Machine Learning Models May Predict Emergency Admissions
Machine learning applied to electronic health records can inform care, service planning
WEDNESDAY, Dec. 5, 2018 (HealthDay News) -- Machine learning models incorporated into electronic health records (EHRs) may predict the risk for emergency hospital admissions, according to a study published online Nov. 20 in PLOS Medicine.
Fatemeh Rahimian, Ph.D., from the University of Oxford in the United Kingdom, and colleagues used longitudinal data from linked EHRs of 4.6 million patients aged 18 to 100 years from 389 practices across England between 1985 and 2015 to develop a machine learning model capable of predicting the first emergency admission within 24 months. The model included 56 variables, including patient demographics, lifestyle factors, laboratory tests, currently prescribed medications, selected morbidities, previous emergency admissions, marital status, and prior general practice visits.
The researchers found that compared with the baseline reference Cox proportional hazards model, the final gradient boosting classifier (GBC) model showed a 10.8 percent higher area under the receiver operating characteristic curve (0.848 compared to 0.74) for prediction of risk for emergency admission within 24 months. Across the risk spectrum, GBC showed the best calibration. The addition of temporal information resulted in substantially improved discrimination and calibration for predicting the risk for emergency admission.
"These findings support the potential of incorporating machine learning models into electronic health records to inform care and service planning," the authors write.