Machine learning maps two decades of knee osteoarthritis biomechanics
A team at Anhui University of Chinese Medicine used traditional methods for mapping research and machine learning to study 3,110 English-language publications about the biomechanics of knee osteoarthritis (KOA) published between 2005 and 2025. Their findings, published in Frontiers in Medicine, show that the field has grown almost ninefold in yearly publications. They also found that research has gradually shifted away from mainly describing how people walk and toward how cells respond to physical forces, how damaged cartilage can be repaired, and how surgery can preserve the knee joint. The scale of this growth is striking.
In 2005, researchers published just 37 papers on this topic. By 2025, the yearly total had risen to 321. An exponential curve fit the yearly publication numbers well, with a coefficient of determination of 0.9425. Growth was not steady every year: the number of papers fell from 248 in 2022 to 236 in 2023. However, the overall increase was clear. Publications from 2025 alone made up more than 10 percent of all the papers studied, showing how quickly research in this field continues to grow.
To collect the papers, the researchers searched the Web of Science Core Collection and Scopus on July 26, 2026, and found 3,786 records. They removed 551 duplicate records and excluded 125 records that lacked keywords or abstracts. This left 3,110 unique publications. Two researchers independently checked the titles, abstracts, and keywords. When they disagreed, a third senior researcher made the final decision.
The team then standardized author names, institutional affiliations, and keywords. They combined singular and plural forms, expanded abbreviations, and merged words with the same meaning. They carried out their analyses in R using the bibliometrix package. The results showed a large gap between countries. The United States led in both the number of publications, with 900 papers or 28.9 percent of the total, and the number of citations, with 40,710.
China ranked second in publication output, with 355 papers, but fifth in total citations, with 4,800. Australia, the United Kingdom, and Germany had the highest rates of international collaboration. Papers involving authors from multiple countries made up 38.9 percent of Australia's output, 36.9 percent of the UK's, and 28.0 percent of Germany's. Among institutions, the University of Melbourne produced the most papers, with 140, while Boston University received the most citations, with 8,554.
The University of Melbourne also led in researcher output. Bennell, Hinman, and Wrigley, all affiliated with Melbourne, ranked highest in the number of papers, with 69, 62, and 50 publications, respectively. Researcher output followed a pattern in which many people published little and a few published a great deal. About three-quarters of the authors contributed only one paper to the collection, while a small group published repeatedly.
The fitted Lotka-type exponent was 2.38, and the Kolmogorov–Smirnov test did not reject this fitted distribution. However, the test did reject the traditional fixed exponent of 2. The researchers also identified nine core journals using Bradford's law, which helps identify journals that publish a large share of work in a field. Osteoarthritis and Cartilage was the leading journal, with 207 papers and 15,525 citations. It was followed by Gait & Posture and the Journal of Orthopaedic Research. An analysis of sudden increases in citations to particular references showed how the main focus of the field changed over the two decades.
From 2006 to 2013, frequently cited studies helped establish the field's basic understanding of knee mechanics. Zhao and colleagues used direct measurements inside the body to confirm the link between the knee adduction moment and the contact force in the inner part of the knee joint. A higher knee adduction moment reflects a greater turning force that can affect how pressure is spread across the joint. Astephen and colleagues showed that the forces and movements of the lower limbs differed at different levels of osteoarthritis severity. Researchers later focused on whether unusual loading could predict changes in joint structure. Bennell's work on dynamic loading on the inner side of the knee and cartilage loss had a citation burst strength of 27.20 from 2012 to 2016. Chehab's five-year study, which linked knee adduction and flexion moments to cartilage changes, had a burst strength of 28.88. After 2017, sudden increases in citations increasingly involved osteoarthritis that developed early or after an anterior cruciate ligament injury. One reference continued to show a citation burst through the end of the analysis period in 2025.
