Development and external validation of a novel prediction model for the TraumaTriage App - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Eur J Trauma Emerg Surg . 2026 Apr 17;52(1):135. doi: 10.1007/s00068-026-03175-8 Search in PMC Search in PubMed View in NLM Catalog Add to search Development and external validation of a novel prediction model for the TraumaTriage App Max Gulickx Max Gulickx 1 Department of Surgery, University Medical Center Utrecht, C04.332, Heidelberglaan 100, Utrecht, 3584 CX The Netherlands Find articles by Max Gulickx 1, ✉ , Robin D Lokerman Robin D Lokerman 1 Department of Surgery, University Medical Center Utrecht, C04.332, Heidelberglaan 100, Utrecht, 3584 CX The Netherlands Find articles by Robin D Lokerman 1 , Rogier van der Sluijs Rogier van der Sluijs 2 Department of Radiology, University Medical Center Utrecht, Utrecht, The Netherlands Find articles by Rogier van der Sluijs 2 , Rinske M Tuinema Rinske M Tuinema 3 Regional Ambulance Facilities Utrecht, Bilthoven, The Netherlands 4 Department of Emergency Medicine, Diakonessenhuis Utrecht/Zeist/Doorn, Utrecht, The Netherlands Find articles by Rinske M Tuinema 3, 4 , Risco van Vliet Risco van Vliet 5 Regional Ambulance Facilities Brabant Midden-West-Noord, ’s- Hertogenbosch, The Netherlands Find articles by Risco van Vliet 5 , Falco Hietbrink Falco Hietbrink 1 Department of Surgery, University Medical Center Utrecht, C04.332, Heidelberglaan 100, Utrecht, 3584 CX The Netherlands 6 Trauma Center Utrecht, Utrecht, The Netherlands Find articles by Falco Hietbrink 1, 6 , Rolf H H Groenwold Rolf H H Groenwold 7 Department of Clinical Epidemiology, Leiden University Medical Center, Leiden, The Netherlands Find articles by Rolf H H Groenwold 7 , Mark van Heijl Mark van Heijl 1 Department of Surgery, University Medical Center Utrecht, C04.332, Heidelberglaan 100, Utrecht, 3584 CX The Netherlands 6 Trauma Center Utrecht, Utrecht, The Netherlands 8 Department of Surgery, Diakonessenhuis Utrecht/Zeist/Doorn, Utrecht, The Netherlands Find articles by Mark van Heijl 1, 6, 8 Author information Article notes Copyright and License information 1 Department of Surgery, University Medical Center Utrecht, C04.332, Heidelberglaan 100, Utrecht, 3584 CX The Netherlands 2 Department of Radiology, University Medical Center Utrecht, Utrecht, The Netherlands 3 Regional Ambulance Facilities Utrecht, Bilthoven, The Netherlands 4 Department of Emergency Medicine, Diakonessenhuis Utrecht/Zeist/Doorn, Utrecht, The Netherlands 5 Regional Ambulance Facilities Brabant Midden-West-Noord, ’s- Hertogenbosch, The Netherlands 6 Trauma Center Utrecht, Utrecht, The Netherlands 7 Department of Clinical Epidemiology, Leiden University Medical Center, Leiden, The Netherlands 8 Department of Surgery, Diakonessenhuis Utrecht/Zeist/Doorn, Utrecht, The Netherlands ✉ Corresponding author. Received 2025 Oct 27; Accepted 2026 Apr 4; Issue date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13090252 PMID: 41995835 Abstract Purpose Accurate pre-hospital trauma triage is essential for optimizing survival and functional outcomes within inclusive trauma systems. A prediction model incorporated into the TraumaTriage App (TTApp) has been shown to improve patient allocation in the Netherlands, but was developed based on a relatively small cohort. This study aims to redevelop and improve the performance and applicability of the TTApp’s prediction model before nationwide implementation. Methods In this prospective multicenter cohort study, data from all trauma patients transported within two emergency medical service (EMS) regions (i.e., Brabant and Utrecht), between February 2015 and October 2019, were used to develop and externally validate a prediction model to identify severely injured adults (Injury Severity Score [ISS] ≥ 16) using routinely available pre-hospital data. A Gradient Boosting Decision Tree (GBDT) algorithm was applied, and model performance was evaluated in terms of discrimination and calibration. Results The development cohort included 51,001 patients (median age 63.8 years, median ISS 9), and the external validation cohort included 29,737 patients (median age 62.1 years, median ISS 9). In external validation, the GBDT model showed excellent discrimination (c-statistic 0.850; 95% CI, 0.837–0.863) and good calibration (calibration-in-the-large 0.009; slope 