AC OBSERVATION

Dr. Anthony Chang

“I don’t care what anything was designed to do. I care about what it can do.”
- Gene Kranz, NASA flight director 

As a senior clinician with a data science and artificial intelligence background and focus, I am in a very privileged position of communicating continuously and concomitantly with both clinicians and data scientists around the world in the domain of AI in healthcare. 

AI has had some limited successes in clinical medicine and healthcare, such as deep learning in medical imaging and protein structure determination, but its panoply of capable tools has not been as widely adopted as we had all hoped in healthcare. Here are some of the reasons (five under the category of “technology” and an additional five under “human”) why AI has had some challenges in healthcare, and how each of these challenges can be neutralized with a potential solution. 

 

Technology 

Healthcare data are often in disarray and often sequestered in different locations.

Healthcare databases are often sequestered (on paper even) in various locations in the healthcare system and are frequently not organized into a database repository (“data archipelago”). 
Solution: Health systems need to set up a data governance structure for the entire organization as well as a data visual map with a template and overlay for each project. 

 

There is an overall lack of appreciation for how inaccurate healthcare data can be.

It is estimated that up to 30% of healthcare data in electronic medical records and other sources are inaccurate and/or incomplete (“print bias”).
Solution: Both data scientists and clinicians need to be involved in data curation and data education to improve accuracy and completeness of healthcare databases. 

 

Better predictions in models do not mean there will be improved patient outcomes.

There is often an understandable fixation on the AUROC metric and yet often hardly any focus on any potential impact on change in health outcome (“scoring runs but not winning games”). 
Solution: Data science projects should be accountable for both real-world deployment and also long term followup for any potential change in patient outcome. 

 

Artificial intelligence is better suited for complicated than for complex situations.

Data science is better suited for deconstructing dimensions that are predictable (engineering and physics) but less ideal for solving problems in nature (weather) or humans (sports or pathophysiology).
Solution: Data science needs to progress beyond deep learning and explore other sophisticated AI methodologies such as reinforcement learning and cognitive architecture.

 

Universal hesitation to share raw healthcare data amongst healthcare institutions.

There is an understandable reluctance amongst clinicians and administrators to share raw patient data for collaborative studies across institutional boundaries (“willing to collaborate but not to share”). 
Solution: Innovative AI-related technologies such as federated and swarm learning as well as synthetic data generation can decrease the absolute necessity to share raw patient data. 

 

Human

Team focus is misplaced on academic publications instead of clinical practice.

The literature is replete with publications that focus more on technical aspects of AI in healthcare but much less on how these AI tools are implemented in real-world practice (the “publication-to-practice” chasm). 
Solution: AI in healthcare projects should start with real world problems that need to be solved with both clinician and data scientist co-champions.  

 

Insufficient awareness and appreciation for each other’s expertise and domains.

While data scientists often overestimate their understanding of many nuances of clinical medicine and healthcare, clinicians often lack full comprehension of health data, databases, and data science.  
Solution: Both clinicians and data scientists should allocate time in each other’s domains as well as collaborate continuously on projects from ideation to implementation.  

 

New technology such as artificial intelligence is slow to adopt in healthcare.

This slow adoption of new technology is perhaps even slower in healthcare with its complex ecosystem and challenging finances. Many health systems still possess fax and copy machines with a heavy reliance on paper.  
Solution: Artificial intelligence can serve as a North Star to inspire a more expedient transformation of an inefficient healthcare ecosystem towards precision medicine. 

 

Excessive concern about regulatory, ethical, accountability, and legal aspects of AI in healthcare.

We need to balance the current asymmetric discussion about the regulatory, ethical, accountability, and legal (REAL) issues so that we also discuss these issues in the context of not having AI. 
Solution: The front line practitioners of AI in healthcare needs to organize a proactive effort to consider these issues and implement measures to maximize the Quintuple Aim.   

 

Not enough access to data science expertise devoted to healthcare.

Even with the emergence of more user-friendly AI tools such as large language models, there remains a dire need for AI expertise and guidance in healthcare. 
Solution: We need to educate and train as many clinicians in the realm of data and data science as early in their careers as possible. In addition, we need to nurture a cohort of clinician-data scientists. 

 

Artificial intelligence in healthcare will need to be reshaped for it to be successful in its adoption. Even with the aforementioned shortcomings, sound solutions can be implemented for all of these challenges. The most valuable resource in healthcare is not artificial intelligence, but rather human intelligence and dedication with passion to transform healthcare. 

These insights and discussions on AI in healthcare and how we can successfully adopt AI will be discussed at the in-person AIMed24 meeting scheduled currently for November 17-19, 2024 at the sublime Caribe Royale resort in Orlando, Florida. 

See you there!

- ACC