PAIN

Artificial intelligence for disparities in knee pain assessment

An algorithm developed through machine learning on existing radiological data from diverse people has the potential to improve the diagnosis and management of knee pain from osteoarthritis and reduce racial disparities in the assessment of knee pain.

Knee pain is one of the most common reasons for seeking medical attention. One of the most common causes of knee pain is osteoarthritis (OA). In a study in 1998, OA was found to be the most prevalent form of arthritis and one of the most prevalent chronic conditions in the USA1. It is estimated that nearly 70 million Americans, about one in every three, are affected by arthritis or musculoskeletal disease2. With the aging of the US population, the burden of OA is expected to increase. Knee OA is associated with major morbidity and is among the leading causes of disability in the USA3,4,5. Arthritis-related activity and work limitations and severe pain disproportionately impact Black patients relative to their effect on white patients. The prevalence of knee OA among older Black people is at least as high as that reported for white people1. However, there are substantial disparities in reporting of knee pain, with Black patients reporting more-severe pain than white patients do. In this issue of Nature Medicine, Pierson et al. report their development of an algorithm that is able to classify OA severity through the use of radiological data in a manner that corresponds to pain experienced by Black patients6.

In the management of kneepain, particularly that from OA, the first line of treatment is physical therapy, weight loss, and pharmacological therapies with pain and anti-inflammatory medications. But when all else fails, surgical intervention, such as knee replacement, is recommended. The indications for surgery include not only pain but also function (pain that limits everyday activities) and quality of life. So it is not surprising that elective knee replacement is one of the fastest-growing elective surgical procedures in the USA and is appropriately a major cost center for the Centers for Medicare and Medicaid, which pays for majority of knee replacements in the country.

However, there are marked disparities in the receipt of knee-replacement surgery. Numerous studies have documented persistent racial differences in the utilization of this surgery3,4,5. Black patients are significantly less likely than are white patients to receive knee-replacement surgery. Disparities in knee-replacement utilization represent one of many racial disparities that exist across various healthcare conditions and settings7,8. The reasons for these disparities are complex and involve patient-level, provider-level and system-level factors. How patients with knee OA are assessed by radiological measures might also be a factor in disparities in the management of knee pain.

Here, a key premise of the Pierson et al. study is that racial variations in the assessment of knee pain in patients with knee OA have both internal sources (joint pathology) and external sources (social stress and isolation, for example)6. One common measure of internal causes of knee pain in patients with knee OA is the Kellgren–Lawrence Grade (KLG), a radiological grading of disease severity developed and validated in a predominantly white population. But even adjusting for objective measures of OA severity, such as the KLG, in expression of pain does not entirely explain differences between Black patients and white patients in their pain level. This is in part because there is little correlation between a patient’s level of pain expression and radiological evidence of disease9.

The authors used deep learning, an advanced form of artificial intelligence, on existing radiological data to develop an algorithmic measure of OA severity linked to pain. Their sample consisted of the 4,172 people (of whom 97% are white) who participated in the Osteoarthritis Initiative, a multi-center, longitudinal, US National Institutes of Health–funded study of participants 47–79 years of age who had or were at risk of developing knee OA. The authors note that, according to the Knee Injury and Osteoarthritis Outcome Score, the median pain for a Black patient was worse than that of 75% of non-Black patients in the sample. Furthermore, Black patients had more-severe OA by KLG. However, adjusting for the KLG (i.e., comparing those with similar KLG scores) did not reduce or eliminate differences between Black people and non-Black people in their pain level. After the authors applied the algorithmic measure they developed, they found marked improvement, over the KLG, in the predictive performance of the measure for grading OA severity, and this measure reduced the unexplained racial disparity in pain. To a lesser extent, the algorithmic measure also performed better than KLG at grading OA severity in other under-served populations, such as those with low income or low education6. The authors suggest that the diversity of the people in the training dataset resulted in improved predictive power of the algorithmic measure. About 20% of the people in the training sample were Black or had low income or low education.

