Prediction of mortality in hip fracture patients

Background: Hip fractures are associated with elevated mortality and require extensive health care resources. As populations grow increasingly older the total amount of hip fractures has also escalated. An increasing amount of people live to be more than 100 years old and this group is expected to reach more than 50 times current estimated levels during the course of this century. As a result, the age span of geriatric hip fracture patients is now almost 50 years, and as the defining characteristic of these patients aside from their age is merely the anatomical region of the fracture, it is a very heterogenous group. Our main aim was to develop and externally validate machine learning models for prediction of mortality after hip fracture to identify which patients are the most at risk. In preparation for this we wanted to explore the value of routine blood samples for prediction of mortality after hip fracture. Furthermore, we wanted to investigate temporal trends of incidence and mortality in the emerging group of centenarian hip fracture patients, and finally train machine learning models for estimations of mortality 1, 3, 6 and 12 months after hip fracture and externally validate these models so they ultimately could be used clinically.

Methods: Studies 1 and 3 of the thesis are based on a database of consecutive hip fracture patients from Bispebjerg University Hospital in Copenhagen, Denmark, queried at different time points for different data. Study 1 consisted of 792 patients and their age, sex and admission blood samples as well as follow up data on mortality. The predictive value of preoperative blood samples was compared using receiver operating characteristics curves (ROC) and univariate and multivariate regression was performed to determine which blood samples were associated with 3- month mortality. Study 2 was a nationwide survey of all hip fractures in Denmark over a 17-year period. A total of 517 centenarian patients had suffered hip fractures and temporal trends of incidence and mortality were analyzed. Furthermore, comparisons to hip fracture patients from the same cohort with the age interval 70 to 99 (n = 124,007) were done concerning incidence, mortality and comorbidities. Study 3 consisted of 1186 patients from the same database as study 1 but from a longer time interval and with a wide range of biochemical and anamnestic data as potential predictors, as well as follow up data on mortality. Three different types of machine learning models, a Random Forest (RF), an extreme gradient boosting (XGB) and a generalized linear model (GLM) were developed for 1-, 3-, 6-, and 12- month mortality using the 10 to 13 most important features selected by the Boruta algorithm. The data was partitioned so that 70% was used for training and 30% was used as a holdout test set, results were compared using the area under the curve (AUC) for receiver operating characteristics (ROC) curves, calibration slope and intercept and decision curve analysis (DCA). Study 4 was based on 5055 hip fracture patients from Karolinska Solna and Huddinge from a 10-year period, data was collected from RIKSHÖFT and from Karolinska Data (KARDA). The models developed in study 3 were deployed and comparisons using AUC and calibration curves and DCA were performed. The best performing models were recalibrated as event rates were lower in the external validation data than in the development data.

Results: Elevated creatinine on admission blood samples was associated with an almost threefold increased 3-month mortality and had an AUC of 0.69 (0.64–0.74). The other blood samples under study had lower AUC values and on multivariate regression only age, creatinine, potassium and albumin remained as risk factors. For the centenarian hip fracture patients, incidence had declined slightly but mortality had remained stable during the study period. The centenarians under study had less registered comorbidities as measured by Charleston comorbidity index (CCI) than the comparison group of younger hip fracture patients. 68 % of centenarian patients had a CCI of 0 versus 46 % in the comparison group, but the mortality was higher. The developed machine learning models had AUC values of approximately 0.8 for all timepoints but the XGB model was the overall best performer as it was more calibrated and had better DCA. An online tool based on the XGB models has been developed for evaluation and educational purposes (https://hipfx.shinyapps.io/hipfx/). For the external validation study, mortality was lower in the Swedish data set than in the data that the models were developed on. The AUC values especially for the longer timeframes was acceptable but lower than in the development study and again, it was the XGB model that performed best overall with values of 0.72, 0.74, 0.75 and 0.77 for 1-, 3-, 6-, and 12-month mortality respectively. In line with the observed difference in event rates models were not well calibrated, so the XGB models were recalibrated using bootstrapped isotonic regression with much improved calibration.

Conclusions: Of the blood samples under study, creatinine was the best predictor of mortality after hip fracture with an AUC of 0.69 which is similar to results in studies evaluating the use of ASA classification, CCI index and Possum score, so admission blood samples could be an interesting addition to prediction models for mortality after hip fracture. Our findings on centenarian hip fracture patients suggest the compression of morbidity theory proposed in many other studies on subjects older than 100 years which could be an important finding as this age group is expected to increase drastically in the coming century. Based on our results, healthcare needs do not seems to increase proportionally with age for the group under study. Furthermore, as at least some of the known predictors for mortality does not seem to be linearly separable, prediction models for mortality in hip fracture populations might benefit from machine learning techniques to model complex interactions between factors. The XGB models developed in the third study of the thesis had the best results for prediction of 1-, 3-, 6- and 12-months mortality after hip fracture. The models performed acceptably in the external validation, especially for the later timepoints, but needed recalibration as the mortality rates were different in the development data and the external validation. The previously developed online tool will be updated with the recalibrated models so that they can be used clinically.

List of scientific papers

I. Value of routine blood tests for prediction of mortality risk in hip fracture patients. Mathias Mosfeldt, Ole B Pedersen, Troels Riis, Henning O Worm, Susanne van der Mark, Henrik L Jørgensen, Benn R Duus, Jes B Lauritzen. Acta Orthop. 2012 Feb;83(1):31-5.
https://doi.org/10.3109/17453674.2011.652883

II. Centenarian hip fracture patients: a nationwide population-based cohort study of 507 patients. Mathias Mosfeldt , Christian M Madsen, Jes B Lauritzen, Henrik L Jørgensen. Acta Orthop. 2019 Aug;90(4):342-347.
https://doi.org/10.1080/17453674.2019.1602386

III. Development and internal validation of a multivariable prediction model for mortality after hip fracture with machine learning techniques. Mathias Mosfeldt; Henrik Løvendahl Jørgensen; Jes Bruun Lauritzen; Karl-Åke Jansson. Calcif Tissue Int. 2024 Apr 16. Online ahead of print.
https://doi.org/10.1007/s00223-024-01208-1

IV. External validation of machine learning models for estimation of mortality 1, 3, 6 and 12 months after hip fracture. Mathias Mosfeldt; Henrik Løvendahl Jørgensen; Jes Bruun Lauritzen; Karl-Åke Jansson. [Manuscript]