Artificial intelligence should be focused on giving clinicians superpowers

What is a good design philosophy when it comes to introducing and applying AI in the clinic? The idea here is to focus on giving providers "superpowers." Rather than approaching AI as a way to replace providers or simply transfer their knowledge into a machine, the real focus should be on augmenting their existing skills and capabilities so that they can do their job even better while having more positive and humane interactions with their patients.

AI in the clinic

If there ever was a bandwagon that’s been jumped on, it’s that of applying artificial intelligence (AI) and machine learning (ML) in healthcare.[1]  The combination of hype, demonstrated usefulness in other fields and perceived potential within the clinic is generating a steady stream of headlines about how AI will transform healthcare to be better, faster and cheaper.

There are many examples where AI is having an impact or being seriously explored as a component of healthcare:

Imaging. Computer algorithms have been applied to medical images for decades now, but the deep learning approaches are showing real promise in detecting disease within images. Significant progress has been made in radiology, oncology and other specialties. This application area is perhaps the most advanced of any where AI is trying to be applied.

Electronic Health Records (EHRs). Extracting actionable information from patient records has been an active area of research and commercialization for decades now. While progress has been made in digitizing patient information and companies like Cerner and Epic cover ~50% of the market for EHR systems, extracting data that is both usable by an AI platform and that is actually useful clinically is still challenging. Regardless, leveraging patient data to provide better care is an area where AI is being rapidly developed and applied, with potentially huge downstream benefits.

Mobile health and telemedicine. Similar to EHR data, the number of sensors nearly all of us carry around can be used to build a rich and personalized picture of our health and behavior. Recent "hospital at home" efforts are a more narrowly focused area that is looking to support a patient’s recovery from procedures at home, rather than in the hospital. As these efforts continue to mature and new ones emerge, AI will undoubtedly play a key role in leveraging this information in clinically-relevant ways.

These are just a few samples of where AI is making its way into healthcare. The role of the physician is obviously pretty key in all of these efforts, primarily for their domain knowledge that is (hopefully) leveraged to help design, train and evaluate AI systems. However, is the best way to approach this problem simply just leveraging provider knowledge and skill and simply transferring it into some other external system? Should the goal of using AI be to replace doctors and nurses?

What is potentially a better design philosophy when it comes to implementing AI in the clinic?

Providers with superpowers

Before replacing people or tasks, we should focus more heavily on enhancing human capabilities. Perhaps some of the best opportunities to improve patient outcomes and the healthcare system at large come from leveraging AI to make skilled people even better at what they do.

This view entails a much more human-centric approach to design than what is commonly seen in AI applications today.[2] This relatively slight change in perspective, however, can have a significant change in how your problem is approached and what solutions are conceived.

Peoples' brains are amazing pattern recognition machines and trained brains are particularly good at finding and linking often subtle pieces of information into a coherent whole. This is exactly what we want a clinician to do for us when they are diagnosing one of our health issues.

Pattern recognition too, is exactly what AI, and deep learning in particular, is good at. But rather than just taking the provider out of the equation, integrating AI into a support role that amplifies their skills or fills gaps in their knowledge will multiply the effectiveness of how well they can do their job.

For instance, extensive training and experience is what underlies physician success at diagnosis, but this task is becoming increasingly challenging with growing amounts of new data types and measurements that are being captured in the clinic or at the home. The rate at which image, genomic, metabolic, wearable sensor and other data is being gathered and put into a form that could actually be used is massive, but transforming it into an actionable form is still a ways off. Controlled clinical trials will be needed to figure out whether or not certain data is actually of benefit, but understanding this potential value will be greatly facilitated by providing better, human-digestible portions, and AI can help do this.

