Skip to main content

Eliminating Bias in Osteoarthritis Treatment Using AI

Meet Eleanor and Eve. They share a two-room unit at a retirement home. One afternoon, while playing a game of shuffleboard, Eleanor and Eve both experience a sharp pain in their right knee. They decide to pay a visit to their doctor, and their situations are assessed with two methods: first, the doctor has them fill out a survey describing their pain, and second, they undergo a radiographic scan of their knee. However, Eleanor’s knee hurts significantly more than Eve’s and she vocalizes this on the survey. Despite this, the doctor concludes that they will receive the same treatment. But something else is different: Eleanor is Black, and Eve is white.

Eleanor is not alone in this situation. Every year, Black patients suffering from knee osteoarthritis (OA) describe experiencing more pain on average than white patients, but their concerns, like in Eleanor’s case, are dismissed due to the Kellgren-Lawrence grade, also known as KLG. The scale is based on radiography, but the categories of the measure are not equidistant. This essentially means that it cannot identify all possible sources of pain, and consequently, nor can it assess each patient’s situation entirely accurately. More specifically, as a recent study by Nature Medicine describes, the scale was “developed decades ago in white British populations.” Now that KLG is used on a more diverse group, its origin can prevent it from accounting for causes of pain specific to people of color. While the specific reasoning for this increased level of pain among Black patients remains unknown, potential reasons include either external factors, such as stress, or simply “more severe osteoarthritis within the knee,” as the study also suggests.

The importance of receiving treatment accurate to one’s personal illness cannot be understated; knee OA occurs in 10% of all men and 13% of all women 60 or older. The potential cost to the over-60 Black population presented by their lack of proper treatment will grow even more troubling as OA grows more common from an increasing aging population and obesity.

However, hope is not lost, as new medical algorithms lead the way to overcoming this problem. “[T]he researchers [who worked on the Nature Medicine study] trained a deep-learning model to predict the patient’s self-reported pain level from their knee x-ray,” writes Karen Hao for MIT Technology Review. What the accuracy of these predictions reveals about pain levels proves hopeful; the algorithms were not only more accurate than KLG, but this ability to correctly predict self-reported pain levels means that there does exist a relation between pain felt by a patient and the information found in their x-ray scan. Going forward, this can validate the self-assessment provided by Black patients experiencing greater pain and grant them access to proper treatment for their ailment.

The promising research serves as a shining example of the benefits of the use of artificial intelligence in health care. By opening the door to the provision of better, more thoughtful care for all, the model used for knee OA represents the potential of artificial intelligence to overcome human limitations and biases, whether systemic or implicit. While the bias found in the actual systems used in healthcare were seen by KLG, a review from 2017 published in Academic Emergency Medicine found across studies “an implicit preference for white patients, especially among white physicians,” writes Aaron E. Caroll for The New York Times. This bias connected to other trends found between the preference towards white patients and decision making, sometimes even for treatments. Moreover, this bias was even spotted in a study by Journal of the American Board of Family Medicine on knee OA in which doctors were found more likely to consider a white patient more medically cooperative than a Black one for the same case of knee OA. Ultimately, this burgeoning technology suggests a bright future for the use of artificial intelligence in healthcare, one that considers more perspectives through the use of more diverse data training sets. 

Perhaps next, algorithms will tackle how Eleanor can finally beat Eve at shuffleboard.


Comments are closed.