and regular-time predictions. The root mean square errors were 6.38 and 4.68 for long-term and short-term predictions, respectively. Histograms of average visiting interval of the 3,159 patients under varied monitoring scheduling rules were obtained. Figure illustrates the histogram of average visiting interval with the above scheduling criteria.

Conclusion: We established a nature-history model for individual depression prediction using EHR. Simulation results show that the model works more accurately for short-term predictions. A simulation-based monitoring system was further proposed to design personalized scheduling tables for patients based on the proposed prediction model.