Two recent knee osteoarthritis studies each see half of the same disease process. One instrumented how people walk, using floor force plates that record the push of the foot against the ground. It found that a large sideways ground reaction impulse predicts cartilage damage two years later. Its outcome, however, was a radiologist's coarse damage score, and its cohort was mostly White and lean. The other study automatically segmented cartilage on resting MRI and compared each person's thickness against healthy people of the same age, sex, body mass index and race. Measured this way, early disease often shows up as cartilage thickening before thinning. That study had no information about how the joint is loaded. We propose to join the two in the cohort that has gait data. We would segment the cartilage automatically on its serial MRIs, convert regional thickness into demographic-matched centile scores, and ask whether mechanical load drives a swell-then-thin trajectory that a monotone damage score cannot represent.
The proposal has two parts. The first is a measurement change: replace "did the damage grade go up" with "how far did this cartilage region move against its demographic peers." The second is a population question: once load and structure are both expressed relative to peers, does a given loading pattern carry the same structural risk in heavier and lighter people, and in women and men? Answering this matters because force-plate thresholds learned in one narrow cohort are already being read as general rules. Centile norming is the cheapest available tool for testing whether they transfer.
Core Claim
Our hypothesis is that, in knees with no or early radiographic osteoarthritis, high lateral ground reaction impulse is associated with an increase in regional cartilage-thickness centile in the loaded femoral regions over the follow-up interval. Thinning would follow only later or at higher baseline damage. A worsening score that counts only damage would therefore record these knees as stable or would misattribute the change. We also hypothesize that the size of this load–structure association differs by body mass index and sex after both sides are peer-normed.
Both claims can fail cleanly. If loaded knees show only centile decline, the thickening story from the Osteoarthritis Initiative (OAI) is not mechanically driven. If the association is flat across subgroups once ground reaction force is simply divided by body weight, the centile machinery buys nothing for the population question.
Where This Sits in Prior Work
The motivating evidence comes from two papers. Costello et al. fit a Super Learner (a weighted ensemble of tree models, boosted trees and penalized regressions) on force-plate waveforms and week-long accelerometry in 947 Multicenter Osteoarthritis Study (MOST) limbs. They predicted two-year medial cartilage worsening graded with MOAKS, a semi-quantitative radiologist score of knee MRI damage. Using causal g-computation, they estimated that high lateral impulse raises worsening risk by 5.5%. They state that the cohort's demographics limit transfer.
Charting Aging Trajectories of Knee Cartilage Thickness segmented OAI cartilage with a 3D U-Net, projected it to thickness maps, and fit generalized additive mixed models on 957 osteoarthritis-free people. These models give smooth age, sex, BMI and race reference curves that turn raw thickness into a centile. The paper reports anterior and lateral femoral thickening preceding thinning. We reuse both its segmentation-to-thickness-map pipeline and its centile models as the measurement instrument.
The difference from each is direct. Costello has load but a coarse, monotone outcome and no demographic norming. The OAI study has a fine, normed outcome but no load. Neither can test whether load causes thickening. The cross-direction table in our file shows no work coupling automated cartilage segmentation, gait loading and longitudinal change.
Two lines of work constrain the design. Segmentation-first pipelines pass upstream boundary errors into everything downstream. That is a known weakness of region-of-interest masking, as noted in the file's discussion of Bo et al. and related work. A U-Net trained on OAI and applied to MOST images is exactly that risk, so segmentation quality becomes a gating experiment rather than an assumption. Second, Gare et al. showed in lung ultrasound that body-habitus tissue becomes a shortcut, with masking raising AUC from 0.6553 to 0.7742. In our setting body mass is not only a shortcut but a genuine mechanical moderator. That is why we norm rather than mask: we keep BMI in the model as a reference variable, not as a pixel region the classifier can exploit.
Tan et al. supply the general principle that grounding predictions in localized pathology reduces demographic gaps. Our regional centiles are a measurement-level analogue of their 67 pathological signs.
Research Program
The first stage builds the instrument. We take the OAI-trained cartilage segmentation and thickness mapping, apply them to MOST baseline and follow-up MRIs, and check agreement in a stratified sample. The checks are manual correction on a subset and concordance of segmented thinning with MOAKS cartilage grades. MOST and OAI are separate studies with their own acquisition protocols. We will document the differences and treat them as an institutional shift. If agreement is poor, we fine-tune on the manually corrected MOST subset and report how much downstream conclusions move with the segmentation version. We then express each femoral and tibial subregion's thickness as a centile using the OAI reference models. Change in centile over follow-up is the primary outcome, kept signed so that thickening and thinning remain separable.
