A new study published in the Journal of Orthopedic Research indicates that an AI-based model trained with basic blood test and laboratory data, as well as basic demographic data, can predict a patient’s risk of death in 1, 5 and 10 years of suffering a hip fracture.
In analysis of 3,751 hip fracture patient records from two hospital database systems at Beth Israel Deaconess Medical Center in Boston, the one-year mortality rate for all patients was 21% and for those aged 80 years or more was 29%. After evaluating 10 different machine learning classification models, the researchers found that the LightGBM model had the most accurate 1-year mortality prediction performance.
Age, blood sugar levels, certain red blood cell characteristics, white blood cell levels, urea nitrogen levels, platelet count, calcium levels, and blood clotting time were the most powerful factors. predictive. Most of these were also among the top 10 features of LightGBM’s 5- and 10-year mortality prediction models.
"Our models show that certain biomarkers may be particularly useful in characterizing the risk of poor outcomes after hip fractures," said corresponding author George Asrian of the University of Pennsylvania.
Hip fracture is a common condition with a high degree of associated morbidity and mortality. More than 300,000 hip fractures occur in the U.S. each year, accounting for more than 40% of fracture-related nursing home admissions and 70% ($12 billion) of direct costs in fracture care. Within 1 year of injury, 20%. 30% of patients die and 50% lose the ability to walk.
The likelihood of fracture, which most commonly occurs after a fall, increases exponentially with age, making it an especially important issue to address in an aging population. Risk factors include age, osteoporosis, increased body mass, decreased visual perception, physical inactivity, muscle weakness, chronic diseases that increase the risk of imbalance, such as arthritis and Parkinson’s disease, and cognitive impairment.
Summary
The purpose of this retrospective study was to evaluate whether mortality after hip fracture can be predicted by a machine learning model trained with basic blood test and laboratory data, as well as basic demographic data. Furthermore, the purpose was to identify the key variables most associated with 1-, 5-, and 10-year mortality and investigate their clinical significance.
Input data included 3,751 hip fracture patient records obtained from the Medical Information Mart for Intensive Care IV database , which provided records from hospital database systems at Beth Israel Deaconess Medical Center. The 1-year mortality rate for all patients studied was 21% and for those over 80 years of age it was 29%.
We evaluated 10 different machine learning classification models and found that LightGBM has the strongest 1-year mortality prediction performance, with an accuracy of 81%, an AUC of 0.79, a sensitivity of 0.34, and a specificity of 0.98 on the test set.
Top-rated features of the 1-year model included:
- Age
- Blood glucose
- Distribution of red blood cells
- Mean corpuscular hemoglobin concentration
- white blood cells
- Ureic nitrogen
- prothrombin time
- Platelet count
- Calcium levels
- partial thromboplastin time
Most of these were also in the top 10 features of the trained 5- and 10-year LightGBM prediction models. Testing these high-ranking biomarkers in new hip fracture patients may help clinicians assess the likelihood of poor outcomes for hip fracture patients, and further research may use these biomarkers to develop a risk score for hip fracture. mortality.
Figure : Bar chart showing feature importance values for the top 10 features using the LightGBM 1-year mortality prediction model trained on the 156 features. Below the graph, percentage differences of the mean values of each variable for the cohort that died within 1 year compared to the cohort that survived 1 year, calculated as ([1]/[0] − 1 × 100%) . MCHC, mean corpuscular hemoglobin concentration; PT: prothrombin time; PTT: partial thromboplastin time; RDW: red blood cell distribution width.
Discussion
Overall, we have shown that it is possible to develop a highly accurate machine learning model that can estimate 1-year mortality in hip fracture patients in this data set. It is possible to isolate most biomarkers and demographic attributes and retrain a prediction model on only the top 10 features to produce a model of comparable accuracy (within 1%).
Many models were tested to identify the most appropriate classification algorithm for the prediction of hip fracture mortality. The LightGBM prediction model performed better with balanced accuracy and stronger AUC values. LightGBM is a boosting algorithm that involves training multiple models in sequence, with each one showing improvement over its predecessor.
As many previous literature sources have indicated, age is an important predictor of mortality in individuals due to impairment of repair mechanisms, immune response (critical for defending against bacterial infection), and mobility.
The second most important variable, blood glucose , was found to be important as well. Although there was no mean difference between patients with 1-year mortality and those who survived >1 year, blood glucose may provide an additional measure of disease burden in the elderly. Those with too much glucose may suffer from diabetes sequelae including poor healing and mobility. Those with too little glucose may experience additional falls due to being in a hypoglycemic state and may be considered a fall risk (further limiting mobility). Markers of hematological health were also useful.
Other notable markers that were included in the top 10 variables were urea nitrogen (a marker of overall kidney function that can be used to determine whether individuals are volume depleted, possibly due to sepsis), PT/PTT, and cell count. platelets, which can determine whether individuals maintain their coagulation capacity and serum calcium, which has been identified as a predictor of mortality in patients with osteoporosis.
Conclusions Hip fractures are a serious event that can lead to death within the first year after the event in approximately 21% of patients of all ages and up to 29% of patients over 80 years of age. While age may be the most important variable predicting poor outcomes in hip fracture patients, the performance of the models trained in this project shows that biomarkers also play an important role in determining risk. MCHC, mean corpuscular hemoglobin concentration; PT: prothrombin time; PTT: partial thromboplastin time; RDW: red blood cell distribution width in patients with hip fractures can help evaluate the mortality risk of patients. LightGBM is a robust and powerful tool for predicting mortality over short and long time periods, allowing simple analysis of the most important input variables. With additional models using larger, well-balanced data sets, it may be possible to develop a formal risk score for hip fracture patients that can be used by clinicians, using LightGBM and the variables described above. |
