Study Finds Google’s Kidney Disease AI to be Less Effective on Women

Study Finds Google’s Kidney Disease AI to be Less Effective on Women in 2019, Google developed an artificial intelligence system that could predict severe kidney injury in patients up to two days in advance. Since kidney-related injuries are a common cause of death among hospitalized patients, the results were met with high hopes for the future. The Department of Veteran Affairs said it would immediately start work to bring the system to the market.

However, a recently-published study has discovered that the AI system is not as effective when applied to females as compared to males. Another big problem is that simply modifying the system’s gender representation and sample size might not be enough to fix the algorithm.

This possible problem in the initial study was acknowledged by the DeepMind and Veterans Affairs researchers. They noted that female patients made up less than 7% of the total patients in the study, and their model performance was very low compared to their male counterparts. However, it must be acknowledged that their findings only took into account patients suffering from early-stage kidney failure.

These results were expanded upon in a more recent study of the deep learning AI led by a research team at the University of Michigan, which found that the model consistently misidentified female patients regardless of the severity of their acute kidney injury.

The team also mentioned that when it added more women to the dataset, the performance discrepancies showed improvement among the general population. The problem persisted, however, when the AI was applied to a group consisting entirely of veterans, despite the enhanced controls for gender.

This points to the fact that the problem isn’t solely the underrepresentation of women, according to the researchers; there are greater underlying issues at play. These might include practice patterns and the varying characteristics of female patients.

About 1.7 million people per year lose their lives due to acute kidney injury, which is not easy to spot, and patients’ conditions often worsen before they receive potentially lifesaving treatment. A recent study found that 28% of those who develop the condition within the VA die within a year, with over 5% dying in the hospital.

In the initial study of Google’s AI system, accurately predicted 90% of the patients in the VA whose kidney function eventually deteriorated to such an extent that they required dialysis. These sorts of results are typically considered breakthroughs in the medical AI space, but it’s often ignored how much more work lies ahead to ensure that the technology works well across different healthcare settings and on different kinds of patients.

Jie Cao, the lead author of the study and a Ph.D. student at the University of Michigan, said, “We still have a long way to go in terms of using these models to change how health care works and to change a patient’s health.”

She added that hasty adoption of Google’s model might result in poor treatment choices by clinicians who might not be aware of its flaws.

Google has announced new initiatives designed to speed up the company’s AI research and development. The company’s researchers have released protocols and open-source code to implement whichever of its two models performs better. Furthermore, Google Cloud is providing a pilot-ready version of the model to healthcare systems worldwide that want to evaluate its potential benefits for their patients.