Suzhou University team uses tibial plateau X-ray parameters plus machine learning to identify complete and incomplete discoid lateral meniscus, reaching validation AUCs of 0.885 and 0.861

The study brought incomplete discoid lateral meniscus (ICDLM) into the analytical framework and reported that it can be identified from the same set of radiographic parameters, with gradient boosting performing best at a validation AUC of 0.861, specificity 0.964, and accuracy 0.856. The text states that prior work focused mainly on CDLM and neglected ICDLM, and the authors describe this as the first time incomplete discoid meniscus was included in the analytical framework, with subtype-specific models provided. Single-center retrospective cohort with only 64 ICDLM cases; performance was assessed in an internal held-out validation set, with AUC 95% confidence intervals estimated by DeLong's method and other metrics from 5,000 stratified bootstrap resamples.

LASSO selected six features within the training set: gender, lateral joint space, height of the lateral tibial spine, lateral slope of the lateral tibial spine, lateral slope of the medial tibial spine, and tibial eminence width/tibial plateau width ratio, on which seven machine-learning models were built. The text notes that conventional single-factor prediction models have accuracy of roughly 50%, whereas this study replaced single-indicator judgment with a multivariable combination, reporting selection at lambda-min = 0.0092. Data were randomly split 7:3 into training and validation sets, with splitting preceding all preprocessing and feature selection to reduce information leakage; the penalty parameter was chosen by 10-fold cross-validation at the minimum cross-validated loss, and the selected feature set was applied unchanged to the validation data.

SHAP analysis identified lateral joint space as the leading contributor for both CDLM and ICDLM prediction; for both subtypes, a wider lateral joint space, a higher tibial eminence width-to-tibial plateau width ratio, a greater lateral slope of the medial tibial spine, a lower height of the lateral tibial spine, and a lower lateral slope of the lateral tibial spine were associated with higher predicted probabilities. The text reports that gender contributed more prominently to CDLM prediction but had limited importance for ICDLM, suggesting it should be interpreted together with the radiographic parameters rather than as an independent causal determinant. SHAP values were estimated with the fastshap package using 100 Monte Carlo simulations, and mean absolute SHAP values plus dependence plots were used to assess feature importance and effect direction.

Calibration and decision-curve analysis showed calibration intercept/slope of 0.671/1.848 and a Brier score of 0.147 for the CDLM SVM model, and 0.258/1.144 with a Brier score of 0.104 for the ICDLM gradient boosting model; net benefit exceeded treat-all and treat-none strategies at threshold probabilities of 0.13-0.50 for CDLM and 0.02-0.50 for ICDLM. On this basis the authors position the models as adjuncts to routine radiographic assessment, identifying patients with a high predicted likelihood of DLM and supporting timely referral for MRI or specialist evaluation. Calibration and decision-curve analyses were performed in the internal validation set; the models showed high specificity with comparatively lower sensitivity, which the authors use to frame them as adjunctive screening rather than confirmatory tools.

The study addresses populations who undergo knee radiography and for whom standardized anteroposterior films are available, aiming to identify patients with a high predicted likelihood of discoid lateral meniscus during routine radiographic assessment and to support timely referral for MRI or specialist evaluation; the authors position the models as high-specificity, comparatively lower-sensitivity adjuncts suited to preliminary risk stratification and referral decisions rather than as replacements for arthroscopy, the reference standard for diagnosis and morphological classification.

This is an incomplete reading: the contents of Figures 1-6, Supplementary Tables 1-4, and Supplementary Figures 1-6 were not loaded, so the specific values of each radiographic parameter across the three groups and the details of between-group comparisons cannot be restated here; in addition, model performance comes only from single-center internal validation, and how the models perform across populations, imaging protocols, and institutions, as well as how model-assisted assessment affects clinical workflow and patient outcomes, remain open questions for prospective research.