Utilizing Machine Learning for Life-Saving (and More Efficient) Inpatient and Outpatient Triage.
The act of triage is a foundational piece of modern medicine dating back to the Holy Roman Empire. During the Napoleonic Wars, French military doctor Dominique Jean Larrey developed modern triage techniques combined with a newfound emphasis on sanitation when dealing with battlefield injuries.
A few hundred years later, triage is now a vital function for modern emergency departments (ED’s) often stretched to capacity limits due to patient volume increases, a rapidly aging population, chronic staff shortages, and a waxing and waning pandemic.
However, one element of triage that hasn’t changed since the 15th century is that it has been exclusively performed by humans. At least, until now.
How machine learning helps improve triage (especially for outpatients)
Taking triage calls can be particularly challenging for humans in a telemedicine or remote health monitoring system (RHMS) environment when patients aren’t physically at the hospital for observation.
But telemedicine and RMHS systems are one of the healthcare industry’s key tools in combating hospital overcrowding.
That’s why researchers have begun applying machine learning (ML) tools to help scale triage activities, both at the hospital and for telemedicine applications. These models can also better support healthcare workers who must typically rely on their own knowledge and experience when performing triage.
Healthcare providers have used other clinical decision support systems (CDSS) since the 1980s: Statistics from 2013 show that 41 percent of U.S. hospitals with access to electronic medical records used a CDSS. However, these systems are often little more than static online databases that don’t react intelligently to specific patient issues and medical histories.
Instead, ML models learn from real-world datasets and experience and use algorithms to “prioritize patients for us rather than expending the time to generate a rating that may not even be consistent between providers.”
Aside from improving triage productivity and accuracy, ML models have the potential to save healthcare providers, payers, and consumers plenty of resources by facilitating more accurate assessments of outpatients. The average cost of a three-day inpatient hospital stay in the U.S. is around $30,000.
Machine learning and patient evaluation: The studies
It’s not just about cost, of course, when patients’ lives hang in the balance. Indeed, a UK-based study showed that treatment delays for ophthalmic conditions resulted in permanent eye damage in more than 70 percent of cases, illustrating the need for fast and accurate triage.
While still in its relative infancy, considerable research has examined the efficacy of ML models on in- and outpatient triage. Here are some of the most compelling recent studies.
Li et al. (2021): The researchers examined how ML-aided triage can help reduce the workload for practitioners when assessing patients for ophthalmology specialist outpatient clinics (SOPC) in Hong Kong.
Among other findings, the study noted that human-in-the-loop machine learning models (which combine a human researcher and ML model) performed better than AI-only or expert-only systems (with an AUC score of .84 compared to .763 and .685, respectively).
Xie et al. (2021): In this case, researchers looked at the effectiveness of an ML-based triage tool for estimating mortality among inpatients compared to other clinical scores.
Researchers developed the ML-based Score for Emergency Risk Prediction (SERP) tool and analyzed ED patients in Singapore from 2009 to 2016. The tool was underpinned by an ML-based scoring framework called AutoScore, developed by the same researchers.
According to Health IT Analytics, the tool provided a more “accurate estimate of a patient’s risk of death” compared to non-ML triage systems, such as Patient Acuity Category Scale, Modified Early Warning Score, National Early Warning Score, Cardiac Arrest Risk Triage, Rapid Acute Physiology Score, and Rapid Emergency Medicine Score.
“SERP had a better prediction rate than existing triage scores and maintained easy implementation and ease of ascertainment in the ED,” wrote Health IT Analytics.
Hadi et al. (2019): This study from the University of Leeds examined the prioritization of outpatients using ML models, first by using a native Bayesian classifier to analyze outpatient medical records and medical device data. The ML model then ingests this data to predict the likelihood of a life-threatening condition.
In particular, this study examines two uplink techniques: Weighted Sum Rate Maximization (WSRMax) and Proportional Fairness (PF). “Using these approaches,” the authors write, “we illustrate the utility of the proposed system in terms of providing reliable connectivity to medical IoT sensors, enabling the OPs to maintain the quality and speed of their connection.”
Miles et al. (2020): This systematic review analyzed the accuracy of ML models in emergency triage of all incoming patients, not just outpatients, using medical literature from Medline, CINAHL, PubMed, and other sources.
The researchers analyzed 92 models using the C-statistic method of gauging efficacy, with neural networks and tree-based methods recording a median score of .81 (values higher than .8 generally indicate a strong model) for patients requiring hospitalization. For those requiring critical care, neural networks had a median C-statistic of .89.
While neural networks seemed to score higher overall, the researchers note that “models derived by logistic regression were more transparent in reporting model performance.”
Fernandes et al. (2020): Finally, this systematic review examined literature from ScienceDirect, IEEE Xplore, Google Scholar, Springer, MedlinePlus, and Web of Knowledge to assess ML-based tools against traditional triage methods.
The researchers found that logistic regression was the most commonly used method for model development. Most applications focused on improving patient prioritization, predicting critical care needs and hospital/ICU admission, length of stay, and mortality.
“In the papers where CDSS were validated in the ED, the authors found that there was an improvement in the health professionals’ decision-making, thereby leading to better clinical management and patients’ outcomes,” the authors wrote. However, they also noted that more than half of the examined studies did not include an implementation phase, making their effectiveness difficult to assess.
CapeStart: The medical machine learning experts
CapeStart’s machine learning engineers, medical subject matter experts, and data scientists work with leading healthcare and research organizations every day to improve medical outcomes through advanced machine learning. Our off-the-shelf and custom ML models, training datasets, and data annotation services are tailor-made to boost the efficiency and efficacy of complex medical research and clinical applications.
Contact us today to schedule a brief discovery session with one of our experts.