HHIT Series Episode 1: Computational medicine and the extraordinary potential of AI and machine learning in healthcare, with Raimond L. Winslow

We won’t have robot doctors in the future - human caregivers will always be in the medical loop. 

“What we're doing with AI is simply trying to provide new kinds of information and insights to caregivers to help them make better decisions.”

Dr Winslow never wants to “discount the intuition, understanding and experience of caregivers. I simply want to amplify and enable it even more through AI and ML methods”.

Meanwhile physicians will need to be trained differently for this coming era of medicine. 

“They will need to be trained to be able to critically evaluate papers that apply AI and machine learning to healthcare. In that sense we'll be developing new hybrid physicians.”

AI and ML in critical care

The “most exciting thing” for Dr Winslow is seeing the impact that AI and machine learning is already having in areas of critical care medicine. 

In the next decade he expects to see applications of machine learning and AI in critical care medicine to help treat patients more effectively, save lives and produce improved outcomes.

Predicting septic shock

Dr Winslow and his team developed an algorithm that looked at data from patients with sepsis and predicted which of those patients would make a transition to septic shock, a life-threatening condition that happens when your blood pressure drops to a dangerously low level post-infection. 

“On average, that algorithm provided about an eight-hour early warning of the impending nature of septic shock.”

The algorithm used information such as the patient’s heart rate, respiratory rate and temperature to make its prediction. In the end the patient’s blood lactate level - the measure of lactic acid in their blood - was the “number one predictor of risk of septic shock”.

The algorithm could predict septic shock when physicians could not, by finding “coincidences and small events, combining them in some strong and powerful nonlinear way to produce a robust signature”.

Dr Winslow thinks of this as “a computational discovery, a new kind of computationally derived diagnosis of what septic shock is”. 

Along with giving such diagnoses, AI might ultimately make specific recommendations to caregivers too.

“Imagine a charge nurse who knows that within the next 24 or 48 hours they are going to have five really sick patients in their unit. Maybe they need to bring in additional caregivers to help those patients. That act alone could help improve the quality of care and outcomes in that particular unit, aside from the direct interventions made by physicians.” 

Challenges

The problem with such prediction algorithms is that they may not translate to general settings, such as large, busy hospitals. Another sepsis algorithm deployed in a multi-centre clinical site in the US performed as poorly as “a random coin toss”, causing concern at the FDA.

“I don't believe that it's possible to publish an algorithm with a set of parameters learned on a set of data that will generalise to everywhere. An important aspect of trials is appreciating this and understanding that the approach to learning that algorithm can be disseminated; but that algorithms may need to be learned on the region's particular patient population. They will always work best when trained on that data. The goal is to deploy platforms that learn from the data that is being collected from the patient demographics being served in that particular hospital system.”