Publication
Predicting the Hospital Length of Stay of Patients with Femoral Neck Fractures
| Summary: | Femoral neck fractures are debilitating and one of the most important reasons for long-standing pain in the elderly population. They are among the most frequent causes of hospital admissions in the elderly. In most cases, they tend to have longer lengths of hospital stay despite the damage it causes in this population due to immobility and isolation. With the predictions pointing at an increase in the elderly population in the years to come, it is also expected that the number of these fractures will increase in parallel with this growth, so it is crucial to find ways to respond to this tendency. This study presents two predictive models that attempt to estimate the length of stay, in time intervals (1-11 days, 12-17 days, 18-16 days, and over 27 days), in hospitals of patients with femoral neck fractures. One model is for making predictions on the day of the admission, and the other is to make predictions three days later (with more information on the patient’s condition). Another goal of this study is to understand the factors that impact the length of stay. Although different algorithms have been tested, the final models were built using XGBoost, as this algorithm presented the best performance in model accuracy and speed of training. Model 1 (day of admission) achieved an accuracy of 0.4, while model 2 (three days after admission) achieved an accuracy of 0.44. The models showed that the factors with a higher predictive impact were the number of diagnoses, the week of admission, and patient age. Although this study’s results are not deployment-grade, they encourage how machine learning can be used in this context. Since this study used only administrative data, further studies could use sociodemographic and clinical data to build models with increased performance. |
|---|---|
| Main Authors: | Horta, Mariana Dias Suspiro Rodrigues |
| Subject: | Femoral Neck Fracture Machine Learning Predictive Model XGBoost LightGBM |
| Year: | 2023 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade Nova de Lisboa |
| Language: | English |
| Origin: | Repositório Institucional da UNL |
| Summary: | Femoral neck fractures are debilitating and one of the most important reasons for long-standing pain in the elderly population. They are among the most frequent causes of hospital admissions in the elderly. In most cases, they tend to have longer lengths of hospital stay despite the damage it causes in this population due to immobility and isolation. With the predictions pointing at an increase in the elderly population in the years to come, it is also expected that the number of these fractures will increase in parallel with this growth, so it is crucial to find ways to respond to this tendency. This study presents two predictive models that attempt to estimate the length of stay, in time intervals (1-11 days, 12-17 days, 18-16 days, and over 27 days), in hospitals of patients with femoral neck fractures. One model is for making predictions on the day of the admission, and the other is to make predictions three days later (with more information on the patient’s condition). Another goal of this study is to understand the factors that impact the length of stay. Although different algorithms have been tested, the final models were built using XGBoost, as this algorithm presented the best performance in model accuracy and speed of training. Model 1 (day of admission) achieved an accuracy of 0.4, while model 2 (three days after admission) achieved an accuracy of 0.44. The models showed that the factors with a higher predictive impact were the number of diagnoses, the week of admission, and patient age. Although this study’s results are not deployment-grade, they encourage how machine learning can be used in this context. Since this study used only administrative data, further studies could use sociodemographic and clinical data to build models with increased performance. |
|---|
Funded activities
No funded projects found.