Kyle N Kunze1, Evan M Polce2, Arpan Patel3, P Maxwell Courtney4, Scott M Sporer3, Brett R Levine3. 1. Department of Orthopaedic Surgery, Hospital for Special Surgery, 535 E. 70th Street, New York, NY, USA. Kylekunze7@gmail.com. 2. School of Medicine and Public Health, University of Wisconsin, Madison, WI, USA. 3. Department of Orthopaedic Surgery, Rush University Medical Center, Chicago, IL, USA. 4. Rothman Orthopaedic Institute at Thomas Jefferson University, Philadelphia, PA, USA.
Abstract
PURPOSE: To develop a novel machine learning algorithm capable of predicting TKA implant sizes using a large, multicenter database. METHODS: A consecutive series of primary TKA patients from two independent large academic and three community medical centers between 2012 and 2020 was identified. The primary outcomes were final tibial and femoral implant sizes obtained from an automated inventory system. Five machine learning algorithms were trained using six routinely collected preoperative features (age, sex, height, weight, and body mass index). Algorithms were validated on an independent set of patients and evaluated through accuracy, mean absolute error (MAE), and root mean-squared error (RMSE). RESULTS: A total of 11,777 patients were included. The support vector machine (SVM) algorithm had the best performance for femoral component size(MAE = 0.73, RMSE = 1.06) with accuracies of 42.2%, 88.3%, and 97.6% for predicting exact size, ± one size, and ± two sizes, respectively. The elastic-net penalized linear regression (ENPLR) algorithm had the best performance for tibial component size (MAE 0.70, RMSE = 1.03) with accuracies of 43.8%, 90.0%, and 97.7% for predicting exact size, ± one size, and ± two sizes, respectively. CONCLUSION: Machine learning algorithms demonstrated good-to-excellent accuracy for predicting within one size of the final tibial and femoral components used for TKA. Patient height and sex were the most important factors for predicting femoral and tibial component size, respectively. External validation of these algorithms is imperative prior to use in clinical settings. LEVEL OF EVIDENCE: Case-control, III.
PURPOSE: To develop a novel machine learning algorithm capable of predicting TKA implant sizes using a large, multicenter database. METHODS: A consecutive series of primary TKA patients from two independent large academic and three community medical centers between 2012 and 2020 was identified. The primary outcomes were final tibial and femoral implant sizes obtained from an automated inventory system. Five machine learning algorithms were trained using six routinely collected preoperative features (age, sex, height, weight, and body mass index). Algorithms were validated on an independent set of patients and evaluated through accuracy, mean absolute error (MAE), and root mean-squared error (RMSE). RESULTS: A total of 11,777 patients were included. The support vector machine (SVM) algorithm had the best performance for femoral component size(MAE = 0.73, RMSE = 1.06) with accuracies of 42.2%, 88.3%, and 97.6% for predicting exact size, ± one size, and ± two sizes, respectively. The elastic-net penalized linear regression (ENPLR) algorithm had the best performance for tibial component size (MAE 0.70, RMSE = 1.03) with accuracies of 43.8%, 90.0%, and 97.7% for predicting exact size, ± one size, and ± two sizes, respectively. CONCLUSION: Machine learning algorithms demonstrated good-to-excellent accuracy for predicting within one size of the final tibial and femoral components used for TKA. Patient height and sex were the most important factors for predicting femoral and tibial component size, respectively. External validation of these algorithms is imperative prior to use in clinical settings. LEVEL OF EVIDENCE: Case-control, III.
Authors: Manoshi Bhowmik-Stoker; Laura Scholl; Anton Khlopas; Assem A Sultan; Nipun Sodhi; Joseph T Moskal; Michael A Mont; Steven M Teeny Journal: Surg Technol Int Date: 2018-11-11