0.952). Sensitivity was 91.3% and 85.2% at specificity thresholds of 50% and 65%, respectively. Conclusions This externally validated prediction model effectively identifies severely injured patients in the pre-hospital setting and represents a first iteration towards enhancing the TTApp. Designed for use in urgent situations, the model allows predictions with incomplete data and shows potential to reduce undertriage rates toward 10% while maintaining acceptable overtriage levels. Supplementary Information The online version contains supplementary material available at 10.1007/s00068-026-03175-8. Keywords: Trauma triage, Prediction model, Pre-hospital care, Machine learning, Severe injury Background Accurate pre-hospital triage of trauma patients is crucial for patient outcomes, as the transportation of a severely injured patient to a lower-level trauma center (i.e., undertriage) is associated with increased mortality and lifelong disabilities [ 1 ]. Conversely, the transportation of mildly or moderately injured patients to a higher-level trauma center (i.e., overtriage) results in increased costs and overutilization of scarce resources [ 2 ]. The Dutch Health Care Institute and the American College of Surgeons Committee on Trauma (ACSCOT) therefore set a maximum undertriage rate of 10% and even 5%, respectively [ 3 , 4 ]. Unfortunately, no inclusive trauma system has been able to adhere to these standards [ 5 ]. Pre-hospital triage is – in most inclusive trauma systems – performed by Emergency Medical Service (EMS) professionals who assess a patient’s need for specialized trauma care at the scene of injury and subsequently determines transportation destination. In the Netherlands, EMS professionals are aided in their on-scene decision making by the field triage criteria of the Dutch National Protocol of Ambulance Services [ 6 ], which is derived from the American Field Triage Decision Scheme [ 4 , 7 ]. These decision schemes, however, provide limited assistance as they showed to be insensitive (i.e., sensitivity 36.2%, specificity 92.6%) in identifying severely injured patients [ 8 ]. We previously developed an algorithm that was potentially able to decrease undertriage to approximately 10%, while maintaining an overtriage rate of 50% [ 9 ]. When integrated in a mobile application (Trauma Triage App [TTApp]) and implemented in the Dutch pre-hospital practice, the use of the TTApp demonstrated to decrease undertriage rates with 5% from 32% to 27%, while maintaining overtriage at approximately 20% [ 10 ]. However, the performance of the TTApp’s prediction model could, potentially, be further improved using machine learning and a larger study population as part of a continuous improvement strategy, as the previous model was developed using a relatively small sample size and conventional modeling technique. In the current study, we aim to redevelop and externally validate a prediction model to identify severely injured patients to improve the performance and applicability of the TTApp before nationwide implementation. Methods Study setting This prospective, multicenter cohort study - performed between February 1, 2015, and October 31, 2019 - was conducted during the implementation study of the TTApp in two EMS regions (Brabant Midden-West-Noord and Utrecht). The corresponding ambulance services fully cover 3 of the 11 inclusive trauma regions in the Netherlands (Traumazorgnetwerk Midden-Nederland, Netwerk Acute Zorg Brabant, and Acute Zorgregio Oost) non-exclusively [ 10 ]. The participating trauma regions comprise 3 higher-level (i.e., level-1) and 18 lower-level (i.e., level 2 or 3) trauma centers. The ambulance services transport around 160,000 trauma patients to a trauma center annually and serve a region of rural, suburban, and urban surroundings in an area of approximately 5000 km2 that accommodates around 4 million people [ 11 ]. Patients All trauma patients aged 16 years or older who were transported by the participating ambulance services to a trauma center in the participating trauma regions, were included. Patients transported to a trauma center in a non-participating trauma region were excluded as data were unavailable, with the transportation to these regions most likely being related to geographical factors. Patients transported by ambulance service Brabant Midden-West-Noord were