**4. AI and ML in emergency room triage and point of care ultrasound**

Artificial intelligence in medical practice is shaping the way clinicians assess, analyze, and diagnose potentially life-threatening conditions, which will significantly impact the delivery of emergency care. The use of AI algorithmic systems may give the tools to possibly overcome previously ingrained limitations in care delivery strategies [29] thus extending the ability of emergency physicians to diagnose and treat acute and critical illnesses. Over the last 10 years, the U.S. Food Drug and Administration (FDA) has approved more than 500 AI and ML devices [30]. Of these, 100+ were radiology applications for devices used during an emergency.

This section will highlight the applications of artificial intelligence (AI), machine learning (ML), deep learning (DL), and convolutional neural networks (CNN) in emergency room triage and their use, specifically in point of care ultrasound testing. The applicability of these technologies has an obvious advantage in emergency medicine as every year the demands on the emergency medicine practitioner increase as the number of emergency room visits grows and physicians are expected to care for more patients with fewer resources. The ability to provide timely efficient and accurate life-saving interventions is crucial, and AI holds the potential to help physicians streamline processes, increase efficiency, and cognitively offload.

#### **4.1 AI impact on emergency room care**

Triage is the prioritization of the sick and injured based on their need for emergency treatment. Traditionally, in the emergency department clinical support staff gather primary patient demographic data, vital signs, and basic information about a patient's initial presenting problem. The patient then undergoes a brief evaluation by a clinician, usually a nurse, to determine the patient's acuity or need for emergent care or resources. Commonly during this process, a patient is assigned an emergency severity index (ESI) score, which is a common triage tool that provides a clinically relevant framework to stratify patients into five groups from one (most urgent) to five (least urgent) based on acuity and resource needs. This system essentially determines who receives care first. Subsequently, clinicians thoroughly assess presenting symptoms, perform appropriate physical examinations, order applicable laboratory studies, imaging studies, and consultations and either discharge the patient to home or admit them to the hospital as indicated [31]. With the growing number of emergency room visits annually and a growing shortage of nurses and emergency medicine practitioners, the ability to provide timely efficient and accurate life-saving interventions is crucial.

Effective triage is of the utmost importance to patient quality of care and outcome, especially as ER capacities are further and further stretched by increased volume and decreased resources, which have led to prolonged ED stays and wait time for care. Although ER wait times are multifactorial, convenient registration and the early identification of impending life-threatening conditions can obviate adverse patient outcomes and decrease mortality. One study that assessed the performance of a deep learning system, PatientFlowNet, in predicting patient flow in emergency departments found that the PatientFlowNet model prediction of patient arrival rates was higher, with substantially more accuracy in predicting treatment and discharge rates than the baseline methods used in the ER. The resulting mean absolute error was 4.8% lower than the leading baseline [32]. Applying AI tools that combine both clinical narratives (symptoms, pain scores, and ESI) and structured data (demographics and vitals), there is potential to positively influence outcomes.

The AI algorithmic tool (TriageGO) recently developed at Johns Hopkins aims to integrate patient medical health records with presenting symptoms, as well as vital signs to further risk stratify patients and predict morbidity and mortality [33]. Additionally, the DNN model with word embedding AI tool, which integrated clinical narratives and structured data, outperformed and better predicts patients' hospitalization and discharge when compared to the rapid emergency medicine score (REMS) [34]. Furthermore, rapid response is paramount with time-sensitive complaints such as chest pains. Goto et al. neural networks AI model predicts whether patients presenting to ER chest discomfort needs urgent revascularization 12-lead EKG. Their AI model detects the presence of specific EKG characteristics not recognized by physicians [35]. Than et al. developed their "MI3 clinical support tool" to predict the likelihood of myocardial infarction (MI) using machine learning which achieved a high AUC (0.963) for diagnosing MI, which outperformed the European Society of Cardiology 0/3-hour pathway [36]. In all these examples, the impact of AI on today's healthcare system has the potential to be transformational.

### **4.2 AI impact on point of care ultrasound**

Over the last twenty years, ultrasound equipment has become more effective, economical, and compact because of this the applications and uses have broadened and the use of ultrasound at the bedside as a modality has become more ubiquitous. This is especially palpable in the world of emergency medicine (EM). In EM, there is an inherent need to arrive at a time-sensitive diagnosis and initiate potentially life-saving treatments, and the use of bedside ultrasound of point of care ultrasound is a crucial tool that facilitates this. POCUS is the medical use of ultrasound (US) technology
