Hybrid model based on fuzzy logic and classification trees for the prediction of mortality of critical trauma patients: Pre-hospital variables from the RETRAUCI registry Authors: Luis Serviá, Mariona Badia, Neus Montserrat, Judit Vilanova, Gabriel Jiménez, Juan Antonio Llompart-Pou, Jesús Abelardo Barea-Mendoza, Mario Chico-Fernández, Javier Trujillano, the Neurointensive Care and Trauma Working Group of the Spanish Society of Intensive Care Medicine (SEMICYUC) Published: 2026 ABSTRACT AbstractBackgroundIn critically injured trauma patients, tools that stratify injury severity and estimate mortality are essential. Fuzzy logic (FL) enables the creation of accurate, interpretable models but requires decision rules, which can be generated using machine learning (ML) techniques like classification trees (CT). Our objective was to develop a hybrid model combining fuzzy logic and classification trees to estimate ICU mortality risk using only prehospital variables.MethodsWe conducted a retrospective study using data from the Spanish Trauma ICU registry (RETRAUCI) from 2015 to 2022. Patients were randomly divided into derivation (DS) and validation sets (VS) (70:30). Candidate variables were those available in the prehospital phase. A hybrid model (HFL) was developed using a Fuzzy Inference System built with the ‘FuzzyR’ library in RStudio (v 2024.04.2) and the Mamdani method. CHAID (Chi-squared Automatic Interaction Detection) classification trees were used to derive the rules. The HFL’s discrimination and calibration were compared with other scores: Revised Trauma Score (RTS); Glasgow Coma Scale, Age, and Arterial Pressure (GAP); Mechanism, Glasgow Coma Scale, Age, and Arterial Pressure (MGAP); Reverse shock index multiplied by Glasgow Coma Scale score (rSIG); and Trauma Rating Index in Age, Glasgow Coma Scale, Respiratory Rate, and Systolic Blood Pressure (TRIAGE).ResultsThe study included 11,030 records, with 7,728 in the DS and 3,302 in the VS, and an overall mortality of 11.1%. Five variables were selected, ordered by importance: GCS, age, systolic blood pressure, respiratory rate, and heart rate. A total of 32 classification rules were generated. The HFL model achieved the highest accuracy in DS and VS, with AUROC of 0.87 (0.86–0.88) and 0.86 (0.83–0.88), and acceptable calibration with intercepts of -0.11 (-0.18 to -0.04) and − 0.19 (-0.32 to -0.06) and slopes of 0.99 (0.94–1.05) and 0.96 (0.83–0.88).ConclusionsOur hybrid model achieves accuracy comparable to commonly used models and provides clear clinical interpretation, with GCS and age as key variables.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12874-026-02792-6.Keywords: Trauma and injury severity score, Intensive care unit, Mortality, Prognostic scoring systems, Fuzzy logic, Classification trees Background Effective management, prognosis, and quality control of critically injured trauma patients require tools capable of stratifying patient severity by estimating their probability of death [1]. These tools rely on variables collected at various stages of care: prehospital, upon ICU admission, during the first 24 h of hospitalization, and throughout the ICU stay [2]. In this study, we focus on the initial phase. The prehospital phase provides an opportunity to quickly assess mortality risk using readily available variables that do not require advanced technology for collection [2, 3]. To develop accurate risk assessment tools, access to large and representative databases is essential [4]. One such resource is the RETRAUCI registry, which includes data from more than 12,000 trauma patients across 52 units, covering all phases of the care process [5]. Most existing models for estimating mortality risk are based on scoring systems that apply fixed cut-off points to variables, categorizing them according to their deviation from normal values [6]. Partial scores for each variable are combined into a global score, which is then linked to the risk of death. These models are traditionally based on statistical methodologies such as logistic regression. A widely used example is the Revised Trauma Score (RTS), which evaluates three variables—Glasgow Coma Score (GCS), systolic blood pressure (SBP), and respiratory rate (RR)—each scored from 1 to 4 based on predefined categories. For instance, a respiratory rate of 10 scores 4, whereas a