Artificial neural networks outperform linear regression in estimating 9-month patient-reported outcomes after upper extremity fractures with increasing number of variables

Abstract

Abstract Objective: To compare performance between linear regression (LR) and artificial neural network (ANN) models in estimating 9-month patient-reported outcomes (PROs) after upper extremity fractures using various subsets of early mental, social, and physical health variables. Methods: We studied 734 patients with isolated shoulder, elbow, or wrist fracture who completed demographics, mental and social health measures, and PROs at baseline, 2–4 weeks, and 6–9 months postinjury. PROs included 3 measures of capability (QuickDASH, PROMIS-UE-PF, PROMIS-PI) and one of pain intensity. We developed ANN and LR models with various selections of variables (20, 23, 29, 34, and 54) to estimate 9-month PROs using a training subset (70%) and internally validated them using another subset (15%). We assessed the accuracy of the estimated value being within one MCID of the actual 9-month PRO value in a test subset (15%). Results: ANNs outperformed LR in estimating 9-month outcomes in all models except the 20-variable model for capability measures and 20-variable and 23-variable models for pain intensity. The accuracy of ANN versus LR in the primary model (29-variable) was 83% versus 73% (Quick-DASH), 68% versus 65% (PROMIS-UE-PF), 66% versus 62% (PROMIS-PI), and 78% versus 65% (pain intensity). Mental and social health factors contributed most to the estimations. Conclusion: ANNs outperform LR in estimating 9-month PROs, particularly with a larger number of variables. Given the otherwise relatively comparable performance, aspects such as practicality of collecting greater sets of variables, nonparametric distribution, and presence of nonlinear correlations should be considered when deciding between these statistical methods.