To be able to tell beforehand whether disease may occur is crucial, especially for a tricky medical condition like cancer. Predictive analytics driven by artificial intelligence (AI) assists physicians to make quicker and better diagnoses in a more cost-saving manner. Besides, AI models also facilitate projections of which patients may not show up for their next appointments or show signs of relapse on treatments. All these, again, allow physicians to intervene ahead of time and promptly stop any adversity. 

Likewise, when feeding an AI algorithm with numerous images of skin lesions, it is able to make sense of an underlying pattern and tell whether a novel lesion image is life-threatening or not. Unlike human, who requires years after years of training to detect a malignancy, AI is able to perform a similar task within the shortest possible time. Medical professionals and healthcare leaders celebrate what AI can achieve but some suggest, this should not be all. 

Sema Sgaier, Adjunct Assistant Professor at Harvard’s T.H. Chan School of Public Health, Affiliate Assistant Professor of Global Health at the University of Washington, and Co-Founder and Executive Director of Surgo Foundation and Francesca Dominici, Clarence Gamble Professor of Biostatistics, Population Health, and Data Science at Harvard’s T.H. Chan School of Public Health, Co-Director of Harvard University’s Data Science Initiative and Member of the National Academy of Medicine, published an op-ed on Harvard Business Review recently, detailing what we might have missed out.  

Searching for disease mechanisms to optimize treatments 

Sgaier and Dominici believe AI should also be trained to tell why diseases occur and underpin their root causes, so that researchers can use the knowledge to create new drugs and treatments and recognize the ideal recipients. In traditional medicine, causal relationships are established only after many rounds of expensive and time-consuming randomized controlled trials, with causal AI algorithms, Sgaier and Dominici thought the process can be performed as efficiently as predictive analytics. 

For example, according to Sgaier and Dominici, GNS Healthcare is partnering with Alliance for Clinical Trials in Oncology, to heighten the survival rate of colorectal cancer. Causal models were created using clinical trial data of over 2000 patients who were administered with two different drugs. This allowed physicians to identify molecular and clinical causal factors that can be served as biomarkers for survival, enabling physicians to create a more targeted treatment for these patients. 

In another example, instead of using causal AI, researchers applied “what if” reasoning to create counterfactual AI models that can foresee how medical conditions would progress as a result of different medical or healthcare interventions. In this case, data of kidney patients undergoing dialysis in the intensive care unit were employed and the level of creatinine, an indication of kidney failure, was closely monitored. Patients were then given individualized schedules of when to receive what type of dialysis based on the developed models.

Social determinants of health 

On top of which, Sgaier and Dominici also believe causal AI algorithms will allow professionals to cater healthcare resources more effectively. For example, the model they have created found that 20% of pregnant ladies in Uttar Pradesh, a Northern Indian state, prefer to give birth at home despite the possible health risks and contradicts popular beliefs, distance to the hospital is not the social determinant to such action. It turned out whether a pregnant lady had transport arrangement upon delivery and trust that the hospitals they were about to go, were the core reasons affecting their choice of giving birth at home. 

As such, Sgaier and Dominici stressed on the importance of moving AI away from disease prediction. They urged medical professionals to partner companies that look into Bayesian networks, structural equation models, and potential outcome frameworks to promote the generation of causal AI technology. To kick start, perhaps professionals can venture into “why” and search for data that can address the question

Author Bio
synthetic gene empathy chinese artificial intelligence data medicine healthcare ai

Hazel Tang

A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.