OBJECTIVES : Identify factors likely to be associated with new mental-health (NM-H) cases PRE-COVID-19 and POST-COVID-19 and determine whether supplemental information could better identify at risk-employees for interventions.

METHODS : Workpartners Research Reference Database (RRDb) includes direct components (medical and prescription claims), indirect components (sick-leave/paid-time-off, short-term disability, family medical leaves [FMLA]), additional demographics, job-related information, and self-reported survey data. Gradient boosting was used to predict NM-H cases in the periods PRE-COVID-19 and POST-COVID-19 and compared for precision using demographic and direct-claims only (DIRECT) and integrated (INTEGRATED) data models. Proportion of days covered (PDC) was used for medication-adherence for non-mental-health conditions.

RESULTS : The PRE-COVID-19: Compared with DIRECT model, the INTEGRATED model identified 18x as many likely events and had increased precision (70% versus 53%, missing 99% and 91% of cases, respectively). Top predictors in the DIRECT model included: #-of-Agency for Healthcare Research and Quality (AHRQ) specific-conditions, medication-adherence, age, gender, #-of chronic conditions, #-of unique medications. Top predictors in the INTEGRATED model included absence days (prior 30 days), self-reported anxiety, absence days (sick-leave/paid-time-off), #-of-AHRQ specific-conditions, medication-adherence, working ≥70 hours/week. POST-COVID-19: the INTEGRATED model identified >3x as many likely events, with improved precision (79% versus 68%, missing 74% and 30% of cases, respectively). Top predictors in the DIRECT model included #-of unique medications, medication-adherence, gender, #-of-AHRQ specific-conditions, #-of chronic conditions, and total costs while the INTEGRATED model’s top predictors were #-of-prescriptions, absence days (sick-leave/paid-time-off), absence days (COVID), absence days (prior 30 days), self-reported anxiety and salary.

CONCLUSIONS : In both PRE-/POST-COVID-19 time periods, INTEGRATED models reduced false positive rates and outperformed the DIRECT models, suggesting new mental health events can be better identified using additional information. The POST-COVID model identified different factors predicting new mental health claims and outperformed the PRE-COVID model. Accurate prediction of triggers can be used to enable policies proactively addressing potential challenges.