BackgroundFemoral shaft fractures are common in major trauma. Early definitive fixation, within 24 hours, is feasible in most patients and is associated with improved outcomes. Nonetheless, variability might exist between trauma centers in timeliness of fixation. Such variability could impact outcomes and would therefore represent a target for quality improvement. We evaluated variability in delayed fixation (≥24 hours) between trauma centers participating in the American College of Surgeons (ACS) Trauma Quality Improvement Program (TQIP) and measured the resultant association with important clinical outcomes at the hospital level.Methods and findingsA retrospective cohort study was performed using data derived from the ACS TQIP database. Adults with severe injury who underwent definitive fixation of a femoral shaft fracture at a level I or II trauma center participating in ACS TQIP (2012–2015) were included. Patient baseline and injury characteristics that might affect timing of fixation were considered. A hierarchical logistic regression model was used to identify predictors of delayed fixation. Hospital variability in delayed fixation was measured using 2 approaches. First, the random effects output of the hierarchical model was used to identify outlier hospitals where the odds of delayed fixation were significantly higher or lower than average. Second, the median odds ratio (MOR) was calculated to quantify heterogeneity in delayed fixation between hospitals. Finally, complications (pulmonary embolism, deep vein thrombosis, acute respiratory distress syndrome, pneumonia, decubitus ulcer, and death) and hospital length of stay were compared across quartiles of risk-adjusted delayed fixation.We identified 17,993 patients who underwent definitive fixation at 216 trauma centers. The median injury severity score (ISS) was 13 (interquartile range [IQR] 9–22). Median time to fixation was 15 hours (IQR 7–24 hours) and delayed fixation was performed in 26% of patients. After adjusting for patient characteristics, 57 hospitals (26%) were identified as outliers, reflecting significant practice variation unexplained by patient case mix. The MOR was 1.84, reflecting heterogeneity in delayed fixation across centers. Compared to hospitals in the lowest quartile of delayed fixation, patients treated at hospitals in the highest quartile of delayed fixation suffered 2-fold higher rates of pulmonary embolism (2.6% versus 1.3%; rate ratio [RR] 2.0; 95% CI 1.2–3.2; P = 0.005) and required greater length of stay (7 versus 6 days; RR 1.15; 95% CI 1.1–1.19; P < 0.001). There was no significant difference with respect to mortality (1.3% versus 0.8%; RR 1.6; 95% CI 1.0–2.8; P = 0.066). The main limitations of this study include the inability to classify fractures by severity, challenges related to the heterogeneity of the study population, and the potential for residual confounding due to unmeasured factors.ConclusionsIn this large cohort study of 216 trauma centers, significant practice variability was observed in delayed fixation of femoral shaft fractures, which could not be explained by differences in patient case mix. Patients treated at centers where delayed fixation was most common were at significantly greater risk of pulmonary embolism and required longer hospital stay. Trauma centers should strive to minimize delays in fixation, and quality improvement initiatives should emphasize this recommendation in best practice guidelines.

Trauma center characteristics by hospital quartile of delayed fixation.

Patient characteristics by hospital quartile of delayed fixation.

Outcomes by hospital quartile for delayed fixation.

Patient characteristics associated with delayed fixation.

Mixed multilevel model for delayed fixation.

IntroductionAccurately predicting patient outcomes is crucial for improving healthcare delivery, but large-scale risk prediction models are often developed and tested on specific datasets where clinical parameters and outcomes may not fully reflect local clinical settings. Where this is the case, whether to opt for de-novo training of prediction models on local datasets, direct porting of externally trained models, or a transfer learning approach is not well studied, and constitutes the focus of this study. Using the clinical challenge of predicting mortality and hospital length of stay on a Danish trauma dataset, we hypothesized that a transfer learning approach of models trained on large external datasets would provide optimal prediction results compared to de-novo training on sparse but local datasets or directly porting externally trained models.MethodsUsing an external dataset of trauma patients from the US Trauma Quality Improvement Program (TQIP) and a local dataset aggregated from the Danish Trauma Database (DTD) enriched with Electronic Health Record data, we tested a range of model-level approaches focused on predicting trauma mortality and hospital length of stay on DTD data. Modeling approaches included de-novo training of models on DTD data, direct porting of models trained on TQIP data to the DTD, and a transfer learning approach by training a model on TQIP data with subsequent transfer and retraining on DTD data. Furthermore, data-level approaches, including mixed dataset training and methods countering imbalanced outcomes (e.g., low mortality rates), were also tested.ResultsUsing a neural network trained on a mixed dataset consisting of a subset of TQIP and DTD, with class weighting and transfer learning (retraining on DTD), we achieved excellent results in predicting mortality, with a ROC-AUC of 0.988 and an F2-score of 0.866. The best-performing models for predicting long-term hospitalization were trained only on local data, achieving an ROC-AUC of 0.890 and an F1-score of 0.897, although only marginally better than alternative approaches.ConclusionOur results suggest that when assessing the optimal modeling approach, it is important to have domain knowledge of how incidence rates and workflows compare between hospital systems and datasets where models are trained. Including data from other health-care systems is particularly beneficial when outcomes are suffering from class imbalance and low incidence. Scenarios where outcomes are not directly comparable are best addressed through either de-novo local training or a transfer learning approach.

Early triage of trauma patients requiring massive transfusion (MT) may help to marshal appropriate resources and improve treatment and outcome. Artificial intelligence (AI) and machine learning (ML) offer theoretical advantages compared to conventional prediction algorithms but have not been thoroughly evaluated in this population. We hypothesized that AI/ML techniques incorporating all available data in a patient’s medical record could achieve similar, if not higher, performance in the prediction of mortality in MT patients as compared to existing models. Patients from the American College of Surgeons Trauma Quality Improvement Project database (TQIP) were retrospectively reviewed. Those receiving ≥ 5 units of red blood cells and/or whole blood within the first four hours of arrival were defined as MT patients. Those receiving ≥10 units were identified as ultramassive transfusion (UMT) patients. ML models were created to predict 6-hour mortality using variables available at different time points, including patient arrival. Of 5,481,046 patients in TQIP from 2017 to 2021, 47,744 received MT and 20,337 of these received UMT. Using only variables available on arrival, MT AUROC was 0.901 [95% CI 0.895–0.910] which increased to 0.943 [95% CI 0.938–0.948] with addition of 4-hour variables. For UMT, arrival AUROC was 0.858 [95% CI 0.846–0.872] and increased to 0.922 [95% CI 0.914–0.931] at 4 hours. ML models reliably predict mortality in both MT and UMT patients. These are the only ML models trained on MT and UMT patients. Future work can focus on prospective implementation of these models with potential direct integration into the electronic medical record. Real-time utilization of comprehensive patient data may enhance clinical decision-making regarding which patients should continue receiving massive transfusion, thus optimizing the allocation of this limited resource.