(P-558) Development and Validation of a Prediction Model for Mortality in Patients with Hip Fracture: Population-based Cohort Study

Background: One-year mortality following hip fracture surgery is on average 30%, however, there is a lack of user friendly prediction tools to guide clinicians and patients on use of appropriate targeted preventive measures.

Objectives: To improve clinical decision-making, we aimed to develop an easy-to-use chart, displaying one-year mortality of hip fracture patients.

Methods: Using the population-based Danish Multidisciplinary Hip Fracture Registry, all patients with a first time hip fracture from 2011-2017 (N=28,791) were identified. Patient-related prognostic factors available at the time of admission were assessed as candidate predictors: Nursing home residency, comorbidity (Charlson Comorbidity Index), frailty (Hospital Frailty Risk Score), basic mobility (Cumulated Ambulation Score), atrial fibrillation, fracture type, body mass index, age, and sex. We examined the association of the individual predictors with one-year mortality with cumulative incidence, univariable logistic regression, and discrimination (area under the ROC curve [AUC]). We fitted a decision tree model (RPART algorithm using Gini index as splitting criterium) on the development cohort (70% of patients) and plotted relative variable importance of each predictor to the reduction in Gini index. Based on this, we selected predictors contributing more than 25% and build the prediction model using four different methods: 1) a decision tree model using RPART, 2) a random forest model, 3) a decision tree using the C5.0 algorithm with 50 boosting iterations, and 4) a multivariable logistic regression model. We assessed discrimination and calibration of all models on the validation cohort (remaining 30% of patients).

Results: All candidate predictors showed association to one-year mortality but discrimination was poor; age and Charlson Comorbidity Index had the best discrimination with an AUC of 0.65 and 0.61. The final model included nursing home residency, Charlson Comorbidity Index, cumulated ambulation score, body mass index and age. All four models had similar discrimination (AUC of 0.74 +/- 0.01) but the logistic regression model had superior calibration. This model predicted one-year mortality risk varying from 5% to 90% depending on the composition of predictors and was used to construct the risk chart.

Conclusions: Using information obtainable at the time of admission, one-year mortality among patients with hip fracture can be predicted with acceptable discrimination and calibration. We present an easy interpretable chart useable in daily clinical practice and provide new insight regarding the interplay between prognostic factors.