Development of a machine learning model for identifying the optimal situation favoring double-level osteotomy over single-level high tibial osteotomy.

Background: This study aimed to develop a machine learning (ML) model to identify the optimal situation wherein double-level osteotomy (DLO) is favored for severe varus knees by analyzing unfavorable outcomes. This study hypothesized that there are the most favorable algorithms and contributing factors for identifying the optimal situation favoring DLO over opening-wedge high tibial osteotomy (OWHTO).

Methods: Data were retrospectively collected from patients who underwent OWHTO (505 knees). Unfavorable outcome parameters were defined as follows: (1) medial proximal tibial angle (MPTA) > 95°, (2) joint line convergence angle (JLCA) > 4° (insufficient medial release), (3) JLCA < 0° (medial instability), (4) recurrence of varus deformity, and (5) lateral hinge fracture. The input data for the ML model included demographic data and preoperative radiological and intra-operative factors. The ML model was used to evaluate overall and to evaluate each unfavorable outcome. Interpretation by the model was performed by SHapley Additive exPlanations.

Results: The unfavorable group had a larger JLCA and MPTA preoperatively than the favorable group in the conventional comparison. The light gradient boosting machine (LGBM) demonstrated the highest AUC of 0.66 and F-1 score of 0.72 among the ML algorithms. In the overall assessment, the preoperative weight-bearing line ratio (WBLR) was the factor that contributed the most, followed by the preoperative JLCA and the ΔWBLR. ΔWBLR and the preoperative JLCA were the contributing factors for each outcome.

Conclusions: The LGBM model was superior in predicting the optimal situations favoring DLO over OWHTO. Preoperative WBLR, preoperative JLCA, and ΔWBLR significantly contributed to the unfavorable outcomes overall and for each outcome in the ML model.