Machine learning techniques to improve decision making for rural medical transports
With scarce resources in rural health critical access hospitals, patients are often transported to tertiary care centers for both primary and critical care. The appropriate triage and transport of these patients is challenging because there is limited evidence to quantify the risk and potential benefits to drive clinical decisions in this setting. Recent, inpatient and pre-hospital settings have developed early warning and risk scores. However, there is a gap in the literature and absence of clinical decision support to guide inter-facility transports from critical access hospitals to tertiary care centers.
The study variables include: length of stay, Medical Early Warning Scores (MEWS), heart rate, respiratory rate and patient temperature for each patient. In addition, weather condition data were included to capture the potential impact that the physical difficulty associated with transport may have had on the transport decision.
The primary dependent variable was tertiary care resource utilization, as measured by length of stay, as an indicator of potentially avoidable transfers. A transport was coded as inappropriate if the patient's length of stay at the receiving medical center was less than 24 hours. This is a less costly proxy than a physician examination of each patient records. However, it permitted the methodology to be developed and initial results derived. It is anticipated that a more nuanced approach to coding the appropriateness will be used and new results obtained in the near future. This cohort study includes all 2,000 patients transferred to a tertiary care center from a critical access hospital in a rural frontier setting during 2015.
A series of machine learning models classified which patients were more likely to be inappropriately transported. Various algorithms were used including logistic regression, decision trees, and neural networks. Models were also combined using a technique called ensembling. The best model, which was based on the ensembling approach, predicted the target with approximately 70% accuracy.
Algorithms were successfully developed to determine which patients were inappropriately transported based only on information available at the time of transport decision. Although the accuracy will need improvement before full implementation, this study provides a foundation for further analysis and highlights significant variation in the outcomes of patients who receive transfers to rural tertiary care centers.
This work provides an example of why and how data enabled rural health care analystics should be used to inform potentially better medical decisions.