Abstract
This study aimed to identify key risk factors for delirium in trauma patients and to develop an interpretable machine learning model using routinely available demographic, clinical, and laboratory data collected at initial trauma center presentation. We analyzed data from 7,806 trauma patients admitted between 2015 and 2023 and constructed an XGBoost-based prediction model evaluated using a hold-out test set. Model interpretability was assessed using Shapley Additive Explanations to quantify feature contributions, threshold effects, and interactions. Delirium occurred in 568 patients (7.3%). The model demonstrated robust predictive performance, with an accuracy of 92.0%, a macro-average AUC of 0.76, and a micro-average AUC of 0.96. SHAP analysis identified age, Injury Severity Score, lactate dehydrogenase, and estimated glomerular filtration rate as the most influential predictors of delirium risk. These variables exhibited clinically meaningful threshold effects, including increased risk above approximately 60 years of age, ISS greater than 15, LDH levels exceeding 350 IU/L, and eGFR below 90 mL/min/1.73 m², as well as notable interactions. Overall, the proposed interpretable machine learning model effectively predicted delirium risk in trauma patients using routine admission data and provides a transparent basis for individualized risk assessment and early prevention strategies in acute trauma care.
Keywords: Delirium, Trauma, Risk prediction, XGBoost, SHAP, Interpretable machine learning
Subject terms: Biomarkers, Diseases, Health care, Medical research, Risk factors
Introduction
Delirium is a common, severe neuropsychiatric syndrome characterized by acute, fluctuating disturbances in consciousness, attention, and orientation. It frequently occurs in medically compromised patients, particu- larly those with chronic illnesses, cancer, or traumatic injuries. Delirium is associated with adverse outcomes, such as increased risk of dementia, shorter survival time1, prolonged hospitalization, extended duration of mechanical ventilation2, and persistent long-term cognitive decline3. These consequences underscore its im- portance as a critical determinant of the prognosis in medically and surgically vulnerable populations.
Physical trauma includes a wide range of injuries, from localized tissue damage to severe systemic com- plications. Blunt or penetrating trauma frequently causes direct injury to the skin, muscles, and vascula- ture, often resulting in hemorrhage, fractures, and organ injury4. Additionally, the systemic consequences of trauma extend beyond local tissue damage and involve inflammatory cascades that can cause shock, systemic inflammatory response syndrome, and multiple organ dysfunction syndrome5.
Delirium is increasingly recognized as a critical neuropsychiatric complication in trauma patients, par- ticularly among older adults and those with head injuries. Its multifactorial etiologies include metabolic disturbances, systemic inflammation, and physiological stress responses. Trauma-associated delirium has been linked to worse clinical trajectories, such as prolonged hospitalization, increased morbidity, and de- layed or incomplete functional recovery6,7. Because of its acute onset and fluctuating course, delirium is often difficult to recognize promptly. However, early prediction of delirium risk may enable timely preventive and therapeutic interventions, potentially reducing poor outcomes and improving recovery. Accurate iden- tification of patients at high risk of delirium is therefore essential for optimizing trauma care. While prior studies have primarily emphasized prediction performance and feature importance, fewer have focused on translating these findings into clinically actionable insights8–10. Accordingly, an interpretable model that provides such threshold-based insights may enhance real-world clinical applicability. This study aims to develop a predictive model for delirium in trauma patients using demographic characteristics and initial laboratory data collected at trauma center presentation. In addition to prediction, we aim to identify the most influential risk factors, explore threshold effects for high-risk stratification, and explore interactions among variables to capture conditional effects that may refine risk assessment. This approach is intended to provide a clinically interpretable framework for early risk stratification and informed decision-making, ultimately improving outcomes in this vulnerable population.
Methods
Participants
This retrospective study was approved by the Institutional Review Board (IRB), and the requirement for informed consent was waived because of the retrospective study design. All methods were performed in accordance with the relevant guidelines and regulations.
We collected data from trauma patients who presented to the Trauma Center of Dankook University Hospital and were hospitalized in any department from January 2015 to February 2023. To ensure independence of observations and avoid potential bias from correlated data, repeated visits were excluded, and only the first visit per patient was included. After excluding individuals with multiple visits, the data set included 17,390 unique patients. The institution is one of the regional trauma centers in the Republic of Korea, serving South Chungcheong Province. The data set includes laboratory findings, trauma-related diagnoses based on the International Statistical Classification of Diseases (ICD-10), Injury Severity Score (ISS), psychiatric consultation status and diagnoses, and demographic data. Inclusion criteria were: (1) hospitalization for at least 2 days; (2) patients with or without referral for psychiatric consultation, with the requirement that any requested consultation was completed during hospitalization; (3) availability of laboratory and demographic data without missing values; and (4) age ≥ 18 years. In addition to repeated visits, patients were excluded if they were younger than 18 years, had missing laboratory or demographic data, or had an incomplete psychiatric consultation after referral. During hospitalization, psychiatrists with at least 2 years of clinical experience conducted psychiatric consultations. Delirium diagnoses were determined based on psychiatrist-led clinical assessment and chart review of consultation records, considering cognitive status, level of consciousness, and clinical course. Patients diagnosed with delirium during these consultations were classified as the delirium group, and those without delirium were assigned to the control group.