included between May 1, 2016, and October 31, 2019, and patients transported by ambulance service Utrecht were included between February 1, 2015, and September 30, 2019. Outcomes The primary objective was to develop a prediction model based on ISS ≥ 16, a universally recognized threshold to define severe injury and the criterion recommended by the ACSCOT for pre-hospital triage evaluation [ 4 ]. Undertriage is defined as a severely injured patient (ISS ≥ 16) transported to a lower-level trauma center, whereas overtriage is defined as a non-severely injured patient (ISS < 16) transported to a higher-level trauma center. Data collection Pre-hospital and hospital data were prospectively collected during the implementation of the TTApp in the TraumaTriage Intervention study [ 10 ], and were linked using a previously developed linkage tool with an externally validated accuracy of 100% (95%-CI; 100.0–100.0) [ 12 ]. Pre-hospital data contained patients’ demographics, vital parameters, description of the injury mechanism, transportation destination, and a free text in which the physical examination and suspected injuries were described. Trained research assistants read all free texts and classified, independent of each other, whether the EMS professional suspected a serious injury (AIS ≥ 2) in the head or neck, thorax, abdomen, pelvic, or extremity region. Hospital data of the three participating trauma regions were prospectively collected by the Dutch Trauma Registry and comprised – among others – all diagnosed injuries within 30-days after trauma, and mortality. Professional data managers of the trauma registries classified injuries using the Abbreviated Injury Score (AIS) and computed ISSs [ 13 ]. Predictors Predictors associated with severe injury were chosen based on clinical reasoning and previous research [ 9 , 10 ]. The maximum number of predictors was limited to ensure practical applicability when incorporated in the TTApp. The prediction model was developed based on the following 12 predictors within (1) patient demographics (i.e., age), (2) vital signs (i.e., systolic blood pressure, oxygen saturation, and Glasgow Coma Scale pre-sedation and intubation), (3) injury mechanism (i.e., high-energy mechanism criteria), (4) the pre-hospital suspicion of serious injuries (i.e., suspicion of serious head or neck injury, thoracic injury, abdominal injury, pelvic fracture, extremity injury, or suspected serious injury in > 1 region), and (5) ambulance dispatch priority (Table 1 ). Table 1. Predictor variables Demographics Age Vital Signs Systolic blood pressure Oxygen Saturation Glasgow Coma Scale Injury mechanism Mechanism criteria* Suspected serious injury in AIS region Head/neck region Thorax region Abdomen region Pelvic region Extremities region Multiple regions (> 1 region) Urgency Ambulance dispatch priority Open in a new tab * Mechanism criteria include high-energy trauma mechanism such as a fall > 2-meter, motor vehicle crash > 32 km/h, or any type of entrapment Missing data Missing data are inherent and unavoidable in pre-hospital triage settings due to the urgent nature of providing trauma care. This often leads to challenges in the practical application of prediction models relying on logistic regression, as these models require complete data input to generate predictions. A GBDT algorithm was therefore used in the current study, as it deals with missing data by default by selecting the split-direction of a node that minimizes training loss (i.e., sparsity-aware split finding), consequently allowing the model to predict the presence of severe injury when certain parameters (e.g., vital signs) are not available [ 14 ]. Statistical analysis R statistical software (version 4.0.3.) was used to perform all statistical analyses [ 15 ]. The XGBoost package was used to develop a GBDT model to predict the presence of severe injury (ISS ≥ 16). A GBDT model is a type of machine learning algorithm which generally provide highly accurate predictions, as it works by combining multiple decision trees in which each subsequential tree is trained with the residual errors of the previous tree, improving its accuracy and creating a robust prediction model (Appendix 1) [ 14 ]. The model was trained in the Brabant region and externally validated in the Utrecht region. The performance of the final model was estimated