The picture now and the changes ahead are different questions.
After an early-arthritis diagnosis, you may wonder how quickly it will progress. Classifying the current X-ray well does not automatically mean predicting the future well. A prediction study must test whether initial information distinguishes outcomes several years later.
This study developed a model to predict radiographic progression over 4–5 years in knees with early OA. It combined information relatively easy to obtain in clinical practice and evaluated performance outside the development data set.
And we did it in one study group, and we tested it in another.
The study included knees with baseline KL grades 0, 1 or 2. A Random Forest model was developed on 5,966 OAI knees and externally validated on 3,392 knees in the independent MOST data set. These counts refer to knees, not people.
External validation tests a model in a separate group not used for development. It is needed because performance on familiar development data may not carry over to new data. It is different from solving the same examination again.
Progression was defined by an increase in radiographic KL grade; a change from 0 to 1 was considered minor and excluded. OAI was assessed at 4 years and MOST at 5 years. Their different follow-up schedules explain the phrase “4–5 years.”
I used X-ray imaging and fitness, symptom dialysis.
The final model selected 6 variables: baseline KL grade, contralateral OA, lateral narrowing grade, BMI, medial narrowing grade and total WOMAC. WOMAC asks about joint pain, stiffness and function.
Selection of these variables does not mean that changing one prevents progression. Predictive information and factors whose treatment benefits patients are different questions. The model also does not rely on a single test alone.
Original Figure 3
Enlarge FigureSHAP analysis of the six predictors shows contributions of baseline KL grade, contralateral OA, joint space narrowing, BMI and WOMAC.
Journal of Orthopaedic Research (2025). DOI: 10.1002/jor.25982 · Original Figure 3
Original Figure 5
Enlarge FigureThe paper shows example inputs and outputs from its web model. The prediction target was radiographic progression over 4–5 years.
Journal of Orthopaedic Research (2025). DOI: 10.1002/jor.25982 · Original Figure 5
External verification AUC was 0.76.
External validation in MOST reported AUC 0.76, accuracy 0.68, and F1 0.56. AUC describes discrimination between knees that progressed and those that did not. Accuracy and F1 are different performance measures, so the three numbers should not be read as interchangeable.
An AUC of 0.76 does not mean “a 76% chance that my arthritis will progress.” Interpreting an individual prediction requires their inputs, the model’s probability calculation and validation in the clinical population. This result does not demonstrate perfect prediction.
It's not pain or surgery.
The target was KL-grade progression over 4–5 years. Radiographic progression, daily symptoms and when surgery is needed may be related, but are different outcomes. This model’s performance should not be described as predicting pain or timing of surgery.
The study suggests ways to design more detailed follow-up. Whether clinical use benefits patients, and whether it performs adequately in other populations and hospitals, require further validation.
Development and validation follow-up periods were not identical. External validation should be read with both its strength—testing a separate group—and differences in group composition and observation conditions in mind.