Definition of input features and outcome
The outcome of the machine learning model in this study was the prediction of delirium in trauma patients. The model used 28 features (Table 1) primarily derived from demographic, clinical, and laboratory data collected at initial presentation to the trauma center, supplemented by the ISS to represent the overall trauma burden. These features incorporate established risk factors such as age, injury severity, and metabolic disturbances9,10. Eight laboratory measurements from blood tests (e.g., CPK, TB, albumin) and the ISS were included as both continuous raw values (with z-score normalization11 applied to numerical features) and dichotomized variables based on clinically validated thresholds. The presence of clinical conditions directly related to each measurement or score (e.g., leukocytosis, anemia, inflammation) was encoded as indicator variables (1 = yes, 0 = no), using clinically validated thresholds. For example, WBC and hemoglobin were analyzed as continuous measures and as binary indicators for leukocytosis or anemia, and ISS was represented as both a continuous integer (1–75) and a binary variable for severe trauma (ISS > 15)12. Categorical variables, such as sex, ICU admission status, and trauma involving specific body regions, were also encoded as indicator variables (1 = yes, 0 = no). Trauma involving specific anatomical regions was encoded as separate variables to reflect the original structure of the clinical data and allow the model to capture variation across injury locations without assuming homogeneity. The variables used in the model were derived from routinely collected demographic information, injury characteristics, and initial laboratory data available at the time of trauma center presentation, along with injury burden reflecting real-world clinical practice and enhancing the practical applicability of the model.
Table 1.
Baseline characteristics by target group.
| Variable | No Delirium (N = 7238) | Delirium (N = 568) | p-value |
|---|---|---|---|
| Age | 51.6 ± 18.1 | 66.1 ± 16.8 | 0.000 |
| Sex | 0.001 | ||
| Male | 5283 (73.0%) | 378 (66.5%) | |
| Female | 1955 (27.0%) | 190 (33.5%) | |
| ISS | 12.0 ± 10.1 | 15.3 ± 9.5 | 0.000 |
| ICU admission | 0.000 | ||
| No | 4222 (58.3%) | 249 (43.8%) | |
| Yes | 3016 (41.7%) | 319 (56.2%) | |
| Injury area – foot | 0.098 | ||
| No | 6937 (95.8%) | 553 (97.4%) | |
| Yes | 301 (4.2%) | 15 (2.6%) | |
| Injury area – head | 0.003 | ||
| No | 4817 (66.6%) | 413 (72.7%) | |
| Yes | 2421 (33.4%) | 155 (27.3%) | |
| Injury area – thorax | 0.041 | ||
| No | 6445 (89.0%) | 522 (91.9%) | |
| Yes | 793 (11.0%) | 46 (8.1%) | |
| Injury area – lower leg | 0.960 | ||
| No | 6536 (90.3%) | 512 (90.1%) | |
| Yes | 702 (9.7%) | 56 (9.9%) | |
| Injury area – neck | 0.560 | ||
| No | 6859 (94.8%) | 542 (95.4%) | |
| Yes | 379 (5.2%) | 26 (4.6%) | |
| Injury area – lower arm | 0.163 | ||
| No | 6781 (93.7%) | 541 (95.2%) | |
| Yes | 457 (6.3%) | 27 (4.8%) | |
| Injury area – upper arm | 0.275 | ||
| No | 6820 (94.2%) | 542 (95.4%) | |
| Yes | 418 (5.8%) | 26 (4.6%) | |
| Injury area – abdominopelvis | 0.075 | ||
| No | 6222 (86.0%) | 504 (88.7%) | |
| Yes | 1016 (14.0%) | 64 (11.3%) | |
| Injury area – hand | 0.010 | ||
| No | 6965 (96.2%) | 559 (98.4%) | |
| Yes | 273 (3.8%) | 9 (1.6%) | |
| Injury area – upper leg | 0.000 | ||
| No | 6760 (93.4%) | 424 (74.6%) | |
| Yes | 478 (6.6%) | 144 (25.4%) | |
| CPK | 461.7 ± 1472.3 | 501.3 ± 894.5 | 0.337 |
| TB | 0.6 ± 0.4 | 0.7 ± 0.7 | 0.007 |
| Albumin | 4.2 ± 0.6 | 3.8 ± 0.6 | 0.000 |
| Hemoglobin | 13.4 ± 2.0 | 12.2 ± 2.0 | 0.000 |
| WBC | 12.4 ± 8.9 | 13.3 ± 6.1 | 0.001 |
| Platelets | 222.0 ± 65.0 | 204.6 ± 71.8 | 0.000 |
| eGFR | 101.8 ± 20.8 | 88.9 ± 23.4 | 0.000 |
| LDH | 328.6 ± 248.4 | 366.8 ± 237.0 | 0.000 |
| WBC factor | 0.001 | ||
| Leukocytosis | 4319 (59.7%) | 378 (66.5%) | |
| Normal | 2919 (40.3%) | 190 (33.5%) | |
| Albumin factor | 0.000 | ||
| Normal | 6409 (88.5%) | 428 (75.4%) | |
| Decreased | 829 (11.5%) | 140 (24.6%) | |
| TB factor | 0.016 | ||
| Increased | 435 (6.0%) | 49 (8.6%) | |
| Normal | 6803 (94.0%) | 519 (91.4%) | |
| eGFR factor | 0.000 | ||
| Normal | 6903 (95.4%) | 505 (88.9%) | |
| Decreased | 335 (4.6%) | 63 (11.1%) | |
| ISS factor | 0.000 | ||
| Normal / Minor | 5028 (69.5%) | 319 (56.2%) | |
| Major Trauma | 2210 (30.5%) | 249 (43.8%) | |
| Hemoglobin factor | 0.000 | ||
| Anemia | 2171 (30.0%) | 319 (56.2%) | |
| Normal | 5067 (70.0%) | 249 (43.8%) | |
Delirium prediction model and interpretation
