Some Straight Talk on A.I. In Health Care

“Artificial Intelligence will help everyone become a better doctor in the future by eradicating waiting time, prioritizing emails, finding relevant information or making hard decisions rational.“

Bertalan Mesko, M.D. – Director, The Medical Futurist Institute
Image credit:

It’s important to separate the hype from the reality in our quest to incorporate A.I. into care delivery.

Image Credit: Sg2, Henry Soch Executive Summit presentation 2017

In a contributed article to Mobihealthnews on July 1st, Dr. Liz Kwo highlights her “Top 10 Use Cases for AI in Healthcare”. This comprehensive overview outlines the current major categories that health care providers are exploring to incorporate AI into the clinical care continuum. I applaud Dr. Kwo for synthesizing this information and share her enthusiasm for the potential of AI in supporting care teams in their daily work.

However, I do think it is important to balance the tremendous potential of the technology with the current reality. There’s an old saying that goes: “We overestimate the impact of technology in the short-term, and underestimate the impact in the long-term.” That certainly applies to AI in health care. The amount of “digital ink” devoted to the topic would fill several data warehouses.

CB Insights conducted a survey at the end of last year and asked which areas of healthcare would be impacted most by AI. The results are shown in the graphic below:

Image credit: CB Insights, 2020

Every year, Gartner publishes their “Hype Cycle” analyses across multiple industries. Their methodology helps us to separate the hype from reality when considering whether (and when) we decide to incorporate AI into clinical workflow. If you are not familiar with the Gartner Hype Cycle model, this is the best book that I’ve found on the subject: “Mastering the Hype Cycle: How to Choose the Right Innovation at the Right Time.”

Below is the Gartner Hype Cycle for Healthcare Providers 2020.

Image credit: Gartner, Inc.

We can see that, according to Gartner, AI is approaching the “Peak of Inflated Expectations”, and is not expected to reach “The Plateau of Productivity” for five to ten years. With that information as our baseline, let’s examine each of Dr. Kwo’s use cases and where we stand today.

1. AI supports medical imaging analysis – AI is used as a tool for case triage. It supports a clinician reviewing images and scans. This enables radiologists or cardiologists to identify essential insights for prioritizing critical cases, to avoid potential errors in reading electronic health records (EHRs), and to establish more precise diagnoses.

Where we are – The best assessment of where we stand in the implementation of AI in imaging can be found in an Aunt Minnie article by Michael Cannavo (akaPACSMAN) here. To quote him: “Unless you perform a large volume of studies that would benefit from the use of a CT algorithm stroke protocol, lung CT algorithm, or others, you may be hard-pressed to justify the purchase of AI software without obtaining additional revenue. There are significant advantages to using AI, but no clear path for how it will pay for itself in hard dollars.”

2. AI can decrease the cost to develop medicines – Supercomputers have been used to predict from databases of molecular structures which potential medicines would and would not be effective for various diseases. By using convolutional neural networks, a technology similar to the one that makes cars drive by themselves, AtomNet could predict the binding of small molecules to proteins by analyzing hints from millions of experimental measurements and thousands of protein structures.

Where we are – While many pharmaceutical companies and academic medical centers are exploring the use of supercomputers to help identify and develop new drug targets, the benefits are as of yet unproven. The introduction of “high throughput screening,” using robots to test millions of compounds rapidly, generated mountains of leads in the early 2000s but notably failed to solve inefficiencies in the research process. When it comes to AI, big pharma is treading cautiously. The technology has yet to demonstrate it can successfully bring a new molecule from computer screen to lab to clinic and finally to market.

3. AI analyzes unstructured data – In many cases, health data and medical records of patients are stored as complex unstructured data, which makes it challenging to interpret and access. AI can seek, collect, store and standardize medical data regardless of the format, assisting repetitive tasks and supporting clinicians with fast, accurate, tailored treatment plans and medicine for their patients instead of being buried under the weight of searching, identifying, collecting, and transcribing the solutions they need from piles of paper formatted EHRs.

Where we are – The author’s assertion that AI can collect, aggregate, and standardize all of this data regardless of the format is much easier said than done. The lack of data standards in healthcare continues to be the key stumbling block to an accurate, longitudinal patient record. And, collecting disparate data sets and placing that data in the proper context for review is a daunting challenge that has yet to be overcome.

4. AI builds complex and consolidated platforms for drug discovery – AI algorithms can identify new drug applications, tracing their toxic potential as well as their mechanisms of action.

Where we are – See the comments under point number two above.

5. AI can forecast kidney disease – In 2019, the Department of Veterans Affairs (VA) and DeepMind Health created a ML tool that can predict Acute Kidney Injury (AKI) up to 48 hours in advance. The AI tool was able to identify more than 90% of acute AKI cases 48 hours earlier than with traditional care methods.

Where we are – The partnership between VA and DeepMind Health continues. Its next target is to identify how this ML tool can be installed in medical units. A user-friendly platform is also targeted to support clinicians in their treatment decisions that would improve the quality of life for Veterans suffering from AKI.

6. AI provides valuable assistance to emergency medical staff – During a sudden heart attack, the time between the 911 call to the ambulance arrival is crucial for recovery. For an increased chance of survival, emergency dispatchers must recognize the symptoms of a cardiac arrest to take appropriate measures. AI can analyze both verbal and nonverbal clues to establish a diagnostic from a distance.

Where we are – Implementing these types of AI-assisted tools in the EMS is an expensive proposition. That, coupled with the fact that most EMS departments are run at the local town or city level, makes it difficult to achieve scale in deploying these tools. Also, one must consider the training requirements for proper implementation as well.

