There is an ongoing debate about the best way to leverage AI’s potential to improve healthcare without compromising patient safety. Is the best use case an assistive function, where the AI aids the physician in making a diagnosis? Or will AI have the greatest impact on healthcare with an autonomous use case, where the AI makes an independent diagnosis?
A new Google study pending publication in the journal Ophthalmology supports an assistive use case for AI to diagnose diabetic retinopathy. While there are some concerns about the study methodology – in particular, using the test set to tune the algorithm – researchers found that AI and physicians working together can be more accurate than either one alone. Ten ophthalmologists were asked to read retinal images under one of three conditions: unassisted, grades only, and grades + heatmap.
Both types of assistance improved some specialists' diagnostic accuracy; however, model assistance also slowed some physicians down. The AI system was also not tested in patient care, which can have a significant impact on the effectiveness of the AI.
Nonetheless, these are exciting findings, largely because other studies have had far less positive results when assessing the diagnostic accuracy of humans assisted by AI. Previous research has shown quality of care can suffer when humans start to rely on the AI too much and make the same mistakes as the AI. For example, according to a multicenter, retrospective study of 43 mammography facilities in the USA, the use of a computer-aided mammogram test was associated with significantly higher false positive rates and lower overall accuracy in screening.
While it’s good that Google’s study suggests that assistive AI can improve a specialist’s accuracy at detecting diabetic retinopathy, the study does not address a major problem our healthcare system faces - many at-risk patients are not seeing an eye care specialist regularly.
A recent Diabetes Care study tracked 298,383 insured patients with diabetes and found nearly 50% had no eye exam over a five-year study period. Only 15.3% met the American Diabetes Association (ADA) recommendations for annual or biennial eye exams. This means that whatever improvement in quality an assistive AI may provide to specialty care would never reach the vast majority of people with diabetes.
To really move the needle in addressing this care gap, we need to increase patient access to care. An autonomous AI that can make a real-time diagnostic assessment without a specialist’s oversight can be placed in primary care and retail clinics so that more patients can receive a high-quality diagnosis, even if they can’t access a specialist.
A 2018 study in Nature Digital Medicine showed an autonomous AI was capable of high diagnostic accuracy at detecting diabetic retinopathy in primary care. This led to the first ever FDA clearance of an autonomous AI system, IDx-DR, that does not require a specialist to interpret the results, making it usable by healthcare providers who are not normally involved in eye care.
There is no doubt AI has enormous potential to improve the quality of healthcare – study after study has shown AI has the ability to outperform humans with consistently high diagnostic accuracy. But AI promises to do so much more than that; it also stands to improve the accessibility and affordability of healthcare.
Why limit AI’s potential by confining it to specialty care? The patients who can benefit the most from this technology are not in the specialist’s office – they are at their neighborhood clinic.