A LONGITUDINAL PATTERN BASED PROGNOSTIC MODEL FOR DEPRESSION MONITORING VIA RULE-BASED METHOD

Purpose:

Depression is a common mental illness with complex and heterogeneous progression dynamics. Risk grouping of depression treatment population based on their longitudinal patterns has the potential to enable cost-effective monitoring policy design. In this study, we aim to identify a set of risk-predictive longitudinal patterns that predict depression progression and segment population into subgroups with different risk implications using rule-based method.

Method:

We applied the rule-based method on an electronic health record dataset of depression treatment population with person-level longitudinal Patient Health Questionnaire (PHQ)-9 scores for assessing depression severity and Charlson comorbidity scores. We first converted the longitudinal measurements in the first 6 months into a set of clinical features that provide statistical summarizations, characterize progression trajectories, and capture longitudinal patterns that derivate from normality. We then applied the rule-based method on these clinical features together with baseline profiles including age and gender to predict whether treated patients will progress to low risk group (PHQ9 score < 10) in the following 6 months. The prediction capability of rule-based model is compared with other predictive models including logistic regression and Support Vector Machine (SVM) using 5-fold cross validation. A set of significant rules are extracted from the rule-based model. The significance of each rule is tested by comparing risk levels in the rule endorsing and un-endorsing groups. Item response theory is further used to characterize the associations between significant rules and the underlying depression progressions and individual risks prediction.

Result:

12 longitudinal rules are identified from the depression treatment population (Table 1). All rules are statistically significant under the Wilcoxon rank sum test. The prediction accuracy is measured by area under the curve (AUC); it is 0.814 using rule-based model, 0.805 and 0.793 in logistic regression model and SVM model respectively.

Conclusion:

We applied the rule-based model on a depression treatment population and identified a set of risk-predictive rules which are able to segment individuals into subgroups with different mechanisms of depression progression based on their longitudinal patterns.