To predict delirium occurrence from clinical data, several approaches can be used, including linear statisti- cal methods (e.g., Elastic Net13, logistic regression14, neural network–based methods (e.g., MLP, TabNet15), and tree-based ensemble methods (e.g., LightGBM16, Random Forest17, XGBoost). In this study, we used the XGBoost (Extreme Gradient Boosting) algorithm18. XGBoost was chosen as the primary predictive model because it has shown strong performance on structured clinical data set19–23, effectively captures nonlinear relationships, and remains robust to class imbalance and missing values. Its regularized objective function reduces overfitting, and the sparsity-aware algorithm allows efficient handling of missing data.
Because delirium cases represented a minority class in the study population, we explicitly considered the potential impact of class imbalance during model development. We did not adopt data-level resampling strategies such as random undersampling or oversampling as the primary approach. Undersampling can substantially reduce the effective sample size by discarding clinically informative observations from the majority class, whereas oversampling or synthetic sample generation methods may distort the original data distribution or generate clinically implausible synthetic samples in medical tabular data24. Therefore, class imbalance was addressed at the model level using a class-weighted XGBoost framework rather than by resampling the training data18.
Additionally, the model is highly scalable for large data sets, and its integration with SHAP (Shapley Additive Explanations)25 enables transparent interpretation of feature importance, individual contributions, and threshold effects. This interpretability allows clinicians to understand the rationale for classifying pa- tients as high risk. Using SHAP values, we quantified each predictor’s contribution to the model output and evaluated its overall influence across the data set.
To further characterize the relationships between predictors and delirium risk, SHAP dependence plots were used to visualize how feature attributions varied across feature value ranges26. In addition, SHAP interaction values were computed to decompose model predictions into main effects and pairwise interaction effects, allowing assessment of whether the contributions of key clinical and laboratory predictors varied according to sex25. This approach enables analysis of conditional feature contributions without imposing linearity or parametric assumptions. Joint effects were examined for predictors with high global importance and established clinical relevance.
Compared with standard regression approaches such as logistic regression, the proposed our approach may offer practical advantages for this study. Conventional regression models typically require prespecified interaction or nonlinear terms, whereas this framework can identify such patterns more flexibly from the data. In addition, SHAP enables intuitive assessment of how predicted risk changes across feature values, including clinically relevant threshold-like patterns25.
Statistical analysis
Model training was conducted on an NVIDIA RTX 3090 GPU, with all training and evaluation pro- cedures performed using Python 3.10.8, PyCaret 3.3.2, Pandas 1.5.8, scikit-learn 1.4.2, and SHAP 0.45.1. To assess the robustness and generalizability of the delirium prediction model based on demographic and initial laboratory measurements from trauma center patients, we used five-fold cross-validation26 during model development. Demographic characteristics, laboratory values, medication use, and clinical outcomes were compared between the training set and test set. Indicator variables were summarized as frequencies and percentages, and continuous numerical features as medians with interquartile ranges. For group comparisons, t-tests were used for continuous variables, and χ2 tests for categorical variables.
Results
Baseline characteristics
Figure 1 presents the study flowchart. Of 17,361 patients who visited the trauma center during the study period, 7,806 met the inclusion criteria and were included in the final analysis: 7,238 in the non-delirium group and 568 in the delirium group. The overall incidence of delirium was 7.3%. The data set was randomly divided into a training set (70%) and a test set (30%) for model development and validation. Patients who developed delirium were significantly older than those without delirium (66.1 ± 16.8 years vs. 51.6 ± 18.1 years) and had a slightly lower proportion of males (66.5% vs. 73.0%). The delirium group also had a higher Injury Severity Score (15.3 ± 9.5 vs. 12.0 ± 10.1; p < .001), indicating more severe trauma. ICU admission was more frequent among patients who developed delirium (56.2% vs. 41.7%). Other variables, such as injury site and laboratory findings, are summarized in Tables 1 and 2.
