Summary

This study represents a critical step toward understanding which variables and pathoanatomic risk factors contribute to patient-reported outcomes after MPFL-R using machine learning to assess multiple variables within a large post-operative PI cohort.

Abstract

Study Aim:
The primary aim of this study is to use supervised machine learning regression models to identify significant variables that influence two-year patient-reported outcomes in recurrent lateral patellofemoral instability (LPI) following medial patellofemoral ligament reconstruction (MPFL-R). A secondary aim is to determine the frequency of these variables in this patient population.

Methods

Prospectively collected data from 921 primary MPFL reconstruction procedures performed by a single surgeon was prepared for this study. 741 patients (80.5%) had completed Banff Patellofemoral Instability Instrument (BPII 2.0) scores at two years post-operative. Ten patient demographic variables and eighteen pathoanatomic variables, as well as surgical techniques and LPI recurrence were examined in relation to BPII scores. Pathoanatomic variables included trochlear dysplasia, trochlear bump height, TT-TG and TT-PCL distances, patellar tilt, Caton Deschamps ratio, Wiberg classification, and patella cartilage lesions and clinical measures of femoral and tibial rotation and the Beighton score.

The research methodology was designed to not only assess relationships between independent variables and the dependent outcome variable of disease-specific quality of life (BPII score) but also to enhance the model's accuracy and robustness. Penalized regression techniques were employed for simultaneous model selection and estimation, and to manage multicollinearity in the predictors.

Results

The patient population was 232 males (25.2%) and 689 females (74.8%), with a mean BMI of 24.4. The mean age at first dislocation was 15.0, and at surgery was 24.2. Bilateral LPI was evident in 452 (49.1%), and 464 (50.4%) had a positive Beighton score. Trochlear dysplasia was high grade in 350 (36.9%) and low grade in 463 (50.3%). There were 33 redislocations (3.6%) at 2-years post-operative.

Multiple regression models were built to assess outcomes, including 2-year post-operative BPII score, change in BPII 2.0 score from pre- to post-operative, and risk of redislocation. Initial analysis demonstrated combinations of variables, including high-grade trochlear dysplasia, elevated TT-TG >21mm, CD ratio >1.43, age >30, presence of patella chondral lesions >grade 2, and pre-operative BPII 2.0 score <17.1 were associated with lower 2-year BPII 2.0 scores.

Ongoing Analysis:
Interaction terms will be introduced in the subsequent phases of the learning model to explore the interdependencies among predictors, providing a deeper understanding of how these variables influence each other. Additionally, a Random Forest model will be implemented to capture potential nonlinear relationships and complex interactions. This will facilitate comprehensive modelling of complex interactions and nonlinear relationships. The use of variable importance measures in the Random Forest model will significantly enhance the interpretability of the results, clearly identifying which predictors are most influential on outcomes.

Conclusion

This study represents a critical step toward understanding which variables and pathoanatomic risk factors contribute to patient-reported outcomes after MPFL-R using machine learning to assess multiple variables within a large post-operative PI cohort. This knowledge will provide essential information about the prevalence and combinations of risk factors in PI.