The second stage norms the exposure. OAI provides no healthy gait reference in our file. We therefore fit the same kind of additive model for lateral and vertical impulse within MOST limbs that stayed radiographically normal and showed no MOAKS worsening. This gives a "load centile" conditional on age, sex and BMI. Race is included only to the extent the cohort supports it; Costello reports the cohort as overwhelmingly White, and we will not pretend otherwise.
The third stage is the analysis. We estimate the dose–response between load centile and regional thickness-centile change with mixed models across subregions. We test for a non-monotone pattern: thickening at moderate baseline damage and thinning at higher damage. We also test for interaction with BMI and sex. To stay comparable with Costello, we re-estimate the marginal effect of high lateral impulse using their g-computation approach, now with the continuous centile outcome.
The fourth stage tests transfer. The OAI centile charts are fit in one cohort and applied to another. We check whether MOST knees with no worsening sit near the 50th centile as expected. Any systematic offset is a direct measurement of cross-cohort bias in the reference, and we report it before interpreting any effect.
Baselines
The naive reference point is a model given only demographics, BMI and baseline MOAKS grade, with no gait and no segmentation. It shows how much of two-year worsening is predictable from who the patient is and how damaged the knee already is. Any later gain must be measured against it, not against chance.
The strongest existing method is the Costello Super Learner on gait and accelerometry features, predicting MOAKS medial worsening. It reports a median AUC of 0.73 (range 0.64–0.80) across 100 held-out test sets in Costello et al.. We will re-implement it on the same 947-limb definition, using 100 repeated splits that we share across all models, with the same learner library and tuning budget. We will report our reproduced number beside theirs. Adding regional centile features at baseline counts as beating it if the median AUC rises with a paired split-wise difference whose 95% interval excludes zero. For our new centile-change outcome there is no reported result on MOST, so comparisons there are internal.
The relevant number from the centile approach is from OAI, not MOST. It is an AUC of 0.81 for radiographic osteoarthritis incidence and 0.98 for KL-grade-2 progression, per Charting Aging Trajectories. No reported result was found on Lacuna for centile models on MOST.
Three ablations isolate the key ingredients. Raw segmented thickness change without centile norming tests whether norming matters. Load expressed as simple body-weight-normalized impulse, instead of the load centile, tests whether our exposure norming matters. The full model with gait removed tests whether load adds anything to structure-only trajectories.
Why not just…?
Why not just add BMI, sex and race as covariates to Costello's model? This is the simplest debiasing step, and for prediction AUC it may work almost as well. Tree ensembles can learn demographic interactions. But covariate adjustment leaves the outcome as a monotone damage grade, so it cannot reveal thickening at all. It also does not put the outcome on a scale where a one-unit change means the same thing for a 45-year-old woman and a 70-year-old man. We will run it as the main AUC comparator, and we expect it to match centile features on worsening prediction. The proposal's claim rests on the signed-trajectory and interaction analyses, not on AUC.
Why not just divide ground reaction force by body weight, as biomechanics already does? This is the real threat to the load-centile half of the proposal. If body-weight normalization alone removes the BMI and sex interaction in the dose–response, our exposure norming is unnecessary, and we will drop it in favor of the standard convention. The ablation above settles this directly. Load centiles earn their place only if residual heterogeneity survives body-weight normalization.
Why not just run the OAI centile trajectory model, which has no gait, on MOST? If structure alone predicts as well, gait is redundant for prediction. The gait-removed ablation measures exactly this. Even then, the mechanistic question remains open: whether load precedes thickening is not answered by a structure-only predictor. Costello's g-computation design is the tool for that question, and we keep it.
Decisive Checks
The project hinges on four results. The first is segmentation agreement on MOST good enough that centile change is not dominated by boundary noise; we will report agreement with manual correction and with MOAKS. The second is a positive, region-specific association between lateral load centile and increasing femoral thickness centile at low baseline damage. Its absence would falsify the mechanical-thickening hypothesis. The third is subgroup heterogeneity in the load effect that persists after body-weight normalization; its absence would collapse the population claim into existing practice. The fourth is a calibration check that non-worsening MOST knees center on the OAI 50th centile. Its failure would itself be a publishable finding about how far demographic reference charts transfer between cohorts.
References
- Gait, physical activity, and tibiofemoral cartilage damage: A longitudinal machine learning analysis in the Multicenter Osteoarthritis Study
- Charting Aging Trajectories of Knee Cartilage Thickness for Early Osteoarthritis Risk Prediction: An MRI Study from the Osteoarthritis Initiative Cohort.
- Interpretable Rheumatoid Arthritis Scoring via Anatomy-Aware Multiple Instance Learning
- The Role of Pleura and Adipose in Lung Ultrasound AI
- Fairer AI in ophthalmology via implicit fairness learning for mitigating sexism and ageism