Fig. 1.
Flow diagram of patient selection for the delirium prediction model.
Table 2.
Baseline characteristics by dataset split.
| Variable | Test (N = 2339) | Train (N = 5467) | p-value |
|---|---|---|---|
| Age | 53.1 ± 18.6 | 52.5 ± 18.3 | 0.197 |
| Sex | 0.534 | ||
| Male | 1708 (73.0%) | 3953 (72.3%) | |
| Female | 631 (27.0%) | 1514 (27.7%) | |
| ISS | 12.4 ± 10.1 | 12.2 ± 10.1 | 0.349 |
| ICU admission | 0.952 | ||
| No | 1338 (57.2%) | 3133 (57.3%) | |
| Yes | 1001 (42.8%) | 2334 (42.7%) | |
| Injury area – foot | 0.305 | ||
| No | 2253 (96.3%) | 5237 (95.8%) | |
| Yes | 86 (3.7%) | 230 (4.2%) | |
| Injury area – head | 0.541 | ||
| No | 1555 (66.5%) | 3675 (67.2%) | |
| Yes | 784 (33.5%) | 1792 (32.8%) | |
| Injury area – thorax | 0.943 | ||
| No | 2089 (89.3%) | 4878 (89.2%) | |
| Yes | 250 (10.7%) | 589 (10.8%) | |
| Injury area – lower leg | 0.826 | ||
| No | 2115 (90.4%) | 4933 (90.2%) | |
| Yes | 224 (9.6%) | 534 (9.8%) | |
| Injury area – neck | 0.143 | ||
| No | 2204 (94.2%) | 5197 (95.1%) | |
| Yes | 135 (5.8%) | 270 (4.9%) | |
| Injury area – lower arm | 0.238 | ||
| No | 2206 (94.3%) | 5116 (93.6%) | |
| Yes | 133 (5.7%) | 351 (6.4%) | |
| Injury area – upper arm | 0.961 | ||
| No | 2205 (94.3%) | 5157 (94.3%) | |
| Yes | 134 (5.7%) | 310 (5.7%) | |
| Injury area – abdominopelvis | 0.177 | ||
| No | 1996 (85.3%) | 4730 (86.5%) | |
| Yes | 343 (14.7%) | 737 (13.5%) | |
| Injury area – hand | 0.427 | ||
| No | 2261 (96.7%) | 5263 (96.3%) | |
| Yes | 78 (3.3%) | 204 (3.7%) | |
| Injury area – upper leg | 0.205 | ||
| No | 2167 (92.6%) | 5017 (91.8%) | |
| Yes | 172 (7.4%) | 450 (8.2%) | |
| CPK | 469.6 ± 1709.0 | 462.4 ± 1305.3 | 0.854 |
| TB | 0.6 ± 0.5 | 0.6 ± 0.4 | 0.891 |
| Albumin | 4.2 ± 0.6 | 4.2 ± 0.6 | 0.160 |
| Hemoglobin | 13.4 ± 2.0 | 13.2 ± 2.1 | 0.012 |
| WBC | 12.3 ± 5.4 | 12.5 ± 9.9 | 0.191 |
| Platelets | 221.5 ± 67.7 | 220.4 ± 64.7 | 0.528 |
| eGFR | 100.9 ± 21.5 | 100.9 ± 21.1 | 0.966 |
| LDH | 328.2 ± 210.3 | 332.7 ± 262.1 | 0.427 |
| WBC factor | 0.690 | ||
| Leukocytosis | 1399 (59.8%) | 3298 (60.3%) | |
| Normal | 940 (40.2%) | 2169 (39.7%) | |
| Albumin factor | 0.181 | ||
| Normal | 2067 (88.4%) | 4770 (87.3%) | |
| Decreased | 272 (11.6%) | 697 (12.7%) | |
| TB factor | 0.643 | ||
| Increased | 140 (6.0%) | 344 (6.3%) | |
| Normal | 2199 (94.0%) | 5123 (93.7%) | |
| eGFR factor | 0.978 | ||
| Normal | 2219 (94.9%) | 5189 (94.9%) | |
| Decreased | 120 (5.1%) | 278 (5.1%) | |
| ISS factor | 0.803 | ||
| Normal / Minor | 1597 (68.3%) | 3750 (68.6%) | |
| Major Trauma | 742 (31.7%) | 1717 (31.4%) | |
| Hemoglobin factor | 0.158 | ||
| Anemia | 719 (30.7%) | 1771 (32.4%) | |
| Normal | 1620 (69.3%) | 3696 (67.6%) | |
Performance of the XGBoost model
The predictive performance of the XGBoost model, trained on the training set, was evaluated on the test set using AUROC curves. At the default threshold of 0.5, the model achieved an overall accuracy of 92.01%, and the AUC values for both the non-delirium group (class 0) and the delirium group (class 1) were 0.76 (Fig. 2). To better characterize model performance under class imbalance, both micro- and macro-average AUCs were evaluated. The macro-average AUC was calculated as the unweighted mean of the AUROC values for each class, reflecting equal importance of delirium and non-delirium patients regardless of class prevalence. In contrast, the micro-average AUC was computed by aggregating prediction outcomes across all patients and therefore primarily reflects performance driven by the majority class. Accordingly, the micro-average AUC for the entire group was 0.96, indicating excellent overall discriminative performance, while the macro-average AUC was lower at 0.76. This difference suggests that the model showed strong predictive ability for the majority class (non-delirium patients), but its discriminative power was limited for the minority class (delirium patients). These findings reflect the effect of class imbalance, indicating that although the model performed well at the population level, it faced challenges in early identification of delirium patients, who are clinically more critical.