and described in terms of discrimination and calibration. Discrimination was determined using the concordance-statistic (c-statistic) and – after plotting the receiver operating characteristic curve (ROC-curve) –undertriage rates (i.e., 1 – sensitivity) were assessed using predefined values for acceptable overtriage rates of 50% and 35% (i.e., 1 – specificity). Calibration was determined in terms of the calibration-in-the-large and calibration slope and was depicted in a calibration plot (Fig. 2 ). Calibration is essential when externally validating a prediction model as it refers to the agreement between the predicted probabilities and observed outcomes, ensuring that the model’s predicted probabilities accurately resemble to the true risk of severe injury in the population it is applied to. A decision curve analysis was performed to assess net benefit across a range of threshold probabilities in the development and validation cohorts [ 16 ]. Fig. 2. Open in a new tab Calibration plot of the prediction model at external validation (using data from the Utrecht region) Results Development of triage algorithm A total of 51,001 patients were included in the Brabant region during the study period, comprising the development data set (Fig. 1 ). The median age of these patients was 63.8 years (IQR; 41.0–79.9), 25,949 (50.9%) patients were male, and the median ISS was 9 (IQR; 4–9). A total of 1238 (2.4%) of the patients were severely injured according to the primary outcome (ISS ≥ 16), and undertriage and overtriage rates were 29.4% and 24.9%, respectively (Table 2 ). Model performance is shown in Table 3 . Development of the prediction model showed a c-statistic of 0.884 (95% CI, 0.874–0.894), with sensitivity of 95.0% and 89.5% at specificity cut-off thresholds of 50% and 65%, respectively. Fig. 1. Open in a new tab Flowchart of patient inclusion in the development and validation study of Trauma Triage Prediction Model Table 2. Baseline characteristics of adult trauma patients who were transported by participating EMS to an Emergency Department Variables Brabant region (development) Utrecht region (validation) n = 51,001 n = 29,737 Demographics Median (IQR) Median (IQR) Age, y (median, IQR) 63.8 (41.0–79.9) 62.1 (38.1–80.0) > 65 (n, %) 24,741 (48.5) 13,802 (46.4) Man, (n, %)( 25,949 (50.9) 14,169 (47.7) ISS (median, IQR) 9 (4–9) 9 (4–10) Pre-hospital vital parameters Median (IQR) Median (IQR) Systolic blood pressure, mmHg 141 (126–160) 142 (127–161) Heart Rate 81 (72–93) 80 (71–90) Respiratory rate 16 (14–18) 16 (14–18) Glasgow Coma Scale 15 (15–15) 15 (15–15) N (%) N (%) Systolic blood pressure < 90mmHg 538 (1.1) 405 (1.4) Heart rate > 110/min 3185 (6.2) 1580 (5.3) Respiratory rate < 10 or > 29/min 724 (1.4) 432 (1.5) Glasgow Coma Scale < 13 1708 (3.4) 1176 (4.0) Revised Trauma Score < 12 2247 (4.4) 1602 (5.4) Mechanism of injury N (%) N (%) Mechanism criteria a 1216 (2.5) 1554 (5.2) Penetrating 289 (0.6) 263 (0.9) Burns or inhalation injury 105 (0.2) 254 (0.9) Suspected serious injury in AIS region N (%) N (%) Head or Neck 6160 (12.1) 5700 (19.2) Face 976 (1.9) 2039 (6.9) Thorax 1985 (3.9) 2011 (6.8) Abdomen 616 (1.2) 639 (2.1) Pelvic 564 (1.1) 319 (1.1) Extremities 11,361 (22.3) 12,835 (43.2) Multiple regions 2475 (4.9) 2668 (9.0) HEMS assistance 1062 (2.1) 225 (0.8) Clinical characteristics N (%) N (%) ISS ≥ 16 1238 (2.4) 920 (3.1) Undertriage b 365 (29.4) 272 (29.6) Overtriage c 12,379 (24.9) 3862 (13.4) Transportation Destination Higher-level trauma center 13,246 (26.0) 4510 (15.2) Lower-level trauma center 37,728 (74.0) 25,227 (84.8) Highest dispatch priority 19,838 (38.9) 10,018 (33.7) Distance closest higher-level trauma center 31.4 (21.3–42.8) 17.6 (9.9–24.4) Pre-hospital intubation 380 (0.7) 327 (1.1) Emergency Intervention 193 (0.4) 150 (0.5) Admission to Intensive Care 1358 (2.7) 1107 (3.7) 24 h mortality 83 (0.2) 76 (0.3) Open in a new tab Abbreviations: ISS, Injury Severity Score a Mechanism criteria include a fall > 2-meter, motor vehicle crash > 32 km/h, or any type of entrapment b Severely injured patient (ISS ≥ 16) transported to a lower-level trauma center c Non-severely injured patient (ISS < 16) transported to a higher-level trauma center Table 3. Performance of Trauma Triage Prediction Model at development and external validation Performance Development External validation Discrimination