GEMRA: Geriatric Emergency Medicine Risk Prediction Model for Return VisitAdmissions

点击翻译按钮获取中文摘要

Summary: At least 400 older adults a day are discharged from US emergency departments (EDs) and within 72 hours experience a return ED visit resulting in hospital admission (RVA). Geriatric RVA have dramatically higher morbidity and mortality than patients admitted to hospital on their initial ED visit. These outcomes, combined with the clinical complexity of geriatric presentations, demonstrate a critical need for clinical decision support (CDS) for ED discharge decisions and improved post-ED care management in older adults. National guidelines recommend that all older adults receive formal risk screening in the ED. Existing geriatric ED risk assessment tools lack predictive validity and are not designed to identify the multifactorial risk of an RVA event within 72 hours after ED discharge. Our long-term goal is to improve the outcomes of older adults using machine learning models for clinical decision support (CDS) in emergency medicine. The goal of this study is to develop and validate a machine learning model that predicts geriatric emergency medicine 72-hour RVA (GEMRA), and can be used as a feasible ED CDS tool. In order to maximize the impact and generalizability of GEMRA across a wide range of US ED environments and populations, the model input variables used will be clinical data collected in the course of normal clinical care, and thus widely available in emergency health records (EHRs). GEMRA will be developed and validated with data from five diverse hospitals across two health systems that span a wide range of demographic, socioeconomic, and ethnic backgrounds. The study will be conducted by a closely collaborating interdisciplinary team that includes emergency medicine, machine learning, and CDS experts, with extensive experience in geriatric emergency medicine research as well as developing and evaluating technological driven interventions to improve post-ED outcomes. Our preliminary work demonstrates that an early machine learning model using 478 clinical data input variables can accurately identify ED patients at high risk of RVA, outperforming an existing, unvalidated traditional RVA risk score that used six clinically derived risk factors. Our specific aims include: (1) Optimize GEMRA through model refinement, validation with retrospective data from unseen populations, as well as explanation of model performance variation across different clinical subgroups; (2) Assess GEMRA's clinical value through prospective validation at three different hospitals, comparing model performance to existing ED geriatric and RVA risk tools, as well as real-time clinician judgment; (3) Engage multidisciplinary stakeholders in the design of both a GEMRA CDS prototype and a complementary multidisciplinary clinical RVA risk assessment workflow; and subsequently evaluate the feasibility of these products in ED clinical practice during a short-term pilot implementation study. Completion of these aims could transform older adult post-ED risk screening, leveraging the computational power and scalability of machine learning to identify patients at risk of early post-ED adverse outcomes. Subsequent implementation of GEMRA CDS would inform risk-mitigating interventions, potentially impacting outcomes in this vulnerable population.