The study's main method was non-negative matrix factorization (NMF), a machine-learning technique that finds hidden themes in large collections of text. The researchers combined the titles, abstracts, and standardized keywords of the papers. They then used term frequency–inverse document frequency weighting, a method that gives words different weights based on how often they appear in a document and how useful they are for distinguishing it from others. They tested models with between 8 and 20 topics. A model with 19 topics provided the best balance between how clearly the topics could be understood and how consistently they could be reproduced. Its mean normalized pointwise mutual information score was 0.4225.
The researchers carefully tested whether the results could be reproduced. They ran the model 30 times on the full collection, each time using a random starting point. They also ran it 50 times on samples containing 80 percent of the papers, with the samples balanced by publication year. They used cosine similarity and the Hungarian matching method to match topic-word patterns across runs. Fourteen of the 19 topics passed both reproducibility tests. The other five were marked as exploratory topics and were interpreted with caution. What did the 19 topics show?
Research on walking patterns and lower-limb movement made up the largest share of the literature. It was followed by exercise rehabilitation, pain and physical function, and studies of how cartilage cells respond to physical forces and how the cartilage's supporting material works. However, changes over time told a more complex story. When the researchers compared the average yearly share of each topic in 2016–2020 with its share in 2021–2025, the biggest increases were in stem cells and cartilage repair, how cartilage cells respond to mechanical forces, and meniscus tears, meniscus displacement, and repair of meniscal roots.
Research on high tibial osteotomy and realignment of the lower limbs also grew, as did research on body fat, the infrapatellar fat pad beneath the kneecap, and inflammation. In contrast, the relative shares of research on walking movements, footwear supports and lateral-wedge insoles, and body weight and obesity became smaller. However, the authors stressed that the yearly topic shares always add up to 100 percent. These changes therefore show how research attention was redistributed within a growing collection of papers, not necessarily that the actual number of papers on those subjects fell. The authors also took care to distinguish recent increases from trends across the full study period.
For example, research on how cartilage cells respond to mechanical forces had gained a larger share recently, even though its overall trend from 2005 to 2025 was downward. Research on stem cells showed both a positive long-term trend and a recent increase. A sensitivity test using normalized soft topic assignments, which allowed a paper to be linked to topics with different degrees of confidence, strongly agreed with the main results based on assigning topics directly. The Spearman correlation was ρ = 0.926, and 17 of the 19 topics showed the same direction of change. A separate model using only Web of Science records also showed strong agreement for 18 of the 19 topics. However, the researchers acknowledged that differences between databases and the restriction to English-language publications may have affected the patterns they found.
The study has several limitations. It left out conference papers, which may mean that some developments in computing and engineering were not fully represented. It also excluded publications in languages other than English, which may have left out research priorities specific to certain regions. The authors cautioned that citation counts should not be treated as direct measures of research quality because newer papers have had less time to receive citations.
The authors recommend that future research focus on models that follow changes over time and are tested using independent data. These models should combine walking measurements, medical imaging, and computer-based modeling at different levels, from whole-joint movement to smaller biological processes. Most importantly, researchers should directly test whether changing the mechanical forces acting on the knee can produce lasting improvements in patients' health and in the structure of the joint.
Subject of Research: A study of published research and its main topics in knee osteoarthritis biomechanics from 2005 to 2025
Article Title: Evolution of knee osteoarthritis biomechanics research, 2005–2025: bibliometric mapping and machine-learning-assisted topic modeling
Article References: Lu, F., Liu, L., Yang, P., Gao, H., Song, H., Wang, G., & Wang, Z. (2026). Evolution of knee osteoarthritis biomechanics research, 2005–2025: bibliometric mapping and machine-learning-assisted topic modeling. Frontiers in Medicine, 13, Article 1968099. https://doi.org/10.3389/fmed.2026.1968099
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1968099
Keywords: knee osteoarthritis, biomechanics, bibliometrics, topic modeling, non-negative matrix factorization, gait analysis, mechanobiology, cartilage repair, meniscus, high tibial osteotomy, citation analysis, research trends
News Source: Teresa Odom. (October 26, 2026). Machine learning maps two decades of knee osteoarthritis biomechanics research. Scienmag.