Are Machine Learning Models Superior to Logistic Regression Models to Predict 30-d Mortality Post-Hip Fracture Surgery?

Abstract Hip fractures represent a significant global health burden, with high mortality rates. Accurate prediction of 30-d postoperative mortality is instrumental to optimize patient care. This study aimed to validate previously developed logistic regression and machine learning (ML) models for predicting 30-d mortality after hip fracture surgery measured from the date of the index operation (counting deaths occurring in-hospital or within 30 d after surgery) and compare their performance against one established calculator. We included 47 276 patients aged ≥65 yr who underwent hip fracture surgery, excluding cancer, atypical, pathologic, and stress fractures. We applied previously derived logistic regression and ML equations from the 2011-2017 NSQIP dataset, to the updated 2018-2020 dataset. We developed 2 models for each, one derived on admission and one taking into account operative course. For each of these models we also developed a full model and a parsimonious counterpart. We assessed discrimination using the Area Under the Curve (AUC) metric. We compared AUCs for models with the DeLong test. We developed an online calculator for our most highly performing model. We also benchmarked our preoperative logistic and ML calculators against a previously published preoperative model. The logistic models demonstrated acceptable to good discrimination with AUCs as follows: preoperative full of 0.756; preoperative parsimonious of 0.748; Postoperative full of 0.829; Postoperative parsimonious of 0.817. ML models maintained good discrimination performance: preoperative full, 0.772; preoperative parsimonious, 0.751; Postoperative full, 0.865; Postoperative parsimonious, 0.838. ML models were slightly superior to logistic regression models. A previously published model by Harris et al. had more variables and demonstrated no significant superiority. Both logistic regression and ML approaches offer clinically useful predictions for 30-d mortality after hip fracture surgery. Prospective validation in other populations and then integration of the parsimonious models into clinical practice may enhance perioperative decision-making and possibly patient outcomes.