A New Random Forest Algorithm-Based Prediction Model of Post-operative Mortality in Geriatric Patients With Hip Fractures

Abstract

Background: Post-operative mortality risk assessment for geriatric patients with hip fractures (HF) is a challenge for clinicians. Early identification of geriatric HF patients with a high risk of post-operative death is helpful for early intervention and improving clinical prognosis. However, a single significant risk factor of post-operative death cannot accurately predict the prognosis of geriatric HF patients. Therefore, our study aims to utilize a machine learning approach, random forest algorithm, to fabricate a prediction model for post-operative death of geriatric HF patients.

Methods: This retrospective study enrolled consecutive geriatric HF patients who underwent treatment for surgery. The study cohort was divided into training and testing datasets at a 70:30 ratio. The random forest algorithm selected or excluded variables according to the feature importance. Least absolute shrinkage and selection operator (Lasso) was utilized to compare feature selection results of random forest. The confirmed variables were used to create a simplified model instead of a full model with all variables. The prediction model was then verified in the training dataset and testing dataset. Additionally, a prediction model constructed by logistic regression was used as a control to evaluate the efficiency of the new prediction model.

Results: Feature selection by random forest algorithm and Lasso regression demonstrated that seven variables, including age, time from injury to surgery, chronic obstructive pulmonary disease (COPD), albumin, hemoglobin, history of malignancy, and perioperative blood transfusion, could be used to predict the 1-year post-operative mortality. The area under the curve (AUC) of the random forest algorithm-based prediction model in training and testing datasets were 1.000, and 0.813, respectively. While the prediction tool constructed by logistic regression in training and testing datasets were 0.895, and 0.797, respectively.

Conclusions: Compared with logistic regression, the random forest algorithm-based prediction model exhibits better predictive ability for geriatric HF patients with a high risk of death within post-operative 1 year.

Keywords: hip fracture; machine learning; mortality; prediction model; random forest.

Figure 1

The procedure of establishing mortality prediction models in this study.

Figure 2

The continuous variables distribution of live group and dead group.

Figure 3

The dichotomous variables of live group and dead group.

Figure 4

The correlation analysis results of all variables.

Figure 5

(A) The boxplot reveals the importance of each of the individual variables in random forest algorithm. Boxplots in green, yellow, and blue were confirmed as important, tentative, and unimportant variables, respectively. (B) Decisions of rejecting or accepting features by random forest in 100 Boruta function runs. (C) Lasso coefficient profiles of all features. (D) The tuning parameter λ (lambda) selection in the Lasso regression model used 10-fold cross-validation by minimum criteria.

Figure 6

The ROC curves of continuous variables, prediction model constructed by random forest algorithm, and traditional logistic regression in training dataset.

Figure 7

The ROC curves of continuous variables, prediction model constructed by random forest algorithm, and traditional logistic regression in testing dataset.