C-statistic, 95%-CI 0.884 (0.874–0.894) 0.850 (0.837–0.863) Sensitivity Sensitivity, at 50% Specificity* 95.0% 91.3% Sensitivity, at 65% Specificity* 89.5% 85.2% Undertriage Undertriage, at 50% Overtriage 5.0% 8.7% Undertriage, at 35% Overtriage 10.5% 14.8% Calibration Calibration-in-the-large 0.009 Calibration Slope 0.952 Overall performance Brier score 0.025 Open in a new tab *Sensitivity at cut-off specificity External validation of triage algorithm A total of 29,737 patients were included in the Utrecht region during the study period, comprising the validation data set. The median age of these patients was 62.1 (IQR; 38.1–80.0), 14,169 (47.7%) patients were male, and the median ISS was 9 (IQR: 4–10). A total of 920 (3.1%) patients were severely injured according to the primary outcome (ISS ≥ 16), and undertriage and overtriage rates were 29.6% and 13.4%, respectively. External validation of the prediction model showed a c-statistic of 0.850 (95% CI, 0.837–0.863) and sensitivity of 91.3% and 85.2%, at specificity cut-off thresholds of 50% and 65%, respectively. The model was well calibrated with a calibration-in-the-large of 0.009 and calibration slope of 0.952. The Brier score was 0.025 indicting good overall predictive performance. (Fig. 2 ). Decision curve analysis showed comparable net benefit between the development and validation cohorts, indicating consistent model performance (Appendix 2.) Improving undertriage The prediction model could potentially improve undertriage rates in the external validation cohort from 29.6% to 8.7% (∆ difference, 20.9%) and to 14.8% (∆ difference, 14.8%) at an overtriage rate of 50% and 35%, respectively. Threshold probabilities with corresponding under- and overtriage rates are shown in Appendix 3. Discussion In this prospective, multicenter cohort study, a novel pre-hospital triage prediction model was developed and externally validated to identify patients with severe injury using a GBDT model. The prediction model demonstrates the potential to further reduce undertriage rates in the derivation and validation regions and was well calibrated. Incorporating the novel prediction model in the TTApp might therefore further improve pre-hospital patient allocation and could increase the TTApp’s applicability, as the prediction model was specifically designed to make predictions in urgent situations. So far, identifying patients in need of specialized trauma care is predominantly based on field triage protocols, which are flowchart like structures unable to distinguish the interdependent relationships of signs and symptoms of severe injury, resulting in limited sensitivity in identifying patients in need of specialized trauma care [ 8 ]. Several studies have developed prediction models to identify patients in need of specialized trauma care, using both traditional statistical approaches and, more recently, machine learning techniques. However, these models often predict specific outcomes (e.g., hemorrhage risk or need for hemorrhage control) or are derived from selected populations (e.g., helicopter-transported patients or those directly transported to higher-level trauma centers), limiting their generalizability to general trauma triage and identification of patients requiring transport to a higher-level trauma center [ 17 – 22 ]. We previously developed a pre-hospital prediction model to select severely injured patients (ISS ≥ 16) at the scene of injury, which was able to potentially reduce undertriage rates to 11.2%, at an overtriage rate of 50% in the development cohort [ 9 ]. When integrated in the TTApp and implemented during a pilot study in approximately 25% of the Dutch trauma regions, undertriage rates decreased from 31.8% to 26.8%, while overtriage rates did not increase (20.4%) [ 10 ]. The previous model was, however, developed based on a relatively small cohort (4950 patients of which 435 patients had an ISS ≥ 16), and was based on a logistic regression model which required complete data availability to provide predictions. The current prediction model was developed using a substantially larger cohort of 51,001 trauma patients (of whom 1238 patients had an ISS ≥ 16) and, when externally validated, yielded undertriage rates of 8.7% and 14.8% at overtriage rates of 50% and 35%, respectively, lower than those reported for our previous TTApp prediction