rate of 9 scores 2, illustrating how small differences can shift a patient into a different risk category [6]. Fuzzy Logic (FL) has been applied in medicine, particularly in control devices and clinical decision-making processes [7–9]. Unlike traditional models with rigid thresholds, FL enables the categorization of continuous variables using overlapping ranges, allowing for a more flexible interpretation. It introduces linguistic labels such as “low,” “normal,” or “high” to describe values like blood pressure [10]. The development of an FL-based model requires collaboration with clinical experts to define appropriate overlap intervals and ensure that the system’s rules remain interpretable [11]. However, establishing a comprehensive set of rules can be complex. In this context, machine learning techniques such as classification trees (CT) are valuable. CTs help generate a manageable number of rules that can be integrated into the FL model while preserving its interpretability [12–14]. Our hypothesis is that a model based on fuzzy logic and decision rules derived from classification trees can achieve satisfactory predictive performance while remaining interpretable. In summary, the aim of this study is to develop a hybrid model that combines fuzzy logic and classification trees to estimate mortality risk in trauma patients admitted to the ICU, using only prehospital variables. The performance of this model will be compared with that of commonly used scoring systems. Methods Study population This retrospective study evaluated prognostic accuracy using prehospital variables collected from critically injured trauma patients included in the RETRAUCI registry between 2015 and 2022 [15]. Records with missing data were excluded, and a complete-case analysis was performed. The RETRAUCI (National Trauma ICU Registry) study includes data from 52 ICUs across Spain. It operates through a web-based electronic database (retrauci.org) and is endorsed by the Neurointensive Care and Trauma Working Group of the Spanish Society of Intensive Care Medicine (SEMICYUC) [16]. The study complied with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Hospital Universitario 12 de Octubre, Madrid (Ref. 12/209). As the analysis was retrospective and based on deidentified data, informed consent was not required. Patients were treated according to Advanced Trauma Life Support (ATLS) principles. This study adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines [17]. Sample size calculation confirmed sufficient data for both model development and validation. Following conventional recommendations for logistic regression-based models, at least 10 events (non-survivors) per predictor variable are required. Assuming an estimated mortality rate of 10% and up to 10 candidate predictors, at least 1,000 patient records were needed for each set (derivation and validation) [18]. The dataset met these requirements. The final study group was randomly divided (70:30) into a derivation set (DS) and a validation set (VS). The DS was used to develop the model. Predictive variables Candidate variables were limited to those available during the prehospital phase, excluding the need for laboratory or imaging studies. The following variables were included: Glasgow Coma Scale (GCS), age (years), systolic blood pressure (SBP, mmHg), respiratory rate (RR), and heart rate (HR). Additional categorical variables included sex, type of trauma (penetrating vs. non-penetrating), and mechanism of injury (road traffic accident, fall, occupational accident, sport-related injury, assault, self-injury, other, or unknown) [15]. The outcome variable was hospital mortality. Hybrid model of fuzzy logic and classification tree (hfl) The HFL model was developed in four stages [9, 19–21]: Fuzzification: In this phase, crisp numerical variables were transformed into fuzzy linguistic variables using membership functions that define overlapping categories. For example, the crisp GCS variable (ranging from 3 to 15) can be transformed into a linguistic variable with four categories: GCS OK (15), Medium (9–14), Low (4–8), and Very Low (3). These classifications are then used to construct membership functions. There is no overlap in extreme values (3 and 15), while Medium and Low categories do overlap, reflecting the fuzzy nature of the model. These fuzzy groupings