Using Electronic Health Records Data to Identify

Using Electronic Health Records Data to Identify Incarcerated Persons at Increased Risk for HIV Acquisition Session: 1213 Presenter: Alex Treacher, PhDTeam and collaborators Development and successful deployment of a high-performance machine learning models requires an interdisciplinary team to maximize both performance, and clinical impact of a predictive model. Ank Nijhawan, M.D. [...] S Kenneth Dobbs NIH Funding: R01 MH 129185Background/Introduction Incarcerated persons experience disproportionately high rates of HIV acquisition. [...] Development of tools to help identify incarcerated individuals with elevated HIV risk can optimize allocation of resources for testing, counseling, and prevention. [...] We developed a predictive risk model to estimate individuals' future HIV risk using electronic health records from the 8th largest US jail. [...] 33,458 Individuals incarcerated 2015-2022 at DCJ with an HIV test result 2,659 HIV diagnosis Underwent chart review 30,785 HIV negative 486 New HIV positive diagnosis during incarcerationEnd to end hyperparameter optimization • Optimized parameters for processing and model development using Bayesian optimization (TPE) Modeling Approach Data split: Stratified Multi- holdout 10-fold cross val [...] regularization Goal: Create an optimized classification model to predict patients with a future HIV+ diagnosis, and use the prediction probability to estimate risk. Model analysis: • Final model performance • Model calibration • Feature importance (PFI) • Risk stratificationResults: Model Performance AUROC .702 Bal Acc .655 Sens .687 Spec .624 18763 12022 342 144 % Pop. [...] Lower 95% CI .64 .50Future Work Evaluation of performance on newly available data Develop/build deployment pipelines Prospective implementation to evaluate performance in identifying people entering the jail who need HIV and STI testing and may be potential PrEP candidates Integrate features available from Parkland Health EHRSummary • Created HIV dataset/cohort from the 8th largest jail in the U [...] • To our knowledge, this is the first HIV prediction model developed for an incarcerated population. [...] • Predictive performance with good calibration is likely in a range that can improve efficiency for HIV prevention resources in jails. • Limitations of the performance are likely due to limited information from jail EHR. [...] • Given the large population of individuals at risk for HIV who pass through US jails, the potential population-level impact of a jail HIV prediction model is substantial and warrants prospective evaluation.