A Machine Learning Health System to Integrate Care for Substance Misuse and HIV Treatment and Prevention Among Hospitalized Patients
The use of an automated method to screen for HIV risk among hospitalized patients with substance misuse has the potential to increase the identification of those at risk to afford them access to testing, prevention, and intervention.
Project Details -
Completed
-
Grant NumberR21 HS028511
-
Funding Mechanism(s)
-
AHRQ Funded Amount$296,576
-
Principal Investigator(s)
-
Organization
-
LocationChicagoIllinois
-
Project Dates09/01/2021 - 08/31/2024
-
Care Setting
-
Medical Condition
-
Population
-
Type of Care
-
Health Care Theme
Substance misuse puts individuals at higher risk for HIV acquisition or transmission. For example, the prevalence of HIV among those who inject drugs is 17 percent compared to just 0.34 percent in the general population. Despite this clear risk, identifying those at risk during hospitalization is suboptimal due to busy care settings and the acuity and severity of a patient’s illness. The use of clinical decision support (CDS) tools informed by supervised machine learning (ML) has the potential to automate screening, removing the burden of screening from busy providers, and increasing the identification of those at risk. Supervised ML uses data for training that is already tagged with correct answers, equivalent to learning in a supervised environment.
This research will develop, train, test, and evaluate an interoperable ML classifier to identify risk for HIV acquisition or transmission among hospitalized patients with substance misuse and assess its real-time performance. The research will take place at Rush University Medical Center in Chicago, adjacent to underserved communities with the highest number of deaths due to heroin overdose. Rush’s Substance Use Intervention Team (SUIT) has a goal to screen all hospitalized patients for substance misuse, intervening with a harm-reduction model based on risk. Based only on the presence of sexually transmitted infections associated with HIV, retrospective data from SUIT showed that six percent of Rush substance use disorder patient encounters had a risk for HIV transmission or acquisition, considered to be an underestimate. A ML classifier has the potential to identify additional numbers by automating the risk screen.
The specific aims of the research are as follows:
- Develop, train, and test an ML classifier with high sensitivity (0.8) and specificity (0.8) to identify risk for HIV acquisition or transmission among hospitalized patients with substance misuse.
- Integrate the ML classifier in the electronic health record (EHR) infrastructure to test predictive validity in real time, and conduct interrupted time series to establish the effect of the classifier.
A ML health system approach will be used, leveraging EHR data, including clinical, social, and behavioral determinants captured in structured data fields and in clinical notes via natural language processing (NLP). Using retrospective data, all encounters of adult patients with substance use disorder and associated HIV risk will be matched in a 1:2 ratio to non-cases. Three supervised ML models will be trained and tested: a baseline model with structured data, an NLP model with unstructured data, and a model with both structured and unstructured data. The best model in terms of efficiency and accuracy will inform the CDS tool. The tool will be integrated within the EHR infrastructure and its predictive validity in real time assessed. The ML classifier will integrate and standardize NLP approaches into an interoperable health information technology platform via Fast Healthcare Interoperability Resources to ensure interoperability and scalability. To test and validate the model, an interrupted time series over 12 months will be conducted in the production environment. The research team hypothesizes that identified cases will increase from the baseline six percent to 10 percent. Once the research is complete, the tool will be available as an open-source tool. Across health systems, this tool may have population-level impact as it facilitates an interoperable and scalable intervention to integrate HIV prevention at the point of care.