were defined by consensus among all authors (LS, MB, NM, JV, GJ, JALL, JAM, MCF, and JT), all with clinical experience, to ensure interpretability and clinical relevance. Fuzzification: In this phase, crisp numerical variables were transformed into fuzzy linguistic variables using membership functions that define overlapping categories. For example, the crisp GCS variable (ranging from 3 to 15) can be transformed into a linguistic variable with four categories: GCS OK (15), Medium (9–14), Low (4–8), and Very Low (3). These classifications are then used to construct membership functions. There is no overlap in extreme values (3 and 15), while Medium and Low categories do overlap, reflecting the fuzzy nature of the model. These fuzzy groupings were defined by consensus among all authors (LS, MB, NM, JV, GJ, JALL, JAM, MCF, and JT), all with clinical experience, to ensure interpretability and clinical relevance. Rule Generation via Classification Tree (CT): To define the rules, we used a methodology based on classification trees, specifically the CHAID type (Chi-squared Automatic Interaction Detection). CHAID was selected as it is appropriate for categorical variables and allows for multi-way splits, unlike CART which only performs binary splits. This approach uses chi-squared statistical tests to determine the partitions, making it particularly suitable for categorical variables. The classification tree selects the final variables included in the model and establishes the classification rules and the hierarchy among the variables. Rule Generation via Classification Tree (CT): To define the rules, we used a methodology based on classification trees, specifically the CHAID type (Chi-squared Automatic Interaction Detection). CHAID was selected as it is appropriate for categorical variables and allows for multi-way splits, unlike CART which only performs binary splits. This approach uses chi-squared statistical tests to determine the partitions, making it particularly suitable for categorical variables. The classification tree selects the final variables included in the model and establishes the classification rules and the hierarchy among the variables. Additionally, the CT helps define the probability of death (PDEATH) groups for each decision rule. 3.Fuzzy Inference System (FIS): A Mamdani-type fuzzy inference system was used to integrate the membership functions and decision rules, producing fuzzy output values for each patient [9, 10, 12]. Fuzzy Inference System (FIS): A Mamdani-type fuzzy inference system was used to integrate the membership functions and decision rules, producing fuzzy output values for each patient [9, 10, 12]. 4.Defuzzification: The defuzzification procedure converts the fuzzy values obtained by the inference model (Mandani) that has combined the membership functions with the decision rules, into PDEATH values (between 0 and 1). Defuzzification: The defuzzification procedure converts the fuzzy values obtained by the inference model (Mandani) that has combined the membership functions with the decision rules, into PDEATH values (between 0 and 1). Once the model was constructed, it was applied to both the derivation and validation sets to calculate mortality probabilities for each patient. Comparison with other models We compared the HFL model with other models that also use prehospital variables. Table 1 lists the variables used in each model: Revised Trauma Score (RTS) [6], Glasgow Coma Scale, Age, and Arterial Pressure (GAP), Mechanism, Glasgow Coma Scale, Age, and Arterial Pressure (MGAP) [22], Reverse shock index multiplied by Glasgow Coma Scale score (rSIG) ((SBP/HR) × GCS score) [23], and Trauma Rating Index in Age, Glasgow Coma Scale, Respiratory Rate, and Systolic Blood Pressure (TRIAGE) [4]. Table 1Variables included in the different risk models for prehospital useGCSHFLRTSGAPMGAPRSIGTRIAGEXXXXXXAGEXXXXSBPXXXXXXRRXXXHRXXAbberviations: GCS Glasgow Coma Score, SBP Systolic Blood Pressure, RR Respiratory Rate, HR Heart Rate, RTS Revised Trauma Score, GAP GCS Age and systolic blood pressure model, GAPM GCS Age, systolic blood pressure and mechanism model, RSIG Reverse Shock Index multiplied by GCS model, TRIAGE Trauma Rating Index in Age, Glasgow Coma Scale, Respiratory rate and Systolic blood pressure, HFL Hybrid Fuzzy Logic model developed in this work Variables