Fig. 2.
ROC curves.
SHAP-based feature contribution
To improve the interpretability of the XGBoost model and identify the most influential predictors of delirium, we applied SHAP. SHAP quantifies the marginal contribution of each input feature to the model’s prediction and is especially appropriate for interpreting complex tree-based models.
Figure 3a displays the SHAP summary plot, which shows the distribution of SHAP values for each feature and illustrates how feature values, color-coded from low to high, affect the model output. Figure 3b presents a bar chart of the mean absolute SHAP values, ranking features by their average contribution to the model’s predictions.
Fig. 3.
(a) SHAP summary plot and (b) Mean SHAP values.
Both visualizations consistently identified Age as the most influential predictor of delirium. SHAP analysis showed that older patients received higher SHAP values, indicating a strong association between advanced age and increased delirium risk. Other highly ranked features were ISS, LDH, eGFR, and ICU admission. Additionally, WBC, platelet count (Plt), and Alb were important contributing variables. The SHAP summary plot indicated that WBC levels and ICU admission positively contributed to the prediction of delirium.
Among injury site variables, trauma to the upper leg, thorax, and head made significant contributions, indicating that injury location may affect a patient’s vulnerability to developing delirium. Sex and Hb also showed moderate influence.
Threshold effects of top predictors
SHAP dependence plots of the leading predictors are presented to illustrate how their contributions to delirium risk varied across value ranges. Top-ranked features, such as age, ISS, LDH, and eGFR, showed identifiable inflection points, indicating that their contributions varied across specific value ranges (Fig. 4).
Fig. 4.
SHAP dependence plots of the predictors of delirium risk.
Age showed a clear positive association with predicted risk, as SHAP values increased markedly beyond approximately 55–60 years, indicating an inflection point where age contributed more strongly to delirium risk. ISS also showed a generally positive relationship, with SHAP values rising especially after a score of 15. A minor dip appeared around ISS 25, likely because of sparse data, but this did not alter the overall trend of higher ISS contributing to increased risk. LDH demonstrated a threshold effect, with values above 300–400 IU/L linked to a sharp rise in SHAP values, suggesting that elevated LDH substantially increased predicted risk. In contrast, eGFR showed a negative association, with higher values corresponding to lower SHAP values. An inflection point occurred around 90 mL/min/1.73 m², after which the contribution plateaued or slightly reversed. Although some variability was present at the extremes, this pattern supports the interpre- tation that preserved renal function protects against delirium risk. Sex-specific interaction patterns for key laboratory predictors were visualized using SHAP interaction value plots for eGFR, albumin, and total bilirubin (Fig. 5).
Fig. 5.
SHAP interaction plots illustrating interactions with Sex.
For eGFR, a clear sex-dependent interaction was observed, despite its overall negative main association with delirium risk. Female patients showed distinctly positive SHAP interaction values below approximately 90 mL/min/1.73 m², indicating a stronger contribution to predicted risk when mildly reduced renal function was present. In contrast, males exhibited nearly flat interaction patterns across the entire range. These results suggest that the association between subnormal renal function and delirium risk was largely driven by female patients. Albumin and total bilirubin also showed meaningful sex-related interactions. For albumin, females had positive SHAP interaction values at lower concentrations, while males showed nearly flat or weakly positive trends, indicating that the adverse effect of hypoalbuminemia on delirium prediction was primarily attributable to female patients. For total bilirubin, a divergent sex-dependent trend was observed: the SHAP interaction values showed a positive correlation in females, whereas a slightly negative correlation was seen in male, suggesting a robust and persistent sex-dependent difference in the relationship between bilirubin levels and delirium risk.
Discussion
Clinical implication
In this study of trauma patients, advanced age, higher injury severity (higher ISS), elevated LDH levels, and reduced kidney function (i.e., lower eGFR) were identified as key predictive features of delirium, consistent with and extending prior literature in acutely ill population27–29.
