Leveraging Machine Learning to Predict Mental Health Referral Follow-up Among US Military Personnel.
Noncompliance with mental health referrals among US military personnel remains a significant barrier to care. Operational deployments and military stressors contribute to mental health challenges, impacting treatment access and increasing costs for the Department of Defense. Identifying service members unlikely to follow through on referrals may enable targeted interventions.
To develop machine learning (ML) models to predict noncompliance with mental health referrals and identify key predictors among active-duty personnel.
This study utilized retrospective data to create predictive models for referral noncompliance.
The study sample consisted of 14,289 active-duty personnel who received mental health referrals through the Periodic Health Assessment (PHA) from 2016 to 2020.
Predictors included demographics, health screenings, medical history, and prior health care utilization. Outcome measures focused on noncompliance within 90 days of referral.
Noncompliance with referrals occurred in 34.0% of the sample. Among predictive models, extreme gradient boosting (XGBoost) achieved the highest performance (AUC ≈ 0.80), with prior health care utilization (eg, previous clinic visits and mental health diagnoses) identified as the strongest predictor, followed by alcohol screening and age.
ML models demonstrated strong potential for identifying at-risk individuals, supporting targeted interventions to improve mental health care follow-up. Future research will emphasize validation and explore mechanisms influencing noncompliance.
To develop machine learning (ML) models to predict noncompliance with mental health referrals and identify key predictors among active-duty personnel.
This study utilized retrospective data to create predictive models for referral noncompliance.
The study sample consisted of 14,289 active-duty personnel who received mental health referrals through the Periodic Health Assessment (PHA) from 2016 to 2020.
Predictors included demographics, health screenings, medical history, and prior health care utilization. Outcome measures focused on noncompliance within 90 days of referral.
Noncompliance with referrals occurred in 34.0% of the sample. Among predictive models, extreme gradient boosting (XGBoost) achieved the highest performance (AUC ≈ 0.80), with prior health care utilization (eg, previous clinic visits and mental health diagnoses) identified as the strongest predictor, followed by alcohol screening and age.
ML models demonstrated strong potential for identifying at-risk individuals, supporting targeted interventions to improve mental health care follow-up. Future research will emphasize validation and explore mechanisms influencing noncompliance.