included in the different risk models for prehospital use Abberviations: GCS Glasgow Coma Score, SBP Systolic Blood Pressure, RR Respiratory Rate, HR Heart Rate, RTS Revised Trauma Score, GAP GCS Age and systolic blood pressure model, GAPM GCS Age, systolic blood pressure and mechanism model, RSIG Reverse Shock Index multiplied by GCS model, TRIAGE Trauma Rating Index in Age, Glasgow Coma Scale, Respiratory rate and Systolic blood pressure, HFL Hybrid Fuzzy Logic model developed in this work Performance of the models Different accuracy values were measured: AUC ROC, Sensitivity (S), Specificity (E), Positive Predictive Value (PPV), and Negative Predictive Value (NPV). The Youden Index was applied to find the optimal cutoff point for each model used. Additionally, using the probabilities calculated in the FL model, the calibration curve was constructed, and the Intercept and Slope values were determined [3]. Good calibration is indicated by an Intercept close to 0 and a Slope close to 1. These measurements were performed in both the Development set (DS) and the Validation set (VS). Statistical analysis Categorical variables were described as percentage and continuous variables as median (interquartile range), as they did not follow a normal distribution (Kolmogorov-Smirnov test). For the comparison between the groups Derivation-Validation and Survivors-Non-survivors, the Mann-Whitney test was used for continuous variables and the chi-square test for categorical variables. A p < 0.05 was considered statistically significant. The statistical calculations were performed in RStudio (version 2024.04.2) [24]. For conducting CHAID-type classification trees, we used the “CHAID” library, and for the FIS model, we employed the “FuzzyR” library [25]. As specific hyperparameters for the models used, we employed the following for the CHAID-type CT (alpha2 = 0.05, alpha3=-1, alpha4 = 0.05, minsplit = 50, minbucket = 25, minprob = 0.01, cross-validation = 10 folds); and for the Mamdani-type FIS (AND method = MIN, OR method = MAX, Implication = MIN, Aggregation = MAX, and Defuzzification=CENTROID). Results The study included a total of 11,030 records (see Fig. 1). A development set (DS) of 7,728 records and a validation set (VS) of 3,302 records were randomly generated. Flow diagram for selection of patients Table 2 presents the distribution of available prehospital variables in the DS and VS. The overall hospital mortality rate was 11.1%. Table 3 displays the DS divided into survivors and non-survivors, highlighting statistically significant differences in all included variables. Table 2Demographic and clinical characteristics of the whole sample and distributed by the derivation and validation sets (n = 11030)VariableALLN = 11,030DERIVATIONN = 7728VALIDATIONN = 3302p-valueAge (years)48 (33–64)49 (33–64)48 (33–63)0.577Age (groups)0.909 < 454900 (44.4)3419 (44.2)1481 (44.9) 45–653627 (32.9)2544 (32.9)1083 (32.8) 65–751375 (12.5)966 (12.5)409 (12.4) > 751128 (10.2)799 (10.3)329 (10.3)Sex (% male)8675 (78.6)6105 (79.0)2579 (77.8)0.171Type of injury0.270 Road Traffic accident4685 (42.5)3294 (42.6)1391 (42.1) Fall3016 (27.4)2106 (27.3)920 (27.9) Occupational accident863 (7.8)622 (8.0)241 (7.3) Sport related637 (5.8)473 (6.1)220 (6.7) Assault599 (5.4)434 (5.6)165 (5.0) Self-injury693 (6.3)447 (5.8)190 (5.8) Others348 (3.2)238 (3.1)110 (3.3) Unknown179 (1.6)114 (1.5)65 (2.0)Mechanism penetrant729 (6.6)526 (6.8)203 (6.1)0.202GCS score15 (10–15)15 (9–15)15 (9–15)0.594GCS groups 155630 (51.0)3962 (51.0)1668 (50.5)0.128 9–142917 (26.4)1997 (25.8)920 (27.9) 4–81502 (13.6)1065 (13.8)437 (3.2) 3981 (8.9)704 (9.1)277 (8.4)SBP (mmHg)121 (105–140)121 (105–140)123 (105–140)0.104SBP groups < 901354 (12.3)953 (12.3)401 (12.1)0.103 90–1208768 (79.5)6167 (79.8)2601 (78.8) > 120908 (8.2)608 (7.9)300 (9.1)RR17 (15–20)17 (15–20)17 (15–20)0.608RR groups < 10456 (4.1)334 (4.3)122 (3.7)0.269 10–3010,239 (92.8)7155 (92.6)3084 (93.4) > 30335 (3.0)239 (3.1)96 (2.9)HR89 (75–105)89 (75–104)89 (75–105)0.424HR groups < 60706 (6.4)510 (6.6)196 (5.9)0.359 60–1209350 (84.8)6529 (84.5)2821 (85.4) > 120974 (8.8)689 (8.9)285 (8.6)Hospital mortality1220 (11.1)883 (11.4)337 (10.2)0.061Values ​​expressed as percentages or median (Interquartile range). GCS: Glasgow Coma Score. SBP: Systolic blood pressure. RR: Respiratory Rate. HR: Heart Rate. p-value: calculated using chi-square test or Mann-Whitney test Demographic and clinical characteristics of the whole sample and distributed by the derivation and validation sets (n = 11030) Values ​​expressed as percentages or median (Interquartile range). GCS: Glasgow Coma Score. SBP: Systolic blood pressure. RR: Respiratory Rate. HR: Heart Rate. p-value: calculated using chi-square test or Mann-Whitney test Table 3Demographic and clinical characteristics of patients in the derivation set distributed by hospital mortality (n = 7728)VariableSURVIVORSN = 6845NON-SURVIVORSN = 883p-valueAge (years)47 (32–61)66 (50–77)< 0.001Age (groups)< 0.001 < 453131 (47.2)188 (21.3) 45–652303 (33.6)241 (27.3) 65–75777 (11.4)189 (21.4) > 75534 (7.8)265 (30.0)Sex (% male)460 (79.8)645 (73.0)< 0.001Type of injury< 0.001 Road Traffic accident3016 (44.1)278 (31.5) Fall1698 (24.8)408 (46.2) Occupational accident572 (8.4)50 (5.7) Sport related429 (6.3)18 (2.0) Assault416 (6.1)18 (2.0) Self-injury397 (5.8)76 (8.6) Others222 (3.2)16 (1.8) Unknown95 (1.4)19 (2.2)Mechanism penetrant492 (7.2)34 (3.9)< 0.001GCS score15 (12–15)6 (3–12)< 0.001GCS groups 153828 (55.9)134 (15.2)< 0.001 9–141810 (26.4)187 (21.4) 4–8839 (12.3)226 (25.6) 3368 (5.4)336 (38.1)SBP (mmHg)121 (107–140)120 (87–149)< 0.001SBP groups < 90721 (10.5)232 (26.3)< 0.001 90–1205637 (82.4)530 (60.0) > 120487 (7.1)121 (13.7)RR18 (15–20)16 (14–20)< 0.001RR groups < 10185 (2.7)149 (16.9)< 0.001 10–306453 (94.3)702 (79.5) > 30207 (3.0)32 (3.6)HR88 (75–103)90 (70–110)0.110HR groups < 60402 (5.9)108 (12.2)< 0.001 60–1205876 (85.8)653 (74.0) > 120567 (8.3)122 (13.8)Values expressed as percentages or median (Interquartile range). GCS: Glasgow Coma Score. SBP: Systolic blood pressure. RR: Respiratory Rate. HR: Heart Rate. p-value: calculated using chi-square test or Mann-Whitney test Demographic and clinical characteristics of patients in the derivation set distributed by hospital mortality (n = 7728) Values expressed as percentages or median (Interquartile range). GCS: Glasgow Coma Score. SBP: Systolic blood pressure. RR: Respiratory Rate. HR: Heart Rate. p-value: calculated using chi-square test or Mann-Whitney test Hybrid fuzzy logic model (hfl) The HFL model was constructed using the DS. The overall model design is illustrated in (Fig. 2). Diagram of the hybrid model of fuzzy logic and classification trees (CHAID). GCS: Glasgow Coma Score. SBP: Systolic Blood Pressure. RR: Respiratory Rate. HR: Heart Rate. PDEATH: Probability of hospital mortality. Model Phases: A- Fuzzification. Input variables and the membership functions used are shown. B- Fuzzy Inference System. Mamdani type. Creation of rules through a methodology based on CHAID classification trees (Chi-square automatic interaction detection). C- Defuzzification. Obtaining PDEATH values for each record The fuzzification phase is detailed in Table 4, which includes the transformation of numeric input variables into linguistic categories and the membership functions used for both inputs and the output variable (probability of death). Triangular or trapezoidal membership functions were used for most variables, while singleton functions were assigned to the extreme GCS values (3 and 15) to reflect their categorical certainty. Table 4Fuzzy database with fuzzy membership functions for mortality risk prediction N VariableROLEValuesFuzzy MembershipFunctions1GCSInputOK (15)SING (15)Medium (14 − 9)TRA (8-12-14-14)Low (8 − 4)TRA (4-4-6-10)Very Low (3)SING (3)2AGE (years)InputYoung (< 45)TRA (0-16-30-45)Middle (45–64)TRA (40-45-60-65)Old (65–75)TRA (60-65-75-80)Very Old (> 75)TRA (75-90-100-100)3SBP (mmHg)InputLow (< 90)TRA (0-0-40-90)Normal (90–120)TRA (70-100-140-170)High (> 120)TRA (140-180-200-200)4RRInputLow (< 10)TRA (0-2-8-14)Normal (10–30)TRA (12-14-18-22)High (> 30)TRA (18-30-60-60)5HRInputLow (< 60)TRA (20-20-50-70)Normal (60–120)TRA (50-70-120-130)High (> 120)TRA (120-140-200-200)6RISKOutputVery Low (< 0.02)TRA (0-0-0.005-0.015)(PDEATH)Low (0.02–0.08)TRI (0.005-0.08-0.12)Medium (0.08–0.16)TRI (0.08-0.12-0.2)High (0.16–0.32)TRI (0.12-0.2-0.24)Very High (0.32–0.48)TRI (0.2-0.4-0.5)Extreme (> 0.48)TRA (0.4-0.6-1-1)Abbreviations: GCS Glasgow Coma Score, SBP Systolic Blood Pressure, RR Respiratory Rate, HR heart Rate, TRI Triangular, TRA Trapezoidal, SING Singleton Fuzzy database with fuzzy membership functions for