A model combining preoperative magnetic resonance imaging features with clinical characteristics showed moderate cross-validated discrimination in classifying meniscal tears as repairable or nonrepairable among patients who subsequently underwent arthroscopy.
Researchers retrospectively analyzed 491 consecutive patients who underwent knee magnetic resonance imaging (MRI) followed by arthroscopic evaluation at Tongji Hospital in Shanghai from January 2018 through December 2023. Eligible patients had a clinically suspected or MRI-confirmed meniscal tear and documentation that the tear was repaired or considered nonrepairable during surgery. Patients with previous surgery on the same knee, concomitant fracture, infection, tumor, or incomplete required MRI sequences were excluded.
Among the 491 patients, arthroscopy classified the evaluated tear as repairable in 261 and nonrepairable in 230.
Clinical and demographic variables were collected separately. Two musculoskeletal radiologists with 8 and 12 years of experience independently assessed standardized MRI features while blinded to the arthroscopic findings. These features included tear morphology and displacement, meniscal extrusion, bone marrow edema, joint effusion, cruciate ligament integrity, and cartilage degeneration. Reader disagreements were resolved by consensus.
Agreement between readers was substantial to almost perfect for the reported imaging measures, including tear displacement, cartilage grade, anterior cruciate ligament integrity, and meniscal extrusion.
The researchers defined repairability intraoperatively according to tear location within the peripheral vascular zone, tissue quality, tear length for amenable vertical longitudinal tears, and the ability to obtain stable reduction and fixation. Tears involving the avascular zone, complex or degenerative patterns, poor tissue quality, or technical unsuitability for fixation were classified as nonrepairable. Three fellowship-trained sports medicine surgeons performed or supervised the procedures.
Following feature selection, the researchers evaluated logistic regression, random forest, gradient boosting machine, and support vector machine models using fivefold stratified cross-validation and grid-search tuning. They also used 200 bootstrap resamples to correct performance estimates for optimism.
Logistic regression had the highest mean area under the receiver operating characteristic curve among the four models, at 0.78. Its optimism-corrected area under the curve was 0.76. At the reported classification threshold, the model had an accuracy of approximately 71%, sensitivity of approximately 73%, and specificity of approximately 68%.
Although the support vector machine produced an area under the curve of 0.77, its accuracy was only 29%, with sensitivity of 23% and specificity of 36%. The discrepancy illustrated that similar ranking discrimination did not translate into useful classification at the evaluated threshold.
Calibration analysis showed a slope of 0.95 and a Brier score of 0.21 for the logistic regression model. In decision-curve analysis, logistic regression, random forest, and gradient boosting showed positive estimated net benefit relative to treat-all and treat-none strategies across portions of the evaluated probability range. Random forest produced the highest estimated net benefit, followed closely by logistic regression, whereas the support vector machine had negative net benefit across most thresholds.
Least absolute shrinkage and selection operator regression selected 13 predictors. However, the main regression table provided complete estimates for only a subset; tear length, meniscal extrusion, and bone marrow edema were included in the adjustment model but not shown, and additional details were referred to a supplementary table.
The accepted manuscript reported adjusted associations between repairability status and anterior cruciate ligament injury, grade 4 cartilage degeneration, body mass index, sex, and tear displacement. However, the direction of these associations could not be reliably interpreted because the reported coefficients and odds ratios were inconsistent with the table heading and narrative description. The directional predictor findings therefore require clarification before they can be applied clinically.
The findings had several limitations. The study was retrospective and conducted at a single institution using one MRI scanner and acquisition protocol. The researchers noted that the cohort reflected their institution’s surgical indications and preferences and that validation in independent populations using different scanners and surgeons was needed.
Because all included patients proceeded to arthroscopy, the results may not extend to patients managed nonoperatively. In addition, the cohort had a mean age of approximately 64 years, and 52% of patients had grade 4 cartilage degeneration, suggesting that performance may differ in younger patients with acute traumatic tears. The researchers did not specifically identify those two factors as limitations.
The reference classification also depended partly on surgeon experience, available implants, and intraoperative judgment. The model used manually assessed imaging features rather than direct automated analysis of MRI scans and underwent internal validation only. It predicted intraoperative repairability rather than healing, functional recovery, reoperation, osteoarthritis progression, or long-term benefit from repair.
The study was available as an unedited accepted manuscript, and the publisher cautioned that errors could remain before final publication. The inconsistency in the regression results may therefore be corrected in the version of record.
The researchers said a structured imaging and clinical model could eventually support communication and preoperative planning, but prospective external validation is needed before clinical implementation.
The authors declared no competing interests.
