At first glance, it seems a bit paradoxical, but AI is probably well-equipped to succeed in healthcare. This is because of the probabilistic AI models that we often use today. These models don’t come up with a single, definitive, or correct answer. Instead, they show us something that’s probably right, though it could be wrong. But why is this acceptable when it comes to our collective and individual health; to matters of life and death?

The short answer is that healthcare involves a lot of guesswork, and AI models are now better at guessing than trained professionals. Of course, this is an oversimplification. However, in healthcare, there is often no single, definitive, right answer. Instead, we deal with probabilities and patterns: if patients have symptom X, they are likely to have disease Y and require treatment Z.

It is only when treatment Z fails that we need to take a closer look. Or if other symptoms lead to a different diagnosis. Healthcare professionals continually collect data from their patients to assess their condition and adjust treatment accordingly. They are learning from collective and individual experience, from an ever-growing body of evidence, and the progress being made.

And they rely on their patients’ compliance. If patients don’t follow their doctor’s advice, the situation can get out of control. But it’s about more than just compliance. Patients will, usually, follow their individual preferences and values, and healthcare professionals must take these into account. Besides, imaging techniques and biomarker diagnostics provide us with a wealth of often ambivalent data.

The real magic

This adds to the complexity of modern medicine. While AI has surpassed humans in its ability to interpret medical data, it still requires human oversight and judgment. Having a good initial estimate available helps human healthcare professionals. The real magic lies in the combination. AI can free up doctors’ time, enabling them to refine their diagnoses and focus on therapy and care.

Almost a decade ago, Geoffrey Hinton famously recommended stopping the training of radiologists because, within five years, AI would outperform them. However, today the Mayo Clinic employs more radiologists than at the time of Hinton’s prediction. The clinic has also established a 40-strong AI team comprising AI scientists, radiology researchers, data analysts, and software engineers.

Overall, the Mayo Clinic is using more than 250 A.I. models, both developed internally and licensed from suppliers. The radiology and cardiology departments are the largest consumers.

In some cases, the new technology opens a door to insights that are beyond human ability. One A.I. model analyzes data from electrocardiograms to predict patients more likely to develop atrial fibrillation, a heart-rhythm abnormality.

A research project in radiology employs an A.I. algorithm to discern subtle changes in shape and texture of the pancreas to detect cancer up to two years before conventional diagnoses. The Mayo Clinic team is working with other medical institutions to further test the algorithm on more data.

AI excels at pattern recognition, which is precisely what much of healthcare involves.

AI and public health

In addition to individual health, there is also public health. First and foremost, this is a statistical entity – it is probabilistic by nature rather than deterministic. Public health deals with populations and cohorts rather than individuals. The focus is on prevention and mitigation, as well as limiting the disease burden.

During the pandemic, we learned that a mismatch hinders public health: although individuals bear the burden of the measures, it is difficult for them to quantify the benefits for themselves. They are not ill, after all, and prevention is not an immediately tangible benefit for the individual. At the same time, the reduction in disease burden for larger cohorts can be very well measured.

Probabilistic AI can help recognise patterns and thus distinguish between more effective and less effective measures. The calculation of probabilities can thus lead to more efficient strategies for combating a future pandemic. The same applies to other non-pharmaceutical interventions, such as preventive medical check-ups.

The VUCA of healthcare

Pandemics and public health in particular, and healthcare in general, are, surprisingly or not, prime examples of VUCA: volatility, uncertainty, complexity and ambiguity.

• The volatility of healthcare is reflected in rapid technological advances, changing regulations, frequent policy changes and unpredictable events such as pandemics or drug shortages.
• Uncertainty arises from unpredictable patient outcomes, evolving medical knowledge and the challenge of predicting disease outbreaks or treatment responses.
• Complexity is inherent in the interconnectedness of healthcare systems, the diversity of patient populations, the multitude of stakeholders and the intricate web of clinical, administrative and regulatory requirements.
• Ambiguity is present in the interpretation of clinical data, the application of new research, and the need to make decisions with incomplete or conflicting information.

Given all this, what can probabilistic AI do to navigate the waters of healthcare?

• Managing Volatility: AI analyses real-time data to help healthcare systems respond quickly to sudden changes, like outbreaks or patient surges.
• Reducing Uncertainty: AI predicts patient outcomes and synthesises data, supporting more confident, evidence-based decisions.
• Handling Complexity: AI processes vast, complex data (like genomics and medical records) and automates admin tasks, making healthcare operations smoother.
• Clarifying Ambiguity: AI provides probabilistic assessments and highlights patterns, helping clinicians make sense of unclear situations.

Of course, AI is not a magic solution to all the problems burdening our healthcare systems. Nor will it replace human professionals such as radiologists. However, the way they work will change, probably for the better:

Dr. Halamka, an A.I. optimist, believes the technology will transform medicine.

“Five years from now, it will be malpractice not to use A.I.,” he said. “But it will be humans and A.I. working together.”

Dr. Hinton agrees. In retrospect, he believes he spoke too broadly in 2016, he said in an email. He didn’t make clear that he was speaking purely about image analysis, and was wrong on timing but not the direction, he added.

In a few years, most medical image interpretation will be done by “a combination of A.I. and a radiologist, and it will make radiologists a whole lot more efficient in addition to improving accuracy,” Dr. Hinton said.

That’s a very human approach to applying AI to an inherently human problem.

Image by Kaja Sariwating / Unsplash.