Critical annotations: AI may be as effective as medical specialists at diagnosing disease

In September 2019, CNN published this article about an AI research study. It buries the lede, seemingly intentionally: the spotlight is on the success of AI tools in diagnosis, whereas the study finds that fewer than 1% of papers on AI tools follow robust reporting practices. In fact, an expert quoted at the end of the article stresses that this is the real message of the study.

Besides, the cover image of the article is a robot arm shaking hands with a human, even though the study is about finding patterns in medical images. These humanoid images give a misleading view of AI, as we’ve described here.

See for yourself where the article goes wrong. The contents of the article are reproduced in their entirety on the left, and our annotations are on the right. Link to the original article.

For more details about the pitfalls, read the blog post which introduces this article here.

AI may be as effective as medical specialists at diagnosing disease

The title implies that AI and medical specialists are competing on an equal footing, while hiding the fact that AI tools only perform a narrow slice of the variety of steps that comprise a diagnosis.

The image of a robot arm shaking hands with a human fails to convey that the AI tools being discussed here having nothing to do with robots—in fact, they are used to find patterns in medical data.

A new scientific review has concluded that artificial intelligence (AI) may be able to diagnose disease as successfully as human healthcare professionals , but a lack of quality studies means the real potential of the technology is unclear.

Similar to the title, this sentence gives the impression that AI tools and medical specialists are competing on an equal footing, instead of highlighting that even in the best case, AI tools can only perform a small part of medical diagnoses.

Researchers carried out the first systematic review of existing research into AI in the health sector and published their findings in The Lancet Digital Health journal.

It focused on an AI technique called deep learning, which employs algorithms, big data, and computing power to emulate human intelligence .

“Emulating human intelligence” is not an accurate description for what deep learning does. It gives readers a false intuition that deep learning algorithms compare with human intelligence.

This allows computers to identify patterns of disease by examining thousands of images, before applying what they learn to new individual cases to provide a diagnosis. Excitement is building around the technology, and the US Food and Drug Administration has already approved a number of AI algorithms for use in healthcare.

AI has been hailed as a way to reduce the workload for overstretched medical professionals and revolutionize healthcare, but so far scientific research has failed to live up to the hype.

Of the 20,500 articles reviewed, fewer than 1% were found to be sufficiently robust, said Professor Alastair Denniston from University Hospitals Birmingham NHS Foundation Trust, UK, which led the research, in a statement.

The article buries this important detail in the 6th paragraph instead of highlighting that a large percentage of studies suffer from robustness issues.

“Within those handful of high-quality studies, we found that deep learning could indeed detect diseases ranging from cancers to eye diseases as accurately as health professionals,” said Denniston.

“But it’s important to note that AI did not substantially out-perform human diagnosis.”

Using data from 14 studies, researchers found that deep learning algorithms correctly detected disease in 87% of cases, compared to 86% for healthcare professionals.

AI was also able to correctly identify those patients free from disease in 93% of cases, compared to 91% for healthcare professionals.

In general, reporting performance numbers in a news article is problematic. It is almost never possible in the space of a news article to convey to a reader what the numbers mean or how they were measured. For instance, the article does not even mention which diseases were being diagnosed using the AI tool. Admittedly, this is not the most egregious example of this pitfall, because the article is using the performanc numbers in a relative comparison and emphasizing that the difference between these numbers is small.

While these results are promising, the researchers say better research and reporting is needed to improve our knowledge of the true power of deep learning in healthcare settings.

This will involve better study design, including the testing of AI in situations that are the same as those that healthcare professionals work in.

“Evidence on how AI algorithms will change patient outcomes needs to come from comparisons with alternative diagnostic tests in randomized controlled trials,” said Livia Faes, from Moorfields Eye Hospital, London, in a statement.

“So far, there are hardly any such trials where diagnostic decisions made by an AI algorithm are acted upon to see what then happens to outcomes which really matter to patients, like timely treatment, time to discharge from hospital, or even survival rates.”

Experts hailed the review while emphasizing the need for further research.

“The big caveat is, in my opinion, that the story is not ‘AI may be as good as health professionals’, but that ‘the general standard of evaluating performance of AI is shoddy,’” said Franz Kiraly of University College London.

Despite an expert taking pains to point out what the real message is, the article buries this information at the far end of the article.

Nils Hammerla of Babylon Healthcare, a company that says it uses AI technology to improve the affordability and accessibility of healthcare, believes more work is needed before AI can reach its full potential.

“Machine learning can have a massive impact on problems in healthcare, big and small, but unless we can convince clinicians and the public of its safety and ability then it won’t be much use to anybody,” he said.

The global market for AI in healthcare is surging and is expected to rise from $1.3 billion in 2019 to $10 billion by 2024, according to investment bank Morgan Stanley.

Hospitals around the world are already making use of the technology, including Moorfields Eye Hospital in London.

Doctors are able to use an algorithm developed by DeepMind, a UK-based AI research center owned by Google, to return a detailed diagnosis in around 30 seconds using Optical Coherence Tomography (OCT) scans.

And AI technology can accurately identify some rare genetic disorders using a photograph of a patient’s face, a ccording to a study published in January.

For these and the following example of AI tools being used in healthcare, there is no space given to potential limitations such as privacy risks, validity issues, and bias. Instead, in presenting only the positive potential of these technologies, the article falls short of presenting additional context of how many AI tools in healthcare have failed in the past.

The AI technology, called DeepGestalt, outperformed clinicians in identifying a range of syndromes in three trials and could add significant value in personalized care .

Similar to the title and opening line, this makes it sound like AI tools and clinicians are competing on a similar footing, whereas AI tools are only able to match clinicians on a narrowly defined task.

Copyright to the original article belongs to CNN, and the article can be read in its original form in full at cnn.com. It is republished here for the purposes of critical commentary.

We re-used the code and the copyright notice above from Molly White et al.’s The Edited Latecomers’ Guide to Crypto to generate our annotated article.