MACHINE-LEARNING IDENTIFIES BEST MEASURES TO PREDICT ACL RECONSTRUCTION OUTCOME. Issue 3 (31st March 2019)

• Record Type: Journal Article
• Authors: Potty, Anish GR; Potty, Ajish S. R.; Punyamurthula, Rithesh; Penna, Sreeram; Benavides, Chris; Chavan, Prithviraj; Mistovich, R. Justin
• Is Part Of: Orthopaedic journal of sports medicine. Volume 7:Issue 3(2019:Mar.)Supplement
• Publication Date: 2019-03-31
• Subjects: Sports medicine -- Periodicals; Orthopedics -- Periodicals; Arthroscopy -- Periodicals; Arthroplasty -- Periodicals; Knee -- Surgery -- Periodicals; 616.7
• DOI: 10.1177/2325967119S00144
• Languages: English
• ISSNs: 2325-9671

Abstract

Introduction: Knee injury and Osteoarthritis Outcome Score (KOOS) is a widely used patient-reported outcome measurement to track recovery after ACL surgery. This study focuses on the function of daily living subscale (KOOS ADL), which is calculated based on 17 questions. By employing machine learning to predict KOOS ADL scores, we sought to better understand the relative importance of the survey questions and thereby identify its most critical components as well as questions that do not adequately predict outcomes. Methods: Pre- and post-operative patient reported KOOS ADL survey responses and outcomes scores following ACL surgery were obtained from the Surgical Outcome System data registry(SOS), an international patient-reported outcomes database sponsored and maintained by Arthrex. Patients with missing KOOS ADL survey responses were excluded from the study. Machine learning (ML) algorithms such as Random Forest and Gradient Boosting were used to identify the most critical survey questions that predict KOOS ADL scores with high accuracy. These decision tree-based algorithms predict patient outcomes using several decision rules and thereby determining the relative value of individual questions at predicting patient deficits (e.g., if patients have "Severe" difficulty in ascending stairs, they are more likely to have globally worse scores than those with difficulty with other tasks). Results: 4996 patients were initially identified. Based on compliance with the survey, 2407, 2407, 1817 and 1193 patients records for pre-surgery, 3 month, 6 month and 1 year post-surgery responses respectively underwent further analysis. The dataset consisted of 53.9% males and 46.1% females. Mean age was 29 (range 11 to 70 years). Results from the ML models indicated that by 6 key questions, over 80% of the variance in KOOS ADL scores could be explained instead of standard 17 survey questions (Table 1). Interestingly, the analysis provided similar accuracy at both 6 months and 1 year. Discussion and Conclusion: Most patients have similar functional deficits that can be captured using a simplified version of the KOOS ADL survey. The abbreviated survey would result in a better patient reporting experience while still obtaining quality data. Additional work on predicting post-surgery scores using ML from pre-surgery responses and other patient information would provide valuable insights; however, predicting outcome scores with high accuracy remains challenging. We advocate for novel methods to identify and measure meaningful data to assist with understanding patient outcomes and thereby proving the true value of orthopaedic interventions on functional status. Table 1. –Questions with high predictive value A1 - Descending stairs A2 - Ascending stairs A3 - Rising from sitting A7 - Putting on socks/stockings A9 - Taking off socks/stockings A16 - Heavy domestic duties