AI could be used not to necessarily to give a diagnosis, but instead provide richer and clearer representations of complex multi-parameter data in a way that will help the clinician leverage their skills more effectively. Perhaps even better is in providing a range of diagnoses with an indication of how likely each is given the data at hand. Such an approach could prod a physician to consider a broader range of potential diagnoses, including ones they may not have considered before. This could be particularly valuable with hard to diagnosis conditions or for physicians lacking extensive background or training in a given area. Such a capability would have a significant impact in smaller rural communities with a smaller pool of clinical expertise to draw from.

AI systems that leverage massive EHR data and track and analyze what has happened to patients as they move through treatments over time or for chronic conditions and could be used to help doctors have more informed discussions with patients as to the potential consequences of a patient's decision - choice X leads to 50% of patients coming back into the hospital while choice Y leads to 95% not having a problem again. These "decision trajectories" are just one of a nearly infinite way EHR data could be leveraged to help clinicians and their patients.

While there are certainly academic and industry groups who are doing the type of work just described, this is relatively rare, with this human-centric approach to designing AI applications being an after-thought. Enhancement or augmentation of clinician capabilities ends up being more of a side-effect or consequence rather than a central design focus or constraint. Shifting to such a focus could make a massive improvement in AI adoption.

Advantages of this strategy

Advantages of a "superpower" approach are significant. First, it truly focuses on a (and in some cases the) key stakeholder, the clinician, who will help to define whether or not a solution is adopted. Having the provider front and center in the design process significantly improves the chance that both the right problem is being addressed while also gaining you an essential advocate that will be necessary for AI solutions to actually be adopted.

While regulatory issues are situation specific, augmenting a providers ability to potentially do a task better is very different than completely replacing them in doing that task. In many support-role applications, it is not necessarily obvious what degree or type of FDA approval is needed. The FDA recently released a "AI/ML-Based Software as a Medical Device Action Plan" and the regulation of AI/ML applications is still in early stages.[3] Regardless, augmentation of a provider's existing skill and knowledge base is distinctly different than trying to replace them.

From the entrepreneur's point of view, this approach is a fantastic way to figure out where and how to implement AI solutions. You are working directly with a primary customer to help identify a real need, making it easier to avoid developing an untested or unvalidated need. You are building relationships with your end users that will help you when it comes to validating whether or not you are on the right track. When it comes to getting it adopted in the clinic, you will need those end-users to be massive advocates for what you are trying to do.

Considering the demands placed on clinicians over the past several years, leveraging AI to support clinicians also provides the opportunity to reduce provider burnout and support a more sustainable work environment. More broadly, providing more time for personal and humane patient-clinician interactions should be a core factor motivating any AI effort.

We all want superpowers

Continuing advances in AI and the data that supports them will produce many opportunities to improve our healthcare systems. Central to the success of delivering on these opportunities is how best to approach getting AI into the clinic - testing and validating the methods, getting them through regulatory approvals, having them accepted and adopted by healthcare workers and patients where they can have their promised effect.

The idea here is that perhaps the best way for us to approach this challenge is to focus on finding places where AI can significantly enhance the effectiveness of our doctors, nurses, physical therapists, and other healthcare providers. By providing these highly-trained and skilled professionals with "superpowers" we have a much better chance of overcoming the significant challenges and roadblocks that are facing any effort at improving our healthcare system and improving patient outcomes. Besides, most of us would probably love to have a superpower or two and who wouldn't want a super-powered doc looking after your health?

References & related reading

Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. 2018. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ digital medicine 1:39.

Davenport T, Kalakota R. 2019. The potential for artificial intelligence in healthcare. Future healthcare journal 6:94–98.

Dias R, Torkamani A. 2019. Artificial intelligence in clinical and genomic diagnostics. Genome medicine 11:70.

Topol E. 2019. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. USA: Basic Books, Inc.

Footnotes

[1] AI is a very general term that includes subdomains like machine learning and deep learning. The issues discussed here apply broadly to the whole field, though the promise and challenges associated with deep learning methods are particularly relevant.

[2] The finding of needs should really be separate and independent from identifying the solution that can solve it. However, many groups entering or already working in this space are starting with their existing focus on AI methods.

[3] https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device