However , important limitations also exist for AI algorithms that users of the tool should understand , to ensure that these algorithms do not serve as a new source of error and safety concern . AI models perform as a function of the underlying data on which the algorithms are trained . Therefore , inherent issues with data may be perpetuated by AI algorithms . The two most concerning issues are bias and generalizability . For bias , a seminal study in Science exhibited that AI models can perpetuate racism in health care . For over-generalizing , since models are trained on a data set from a particular population and validated on their performance for that data set , it is important to understand that the application of that AI model to your own institution ’ s population , or the general public , may not perform as well as with the original trained population . Therefore , it is important to recognize that the performance of AI algorithms may deteriorate when used on different populations and that the validity may totally be lost if the models are overfitted to the training population , or if your population is significantly different .
For those makers of AI solutions , there are also a couple of lessons to consider specific to health care . The first is the importance of building trust between physician and technology . If an AI algorithm isn ’ t trusted it won ’ t be used . Physicians have a healthy skepticism of new technology . Therefore , creating algorithms that don ’ t just output a result , but also are transparent about the underlying training data and technology - that have explanation built-in - is critical to their adoption . For example , in radiology some algorithms just spit out a diagnosis without any supporting information . Compare this with an algorithm that shows you arrows or a heat map of where the finding is that informs the diagnosis . It is clear that the latter is more likely to be trusted and , therefore , utilized .
Another critical aspect to understand is clinical workflow and how the solution fits into this workflow and does not disrupt it . Physicians are incredibly busy and hum along to an existing workflow , ideally designed to be efficient . Introduce a new technology , and no matter how great the benefit , if there is an additional workstation or log-in involved , forget about a clinician going out of their way to access this system .
And finally , for those looking to sell AI algorithms into our practices , please know who the customer is . Health care is complex , and it can often be difficult to understand who the purchaser might be and who the ultimate end user might be . But lead with the value proposition , both clinical and administrative , and make sure not to make the same mistake that was abundant in radiology : antagonizing the customer . AI companies hyped up the idea that their algorithms could outperform a physician and , as a result , replace physicians without understanding that the decision-maker about technology purchase or adoption often was the physician themselves . Telling your customer that you are going to put them out of a job is not a recommended strategy for sales success ( not to mention , the combination of AI with physician always outperforms either one alone ). care . This is what is sexy , splashy and makes headlines . However , AI is really best used on tasks that are repetitive and repeatable that humans either do not like to do or better yet , simply are not great at . It isn ’ t good enough that we employ AI that does what humans do well , just incrementally better . Therefore , it is the backend functions and use cases that might have more value than the clinical-facing . Tedious administrative tasks , such as practice billing , coding and claims submissions , might become more automated , more accurate and efficient , improving collections and reducing practice administrative costs .
Perhaps the most important takeaway with the development of AI in health care is the critical role that physicians should play as key stakeholders . AI developers need to understand clinical value , the complexities of health care , workflow and real-world practice . A key example of this is seen in the challenges and poor user interface with EHRs , which had little clinician input and mostly input from administrators as a billing function , first and foremost . This is in comparison with PACS systems , which have made radiology more efficient and accurate , because of its development in conjunction with clinicians . As important is the role of our medical professional organizations engaging in AI development . We ’ ve seen such leadership with the American College of Radiology and its Data Science Institute , and the AMA ’ s Physician Innovation Network to ensure that the clinical perspective remains a key component of technology development . This engagement helps put the guardrails on patient safety , workflow efficiency , clinical value and user needs such that technology remains a tool that makes us better rather than a threat or hindrance to practice .
References :
ARTIFICIAL INTELLIGENCE
1
Wachter R . ( 2015 ) The Digital Doctor : Hope , Hype , and Harm at the Dawn of Medicine ’ s Computer Age . McGraw Hill .
2
Allen B , Agarwal S , Coombs L , et al . 2020 ACR Data Science Institute Artificial Intelligence Survey . Journal of the American College of Radiology . https :// doi . org / 10.1016 / j . jacr . 2021.04.002
3
Obermeyer Z , Powers B , Vogeli C , Mullainathan S . Dissecting racial bias in an algorithm used to manage the health of populations . Science . 2019 Oct 25 ; 366 ( 6464 ): 447-453 .
4
Futoma J , Simons M , Panch T , et al . The myth of generalizability in clinical research and machine learning in healthcare . The Lancet Digital Health . 2020 Sep 1 ; 2 ( 9 ): E489-92 .
Dr . Ding practices radiology at UofL and is a medical director at Humana . He is a former consultant to Google on their AI and machine-learning algorithms in health care .
AI offerings often focus on the clinical-facing side of health
APRIL 2022 21