Abstract
Objectives:
Patellar instability has multiple anatomic risk factors including trochlear dysplasia, patella alta and tuberosity lateralization. However, the exact contribution of each factor on patellar stability has not been clearly delineated due to the fact that 1) each abnormality exists along a spectrum that varies between individuals, 2) multiple measurements exist to describe each risk factor, and 3) the additive effects of these morphological abnormalities are not yet understood. The advent of modern machine learning techniques allows us the opportunity to extract predictive insights by detecting complex patterns underlying health data. We aimed to apply machine learning to differentiate between knee MRI measurements of patients with and without patellar instability to determine the complex relationships across variables that contribute to patellar instability.
Methods:
Utilizing an institutional database, knee MRIs of patients between the ages 18 to 40 at the time of imaging with a diagnosis of recurrent patellar instability were identified. Age and sex matched controls were selected from a database of MRIs with a diagnosis of meniscal tear for comparison. 26 standard measurements that have been used to describe trochlear dysplasia, patella alta, and tuberosity lateralization were performed on each knee, including bony and cartilaginous landmarks for each measurement when applicable. Using these measurements, 3 categories of machine learning methods were performed, including interpretable model-based methods (linear model, decision tree, random forest, and gradient boosted tree), non-interpretable model-based methods (neural network, and support vector machine [SVM]), and instance-based method (k-nearest neighbor [KNN]). A split validation method was used to divide the entire dataset into train, validation, and test subsets with an 80:10:10 split-ratio.
Results:
130 knees with patellar instability (47M, 83F) were included in this study and compared with 130 age- and sex-matched control knees. Table 1 summarizes the performance of each method on the test subset. The decision tree achieved the highest overall accuracy among the interpretable model-based methods, followed by SVM, and KNN among non-interpretable model-based and instance-based methods, respectively. Figure 1 depicts the decision tree model, demonstrating that bisect offsect measurement >72.1%, followed by length of patello-trochlear overlap <= 1.35mm, and cartilaginous trochlear depth <= 3.18mm predicted the presence of patellar instability with an accuracy of 84.62%.
Conclusions:
We present a novel approach to identify patellar instability from MRI measurements, in which an interpretable model-based machine learning method (decision tree) achieved the highest overall performance. Furthermore, patellar lateralization as reflected by bisect offset, patella alta as reflected by patellotrochlear overlap, and cartilaginous trochlear depth indicating dysplasia were found to identify knees with patellar instability. Application of this model and the utilization of uniform measurements can potentially improve the diagnostic accuracy of patellar instability from MRI images. Prospective clinical studies are recommended to further validate this model for its utility in the prediction of patellar instability.
Diagnostic Performance of machine learning methods on the test subgroup
Decision tree model