Which Osteochondritis Dissecans Subjects Will Heal Non-Operatively? An Application of Machine Learning Methods to the ROCK Cohort.

Objectives: Osteochondritis dissecans (OCD) is a focal idiopathic alteration of subchondral bone and/or its precursor with risk for instability and disruption of adjacent articular cartilage that may result in premature osteoarthritis. There are limited evidence-based guidelines to predict which lesions will heal with non-operative treatment. Therefore, this study primarily aims to design and train a classification algorithm to determine whether a patient with OCD of the knee will heal with non-operative treatment from intake visit characteristics. Methods: The Research in OsteoChondritis of the Knee (ROCK) cohort is the largest repository of OCD subjects, which has collected data on over 1400 subjects with OCD knee lesions. It is a longitudinal study at 23 participating institutions. It is registered with ClinicalTrials.gov (NCT02771496). The cohort size and fine level of detail permit optimal algorithm development and allow the present study to consider whether anatomic, image-based, or demographic factors play a role in a lesion's nonoperative healing capacity. The ROCK cohort included all patients seeking care for OCD lesions at any of the participating institutions, except those specified by the following exclusion criteria: (1) diagnosis of a focal chondral defect, (2) patients 26 years or older at the time of enrollment, (3) patient records missing data regarding OCD lesion location, (4) patient records with an incomplete or unverified screening form, and (5) patient records with an incomplete or unverified initial visit form. Subjects were further excluded from the study if data were missing for any feature required for model development. Additional inclusion criteria were that each patient meets the necessary definitions for failure or success of nonoperative management. Failure of nonoperative management was defined as the crossover from nonoperative management to surgery at any point at or beyond three-month follow-up. Successful healing was defined as complete healing on imaging with full return to sports participation. To replicate previous studies examining the healing potential of OCD lesions with logistic regression (Wall et al., JBJS, 2008; Krause et al., AJSM, 2013) a multivariate logistic regression model was developed to ascertain the effects of age, normalized lesion dimensions, and isolated pain or mechanical symptoms on the likelihood of the success of nonoperative management. A second model added sex, race, lesion plane on sagittal and coronal MRI, and lesion anatomical location (lateral femoral condyle, medial femoral condyle, or trochlea) as additional features. Next, a suite of machine learning algorithms, including a random forest classifier, a neural network, linear and radial-kernel support vector machines, a k-nearest-neighbor classifier, and a generalized boosted classifier, was developed using all available features. Each model type was hyperparameter tuned with five-fold cross-validation repeated ten times to select the best model. The raw classification accuracy and area under the receiving operating curve (AUC) were recorded. Results: The current study population includes 81 subjects. The average age in the study cohort was 12.10 years and 38.3% of subjects were female. 72.8% of lesions occurred in the medial femoral condyle, 18.5% in the lateral femoral condyle, and 6.2% in the trochlea. A full description of relevant cohort characteristics is included in Table 1. The logistic regression model developed by previous studies had a cross-validated accuracy of 65.3% and an AUC of 0.645. Only normalized lesion width was associated with an increased likelihood that a lesion would heal nonoperatively (OR = 1.16, CI = 1.04-1.31, p = 0.013). The extended regression model had a cross-validated accuracy of 71.2% and an AUC of 0.750. In this model, coronal lesion location in the lateral or medial-most zone (OR = 0.02, CI = 0.00-0.31, p = 0.012) and sagittal lesion location in the posterior zone (OR = 0.03, CI = 0.00-0.018, p < 0.001) on MRI were associated with an increased likelihood of successful nonoperative treatment. By contrast, increased normalized lesion width was associated with an increased likelihood of nonoperative failure (OR = 1.67, CI = 1.30-2.32, p<0.001). Table 2 compares the effect sizes, odds ratios, and significance for predictors included in the two logistic regression models. A hyperparameter-tuned generalized boosted classifier had the highest cross-validated accuracy and AUC of any model at 74.8% and 0.762, respectively. Regarding the importance of predictor variables: normalized lesion width was the most important variable in the generalized boosting model, followed by lesion location in the posterior sagittal zone, age, the presence of mechanical symptoms, and normalized lesion length. Other variables were not used to determine whether lesions would heal nonoperatively. Table 3 compares the performance characteristics of the three models. Conclusions: Machine learning models can predict which OCD lesions will heal with nonoperative management. Canonical models such as logistic regression can determine risk factors for nonoperative failure and serve as a clinician decision-aid with acceptable sensitivity and specificity profiles following the first encounter with a patient. Advanced machine learning algorithms, such as generalized boosted classifiers, can produce better predictive accuracy for clinical use. Normalized lesion width, lesion location in the posterior sagittal zone, patient age, the presence of mechanical symptoms, and normalized lesion length were the most important variables for successful lesion classification in the best-performing model. In the future, model performance should be tested on new, prospective patient data and compared to clinician decision-making to validate its performance. [ABSTRACT FROM AUTHOR]