Older age has long been recognized as a strong risk factor for delirium30. Our findings reaffirm this association in the trauma population, showing that older patients are particularly vulnerable, potentially because of reduced cognitive reserve, increased frailty, and a higher burden of comorbidities30–32. This result highlights the importance of early delirium screening and prevention strategies in older trauma patients. The association between delirium and injury severity (i.e. elevated ISS) supports the hypothesis that systemic physio- logical stress and inflammatory response are key factors in delirium pathogenesis. Severe trauma likely causes widespread inflammation and neuroendocrine dysregulation, which may contribute to acute brain dysfunction27,33. Identifying ISS as a correlate of delirium highlights the need for close monitoring of patients with high trauma burden, regardless of age. LDH, a marker of tissue damage and cellular stress, showed a significant association with delirium. Elevated LDH levels may indicate underlying hypoxia, ischemia, or systemic inflammation, all implicated in the neurobiology of delirium34,35. Although LDH is not traditionally considered a neuropsychiatric biomarker, these findings suggest its potential as an accessible biochemical indicator of delirium risk in trauma patients. Further research is needed to examine the mechanistic links between systemic metabolic stress and neurocognitive outcomes. Impaired renal function, indicated by reduced eGFR, was another significant predictor of delirium. This finding aligns with previous studies showing that both acute kidney injury and chronic renal dysfunction increase the likelihood of delirium, possibly because of the accumulation of neurotoxic metabolites, fluid–electrolyte imbalances, and altered drug metabolism36–38. These results suggest that careful management of renal function may be a modifiable target for delirium prevention in trauma care.
Beyond establishing these variables as important predictors, SHAP dependence plots were used to further examine how delirium risk varies across their value ranges. These analyses suggest that predicted risk increases more prominently beyond certain levels that may be clinically informative. Specifically, delirium risk rose more steeply in patients aged approximately 55 years or older, in those with ISS values exceeding 15, in patients with markedly elevated LDH (approximately > 300–400 IU/L), and in those with reduced renal function, particularly when eGFR fell below approximately 90 mL/min/1.73 m². Although these values should be interpreted as ranges associated with heightened predicted risk rather than fixed clinical thresholds, they may help clinicians identify trauma patients who warrant closer monitoring for delirium during the acute phase of hospitalization.
We also explored whether the contribution of key predictors differed by sex, given prior evidence of sex-specific heterogeneity in delirium risk factors. SHAP interaction analyses suggested that reduced eGFR, lower albumin levels, and elevated total bilirubin contributed more strongly to predicted delirium risk in female patients, whereas these effects were less pronounced in males. These patterns might be interpreted through the lens of potential sex-based physiological differences. For instance, the steeper risk increase in females at mild renal impairment may relate to reported sexual dimorphism in kidney functional reserve or hemodynamic stress responses39. Similarly, previous research has suggested that higher bilirubin levels could be associated with cognitive impairment in a sex-dependent manner, a relationship potentially influenced by sex hormones40. Although exploratory, these findings indicate that similar degrees of physiological stress may translate into different levels of delirium vulnerability by sex.
Overall, our study contributes to the growing evidence that delirium in trauma patients is a multifactorial phenomenon influenced by age-related vulnerability, injury burden, metabolic stress, and organ dysfunction. By combining global feature importance with dependence and interaction analyses, our findings further suggest that routinely available clinical variables may help identify patients who are at heightened risk. Early identification of high-risk patients using variables such as ISS, LDH, and eGFR may enable timely intervention, closer monitoring, and potentially improved clinical outcomes.
Future improvements
This study has several limitations. First, the analysis used retrospective data from a single tertiary trauma center, which may limit the generalizability of the results. Differences in patient demographics, clinical practices, and delirium management protocols across hospitals could affect model performance. Therefore, external validation using independent datasets, particularly from multiple centers, is essential to assess the robustness and generalizability of the proposed model. Furthermore, prospective validation studies are warranted to evaluate the model’s performance in real-world clinical settings and to confirm its potential for clinical implementation. Second, delirium diagnosis relied on consultation records rather than standardized screening tools, which may have resulted in underdiagnosis or misclassification of mild or transient cases. Future research should use validated instruments, such as the Confusion Assessment Method (CAM), for consistent.
case identification. In particular, distinguishing delirium from the neurocognitive effects of traumatic brain injury (TBI) can be challenging, especially in patients with head injury, and some degree of diagnostic overlap may have occurred. Third, although SHAP-based analyses revealed threshold effects in the relationships between key predictors and delirium risk, these findings should be interpreted with caution. While such effects may provide clinically informative insights for early risk stratification, they do not represent validated or definitive clinical cut-off values. Further studies are required to establish their robustness and clinical applicability. Fourth, several potentially important variables—such as sedative use, medication exposure, pre-existing cognitive impairment, and environmental factors (e.g., sleep disturbance, ICU noise)—were not available in the dataset. These unmeasured confounders might have affected delirium risk but could not be controlled for in this analysis. Finally, the model retained clinically relevant baseline variables to enable interpretation of predictor importance and interactions, although this may limit direct bedside usability. Future studies should explore simplified models using fewer readily available variables. In addition, the present model was trained only on baseline variables without considering longitudinal changes during hospitalization. Although certain predictors—such as the retrospective nature of ISS, the non-routine use of LDH, or the broad threshold of eGFR < 90—may present practical challenges, our findings highlight their collective value as indicators of ‘total trauma burden’ and physiological stress within a multivariate framework. Furthermore, as delirium typically manifests several days after admission, these initial data points provide a critical temporal window for early risk stratification. Therefore, future research is warranted to incorporate time-series variables and explore more rapidly assessable surrogates to enhance the model’s real-time clinical relevance.
