The patients in this study were treated at a trauma center in southern Thailand. The subjects were <15 years of age when enrolled from January 1, 2009, to December 31, 2020. However, patients who died before arrival, died within the first 24 hours following injury, or had unavailable cranial computed tomography were excluded. The clinical information and imaging finding of the children were collected with a structured form, including age, sex, signs, symptoms, GCS score, and mechanism of injury. The severity of TBI was classified based on GCS score as follows: mild TBI (GCS scores 13–15), moderate TBI (GCS scores 9–12), and severe TBI (GCS scores 3–8) [2,3]. The co-existing intracranial injuries including epidural hematoma, contusion, intraventricular hemorrhage (IVH), and SAH were evaluated for information from the cranial region. In addition, the degree of midline shift and obliteration of the basal cistern were estimated by two neurosurgeons.

Using the King’s Outcome Scale for Childhood Head Injury (KOSCHI) classification, the functional outcome in the present study was assessed at hospital discharge and 6-month follow-up [15]. In detail, KOSCHI was categorized into favorable (good recovery and moderate disability) and unfavorable outcomes (severe disability, vegetative state, and death) for binary classifier proposes [3].

Clinical characteristics and imaging findings were calculated from descriptive statistics and reported as percentages for categorical data and mean±standard deviation (SD) for continuous variables. Predictors were estimated using binary logistic regression analysis for predictive model development. The candidate predictors were identified at a P<0.10. Thus, the candidate variables were entered into multivariable analysis for the final model based on the backward elimination method. Any P-values <0.05 were considered statistically significant, and the Akaike information criterion (AIC) with clinical relevance was used for model selection [16,17].

Assessment of the predictive model’s performance consisted of two domains, calibration and discrimination. For calibration, the Hosmer-Lemeshow goodness-of-fit (GOF) test and the calibration plot were performed, and a GOF test P-value ≥0.05 indicated good calibration of the model [18]. Discrimination of the model refers to the predictability of a model to differentiate between binary classifiers. The concordance statistic index (c-index) or AUC was measured to indicate discriminatory ability [18,19]. Consequently, internal validation was performed to observe the overfitting problems of the model. Resampling techniques were used for both 1,000 bootstrapping and five cross-validations in the present study. The results of internal validation were reported as the optimism-corrected c-index for both techniques [19-21]. Finally, the predictive model exhibition was utilized following model performances and internal validation as a two-dimensional nomogram. The statistical analysis was performed using R version 4.4.0 software (R Foundation, Vienna, Austria).

Demographic information is listed in Table 1. The mean age of the patients was 141.8 months (SD, 48.0 months); 20.4% of the present cohort were children <5 years of age. More than two-thirds (64.1%) were male, and motorcycle accident was the most common mechanism of injury. In addition, road traffic injury was the major cause of TBI, comprising 84.5% of all cases. Common findings from the physical examination were scalp injury, loss of consciousness, and amnesia, and 9.7% of cases experienced hypotension episodes at the emergency department. Furthermore, early posttraumatic seizure occurred in 5.8% of the present study cohort. Lower extremity injury was the most common associated injury in the present cohort, and isolated TBI was observed in 75.7% of cases. In addition, hypotension was not significantly associated with multiple organ injuries (chi-square test, P=0.74)

Imaging findings after cranial computed tomography are shown in Table 2. All cases of SDH were acute, and hematomas were commonly located in the frontal area, tentorium cerebelli/falx cerebri, and temporal region. The common co-existing injuries were cerebral contusion (37.9%), calvaria skull fracture (33.0%), and SAH (30.1%). Furthermore, obliteration of the basal cistern and midline shift ≥5 mm were observed in 10.7% and 6.8% of cases, respectively. In the present study, surgical treatment was performed in 14.6% of subjects. Decompressive craniectomy was performed in 7.8% of subjects and intracranial monitoring in 1.9% of all cases. One child underwent burr hole due to bilateral chronic SDH. The mean length of hospital stay in the present cohort was 10.4 days (SD, 7.4 days), and the mortality rate at hospital discharge was 8.7%. Based on the KOSCHI at hospital discharge, good recovery, moderate disability, and severe disability were found in 76.7%, 5.8%, and 7.8%, respectively, of patients. When the 6-month follow-up KOSCHI was administered, good recovery, moderate disability, severe disability, and death were found in 77.7%, 4.9%, 8.7%, and 8.7% of subjects, respectively. Therefore, 17.5% of the present study cohort had an unfavorable outcome at hospital discharge, as shown in Table 3.

Clinical variables were analyzed using univariate logistic regression as shown in Table 4. The following 12 variables were candidate predictors: age group, hypotension, amnesia, GCS score, pupillary light reflex, IVH, SAH, thickness of epidural hematoma, lateralization of SDH, obliteration of basal cistern, midline shift, and treatment. Therefore, candidate predictors were analyzed using multivariable analysis. Following the backward elimination procedure, the final model with the lowest AIC consisted of hypotension (odds ratio [OR], 9.4; 95% confidence interval [CI], 2.0–42.9), GCS score of 3–8 (OR, 8.2; 95% CI, 1.7–38.9), a fixed pupil in one eye (OR, 4.8; 95% CI, 2.6–8.8), and fixed pupils in both eyes (OR, 3.5; 95% CI, 1.6–7.1). In addition, a midline shift ≥ 5 mm (OR, 1.1; 95% CI, 0.62–10.73) and co-existing IVH (OR, 6.5; 95% CI, 0.003–26.1) were also included in the final model, as shown in Figure 1.

The domain of the model discrimination was estimated using the Hosmer-Lemeshow GOF test with a P-value of 0.99 and a c-index of 0.971, indicating good calibration. The domain of the model calibration was evaluated based on the calibration plot, which was the logistic calibration (intercept, −3.790; slope, 7.286), as shown in Figure 2. For internal validation, the overfitting performance of the model was evaluated using 1,000 bootstrapping and five cross-validations. The c-index values of bootstrapping and cross-validation were 0.971 and 0.971, respectively.

The c-index of the predictive model was outstanding without overfitting. Therefore, the final predictive model was finalized into a nomogram for predicting the unfavorable outcomes of new patients at the 6-month follow-up, as shown in Figure 3.