mortality risk prediction Abbreviations: GCS Glasgow Coma Score, SBP Systolic Blood Pressure, RR Respiratory Rate, HR heart Rate, TRI Triangular, TRA Trapezoidal, SING Singleton The final predictive variables were selected using the CHAID methodology. Although sex, type of trauma (penetrating vs. non-penetrating), and mechanism of injury were included in the initial classification tree analysis, they were not selected by the algorithm to form part of the final rule set. Supplementary Figure S2 shows the analysis of the trauma mechanism variable. Neither the original eight RETRAUCI categories nor the three-level risk grouping based on injury mechanism were included in the final model. The CHAID tree generated 32 rules (see Fig. 3). The most influential predictor at the first decision level was GCS, followed by age at the second level. Systolic blood pressure (SBP) emerged as the key variable at the third level, while respiratory rate (RR) and heart rate (HR) contributed to further stratification at the fourth level. Supplementary Figures S3–S7 display the complete classification tree and the resulting decision rules. Rules incorporated into the fuzzy inference system obtained through the CHAID classification tree The classification tree also defined probability of death (PDEATH) ranges associated with each rule. The resulting probability distribution was skewed towards lower values, consistent with other mortality prediction models. Comparison with other prehospital models Table 5 summarizes the predictive performance of the HFL model compared to established prehospital scores in both the DS and VS. Table 5Comparison of the performance of the models in predicting mortality. Derivation and validation setsHFLRTSGAPMGAPRSIGTRIAGEAUROCD0.870.790.840.830.750.85V0.860.780.830.820.740.8495% CID0.86–0.880.77–0.810.83–0.860.82–0.850.73–0.770.84–0.87V0.83–0.880.75–0.800.81–0.850.80–0.840.71–0.770-82-0.87Cut-offD1371819134V1371820124Se (%)D797277716977V827276696777Sp (%)D807474797377V767674797776PPV (%)D332627302530V282525232427NPV (%)D979596959496V979696959596Abbreviations: D Derivation set, V Validation set, AUROC Area Under the ROC curve, CI Confidence interval, Cut-off Cutoff point (Youden’s index), Se Sensitivity, Sp Specificity, PPV Positive predictive value, NPV Negative predictive value, RTS Revised Trauma Score, GAP GCS Age and systolic blood pressure model, GAPM GCS Age systolic blood pressure and mechanism model, RSIG Reverse Shock Index multiplied by GCS model, TRIAGE Trauma Rating Index in Age, Glasgow Coma Scale Respiratory rate and Systolic blood pressure, HFL Hybrid Fuzzy Logic model developed in this work Comparison of the performance of the models in predicting mortality. Derivation and validation sets Abbreviations: D Derivation set, V Validation set, AUROC Area Under the ROC curve, CI Confidence interval, Cut-off Cutoff point (Youden’s index), Se Sensitivity, Sp Specificity, PPV Positive predictive value, NPV Negative predictive value, RTS Revised Trauma Score, GAP GCS Age and systolic blood pressure model, GAPM GCS Age systolic blood pressure and mechanism model, RSIG Reverse Shock Index multiplied by GCS model, TRIAGE Trauma Rating Index in Age, Glasgow Coma Scale Respiratory rate and Systolic blood pressure, HFL Hybrid Fuzzy Logic model developed in this work Supplementary Figure S8 shows the estimated mortality probabilities according to each model. As expected, lower scores in the RTS, GAP, MGAP, and RSIG models were associated with higher mortality, whereas higher scores in TRIAGE and the HFL probabilities were directly associated with increased mortality risk. Notably, the RSIG model demonstrated poor calibration. Figure 4 presents the calibration curves for the HFL model in both the DS and VS, demonstrating good overall calibration, with a slight decrease in performance observed in the validation set. Calibration curves of the HFL model. Based on derivation or validation set Discussion Our Hybrid Fuzzy Logic (HFL) model demonstrates acceptable levels of predictive accuracy, and in some instances, outperforms commonly used prehospital trauma models. A particularly appealing feature of the HFL model is the necessary involvement of clinical experts throughout its development. For instance, during the definition of membership functions, clinicians readily acknowledged the limitations