Authors (including presenting author) :
Wong RMY(1), Chui ECS(1), Chau WW(1), Ko KSY(1), Lai ICH(1), Mok PKL(1), Law SW(1)
Affiliation :
(1)Orthopaedics and Traumatology, Chinese University of Hong Kong
Introduction :
Accurate prediction of hospital length of stay (LOS) alleviates the ever-tightened resource budget and human resources in hospitals. Identifying key factors affecting the length of stay (LOS) of fragility fracture patients is crucial for hospital administrators to better predict hospital bed usage and plan staffing arrangements ahead of hospitalization climax. Artificial intelligence (AI) and machine learning (ML) techniques are well-received techniques in predictive medicine in the recent years. We speculate the accuracy in predicting hospital LOS for fragility fracture patients would be achieved through AI and ML.
Objectives :
The primary aim of the study is to develop a predictive machine learning model of the LOS of geriatric hip fracture patients, based on metrics, with patient demographic features and scores from rehabilitation progress reports. This allows the understanding of the impact of patient medical complexity and delays to surgery on the LOS.
Methodology :
ML analysis was performed on 7605 fragility fracture patients from an integrated rehabilitation service hospital from year 2010 to 2020. Predictive models for outcome prediction were created, and a total of 22 clinical features were generated. Classification models were trained to predict whether LOS exceeds 21 days. Models were built using artificial neural networks (ANN) or using ensemble modelling from various gradient boosting frameworks, including LightGBM, Catboost, XGBoost, etc. Shapley additive explanation (SHAP) was used based on black-box ML models to measure the predictive importance of features.
Result & Outcome :
Each ML technique demonstrated similar accuracy for predicting LOS on the 2 classes. An ensemble classifier achieved an F1 score of 0.695 and an accuracy score of 72.0%. A custom ANN model was trained to achieve an F1 score of 0.476 and an accuracy score of 71.1%. Using SHAP, the 4 most important features for the outcome prediction were found to be admission Modified Barthel Index (MBI) score, age, admission Montreal Cognitive Assessment (MoCA) score and patient source residence.