Conclusion
Our interpretable machine learning model demonstrates the potential to predict delirium in trauma patients using routine clinical data. While the model showed promising predictive accuracy, it is important to note that this study was conducted at a single center using retrospective data, which may limit the generalizability of the findings. Future prospective, multi-center studies are needed to validate the clinical utility of this framework. By providing SHAP-based interpretations, this approach could assist clinicians in early risk stratification, although it should be used as a supplementary tool rather than a definitive diagnostic measure.
Author contributions
SS contributed to methodology design, formal analysis, visualization, drafting of the original manuscript, and manuscript revision. SBL and JJL contributed to data collection and data curation. DK contributed to study conceptualization, methodology design, and critical revision of the manuscript. SIC supervised the overall research process. All authors read and approved the final manuscript.
Funding
This research was supported by the Bio&Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (RS-2023-00220408). This work was also supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) – ICT Challenge and Advanced Network of HRD (ICAN) grant funded by the Korea government (Ministry of Science and ICT) (IITP-2026-RS-00437027).
Data availability
The datasets used and/or analyzed during the current study are not publicly available due to institutional regulations and the protection of patient privacy, but are available from the corresponding author on reasonable request.
Declarations
Competing interests
The authors declare no competing interests.
Ethics approval and consent to participate
This study was approved by the Institutional Review Board of Dankook University Hospital (approval number: Dankook university hospital 2026-02-006). The requirement for informed consent was waived due to the retrospective nature of the study.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Sujong Shin, Seok Bum Lee and Jung Jae Lee contributed equally to this work.
Contributor Information
Dohyun Kim, Email: dohyun.kim@dankook.ac.kr, Email: uptothebest@gmail.com.
Sang-Il Choi, Email: choisi@dankook.ac.kr.
References
- 1.Rockwood, K. et al. The risk of dementia and death after delirium. Age ageing. 28(6), 551–556 (1999). [DOI] [PubMed] [Google Scholar]
- 2.Salluh, J. I. et al. Outcome of delirium in critically ill patients: systematic review and meta-analysis. bmj350 (2015). [DOI] [PMC free article] [PubMed]
- 3.Goldberg, T. E. et al. Association of delirium with long-term cognitive decline: a meta- analysis. JAMA Neurol.77(11), 1373–1381 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Goffette, P. P. & Laterre, P. F. Traumatic injuries: imaging and intervention in post-traumatic complications (delayed intervention). Eur. Radiol.12(5), 994–1021 (2002). [DOI] [PubMed] [Google Scholar]
- 5.Lenz, A., Franklin, G. A. & Cheadle, W. G. Systemic inflammation after trauma. Injury38(12), 1336–1345 (2007). [DOI] [PubMed] [Google Scholar]
- 6.Lat, I. et al. The impact of delirium on clinical outcomes in mechanically venti- lated surgical and trauma patients. Crit. Care Med.37(6), 1898–1905 (2009). [DOI] [PubMed] [Google Scholar]
- 7.Marquetand, J. et al. Delirium in trauma patients: a 1-year prospective cohort study of 2026 patients. Eur. J. Trauma Emerg. Surg.48(2), 1017–1024 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Jauk, S. et al. Risk prediction of delirium in hospitalized patients using machine learning: an implementation and prospective evaluation study. J. Am. Med. Inf. Assoc.27(9), 1383–1392 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Wassenaar, A. et al. Multinational development and validation of an early prediction model for delirium in ICU patients. Intensive Care Med.41(6), 1048–1056 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Van den Boogaard, M. et al. Development and validation of PRE-DELIRIC (PREdiction of DELIRium in ICu patients) delirium prediction model for intensive care patients: observational multicentre study. Bmj 344 (2012). [DOI] [PMC free article] [PubMed]
- 11.Kuhn, M. et al. Applied Predictive Modeling 26 (Springer, 2013).
- 12.Baker, S. P., o’Neill, B., Haddon, W. Jr & Long, W. B. The injury severity score: a method for describing patients with multiple injuries and evaluating emergency care. J. Trauma. Acute Care Surg.14(3), 187–196 (1974). [PubMed] [Google Scholar]
- 13.Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. J. Royal Stat. Soc. Ser. B: Stat. Methodol.67(2), 301–320 (2005). [Google Scholar]
- 14.Hosmer, D. W. Jr, Lemeshow, S. & Sturdivant, R. X. Applied Logistic Regression (Wiley, 2013).
- 15.Arik, S. Ö. & Tabnet, P. T. Attentive interpretable tabular learning. Proc. AAAI Conf. Artif. Intell.35, 6679–6687 (2021).
- 16.Ke, G. et al. Lightgbm: A highly efficient gradient boosting decision tree. Adv. Neural Inform. Process. Syst.30 (2017).