of rigid cutoff values, which often conflict with real-world decision-making in critical care (e.g., defining hypotension using a fixed threshold). The fuzzy logic approach aligns well with how clinicians perceive and manage uncertainty in practice [8, 19]. Furthermore, the interpretability of the model’s decision rules addresses one of the major criticisms of predictive models: the “black box” nature. While an exhaustive rule generation process could have resulted in over 400 rules, which would significantly reduce interpretability, the HFL model maintains a concise and clinically sensible structure [11]. As illustrated in Fig. 3, the derived rules are intuitive. For example, rules 1, 24, and 32 combine GCS scores with extreme age ranges, highlighting clear risk associations. Similarly, the presence of hypotension worsens the prognosis (rules 4, 6, 14, 27), while unusually high SBP values (rule 30) may indicate patients nearing brain death. Extremes of respiratory rate (rules 25 and 31) and abnormal heart rate values further refine the risk profile. The use of CHAID classification trees contributes to model transparency and interpretability [26]. Hybrid models, combining fuzzy logic with decision trees, have shown utility in other domains related to decision-making and risk assessment [27–29]. The CHAID algorithm automatically selected relevant variables and established their predictive hierarchy. In our model, GCS and age emerged as the most influential predictors — a finding consistent with widely accepted prehospital risk models [30, 31]. Interestingly, sex was not selected, despite some studies suggesting increased mortality in females. Still, sex is infrequently included as a standalone risk factor in trauma scoring systems [32]. We had anticipated the inclusion of injury type as a predictive variable. However, our data — similar to the German DGU registry — show a lower incidence of penetrating trauma, often considered less severe in our context. This may be attributed to the rarity of firearm-related injuries in our setting [30, 33, 34]. By contrast, data from other regions (e.g., the UK) report higher mortality and a greater prevalence of penetrating injuries, highlighting the challenge of generalizing predictive models across different populations. Another important consideration is that trauma databases apply varying inclusion criteria, which limits comparability. For example, the TRIAGE model, developed using a large Japanese trauma dataset, excludes HR as a variable and uses ISS ≥ 9 as an inclusion criterion. Our model, on the other hand, includes HR and is based on ICU admission, reflecting a more critical patient population [4]. Limitations Our study has several limitations. Although the database is extensive, a larger sample might reveal additional patterns or decision rules. To ensure an acceptable generalization of the model, an external validation using a different dataset from RETRAUCI is needed. We also acknowledge that other machine learning methodologies (e.g., ensemble models like XGBoost or Random Forest) may yield higher accuracy, albeit at the expense of interpretability [35]. Moreover, alternative rule generation techniques could be explored [26]. Incorporating additional prehospital variables — such as oxygen saturation, frailty scores, prior comorbidities, and prehospital care type — may improve model performance [36, 37]. However, we intentionally prioritized variables that are universally available, regardless of location or trauma mechanism, to ensure broad applicability. In a future version of RETRAUCI, more variables related to prehospital care and comorbidities will be included. The main technical limitation for clinical implementation is the need for specialized software and the complexity of integrating the model into clinical systems [25]. We are currently developing a user-friendly web-based graphical interface to facilitate model use by clinicians. Fuzzy logic offers an alternative conceptual framework for risk modelling. It departs from the rigidity of fixed cutoffs and complements the interpretability of classification trees. Importantly, the model not only functions as a predictive tool but also as a representation of knowledge, illustrating how input variables relate to outcomes — a quality often lacking in traditional models. We emphasize the