STAT recently published an article demonstrating how researchers at Google showed than an artificial intelligence system had the ability to predict acute kidney injury up to 48 hours in advance. Acute kidney injury, which kills approximately 1.7 million around the world annually, is difficult to recognize and can cause patients to “deteriorate rapidly”.
In 2019, the Department of Veterans Affairs found the results promising and opted to begin implementing the system. Recently, a new study has shown that a replica of the AI system, which was trained predominantly on male veterans, does not perform as well on women. The study overestimated the risk for women in certain circumstances and was “less accurate in predicting the condition for women overall”. The results emphasize the need to train AI models on diverse groups of patients and test them among diverse local populations with demographic differences. Particularly, variations in health care delivery may impact the system’s accuracy. It is crucial to ensure that AI works equally well across health care settings and with different patients, “regardless of their race, income level, or gender”.
Google’s AI system, developed by the Alphabet research unit DeepMind Health (since merged with Google’s health division) analyzes patient data to reveal which patients would develop the condition up to two days in advance. In the initial study, the system accurately predicted 9 out of 10 patients in the VA facility whose kidney function declined to the point of needing dialysis. However, as mentioned, the AI models are prototypes “in need of further refinement and evaluation for effectiveness and fairness”. Google has stated it is now taking steps to advance the prototype, with researchers publishing protocols for implementing the “better-performing of its two models, along with open source code” hoping that other clinical researchers are able to build upon their initial findings.
