focused on one small niche ( thank goodness ), computers evolve in an environment of hyper-specialization , often leaving behind general skillsets needed to make distant connections . This limitation in AI hints at the irreplaceability of humans .
While digital computers have superior processing speed on basic tasks ( 10 billion per second versus less than 1,000 / second for humans ) and precision ( error rates 1 in 4.2 billion versus 1 in 100 ), chimpanzees and human toddlers can beat computers in many ways . Computers have mostly serial information processing , but the human brain is both serial and “ massively parallel .” The input / output for each unit in a computer is 1-3 ; in the brain , 1-1,000 . This parallel arrangement helps us aggregate signals from different neurons to improve our precision . Finally , humans have both digital and analog signaling at our disposal , versus only digital for computers . This gives us continuous variables and allows for a wider range of information processing . Just as we compare brains to computers , computer scientists model computers on brains : deep learning was modeled on the mammalian visual system .
Despite its drawbacks , AI has incredible strengths . ML offers cost-saving and patient care optimizing applications in radiology , dermatology , neurosurgery , ophthalmology and many other specialties . These benefits extend to the ED as well . With shrinking working memory , perhaps we could use a computer to delegate tasks and minimize cognitive overload . Hey Siri , catch me up on my patient charting duties , I am going to find something healthy to eat . While machines can ’ t autonomously complete our dreaded professional requirements , many studies in the ED or related to emergency care show promise .
A scoping review of AI in EM cited 150 publications , with only three prospective controlled trials . One fourth of them aimed to improve diagnosis , 13 % focused on imaging , 11 % were performed out-of-hospital and 16 % compared AI to humans . We got beat by AI in about half of those studies , outperformed in various individual tasks : predicting out of hospital cardiac arrest / MI , identifying hyperkalemia , determining triage risk stratification , identifying participants for research enrollment , predicting wound infection , assessing overall mortality and visual reading of medical imaging . One study that let us combine powers with AI led to improved accuracy in detecting MI on ECG .
The operations of emergency care can be categorized into three buckets : what happens before a patient arrives , during their ED stay and after discharge . In prehospital medicine , ML can help with ED volume predictions , computer assisted dispatch and wearables for falls and seizures . ML models have shown promise in assigning
The future is already here – it ’ s just not evenly distributed .
-William Gibson , The Economist , December 4 , 2003
ARTIFICIAL INTELLIGENCE triage levels and predicting which patients will go on to develop critical illness ( sepsis , cardiac complications in chest pain ) while in the hospital . We continue to use ML to search for the tool ( SIRS , MEWS , NEWS , qSOFA ) that would identify patients who might decompensate the moment they arrive at the hospital .
AI could improve efficiency in many aspects of care in the ED .
EM physicians interpret medical imaging on many patients during every shift – AI can help . Algorithms applied to reading of head CTs have reached sensitivities of 94-100 % and negative predictive values of 99 %. ML applied to MRI and X-ray images can quickly diagnose bony injuries of the spine , wrist , hand and ankles . Deep learning software can find pneumonia on chest X-ray and subdural hematomas on head CT . AI can review images of retinal scans and find diabetic retinopathy . It ’ s tough to envision a robot using the EM physician ’ s favorite tool , the ultrasound . While ultrasound images could lend themselves to digital AI reading , for now humans still must obtain the images .
AI ’ s subfield of natural language processing ( NLP ) can save us time , stress and maybe even bad patient outcomes . Think about the online bot you “ chat ” with on Amazon . Could patients who believe they have a medical emergency chat with AI prior to coming to the ED ? Studies have used NLP to basically eavesdrop on a doctor-patient encounter and draft a note based on the conversation ( but don ’ t forget that most communication is nonverbal ). NLP has implications for narrative medicine – letting our patients tell their stories . Would NLP jeopardize the recording of nuanced conversation between patient and doctor , or would it better capture the words of the patient that the doctor may miss or deemphasize ? This distinction is pivotal .
NLP models have reasonable accuracy in detecting sepsis , acute appendicitis and influenza using notes written by ED providers . AI textual analysis of the EMR can predict illness trajectories , potential medical interventions and readmission likelihood . Research has shown AI ’ s ability to predict in-hospital mortality , cardiac arrest 30- day mortality after STEMI , defibrillation success for out of hospital cardiac arrest , and even outperform the Trauma and Injury Severity Score to predict need for life-saving interventions in trauma patients . ML systems can monitor respirations of COPD patients to predict exacerbations . It can also help predict deterioration in asthmatic children a week before symptom onset .
Machine learning in monitoring helped develop the compensatory reserve index ( CRI ) to detect changes in finger arterial blood pressure that precede cardiovascular instability . Another ML model gave an average of 17 minutes advanced notice before ICU patients
( continued on page 8 ) APRIL 2022 7