need for diverse methodological approaches in clinical risk modelling. While clinical judgment remains fundamental, it should be enhanced — not replaced — by interpretable tools co-designed with healthcare professionals [38]. For decision-making, clinical professionals are more likely to accept interpretable tools that align with their usual practice [39]. Conclusions In conclusion, we can state that the hybrid model we developed performs well and achieves accuracy comparable to or superior to commonly used models. The rules generated by the classification tree have a clear clinical interpretation. The most important variables in our model are GCS and age. As a representation of knowledge, FL models serve as a bridge between methodology and routine clinical practice. Any risk calculation method should be implementable as fuzzy logic, which opens a broad and interesting field of research [20]. Supplementary Information Supplementary Material 1: Figure S1: TRIPOD Checklist: Prediction Model Development and Validation. Figure S2: Classification tree (CHAID) with the variable "type of trauma”. Figure S3: First partition of the CHAID tree, differentiating risk groups by GCS variable categories. Figure S4: Classification rules for the GCS group with a score of 15 points. Figure S5: Classification rules for the GCS group in the Medium category. Figure S6: Classification rules for the GCS group in the Low category. Figure S7: Classification rules for the GCS group with a score of 3 points. Figure S8: Probability of death according to the different scores of the models analysed. Acknowledgements Not applicable. Abbreviations Area under the ROC curve Chi-square automatic interaction detection Classification Tree Deutsche Gellschaft für Unfallchirurigie Derivation set Fuzzy Inference System Fuzzy Logic model Glasgow coma scale, Age, and systolic blood Pressure Mechanism, Glasgow coma scale, Age, and systolic blood Pressure Glasgow Coma Scale Hybrid Fuzzy Logic model Heart Rate Intensive Care Unit Injury severity score Negative predictive value Probability of hospital mortality Positive predictive value Registro nacional de TRAuma ingresados en UCI Respiratory Rate Reverse Shock Index multiplied by Glasgow Coma Scale score Revised Trauma Score Systolic Blood Pressure Sociedad Española de Medicina Intensiva, Crítica y Unidades Coronarias Trauma Rating Index in Age, Glasgow Coma Scale, Respiratory Rate, and Systolic Blood Pressure Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Validation set Authors' contributions LS, MB, and JT conceived and supervised the project. LS, JT, MB, NM, JV and GJ analysed the data, interpreted the findings, and drafted the manuscript. JAL, JAB, and MC designed and collected data study. All authors reviewed the manuscript and approved the final version. Funding This work has been funded by the Carlos III Health Institute (ISCIII), PI25/00859, co-funded bythe European Union. RETRAUCI was supported by a grant for the development of an electronic web-based system awarded to Dr. Chico-Fernández (Fundación Mutua Madrileña, reference number AP117892013). Data availability No datasets were generated or analysed during the current study. Ethics approval and consent to participate Ethics Committee approval for the registry was obtained (Hospital Universitario 12 de Octubre, Madrid: 12/209). Due to the retrospective analysis of de-identified collected data, informed consent was not obtained for this study. Consent for publication Not applicable. 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 Associated Data Supplementary Materials Supplementary Material 1: Figure S1: TRIPOD Checklist: Prediction Model Development and Validation. Figure S2: Classification tree (CHAID) with the variable "type of trauma”. Figure S3: First partition of the CHAID tree, differentiating risk groups by GCS variable categories. Figure S4: Classification rules for the GCS group with a score of 15 points. Figure S5: Classification rules for the GCS group in the Medium category. Figure S6: Classification rules for the GCS group in the Low category. Figure S7: Classification rules for the GCS group with a score of 3 points. Figure S8: Probability of death according to the different scores of the models analysed. Data Availability Statement No datasets were generated or analysed during the current study.