Using Machine Learning to Identify Patients at Risk of Acquiring HIV in an Urban Health System.

BACKGROUND: Effective measures exist to prevent the spread of HIV. However, the identification of patients who are candidates for these measures can be a challenge. A machine learning model to predict risk for HIV may enhance patient selection for proactive outreach. SETTING: Using data from the electronic health record at Parkland Health, 1 of the largest public healthcare systems in the country, a machine learning model is created to predict incident HIV cases. The study cohort includes any patient aged 16 or older from 2015 to 2019 (n = 458,893). METHODS: Implementing a 70:30 ratio random split of the data into training and validation sets with an incident rate <0.08% and stratified by incidence of HIV, the model is evaluated using a k-fold cross-validated (k = 5) area under the receiver operating characteristic curve leveraging Light Gradient Boosting Machine Algorithm, an ensemble classifier. RESULTS: The light gradient boosting machine produces the strongest predictive power to identify good candidates for HIV PrEP. A gradient boosting classifier produced the best result with an AUC of 0.88 (95% confidence interval: 0.86 to 0.89) on the training set and 0.85 (95% confidence interval: 0.81 to 0.89) on the validation set for a sensitivity of 77.8% and specificity of 75.1%. CONCLUSIONS: A gradient boosting model using electronic health record data can be used to identify patients at risk of acquiring HIV and implemented in the clinical setting to build outreach for preventative interventions.

Real-Time Risk Tool for Pharmacy Interventions.

Background: Adverse drug events (ADEs) result in excess hospitalizations. Thorough admission medication histories (AMHs) may prevent ADEs; however, the resources required oftentimes outweigh what is available in large hospital settings. Previous risk prediction models embedded into the Electronic Medical Record (EMR) have been used at hospitals to aid in targeting delivery of scarce resources. Objective: To determine if an AMH scoring tool used to allocate resources can decrease 30-day hospital readmissions. Design Setting and Participants: Propensity-matched cohort study, Medicine/Surgery patients in large academic safety-net hospital. Intervention or Exposure: Pharmacy-conducted AMHs identified by risk model versus standard of care AMH. Main Outcomes and Measures: A total of 30-day hospital readmissions and inpatient ADE prevention. Results: The model screened 87 240 hospitalizations between June 2017 and June 2019 and 4027 patients per group were included. There were significantly less 30 day readmissions among high-risk identified patients that received a pharmacy-conducted AMH compared to controls (11% vs 15%; P = 0.004) and no significant difference in readmission rates for low-risk patients. While there was significantly higher documentation of major ADE prevention in the pharmacy-led AMH group versus control (1656 vs 12; P < 0.001), there was no difference in electronically-detected inpatient ADEs between groups. Conclusions: A risk tool embedded into the EMR can be used to identify patients whom pharmacy teams can easily target for AMHs. This study showed significant reductions in readmissions for patients identified as high-risk. However, the same benefit in readmissions was not seen in those identified at low-risk, which supports allocating resources to those that will benefit the most.

Parkland Trauma Index of Mortality: Real-Time Predictive Model for Trauma Patients.

OBJECTIVE: Vital signs and laboratory values are used to guide decisions to use damage control techniques in lieu of early definitive fracture fixation. Previous models attempted to predict mortality risk but have limited utility. There is a need for a dynamic model that captures evolving physiologic changes during a trauma patient's hospital course. METHODS: The Parkland Trauma Index of Mortality (PTIM) is a machine learning algorithm that uses electronic medical record data to predict mortality within 48 hours during the first 3 days of hospitalization. It updates every hour, recalculating as physiology changes. The model was developed using 1935 trauma patient encounters from 2009 to 2014 and validated on 516 patient encounters from 2015 to 2016. Model performance was evaluated statistically. Data were collected retrospectively on its performance after 1 year of clinical use. RESULTS: In the validation data set, PTIM accurately predicted 52 of the sixty-three 12-hour time intervals within 48 hours of mortality, for sensitivity of 82.5% [95% confidence interval (CI), 73.1%-91.9%]. The specificity was 93.6% (95% CI, 92.5%-94.8%), and the positive predictive value (PPV) was 32.5% (95% CI, 25.2%-39.7%). PTIM predicted survival for 1608 time intervals and was incorrect only 11 times, yielding a negative predictive value of 99.3% (95% CI, 98.9%-99.7%). The area under the curve of the receiver operating characteristic curve was 0.94.During the first year of clinical use, when used in 776 patients, the last PTIM score accurately predicted 20 of the twenty-three 12-hour time intervals within 48 hours of mortality, for sensitivity of 86.9% (95% CI, 73%-100%). The specificity was 94.7% (95% CI, 93%-96%), and the positive predictive value was 33.3% (95% CI, 21.4%-45%). The model predicted survival for 716 time intervals and was incorrect 3 times, yielding a negative predictive value of 99.6% (95% CI, 99.1%-100%). The area under the curve of the receiver operating characteristic curve was 0.97. CONCLUSIONS: By adapting with the patient's physiologic response to trauma and relying on electronic medical record data alone, the PTIM overcomes many of the limitations of previous models. It may help inform decision-making for trauma patients early in their hospitalization. LEVEL OF EVIDENCE: Prognostic Level I. See Instructions for Authors for a complete description of levels of evidence.