Developing a machine learning tool to help the most vulnerable following a hip fracture
Researchers have trained a machine learning tool to identify those at greatest risk of mortality in the year following a hip fracture using their age and features easily obtained from basic lab and blood tests.
A machine learning tool that can accurately predict the risk of mortality within 1 year of a hip fracture has been successfully developed. This follows a retrospective study of almost 4,000 first-time hip fracture records from in-hospital database systems at Beth Israel Deaconess Medical Center (Boston, USA). This study identified age as the most significant predictor of mortality, followed by nine features that can be identified in basic blood and lab test data. These results could revolutionize hip fracture treatment and aftercare by helping to identify those at the highest risk of death following hip fracture, and who would benefit from a greater degree of medical intervention.
The risk of hip fracture increases exponentially with age and can severely impact mobility and quality of life. Within a year of fracture, 20-30% of patients die, while 50% lose the ability to walk. Identifying those at greatest risk of deterioration is a key strategy to prevent loss of life and burden on healthcare services.
In the study, different machine learning categorization models were trained to identify the risk of mortality following fracture in a training set of 3,000 patient records, taking into consideration 156 features. The vast majority of these features were from basic blood and lab test results, while some were demographic features such as age or marital status. Microsoft’s LightGBM machine learning framework was the most accurate model when predicting 1-year mortality based on these features, out of the ten machine learning categorization models tested.
Following this training step, the LightGBM model was tested on a data set of 751 patient records, returning an accuracy of 81% when it assessed 1-year mortality prediction. The 156 features that contributed to this prediction were then ranked and compared. While age was decisively the most important feature, the next nine features were all biomarkers obtained from the blood and lab test results. These biomarkers included glucose and platelet count. When 1-year mortality prediction was based on just these ten most important features, the accuracy of the LightGBM model dropped by just 1%. This demonstrates that these ten features are sufficient enough alone in predicting 1-year mortality with a high degree of accuracy. Most of these features were also in the top ten features when predicting 5- and 10-year mortality with LightGBM models.
“Our models show that certain biomarkers can be particularly useful in characterizing the risk of poor outcomes following hip fractures,” explained the corresponding author George Asrian (University of Pennsylvania; PA, USA). The results of this study demonstrate that machine learning tools can be used to spot patterns relevant to health that cannot be identified by humans alone and have the potential to greatly improve mortality after hip fracture.
