Polytrauma patients : epidemiology and outcome

<h4>Background</h4><p dir="ltr">Trauma is a leading cause of death and disability worldwide, particularly among younger populations, but increasingly also among the elderly. Reliable data are crucial both for epidemiological understanding and for guiding improvements in trauma care. Trauma registers such as the Swedish Trauma Register (SweTrau) have therefore become vital resources as they can provide standardized information on injury mechanisms, severity, management, and outcomes. This thesis utilizes SweTrau data to address epidemiological questions and to evaluate the potential of machine learning methods for outcome prediction.</p><h4>Methods</h4><p dir="ltr">All four studies used data collected by SweTrau. Study 1 was a retrospective cohort study of patients with and without pelvic fractures. Baseline characteristics, injury severity, physiological status, and mortality at 30 days and one year after injury were compared between the two groups. Univariate logistic regression was used to identify crude mortality associations. Potential confounders of these associations were assessed using the change-in-estimate method, and variables identified as confounders were subsequently included together with pelvic fracture in a multivariable logistic regression model. Study 2 was a descriptive epidemiological study examining the distribution of musculoskeletal injuries among trauma patients. Studies 3 and 4 were both based on a machine learning (ML) approach. In Study 3, three different ML models were applied to predict mortality, while in Study 4, they were used to predict admission to intensive care unit (ICU) and hospital length of stay (LOS). Results were evaluated with C-statistics, calibration curves, and decision curve analysis.</p><h4>Results</h4><p dir="ltr">Across the four studies, more than 37,000 trauma patients from the Swedish national trauma register were analyzed. In study 1, pelvic fracture was found to be associated with a higher crude mortality compared to patients without pelvic fracture (30-day mortality 9% vs. 4%). However, after adjustment for confounders including age, circulatory shock, severe head injury, and overall injury severity, a pelvic fracture was not a risk factor for mortality, suggesting they reflect injury burden rather than uniquely drive mortality. Study 2 demonstrated that musculoskeletal injuries in trauma were highly prevalent, affecting 41% of all trauma patients, with fractures representing the vast majority of the musculoskeletal injuries. The spine was the most frequently injured region, followed by upper and lower extremities, respectively. Patients with musculoskeletal injuries showed higher Injury Severity Score (ISS), longer hospital stay and slightly increased mortality. Distinct patterns were observed across injury mechanisms: traffic accidents dominated, while penetrating trauma showed clear associations with extremity injuries. In study 3, three ML methods were compared with the Trauma and Injury Severity Score (TRISS) for mortality prediction in 9,208 severely injured trauma patients. All tested ML models, particularly the extreme Gradient Boosting (XGB) model, outperformed TRISS, achieving an Area Under Curve (AUC) of 0.91 (95% CI: 0.88-0.93) versus 0.85 (95% CI: 0.82-0.88) for TRISS. The most important predictors identified for mortality were age, Glasgow Coma Scale (GCS), base excess, New Injury Severity Score (NISS), severity of head and thoracic injuries, systolic blood pressure, and American Society of Anaesthesiologists (ASA) class. The ML models also demonstrated better calibration and higher clinical utility than TRISS. In study 4, the ML approach was extended to the prediction of ICU admission and LOS in 9,056 severely injured trauma patients. The XGB model achieved excellent performance for ICU admission with AUC 0.85 (95% CI: 0.84-0.87), but only moderate accuracy for LOS prediction with AUCs between 0.64 and 0.71 depending on the category. The models were implemented in an online tool for individualized estimation of ICU needs and LOS.</p><h4>Conclusion</h4><p dir="ltr">Together, the four papers demonstrated that trauma outcomes are influenced by injury patterns, physiological status, and comorbidities. They further showed how insights into these factors can be leveraged into predictive models that outperform traditional statistical methods for trauma prediction.</p><h3>List of scientific papers</h3><p dir="ltr">I. The pelvic Fracture - an indicator of injury severity or a lethal fracture? <b>Jonas Holtenius</b>, Peyman Bakhshayesh, and Anders Enocson. Injury, Volume 49, Issue 8, August 2018, Pages 1568-1571. <a href="https://doi.org/10.1016/j.injury.2018.06.016" rel="noreferrer" target="_blank">https://doi.org/10.1016/j.injury.2018.06.016</a></p><p dir="ltr">II. Musculoskeletal injuries in trauma patients: a Swedish nationwide register study including 37,266 patients. <b>Jonas Holtenius</b>, Hans E Berg, and Anders Enocson. Acta Orthopaedica, 2023; 94: 171-177. <a href="https://doi.org/10.2340/17453674.2023.11960" rel="noreferrer" target="_blank">https://doi.org/10.2340/17453674.2023.11960</a></p><p dir="ltr">III. Prediction of mortality among severely injured trauma patients: A comparison between TRISS and machine learning-based predictive models. <b>Jonas Holtenius</b>, Mathias Mosfeldt, Anders Enocson, and Hans E Berg. Injury, Volume 55, Issue 8, 111702 August 2024. <a href="https://doi.org/10.1016/j.injury.2024.111702" rel="noreferrer" target="_blank">https://doi.org/10.1016/j.injury.2024.111702</a></p><p dir="ltr">IV. Development of a new tool for prediction of hospital length of stay and intensive care needs in trauma patients using Machine Learning. Mathias Mosfeldt, <b>Jonas Holtenius</b>, Hans E Berg, Anders Enocson. [Submitted]</p>