- 17.Breiman, L. Random forests. Mach. Learn.45, 5–32 (2001). [Google Scholar]
- 18.Chen, T. & Guestrin, C. Xgboost: A scalable tree boosting system. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining 785–794 (2016).
- 19.Zhou, L., Zhu, Q., Chen, Q., Wang, P. & Huang, H. Predicting hospital outpatient volume using XGBoost: a machine learning approach. Sci. Rep.15(1), 17028 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Mochurad, L., Babii, V., Boliubash, Y. & Mochurad, Y. Improving stroke risk prediction by integrating XG- Boost, optimized principal component analysis, and explainable artificial intelligence. BMC Med. Inf. Decis. Mak.25(1), 63 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
- 21.Fernández-Blázquez, M. A. et al. XGBoost models based on non imaging features for the prediction of mild cognitive impairment in older adults. Sci. Rep.15(1), 29732 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Hidayaturrohman, Q. A. & Hanada, E. Impact of data pre-processing techniques on XGBoost model perfor- mance for predicting all-cause readmission and mortality among patients with heart failure. BioMedIn- formatics. 4(4), 2201–2212 (2024). [Google Scholar]
- 23.Atias, D. et al. Machine learning in epidemiology: an introduction, comparison with traditional methods, and a case study of predicting extreme longevity. Ann. Epidemiol. (2025). [DOI] [PubMed]
- 24.Li, Y., Yang, Y., Song, P., Duan, L. & Ren, R. An improved SMOTE algorithm for enhanced imbalanced data classification by expanding sample generation space. Sci. Rep.15, 23521 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Lundberg, S. M. & Lee, S. I. A unified approach to interpreting model predictions. Adv. Neural Inform. Process. Syst.30 (2017).
- 26.Kohavi, R. others. A study of cross-validation and bootstrap for accuracy estimation and model selection. Ijcai14, 1137–1145 (1995). [Google Scholar]
- 27.Bryczkowski, S. B., Lopreiato, M. C., Yonclas, P. P., Sacca, J. J. & Mosenthal, A. C. Risk factors for delirium in older trauma patients admitted to the surgical intensive care unit. J. Trauma. Acute Care Surg.77(6), 944–951 (2014). [DOI] [PubMed] [Google Scholar]
- 28.Duceppe, M. A. et al. Modifiable risk factors for delirium in critically ill trauma patients: a multicenter prospective study. J. Intensive Care Med.34(4), 330–336 (2019). [DOI] [PubMed] [Google Scholar]
- 29.Maldonado, J. R. Neuropathogenesis of delirium: review of current etiologic theories and common path- ways. Am. J. Geriatric Psychiatry. 21(12), 1190–1222 (2013). [DOI] [PubMed] [Google Scholar]
- 30.Jones, R. N. et al. Does educational attainment contribute to risk for delirium? A potential role for cognitive reserve. Journals Gerontol. Ser. A: Biol. Sci. Med. Sci.61(12), 1307–1311 (2006). [DOI] [PubMed] [Google Scholar]
- 31.Cizginer, S. et al. The cognitive reserve model in the development of delirium: the successful aging after elective surgery study. J. Geriatr. Psychiatr. Neurol.30(6), 337–345 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Bellelli, G. et al. Is delirium the cognitive harbinger of frailty in older adults? A review about the existing evidence. Front. Med.4, 188 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Angles, E. M. et al. Risk factors for delirium after major trauma. Am. J. Surg.196(6), 864–870 (2008). [DOI] [PubMed] [Google Scholar]
- 34.Yan, W. et al. Lactate dehydrogenase can be used for differential diagnosis to iden- tify patients with severe polytrauma with or without chest injury—A retrospective study. Plos one. 19(8), e0308228 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Ondruschka, B., Schuch, S., Pohlers, D., Franke, H. & Dreßler, J. Acute phase response after fatal traumatic brain injury. Int. J. Legal Med.132(2), 531–539 (2018). [DOI] [PubMed] [Google Scholar]
- 36.Pang, H., Kumar, S., Ely, E. W., Gezalian, M. M. & Lahiri, S. Acute kidney injury-associated delirium: a review of clinical and pathophysiological mechanisms. Crit. Care. 26(1), 258 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Coppolino, G. et al. Delirium accompanies kidney dysfunction in hospitalized elderly patients. J. Gerontol. Geriatr.68, 24–30 (2020). [Google Scholar]
- 38.De Sousa, A. Psychiatric issues in renal failure and dialysis. Indian J. Nephrol.18(2), 47–50 (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.van Eeghen, S. A. et al. Unraveling sex differences in kidney health and CKD: a review of the effect of sex hormones. Clin. J. Am. Soc. Nephrol.20(2), 301–310 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.He, S. et al. Sex differences between serum total bilirubin levels and cognition in patients with schizophrenia. BMC psychiatry. 21(1), 396 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets used and/or analyzed during the current study are not publicly available due to institutional regulations and the protection of patient privacy, but are available from the corresponding author on reasonable request.