Introduction

Machine Learning to Predict Mortality Rates

We are quite sure that most of us have spent innumerable hours at an hospital emergency room, waiting to get admitted or waiting to meet with a doctor. However, there are times when patients who do not need any urgent or critical care are called in before. This is triage. Triage is a procedure by which medical providers can categorize and sort incoming patients depending on the resource availability and also patient requirements. Early prediction of patient mortality risks can guarantee to reduce the mortality by assuring effective and efficient resource allocation and treatment planning.

How can Machine Learning Help

To effectively use Artificial Intelligence (AI) in chronic disease management, providers must have access to high-quality and accurate data. According to University of Pennsylvania Medical Center (UPMC) Chief Health Care Data and Analytics Officer Oscar Marroquin, MD, data scientists need access to real-world data to gain insights to be integrated within clinical programs.

In a specific study, researchers tried to compare an interpretable ML triage tool that could predict mortality functions in a cohort of patients who were admitted to the hospital from the Emergency Department (ED) versus the regular triage scores.

Identifying the characteristics of patients that drive the disease across large patient numbers is essential, as it could enable hospitals and providers to forecast disease trajectories and assign crucial resources. Initial efforts to create ML models for this purpose have been restricted by small sample sizes, shortage of generalization to diverse populations, differences in feature missingness, and probability for bias.

ED triage is a multifaceted and complex clinical judgment-based process to understand a patient’s potential and probability of survival and accessibility of medical resources. While scoring tools are vital in risk stratification, the current score systems have displayed various limitations.

We have to use the real-world data, that is the patient-level data and not the one that is collated through conventional randomized clinical trials. On applying the right analytics to that data, we can gain the insights that we require in order to steer our organization towards inventing new real-world based evidence,” states Marroquin.

Researchers have also developed Score for Emergency Risk Prediction (SERP) using a general-purpose ML-based scoring framework termed as AutoScore. By using the ML tool, researchers could offer an accurate estimate of a patient’s risk of death. The study evaluated all ED patients between the dates January 1, 2009, and December 31, 2016.

To understand mortality outcomes after emergency admissions, SERP has been compared to various other triage systems comprising Patient Acuity Category Scale, Cardiac Arrest Risk Triage, National Early Warning Score, Modified Early Warning Score, Rapid Acute Physiology Score, and Rapid Emergency Medicine Score.

Electronic Health Record (EHR) information was pulled out from the SingHealth Electronic Health Intelligence System. The patient data was de-identified, and the list of comorbidities were obtained from discharge records and hospital diagnosis.

Using statistical analysis, the study showed that SERP had an efficient estimate rate than the prevailing triage scores and maintained easy implementation and coupled with helping with easing of uncertainty in the ED. Researchers state that this method of mortality prediction could become widely accepted in the healthcare community.

Conclusion

SERP scores have ability and potential to be broadly used and validated in various circumstances and health care settings. Following the clinical application of SERP in ED triage, additional tailored scores can be derived in various clinical areas through the Machine Learning-based AutoScore framework in future.