AI is surfacing in healthcare in three primary areas: medical imaging, electronic health records (EHRs) and genomics, in the view of Miriam Huntley, Ph.D and CTO and Day Zero Diagnostics, speaking at a panel on the State of AI in Medicine at the recent AI World Conference in Boston. The progress in medical imaging is […] 
AI is surfacing in healthcare in three primary areas: medical imaging, electronic health records (EHRs) and genomics, in the view of Miriam Huntley, Ph.D and CTO and Day Zero Diagnostics, speaking at a panel on the State of AI in Medicine at the recent AI World Conference in Boston.
The progress in medical imaging is substantial, with lots of high-quality image data available. “”We have real data from which we can do real learning,” said Huntley. The way images can be scanned and searched using deep learning and machine vision has been a “real revolution,” she said.
EHRs have penetrated far into the medical provider community in the US. By 2013, 80% of physicians were using EHRs. “The next real challenge with EHR is the quality of data. It can contain arbitrary grammar and spelling mistakes” for instance, she said. Huntley described the results for EHR as mixed.
Given genomic data on a patient, the physician is in a position to make predictions for disease risk, cancer presence and cancer treatment. Scientists have cracked the genetic code for human life, she suggested. And while the cost was once $100,000 for human genome data, today is may cost $1,000 per patient. Still the work is in an early stage.
“Most models used in genomics are simplistic. If you don’t have a lot of data, you can’t do a complex model,” Huntley said. The hard part is not the structure of data but the quality, and while costs are still high, they are dropping; so the scorecard is mixed on AI and genomics.
Day Zero is focused on bacterial infections. “Our mission is to diagnose bacterial infection on day zero,” meaning the infection was previously unknown. “We use genome sequencing; within five hours, we have an answer,” she said.
The firm combines genome sequencing and machine learning in its work. “We expect new models of machine learning to be developed, to give us information we have not seen before,” Huntley said.
How Practicing Medical Professionals and Data Scientists Collaborate
The AI World audience heard from practicing medical professionals in its AI for Healthcare Summit. Sareen Shah, in Pediatric Critical Care, Children’s Hospital of LA, said he is “researching AI applications as they pertain to general pediatric patients.” He works closely with David Ledbetter, data scientist at Children’s Hospital of LA, and chair of the Summit program. “Being on the technical side, I enjoying grilling the clinicians and learning from them,” Ledbetter said.
Yindalon Aphinyanaphongs, MD, PhD, Asst. Professor at the NYU School of Medicine, said he works on getting clinicians to make use of available AI tools. “Why do models not get adopted? That’s the crux of what my group does,” he said during a panel on Healthcare in the Trenches. “We are training the next generation of clinicians.” Like Children’s Hospital of LA, his NYU group also features a close collaboration of data scientists and clinical care professionals.
This transition toward wider adoption of AI in healthcare is marked by a number of “pain points” – including: models don’t work between different institutions; a divide exists between operations and academic orientations; evidence of AI effectiveness is limited (so far); access to hospital data is limited clinically and operationally; and data quality needs to be validated.
At NYU, Dr. Aphinyanaphongs has the backing of the executive team, which he said has made a big difference. “We are a team of data scientists and engineers who provide the capacity for predictive modeling for specific projects, from inception through development, deployment and evaluation,” he said.
In a case study example, he described the treatment path incorporating AI applied to a patient with a certain type of heart failure. The treatment goal was to reestablishing oxygen in the blood as soon as possible. “The goal is to implement and deploy a model to respond to the predicted diagnoses, of the specific heart condition, and have the patient leave the hospital in 7 to 10 days, “ he said. “The predictive model helps to identify the heart condition more quickly,” he said, and the path treatment is expedited by for example pre-approval to bring in specialists. He called the treatment approach “our congestive heart failure pathway.”
Challenges still to overcome include improving the coupling of model predictions and specific interventions, and refining the model performance based on the results of the interventions.
During the Q&A, Shah said, “EHRs have been a blessing and a curse. Most physicians now spend a lot more time in front of the computer, and use a lot more clicks to get where we want to go. This is not what we thought we would be doing when we were training to be physicians,” he said.
Describing his work, he said, “In a 28-hour shift, I make 300 to 500 decisions. I have limited brain space for each,” he said, adding that a “smarter” way to present data would help him. For example, he likes to know the patient’s heart rate over time. “Being able to present data the way you want it is very helpful,” he said.
Dr.Aphinyanaphongs said the underlying purpose of EHRs is often for medical billing, thus the related workflow is optimized for that. For a medical professional, this format can present limitations and uncertainty about how to present the data needed to power AI medical applications.
Learning Computer Science, Medicine to Help Communicate
Moderator David Ledbetter asked the panelists what tools technologists need to provide to help physicians? Shah said clinicians need to have some exposure to computer science, to help “find the shared language between the medical field and the computer science field. We each have reasons for using specific jargon. A lot of training goes into being a medical professional. It’s helpful to simplify the technical parts if you want to get more clinicians involved. Keep the data you present simple. Tell them why it’s important to them. Clinicians have information overload.”
Dr. Aphinyanaphongs said NYU has a curriculum designed to help clinicians come up to speed in what they need to know in computer science, and technologists to come up to speed in medicine and biology. “We try to get everyone on the same level,” he said.
Ledbetter asked how the gap can be bridged between data scientists and clinicians. Sareen, a clinician, said he started going to data science meetings, gaining more “clout” with the data scientists as a result, which he found helpful. “It’s a two-way street,” he said. And he in turn has seen more data scientists going to monthly medical research meetings. “If they integrate with our academic workflow, it helps both sides,” he said
— By John P. Desmond, AI Trends Editor