Machine Learning for Risk Stratification in the Emergency Department (MARS-ED) (MARS-ED)

Rationale Identifying emergency department (ED) patients at high and low risk shortly after admission could help decision-making regarding patient care. Several clinical risk scores and triage systems for stratification of patients have been developed, but often underperform in clinical practice. Moreover, most of these risk scores only have been diagnostically validated in an observational cohort, but never have been evaluated for their actual clinical impact. In a recent retrospective study that was conducted in the Maastricht University Medical Center (MUMC+), a novel clinical risk score, the RISKINDEX, was introduced that predicted 31-day mortality of sepsis patients presenting to an ED. The RISKINDEX hereby also outperformed internal medicine specialists. Observational follow-up studies underlined the potential of the risk score. However, it remains unknown to what extent these models have any beneficial value when it is actually implemented in clinical practice. Objective To determine the diagnostic accuracy, policy changes and clinical impact of the RISKINDEX as basis to conduct a large scale, multi-center randomised trial. Study design The MARS-ED study is designed as a multi-center, randomized, open-label, non-inferiority pilot clinical trial. Study population Adult patients who are assessed and treated by an internal medicine specialist in the ED of whom a minimum of 4 different laboratory results (hematology or clinical chemistry, required for calculation of ML risk score) are available within the first two hours of the ED visit. Intervention Physicians will be presented with the ML risk score (the RISKINDEX) of the patients they are actively treating, directly after assessment of regular diagnostics has taken place. Main study parameters Primary - Diagnostic accuracy, policy changes and clinical impact of a novel clinical risk score (the RISKINDEX) Secondary Policy changes due to presentation of ML score (treatment policy, requesting ancillary investigations, treatment restrictions (i.e., no intubation or resuscitation) Intensive care (ICU) and medium care (MC) admission Length of admission Mortality within 31 days Readmission Patient preference Feasibility of novel clinical risk score

Routine clinical care. Physicians will actively be asked to self-report their clinical impression of each included patient and policy will be monitored.

Routine clinical care. Physicians will actively be asked to self-report their clinical impression of each included patient and policy will be monitored. In the intervention group, physicians will be presented with the RISKINDEX. Subsequently, self-report will again be initiated to evaluate the physicians' response to the ML score and possible policy changes due to the intervention.

Presentation of RISKINDEX to the physician after approximately 2 hours. The ML RISKINDEX is a prediction model based on laboratory data from the ED. It is based on date of birth, sex and at least four laboratory data which are sampled within the first two hours of the ED visit. Laboratory data that are used as input include samples that are commonly drawn in patients that require treatment from an internal medicine physician, such as urea, albumin, C-reactive protein (CRP), lactate and bilirubin.

Discriminatory performance of ML risk score to predict 31-day mortality. This will be calculated using an area under the receiver operating characteristic curves (AUC).

Policy changes after presentation of RISK-INDEX. This will be assessed by a filled out questionnaire by the physician where they state whether a policy change has been made as a result of the RISK-INDEX outcome.

Inclusion Criteria: Adult, defined as ≥ 18 years of age Assessed and treated by an internal medicine specialist (gastroenterologists included) in the ED Willing to give written consent, either directly or after deferred consent procedure (see section 11.2). Exclusion Criteria: <4 different laboratory results available (hematology or clinical chemistry) within the first two hours of the ED visit (calculation ML prediction score otherwise not possible) Unwilling to provide written consent, either directly or after deferred consent procedure (see section 11.2).

van Doorn WPTM, Stassen PM, Borggreve HF, Schalkwijk MJ, Stoffers J, Bekers O, Meex SJR. A comparison of machine learning models versus clinical evaluation for mortality prediction in patients with sepsis. PLoS One. 2021 Jan 19;16(1):e0245157. doi: 10.1371/journal.pone.0245157. eCollection 2021.

Explainable Machine Learning models for Rapid Risk Stratification in the Emergency Department: A multi-center study