Abstract Body

Disengaging from care is associated with all-cause mortality in people living with HIV (PLWH); therefore, tools that can assist with the early identification of persons at-risk for falling out of care are needed. The electronic health record (EHR) serves as an invaluable platform to create such tools. We present an EHR-based predictive model for early identification of persons at high-risk of falling out of HIV care.

We analyzed EHR-data from HIV clinic patients in care at Duke between Jan 2014 and Dec 2018. Clinical data on 38 candidate variables (demographics, sexually transmitted disease (STD) diagnoses, substance abuse, mental health and healthcare utilization patterns) were collected for possible inclusion in the model. The outcome of interest, disengaging from care, was defined as failure to attend an HIV clinic appointment for ? 12 months after documentation of clinic encounters prior to the period of absence. LASSO regression was used in selecting features for a logistic model. The model performance criterion used for LASSO was area under the receiver operating characteristic curve (AUC). A non-regularized logistic model was then fit using the LASSO-selected predictors to verify performance on the testing set. AUC was calculated to assess model performance, and estimated coefficients were transformed to adjusted odds ratios.

2301 subjects (mean age 47 years; 72% male; 58% Black) were included in analysis. Predictor variables positively associated with disengaging from care include number of positive gonorrhea or chlamydia tests; number of recent syphilis tests (OR, 95% CI: 1.38, 1.23-1.56); ever being diagnosed with syphilis; a positive amphetamines test; and schizophrenia diagnosis (OR, 95% CI: 2.83, 1.18-6.78). Predictors negatively associated with disengagement from care are higher age (OR, 95% CI: 0.98, 0.96-0.99); number of recent gonorrhea/chlamydia tests taken (OR, 95% CI: 0.77, 0.73-0.81); 2+ emergency visits; 1+ hospital admissions; 2+ hospital admissions (OR, 95% CI: 0.2, 0.06-0.67); depression diagnosis (OR, 95% CI: 0.63, 0.45-0.87); and visiting the HIV clinic in each half of the year (OR, 95% CI: 0.2, 0.15-0.27). A logistic regression model using these predictors for the testing set results in an AUC of 0.78.

A model including demographics, STD diagnoses, mental health diagnosis, and healthcare utilization behavior variables effectively identifies PLWH at-risk for care disengagement.