*Unlocking the Potential of Artificial Intelligence (AI) for Healthcare DOI: http://dx.doi.org/10.5772/intechopen.111489*

for the bedside evaluation of acute or critical medical conditions. It is utilized for diagnosis, the guidance of procedures, monitoring of certain pathologic states, and as an adjunct to therapy. It has also demonstrated its utility as an adjunct in the resuscitation of the critically ill.

POCUS examinations are typically performed, interpreted, and integrated into care by the treating physician in real-time at the bedside making it distinct from traditional radiology-based applications [37]. Instead of performing a systems-based study designed to interrogate a particular anatomic area, POCUS seeks to help answer specific clinical questions that are often binary in nature (e.g. is there free fluid in the peritoneum, is there a pneumothorax, is there hydronephrosis, etc.). An additional factor that differentiates POCUS from the traditional use of medical ultrasound is the fact that POCUS practitioners are inherently diverse in their training and their ability. Ultrasound image acquisition is a user-dependent skill, and both because of this as well as the binary nature that drives POCUS use at the bedside, POCUS is an area that is ripe for the application of artificial intelligence (AI) and deep learning (DL) [37].

The use of DL in POCUS is varied as the model used depends on the problem it is trained to solve [37]. For example, DL in POCUS has been already used to help identify structures [37], for image enhancement [38], and for the classification of images [38]. In each different application, depending on the clinical question, the POCUS operator would "only need to provide an image, and the trained DL model would be able to immediately return the desired output, whether it be the outline of an organ, an enhanced US image, or the classification of the US image along with a confidence score [37]." The ability of DL application allows the practitioner to cognitively offload some elements of image acquisition and interpretation, and thus be able to concentrate more on real-time application and direct patient care [37]. The advantage of 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.

AI and DL have demonstrated utility in several cardiac studies (e.g., estimation of ejection fraction, calculation of IVC caliber and collapsibility to predict fluid responsiveness, and the identification of cardiac tamponade), as well as pulmonary applications (AI-enhanced lung ultrasound in discriminating viral and bacterial pneumonia, estimation of size of pneumothorax based on location of lung point, and prediction of antibiotic response from US lung images using DL). These applications expand the EM practitioner's ability to risk stratify and implement treatment.

There is further potential, as AI ability evolves, to eventually achieve "real-time" image interpretation. This could in theory expand the number of POCUS practitioners beyond the ranks of physicians or EM-trained clinicians to first responders, EMTs, and those responding to mass-casualty events or real-time disasters. The ability to use POCUS in the "field" by untrained or novice user will allow those on site to potentially diagnose fractures, abdominal/thoracic free-fluid or hemorrhage, pneumothorax, or even cardiac standstill thus optimizing the triage response and subsequent allocation of resources. A similar conclusion can be drawn from those practicing in the global health realm, which is traditionally a lower-resource practice environment.

However, despite the obvious advantages, there are some limitations to the use of AI in POCUS. Imaging modalities, such as CXR, CT, and MRI, have standardized imaging protocols that are archived for later use/review leading to the construction of large persistent imaging datasets for AI to "mine." POCUS images and videos, on the other hand, are acquired and interpreted at the bedside, and findings are immediately applied with variable storage/archiving protocols depending on time limitations, patient acuity, machine capabilities, and institutional guidelines. Additionally, the large variation in POCUS user skill level, the order in which images are acquired, and the image acquisition technique create a great deal of "noise" or randomness which further complicates the building of large, standardized ultrasound datasets. Despite, this as AI advances and DL modeling and the creation of CNN becomes more sophisticated pathways are being found to navigate these limitations.

The impact of AI on today's emergency room can be transformational from its effects on triage to disease diagnosis and detection. AI can reintegrate and augment ER staff rather than replace the human workforce by decreasing the work burden and improving clinical outcomes.