model and other published pre-hospital triage prediction models. Advantageously, the novel model was specifically designed to make predictions in urgent situation without the necessity of complete data availability, as GBDT models deal with missing data by default. The use of these Machine Learning models allows the TTApp to make real-time predictions and continuously update its recommendation when new data becomes available, as predictors such as dispatch priority and vital signs can be automatically gathered from the digital EMS records. This reduces the need for manual input and prevents deriving EMS professionals from their routine of care and comprehensively increases the TTApp’s applicability. Due to both the improved performance and pre-hospital applicability, incorporating the novel prediction model in the TTApp is expected to further improve pre-hospital patient allocation and reduce undertriage. The prediction model allows EMS regions to regulate patient allocation to region specific acceptable overtriage rates to prevent extensive shifts in patient allocation when the TTApp is integrated in the Dutch pre-hospital practice. Strengths and limitations This study has several strengths. First, the prospective data collection and inclusion of all adult trauma patients transported by two ambulance services to both higher- and lower-level trauma centers in three trauma regions is a strength of this study, as it allowed us to develop a prediction model within a large generalizable trauma population and externally validate this model in an adjacent EMS region. Second, the use of a GBDT model is a great strength of this study, as these models generally provide highly accurate predictions and can handle missing data by default unlike logistic regression models. The use of these Machine Learning models is, due to these features, especially ideal in pre-hospital triage settings as it (1) decreases the need for manual input, (2) allows the model to perform when certain parameters (e.g., vital signs) are not measurable or filled in, and (3) is subsequently able to provide real-time predictions. This study also has some limitations. First, the pre-hospital suspicion of serious injuries was determined by trained research assistants who read all pre-hospital free texts and is in future use of the TTApp subjective to the EMS professionals’ primary assessments at the scene of injury. Extensive instructions on the practical use of the TTApp will be conducted prior to the nationwide implementation to ensure consistency and adequate use of the app. Second, while the use of a GBDT model improves the performance and applicability of the novel prediction model, limitations related to missing data remain. In future clinical settings, both the extend and patterns of missing data may differ from those observed in this study, which could impact the model’s performance and generalizability [ 23 ]. Although the model can generate predictions with incomplete input, extensive missingness may be associated with reduced predictive precision. The impact of varying degrees of data completeness on model performance will therefore be examined during implementation. Third, our prediction model was designed to identify severe injury, defined as an ISS ≥ 16. This outcome is widely used to guide the need for transfer to a higher-level trauma center and is recommended by the ACSCOT for evaluating pre-hospital triage. However, prior research suggests that ISS may not fully align with a patient’s actual need for early critical resources or interventions, potentially limiting its effectiveness as sole criterion for triage decision-making [ 24 ]. Future iterations of the TTApp may therefore incorporate alternative outcomes, such as early critical resource use [ 25 ]. The utilization of the TTApp facilitates the start of precision-medicine in pre-hospital trauma triage, and is intended to be implemented in the Netherlands in the coming years [ 3 ]. Future research should evaluate the accuracy of pre-hospital patient allocation following the TTApp’s implementation and continue the improvement strategy as new data becomes available. Conclusion A novel prediction model was developed as a first iteration of updating the TTApp’s performance and demonstrate to surpass the previous model and potentially improve undertriage rates towards 10% at acceptable overtriage rates. The prediction model was specifically designed to make predictions in urgent situations without the necessity of complete data input. Incorporating this model in the TTApp might further improve pre-hospital patient allocation and initiates a continuous improvement strategy. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (15.1KB, docx) Supplementary Material 2 (19.2KB, docx) Supplementary Material 3 (36.5KB, docx) Supplementary Material 4 (66.9KB, docx) Acknowledgements The authors thank all participating members of the Pre-hospital Trauma Triage Research Collaborative and the research staff of the participating Emergency Medical Services and trauma regions. Author contributions All authors contributed to the study conception and design. Materialpreparation, data collection and analysis were performed by Max Gulickx and Robin D.Lokerman. The first draft of the manuscript was written by Max Gulickx and all authors commented on previous versions of the manuscript. All authors read and approved the finalmanuscript. Funding This study was partly funded by grants from the Netherlands Organisation for Health Research and Development (ZonMw) and the Innovation Fund Health Insurers. The funding sources had no influence on the study design, data collection, statistical analysis, interpretation of data and writing of the report. Data availability The data that supports the findings of the current study is not publicly available due to its sensitive nature but is available upon a reasonable request that needs to be approved by the participating Emergency Medical Services and trauma regions, provided that appropriate ethical approval is sought. R-scripts are available upon request. Declarations Ethical approval The Medical Ethical Committee of the University Medical Center Utrecht decided that the Medical Research Involving Human Subjects Act did not apply to this study (reference number: 20/500747). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. MacKenzie EJ, Rivara FP, Jurkovich GJ, et al. A national evaluation of the effect of trauma–center care on mortality. N Engl J Med. 2006;354:366–78. 10.1056/NEJMsa052049. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Newgard CD, Staudenmayer K, Hsia RY, et al. The cost of overtriage: more than one–third of low–risk injured patients were taken to major trauma centers. Health Aff (Millwood). 2013;32(9):1591–9. 10.1377/hlthaff.2012.1142. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Zorginstituut Nederland. Advies bevordering implementatie multitraumanorm - verder omdat het beter is. Published March 28, 2023. https://www.zorginstituutnederland.nl/publicaties/adviezen/2023/03/28/advies-multitraumanorm . Accessed 13 Jun 2023. 4. American College of Surgeons Committee on Trauma. Resources for the optimal care of the injured patient. Available from: https://www.facs.org/quality-programs/trauma/quality/verification-review-and-consultation-program/standards/ . Accessed 13 Jun 2023. 5. van Rein EAJ, van der Sluijs R, Houwert RM, et al. Effectiveness of prehospital trauma triage systems in selecting severely injured patients: is comparative analysis possible? Am J Emerg Med. 2018;36(6):1060–9. 10.1016/j.ajem.2017.11.016. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Ambulancezorg Nederland. Landelijk protocol ambulancezorg. https://www.ambulancezorg.nl/themas/kwaliteit-van-zorg/protocollen-en-richtlijnen/landelijk-protocol-ambulancezorg . Accessed 13 Jun 2023. 7. Newgard CD, Fischer PE, Gestring M, et al. National guideline for the field triage of injured patients: Recommendations of the National Expert Panel on Field Triage, 2021. J Trauma Acute Care Surg. 2022;93(2):e49–60. 10.1097/TA.0000000000003627. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Voskens FJ, van Rein EAJ, van der Sluijs R, et al. Accuracy of prehospital triage in selecting severely injured trauma patients. JAMA Surg. 2018;153(4):322–7. 10.1001/jamasurg.2017.4472. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. van Rein EAJ, van der Sluijs R, Voskens FJ, et al. Development and validation of a prediction model for prehospital triage of trauma patients. JAMA Surg. 2019;154(5):421–9. 10.1001/jamasurg.2018.4752. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Lokerman RD, van Rein EAJ, Waalwijk JF, et al. Accuracy of prehospital triage of adult patients with traumatic injuries following implementation of a trauma triage intervention. JAMA Netw Open. 2023;6(4):e236805. 10.1001/jamanetworkopen.2023.6805. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Ambulancezorg Nederland. Tabellenboek 2018. Available from: https://www.ambulancezorg.nl/sectorkompas/sectorkompas-2018 . Accessed 13 Jun 2023. 12. van der Sluijs R, Lokerman RD, Waalwijk JF, et al. Accuracy of pre–hospital trauma triage and field triage decision rules in children (P2–T2 study): an observational study. Lancet Child Adolesc Health. 2020;4(4):290–8. 10.1016/S2352-4642(19)30431-6. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Gennarelli TA, Wodzin E, Association for the Advancement of Automotive Medicine. Abbreviated Injury Scale 2005: Update 2008. Barrington, IL: Association for the Advancement of Automotive Medicine; 2008. [ Google Scholar ] 14. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. arXiv. Published March 3, 2016. Available from: https://arxiv.org/abs/1603.02754 . Accessed 15 Apr 2025. 15. R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2021. [ Google Scholar ] 16. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Mak. 2006;26(6):565–74. 10.1177/0272989X06295361. [ DOI ] [ PMC free article ] [ PubMed ] 17. Morris R, Karam BS, Zolfaghari EJ, et al. Need for emergent intervention within 6 hours: a novel prediction model for hospital trauma triage. Prehosp Emerg Care. 2022;26(4):556–65. 10.1080/10903127.2021.1984360. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Dinh MM, Bein KJ, Oliver M, et al. Refining the trauma triage algorithm at an Australian major trauma centre: derivation and internal validation of a triage risk score. Eur J Trauma Emerg Surg. 2014;40(1):67–74. 10.1007/s00068-013-0292-4. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Weidman AC, Malakouti S, Salcido DD, et al. A Machine Learning Trauma Triage Model for Critical Care Transport. JAMA Netw Open. 2025;8(6):e259639. 10.1001/jamanetworkopen.2025.9639. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Chen Q, Qin Y, Jin Z, et al. Enhancing Performance of the National Field Triage Guidelines Using Machine Learning: Development of a Prehospital Triage Model to Predict Severe Trauma. J Med Internet Res. 2024;26:e58740. 10.2196/58740. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Frock A, Robbins JT, Vital-Lopez FG, et al. A case study of AI-enabled software as a medical device cleared by the FDA for assessing hemorrhage risk index (APPRAISE-HRI) after trauma. NEJM AI. 2025;2(11):e2401170. 10.1056/aics2401170. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Gauss T, James A, Colas C, et al. Comparison of machine learning and human prediction to identify trauma patients in need of hemorrhage control resuscitation (ShockMatrix study): a prospective observational study. Lancet Reg Health Eur. 2025;55:101340. 10.1016/j.lanepe.2025.101340. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Groenwold RHH. Informative missingness in electronic health record systems: the curse of knowing. Diagn Progn Res. 2020;4:8. 10.1186/s41512-020-00082-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Baxt WG, Upenieks V. The lack of full correlation between the Injury Severity Score and the resource needs of injured patients. Ann Emerg Med. 1990;19(12):1396–400. 10.1016/S0196-0644(05)82606-X. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Lerner EB, Willenbring BD, Pirrallo RG, et al. A consensus–based criterion standard for trauma center need. J Trauma Acute Care Surg. 2014;76(4):1157–63. 10.1097/TA.0000000000000189. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (15.1KB, docx) Supplementary Material 2 (19.2KB, docx) Supplementary Material 3 (36.5KB, docx) Supplementary Material 4 (66.9KB, docx) Data Availability Statement The data that supports the findings of the current study is not publicly available due to its sensitive nature but is available upon a reasonable request that needs to be approved by the participating Emergency Medical Services and trauma regions, provided that appropriate ethical approval is sought. R-scripts are available upon request. Articles from European Journal of Trauma and Emergency Surgery are provided here courtesy of Springer ACTIONS View on publisher site PDF (946.9 KB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top