Explore UCD

UCD Home >

Peter MacMahon

HHIT Series Episode 12: "Generative AI in the ED - Developing tools to predict the future", with Peter MacMahon

On the Human Health, Impact and Technology webinar series on May 22nd 2024, Professor Peter MacMahon, Consultant Radiologist at the Mater Hospital and a Clinical Professor at University College Dublin, spoke to host Professor Patricia Maguire about "Generative AI in the ED - Developing tools to predict the future".

Career Snapshot

Peter was a medical student at University College Dublin before starting his career in the Radiology training scheme at the Mater Hospital, Dublin. After completing that training he moved to Boston, and worked at Massachusetts General Hospital, a centre of excellence in the US. There he specialized in emergency radiology, imaging trauma, stroke and different emergency conditions, along with a fellowship in musculoskeletal imaging and intervention, looking at sport injuries, bone tumors and other aspects. He returned to the Mater Hospital in 2012 as a consultant


Spearheading the use of AI

In 2015, his interest in the role and future of artificial intelligence (AI) in healthcare began to take shape. Given his experience in emergency medicine, a field characterized by its time-sensitive nature and the necessity for rapid, complex diagnoses, he started exploring how AI could assist in processing large amounts of data to aid in making critical clinical decisions. One of the initial areas of focus was the management of stroke, a condition that involves complex imaging studies. He spearheaded the implementation of an AI software solution designed to analyse imaging data, thereby facilitating decision-making for patient care in the emergency department (ED).

Accelerating Care in the ED

The Mater Hospital employs various AI vendors and software solutions for continuous patient care enhancement, particularly in stroke management. When patients with potential acute stroke present at the ED, clinicians must quickly differentiate between hemorrhagic and ischemic strokes.

"We perform extensive upfront imaging for these patients, aiming to get a patient from the ED door to the scanner within 10 minutes. Time is brain— the faster we achieve this, the more brain cells we can preserve."

Within two minutes of the scan, AI provides results to emergency physicians, detecting subtle blood traces and blood vessel blockages among up to 2000 images. While the AI assists in decision-making, it does not operate autonomously. It supports the radiologist in forming an accurate prognosis, thereby speeding up the diagnostic process and facilitating timely decision-making. By deploying the AI software on every patient scan continuously, the hospital has accelerated care delivery and increased clot detection rates by 100%, identifying clots that might be missed, in patients with conditions such as cancer for example. It also flags potential bone fractures in trauma patients, improving diagnostic accuracy and speed.


Future possibilities for Generative AI within the ED

In Ireland, there is a lack of access to MRI scans, particularly in EDs. No ED in the country is equipped with an MRI scanner due to the high costs of purchase and staffing, as well as the spatial constraints in busy, relatively small EDs. This situation contrasts with practices in many other parts of the world. Nonetheless, patients presenting in the ED frequently require MRI scans.

For instance, if patients arrive at a hospital without an MRI scanner, especially at night, they may need to be transferred across the country for an MRI, which is particularly common for spinal injuries. The Mater Hospital, a national spinal injuries unit, provides 24/7 spinal care and accepts such patients, performing necessary MRI scans even at night. If the scan is negative, patients return to their original hospitals. This process requires significant resources for transfers and maintaining the Mater's 24/7 service.


Generative AI to predict a future MRI scan

At the Mater Hospital, his team are investigating the use of Generative AI to predict future MRI scan outcomes based on CT scan data. This research aims to determine if CT scans contain sufficient information to accurately predict MRI results, potentially reducing unnecessary overnight patient transfers and allowing some patients to wait until morning for an MRI.

The research involves training a model with data from patients who presented at the ED with suspected acute spinal injuries and underwent both CT and MRI scans. Using this dataset, the model was trained to predict MRI outcomes based on CT scan data. Preliminary findings indicate that CT scans provide enough data to make accurate MRI predictions. For instance, in cases of Cauda Equina syndrome (disc herniation), the AI can analyse a CT scan and predict the MRI outcome within two minutes. This prediction can either provide high reassurance that there is no immediate issue or indicate a severe condition requiring urgent transfer to the spinal unit.

Medical Imaging Data

An essential aspect of AI research lies not in the models themselves but in the quality and representativeness of the data. It is crucial that the data encompasses a broad population spectrum and is properly annotated, especially in medical imaging, where the model needs clear instructions on what to identify. This process necessitates collaboration between experts and programmers to determine relevant data.

At the Mater Hospital, there is a unique collaborative scenario where an AI research programmer works closely with clinicians from the outset. The programmer acquires data, develops the model, and consults with clinical experts like Peter when encountering challenges. Radiology scans generate the data, and any project roadblocks typically stem from data issues, requiring either modification or additional data. This symbiotic relationship between data managers and domain experts ensures high-quality output, as effective AI models rely on robust data.

Impact of AI in Radiology

AI can be deployed in various aspects of radiology, with significant potential in X-ray imaging due to its high frequency. Although X-rays are not typically life-and-death tests, their sheer volume means that utilising AI in this area could have a substantial impact.

At the Mater Hospital, AI is being made available in frontline care for X-rays, allowing physicians and nurse practitioners to directly access AI results. While radiologists remain the primary gatekeepers of AI outputs for CT scans, frontline staff can now obtain AI-generated X-ray results without intermediary steps.  This system delivers relatively definitive conclusions: negative for fracture, positive for fracture, or indeterminate. With the Mater performing at least 300 bone X-rays weekly, the implementation of AI in this capacity is poised to significantly enhance diagnostic efficiency and impact patient care.

Automation Bias

As radiologists spearhead the integration of AI into clinical settings, foundational training in AI terminology, model functionality, and data significance is essential. Equally important is training on the risks of AI deployment, particularly "automation bias," where humans tend to trust machine results too readily. The team at the Mater Hospital has observed that negative AI results often prompt undue trust. However, clinicians should maintain a healthy skepticism and continue to verify cases as rigorously as they traditionally would.

AI, like humans, is not error-free and will inevitably miss some abnormalities, just as humans do. Two humans reading a scan yield better results than one and similarly a human working alongside AI can be more effective than either working alone. However, instances of automation bias have been observed where both AI and the clinician missed an abnormality. On review, clinicians express disbelief that they missed something they would ordinarily have detected, underscoring the influence of AI bias. It is essential to learn how to incorporate AI into clinical practice without placing undue trust in its results, recognizing that AI will make errors and may fail to detect certain issues.