There are important limitations to consider in interpreting the findings of this innovative study. The authors point out, for example, that they did not evaluate the role of pain in surgical decision-making. After all, pain is not the only indication for surgical intervention. Therefore, applying the algorithm may not affect the number of surgeries performed in general. The study also did not assess if the algorithmic measure was able to predict patient outcomes after surgery.

In their motivation for the study, the authors suggest that racial differences in pain level might explain racial disparities in the utilization of knee-replacement surgery. However, this interpretation needs to be considered with a caution. A key factor for receiving a surgical recommendation for replacement is patient willingness. Black patients have lower preference for joint replacement than do white patients10. This is in part because Black patients have concerns about surgical outcomes11. There are also data showing that Black patients are more likely to receive joint-replacement surgery in low-quality hospitals in the USA12. Therefore, whether better assessment of pain will address disparities in the use of elective joint-replacement surgery remains to be determined.

Those limitations notwithstanding, Pierson et al. bring to the fore a highly innovative approach6 to one of the most important and difficult measurements in clinical medicine. Using modern technologies such as machine learning promises to improve the diagnosis and management of not only pain from OA but also other conditions and symptoms in the practice of medicine. As the authors suggest, this particular algorithmic measure of knee pain might be used as part of a tool kit for decision-making. It could supplement the clinicians’ assessment of pain and in due course may even serve as a component of decision aids for surgical decision-making.

References

  1. 1.

    Lawrence, R. C. et al. Arthritis Rheum. 41, 778–799 (1998).

    CAS  Article  Google Scholar 

  2. 2.

    Altman, R. D., Hochberg, M. C., Moskowitz, R. W. & Schnitzer, T. J. Arthritis Rheum. 43, 1905–1915 (2000).

    Article  Google Scholar 

  3. 3.

    Sharkness, C., Hamburger, S., Moore, R. & Kaczmarek, R. Public Health Rep. 106, 70–74 (1993).

    Google Scholar 

  4. 4.

    Wilson, M. G., May, D. S. & Kelly, J. J. Ethn. Dis. 4, 57–67 (1994).

    CAS  PubMed  Google Scholar 

  5. 5.

    Hoagland, F. T., Oishi, C. S. & Gialamas, G. G. Ann. Rheum. Dis. 54, 107–110 (1995).

    Article  Google Scholar 

  6. 6.

    Pierson, E., Cutler, D., Leskovec, J., Mullainathan, S. & Obermeyer, Z. Nat. Med. https://doi.org/10.1038/s41591-020-01192-7 (2020).

  7. 7.

    Whittle, J., Conigliaro, J., Good, C. B. & Lofgren, R. P. N. Engl. J. Med. 329, 621–627 (1993).

    CAS  Article  Google Scholar 

  8. 8.

    Hannan, E., Kilburn, H. J., O’Donnell, J. F., Lukacik, G. & Shields, E. P. Med. Care 29, 430–441 (1991).

    CAS  Article  Google Scholar 

  9. 9.

    Bedson, J. & Croft, P. R. BMC Musculoskelet. Disord. 9, 116 (2008).

    Article  Google Scholar 

  10. 10.

    Ibrahim, S. A., Siminoff, L. A., Burant, C. J. & Kwoh, C. K. Med. Care 40, I44–I51 (2002).

    Article  Google Scholar 

  11. 11.

    Ibrahim, S. A., Siminoff, L. A., Burant, C. J. & Kwoh, C. K. Arthritis Rheum. 46, 2429–2435 (2002).

    Article  Google Scholar 

  12. 12.

    Cai, X., Cram, P. & Vaughan-Sarrazin, M. Clin. Orthop. Relat. Res. 470, 1185–1193 (2012).

    Article  Google Scholar 

Download references

Acknowledgements

S.A.I. is supported in part by a K24 Mid-Career Development Award from the National Institute of Arthritis and Musculoskeletal and Skin Diseases (grant K24AR05529).

Author information

Affiliations

Authors

Corresponding author

Correspondence to Said A. Ibrahim.

Ethics declarations

Competing interests

The author declares no competing interests.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Ibrahim, S.A. Artificial intelligence for disparities in knee pain assessment. Nat Med 27, 22–23 (2021). https://doi.org/10.1038/s41591-020-01196-3

Download citation

Search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing