Will Artificial Intelligence Transform Translational Medicine

The news that IBM was exploring a potential sale of Watson Health business (1), which was created to “help solve some of the world's most pressing health challenges through data, analytics and artificial intelligence” (2), sent shock waves through the business and health care communities. My first introduction to IBM’s Watson was when the artificial intelligence (AI) computer ran up the score on Ken Jennings and Brad Rutter, who were 2 of the most successful human participants to play Jeopardy. I watched in dismay as the canny computer hit the $1,200 Daily Double, allowing Watson to win $17,973 more than Jennings, thereby eliminating any chance that Jennings had of mounting a late come back during Final Jeopardy. I would have thought that after winning at Jeopardy, the problems presented by health care would be no match for Watson. Indeed, soon thereafter, IBM began building AI products in genomics, medical imaging, and cancer treatments, from drug discovery to clinical trial recruitment. However, ultimately, IBM lacked the technological and scientific expertise to implement their vision (3). What can we learn from this?

The promise of personalized medicine, where “omics,” cardiovascular imaging, and machine learning algorithms can be used by physicians to accurately and precisely identify which patients will benefit from specific therapies is still a promissory note. Although AI and machine learning have been used to increase the success of translational science in drug discovery, imaging, and genomic medicine, the current limitations with respect to accuracy, reproducibility, and governance of AI will need to be sorted out before this technology can be implemented routinely in health care delivery.

Big Data Bigger Problems

One of the major problems with large health care datasets is they lack a common infrastructure, so it can be difficult to harmonize data across different platforms. Moreover, large datasets such as electronic health care records, insurance claims, and pharmacy are often warehoused in different locations in the cloud. This type of fragmentation can decrease the comprehensiveness of large datasets and increase the likelihood of introducing errors in AI training. There are also inherent risks involving bias and inequity in AI systems. For example, if AI systems train on datasets that reflect an underlying racial or sex bias in the health care data, then the AI algorithms that are developed may exacerbate the underlying inequities that already exist in health care systems. As 1 example, AI algorithms for interpreting chest x-rays that were trained on sex-imbalanced datasets have significantly lower performance when interpreting chest x-rays for the underrepresented sex (4). Currently, there are no policies or procedures in place to mitigate the introduction of bias into AI programs, nor do we have laws in place that will allow us to adequately prevent discrimination in AI systems. Most AI products that are submitted to the Food and Drug Administration (FDA) are reviewed through the 501(k) pathway, which requires proof of substantial equivalence to an existing device, rather than evidence that the new device will improve patient care. Moreover, manufacturers are not required to specify the size of the training set nor whether there was an independent validation set. Although the FDA has issued an action plan for assessing Artificial Intelligence/Machine Learning-Based Software as a Medical Device (5), according to a recent STAT survey of 161 FDA-approved AI products, <45% of companies publicly disclosed the amount of patient data that was used to validate their device performance, <9% provided information on sex, and <5% reported the race of their study population (6).

Despite these difficulties, Google, Amazon, and Microsoft are all aggressively expanding into the health care business. These companies have the requisite resources and teams of data scientists that will be required to build the evidence base to support the various potential applications of AI in health care. However, if the past is prologue, the successful application of AI to cardiovascular medicine will also require a closer collaboration between patients, health care workers, health care organizations, and government agencies, lest we get a second opinion from a black box that will jeopardize our health care. As always, we welcome your thoughts on who you think should replace the incomparable Alex Trebek as the host on Jeopardy, or alternatively the role of AI in health care, either through social media (#JACC:BTS) or by e-mail (jaccbts@acc.org).