7. AI contributes to cancer research and treatment, especially in radiation therapy. In some cases, radiation therapy can lack a digital database to collect and organize EHRs, making the study and treatment difficult. To assist clinicians in making informed decisions regarding radiation therapy for cancer patients, a platform has been developed that collects the relevant medical data of patients, evaluates the quality of care provided, optimizes treatments, and offers specific oncology outcomes, data, and imaging.

Where we are – In a classic case of hype versus reality, consider IBM’s Watson and cancer care. Watson’s entry into cancer care and interpretation of cancer genomics was, just like its appearance on Jeopardy!, highly hyped, with overwhelmingly positive press coverage and little in the way of skeptical examination of what, exactly, Watson could potentially do and whether it could improve patient outcomes. An article in STAT looked at Watson for Oncology’s use, marketing, and actual performance in hospitals around the world, interviewing dozens of doctors, IBM executives, and artificial intelligence experts and concluded that IBM released a product without having fully assessed or understood the challenges in deploying it and without having published any papers demonstrating that the technology works as advertised, noting that, as a result, “its flaws are getting exposed on the front lines of care by doctors and researchers who say that the system while promising in some respects, remains undeveloped.” Quoting the STAT authors:

Perhaps the most stunning overreach is in the company’s claim that Watson for Oncology, through artificial intelligence, can sift through reams of data to generate new insights and identify, as an IBM sales rep put it, “even new approaches” to cancer care. STAT found that the system doesn’t create new knowledge and is artificially intelligent only in the most rudimentary sense of the term.

STAT – Casey Ross & Ike Swetlitz

8. AI uses data collected for predictive analytics – Turning EHRs into an AI-driven predictive tool allows clinicians to be more effective with their workflows, medical decisions, and treatment plan. NLP and ML can read the entire medical history of a patient in real-time, connect it with symptoms, chronic affections, or an illness that affects other family members. They can turn the result into a predictive analytics tool that can catch and treat a disease before it becomes life-threatening.

Where we are – The buzzword fever around predictive analytics will likely continue to rise and fall. Unfortunately, lacking the proper infrastructure, staffing, and resource to act when something is predicted with high certainty to happen, we fall short of the full potential of harnessing historical trends and patterns in patient data. In other words, without the willpower for clinical intervention, any predictor – no matter how good – is not fully utilized.

9. AI accelerates the discovery and development of genetic medicine – AI is also used to help rapidly discover and develop medicine with a high rate of success. Genetic diseases are favored by altered molecular phenotypes, such as protein binding. Predicting these alterations means predicting the likelihood of genetic diseases emerging. This is possible by collecting data on all identified compounds and on biomarkers relevant to specific clinical trials.

Where we are – If you live in the U. S. you’ve undoubtedly seen various cancer treatment centers talking about their personalized therapy plans, and especially how they’ll tailor things to your DNA sequence and so on. You would get the impression that we have an arsenal of specifically targeted cancer therapies, waiting for patients to get their tumors sequenced so they can be paired with the optimal treatment. That’s not true. I wish it were, but it just isn’t.

There are estimates that only about 15% of patients total are currently even eligible (under FDA guidelines) to have their tumors sequenced in the hope of matching with a targeted therapy. About one-third of those may actually benefit from the process in the end. This is not exactly what you’d expect if all you knew about this stuff was what you heard on TV. The thing is, that’s actually a significant advance because the number used to be zero in both categories. We really are making progress, and the people who can benefit really can benefit. It’s just that there aren’t nearly as many of them as we’d like, not yet.

10. AI supports health equity – Those responsible for applying AI in healthcare must ensure AI algorithms are not only accurate but objective and fair.

Where we are – Unfortunately, as has been demonstrated in many of the AI algorithms that exist today, we cannot assume that all relevant factors were applied in the training set of the AI algorithms. The medical datasets openly available for use by AI researchers are notoriously biased, especially in the US. It’s not a secret: Healthcare data is extremely male and extremely white, which has real-world impacts. (For a deeper discussion on this topic, check out Amber M. Hamilton’s article in Slate’s Future Tense Silicon Valley Pretends That Algorithmic Bias Is Accidental. It’s Not.) Questions that need to be addressed include: Is the selection of the training factors evidence-based? Are race, gender, and ethnicity data included in the training data set?

Adding a link here to an excellent, insightful blog post from John Halamka, M.D. titled “Learning from AI’s Failures. I love his concluding paragraph:

If we are to learn from AI’s failures, we need to evaluate its products and services more carefully and develop them within an interdisciplinary environment that respects all stakeholders.

John Halamka. M.D., President, Mayo Clinic Platform

AI holds great potential in improving care delivery by optimizing the use of scarce resources and eliminating repetitive and non-value-adding work for the care team. However, AI adoption in healthcare continues to have challenges, such as a lack of trust in the results delivered by an ML system and the need to meet specific requirements. It is essential to take a realistic approach when considering the implementation of AI in your organization. At this time, the best Return on Investment (RoI) case for AI involves operational use cases (e.g., bed management, staffing management, supply chain optimization, etc.) Certain clinical use cases like imaging, dermatology, and pathology can improve workflow and prioritize work lists for those disciplines.

I listed seven ways to prepare for incorporating AI in health care organizations during one of my presentations in 2017. I think they are still relevant today.

Image credit: Sg2, Henry Soch Executive Summit presentation 2017

If you are interested in digging a bit deeper into the topic of AI in health care, I highly recommend this online learning course from my friend and colleague Tom Giordano. His “Plain and Simple” series of courses are excellent.

Image credit: Quovadis Learning Systems

4 thoughts on “Some Straight Talk on A.I. In Health Care

Leave a Reply