Rheumatology Network interviewed Nicholas Brisson, PhD, to discuss his recent study, “Association of machine learning based predictions of medial knee contact force with cartilage loss over 2.5 years in knee osteoarthritis." The study discovered that machine learning was able to develop an equation for predicting medial tibiofemoral contact force (MCF) peak, which supports the theory that mechanical loading directly contributes to disease progression.
In this Q&A, Rheumatology Network interviewed Nicholas Brisson, PhD, to discuss his recent study, “Association of machine learning based predictions of medial knee contact force with cartilage loss over 2.5 years in knee osteoarthritis,” originally published in Arthritis and Rheumatology. The study discovered that machine learning was able to develop an equation for predicting medial tibiofemoral contact force (MCF) peak, which correlated to medial tibial cartilage volume loss for patients with knee osteoarthritis. This supports the theory that mechanical loading directly contributes to disease progression and underscores the notion that knee load predictions can be obtained without much more resource-intensive strategies. Ultimately, reducing MCF magnitude may help curb the structural disease progression that is associated with mechanical loading in this specific patient population.
Rheumatology Network: Can you provide a bit of background on the study?
Nicholas Bisson, PhD: It is commonly believed that forces inside the knee (particularly contact forces on the medial aspect of the knee) are related to degeneration of joint tissues, such as cartilage; however, no research has explicitly tested this hypothesis. Thus far, studies have examined the relationship between surrogate measures of forces inside the knee and cartilage degeneration. A goal of the research community is to confirm that contact forces inside the knee are in fact related to joint degeneration. The major barrier to accomplishing this goal is that forces inside of the knee cannot be measured in people with native knees; that is, knees that have not been replaced with an implant.
To accomplish this objective, we first generated a machine learning based equation to predict the forces inside the knee. This equation was made possible by unique data from people with force-measuring knee replacements available at the Julius Wolff Institute, Charité – Universitätsmedizin Berlin in Germany. We then used this equation to predict knee forces in participants with knee osteoarthritis who completed a 3-year longitudinal study at McMaster University, Canada.
Ultimately, we showed that higher predicted medial knee contact forces led to greater cartilage degeneration over 2.5 years in these people with knee osteoarthritis.
RN: What sparked your interest in using machine learning to predict the medial tibiofemoral contact force?
NB: Machine learning and statistical analysis often use the same or similar tools, but with a different goal. Machine learning is used for making the best predictions (in other words, predicting an outcome from input variables), whereas statistics is used to understand or infer relationships between input variables and an outcome.
In this study, we used machine learning because it’s the tool of preference to be able to predict an outcome – in this case, the force inside the knee (ie, medial tibiofemoral contact force). In fact, machine learning was the key to allowing us to address a major gap in the field: predicting knee contact forces in people with native knees. Subsequently, we used traditional statistics to infer that the predicted knee contact force was in fact related to cartilage degeneration in persons with knee osteoarthritis.
RN: Did the results of the study surprise you?
NB: In hindsight the results make a lot of sense. Three parameters were important in predicting the medial knee contact force: the vertical knee reaction force, walking speed, and the knee adduction moment. The vertical knee reaction force makes sense because it measures axial (up-and-down) forces applied onto the knee. Gait speed makes sense because it is related to muscle forces. That is, the faster you walk, the harder your muscles will contract and the higher the force will be across your knee. Finally, the knee adduction moment is a surrogate for the amount of force applied to the medial aspect (inside) of the knee relative to the lateral aspect (outside) of the knee. The goal of our work was to predict the forces on the medial aspect (inside) of the knee, so it makes sense that this measurement was important.
For predicting cartilage degeneration, this finding wasn’t surprising but was really exciting! For the first time, we have some data showing that contact forces on the medial aspect of the knee are related to cartilage loss over time in persons with osteoarthritis. This confirms a long-held belief by the research community. We hope that others will use our equation (and also develop new ones) to further confirm our findings and provide new insights on this topic.
RN: What is the greater impact of these findings?
NB: We think there are 2 impactful features of this work. First, the equation to predict medial contact forces is relatively easy to use, enabling its use by others to easily predict these forces. We hope that other researchers will use this equation to help move the field forward to better understand how knee contact forces are related to degeneration of joint tissues, for instance, cartilage loss.
Second, we can focus on the parameters in this equation to try to reduce knee joint loading. Ideally, biomechanical interventions can target these parameters to reduce the medial knee contact force and curb the deleterious effects of knee biomechanics on cartilage loss in persons with knee osteoarthritis. We believe that successful interventions to reduce the knee contact force will likely be achieved by reducing the knee adduction moment and/or the vertical knee reaction force.
RN: Were there any strengths or limitations you’d like to discuss?
NB: A major strength of this work was the integration of 2 fairly rare datasets: (1) longitudinal data documenting cartilage loss in persons with knee osteoarthritis, and (2) direct measurement of forces inside the knee. By integrating these data, and using machine learning, we were able to discover a new strategy to facilitate representations of medial knee contact forces from more accessible technologies to continue advancing the field. Also, this work was the first to provide direct evidence that the medial knee contact force is positively related to loss of medial tibial cartilage volume in persons with knee osteoarthritis, supporting the idea that mechanical loading is a key contributor to structural disease progression.
[A limitation of the study was that] the equation to predict the medial knee contact force was derived using gait data from only nine patients with total knee replacements; larger sample sizes would be ideal to improve the accuracy of our prediction equation. Also, the knee osteoarthritis sample comprised mostly older, overweight women with osteoarthritis diagnosed with x-ray. Thus, the extent to which the association between the predicted medial knee contact force and cartilage loss can be extended to other populations is uncertain.
RN: Does your team plan on doing any further research on this topic?
NB: This is the first attempt to use machine learning to predict knee contact forces. We believe that accurate and reliable measures of knee contact forces are crucial for understanding osteoarthritis pathogenesis. We plan to continue leveraging technological advancements and improved understanding of the musculoskeletal system to improve our predictions of joint contact forces. Our experience with machine learning, musculoskeletal modelling and medical imaging puts us in a unique position to continue pushing this frontier.
RN: Is there anything else you’d like our audience to know?
NB: This work would not have been possible without international collaboration. We were fortunate to have contributions from the Julius Wolff Institute, Charité – Universitätsmedizin Berlin in Germany, and from McMaster University and the University of Waterloo in Canada. Also integral to the completion of this work were the study participants, data, resources, and expertise from each institution.