Mark Alan Fontana1,2, Douglas E. Padgett 1, Catherine H. MacLean1
1Center for the Advancement of Value in Musculoskeletal Care, Hospital for Special Surgery, New York, NY, USA; 2Department of Healthcare Policy and Research, Weill Cornell Medical College, Cornell University, New York, NY, USA
Identifying patients at risk of not achieving minimally clinically important changes (MCICs) in PROMs after total joint arthroplasty (TJA) is important for better allocating resources toward monitoring patients and may aid in decision support. However, the ability of such predictive models to work across different PROMs, data sources, and time horizons is unknown.
We applied a machine learning (ML) algorithm, logistic LASSO, to hip and knee registry data from a high-volume facility to predict 2-year MCICs in SF-36 physical (PCSs) and mental component scores (MCSs). We derived models that incrementally incorporated information available: (1) before the decision to have surgery, (2) before surgery, (3) before discharge, and (4) after discharge. We evaluated performance with area under the receiver operating characteristic (AUROC) statistics using a hold-out sample of registry patients not used in model creation. We further tested whether these models could predict 6-month MCICs in PROMIS-10 PCSs and MCSs in a validation sample from our EMR.
12,203 registry patients had valid baseline and 2-year scores. AUROCs for predicting 2-year SF-36 PCS MCICs at the four time points were: 0.67, 0.74, 0.74, and 0.75. For MCS MCICs these were: 0.54, 0.88, 0.88, and 0.88. The EMR validation sample included 1,087 patients. Reusing the registry models, AUROCs for predicting patients’ 6-month PROMIS-10 PCS MCICs at the four time points were: 0.56, 0.63, 0.63, and 0.65. For MCS MCICs these were: 0.50, 0.78, 0.78, and 0.79.
ML algorithms applied to registry data can predict 2-year post-surgical SF-36 PCS and MCS MCICs. Applying these models to EMR data to predict 6-month PROMIS-10 MCICs retains some, but not all, of their predictive power. Across PROMs, data sources, and time horizons, information available before surgery, namely baseline PROMs, yielded the largest gain in predictive power; including available post-surgical information yielded negligible improvement.
Machine Learning, Prediction, Proms, Minimally Clinically Important Change, MCIC, Total Joint Replacement, Total Joint Arthroplasty, Hip, Knee, Robustness, Validation, AUROC