Development and internal validation of the Goliat score to predict 48-hour complications after minor/moderate traumatic brain injury in the emergency department: a single-center cohort study

Abstract

Background

Most patients with non-severe traumatic brain injury (TBI) do not develop complications after 24 h of hospital observation. Identifying low-risk patients could enable safe discharge and follow-up—potentially via telemedicine—while avoiding unnecessary cranial CT scans, particularly in older adults.

Methods

We performed a single-centre, consecutive cohort study (tertiary ED, Spain; 1 June 2019–31 December 2020). Adults with minor/moderate TBI (GCS 13–15) were included. The primary outcome was any clinically relevant complication within 48 h (neurological, cardiorespiratory, or death), ascertained by a structured 48-h telephone follow-up plus electronic health record review by assessors independent from clinical care. Candidate predictors (≤ 6 h) comprised baseline factors (age, sex, Charlson comorbidity index, prior anticoagulation), injury-severity signals (polytrauma-code activation, pupil abnormality, fluctuating GCS), physiology/haematology (systolic blood pressure, platelet count), and serum S100 (< 6 h). A multivariable logistic regression with prespecified functional forms (splines/transforms) and two interactions (sex×log[S100]; hypertension×SBP) was fitted. Internal validation used bootstrap/LOOCV with optimism correction and uniform shrinkage. Performance was assessed by AUC, Brier score, calibration (slope/intercept, plots), and decision-curve analysis (DCA). We defined a risk-stratified pathway at ~ 6 h (discharge vs. short observation vs. admission).

Results

Median age was 84 years (IQR 73–88), and 84.6% were ≥ 65 years. Acute complications occurred in 155 patients (29.5%). Independent predictors were age, sex, platelet count, systolic blood pressure, history of hypertension, S100B level at 6 h, anticoagulant treatment, and any high-risk clinical event (fluctuating GCS, moderate severity, pupil alteration, TBI code activation, anticoagulant reversal, polytrauma). Pathological CT findings did not show a significant contribution to the predictive model (LRT p = 0.10) and were not used. The Goliat score achieved an AUC of 0.72 (95% CI 0.68–0.77), good calibration (Hosmer–Lemeshow p = 0.18), and a Brier score of 0.195. Performance was comparable across age (< 65: AUC 0.70; ≥65: AUC 0.73) and sex subgroups. Using the Youden cut-off, 118 (76.1%) were predicted at high risk (sensitivity), while in those ones without complications 220 (59.5%) were classified at low risk (specificity). In total, of the 257 (49.0%) patients predicted at low risk, 37 had experienced complications.

Conclusions

The Goliat score is a novel clinical prediction tool for estimating the risk of acute complications in non-severe TBI that does not require cranial CT findings, potentially reducing unnecessary radiation exposure and ED resource use. By integrating easily obtainable variables and the biomarker S100B it could support safe discharge decisions, although requires prospective external validation before clinical adoption. By linking predicted risk to a risk-stratified clinical pathway, the score prioritizes safe discharge and targeted observation rather than routine CT reduction.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12873-025-01457-9.

Background

Traumatic brain injury (TBI) is a very common reason for consultation in emergency services [1]. It represents a global concern, classified as a “silent epidemic” due to its high incidence and significant impact on public health [2, 3].

TBI is any alteration in brain function or other evidence of brain pathology caused by an external force [4]. These can be classified by severity using a consciousness level assessment scale (Glasgow Coma Scale (GCS)) into mild, moderate, and severe. Mild injuries account for about 80% of cases [5], and even these can lead to debilitating symptoms [6, 7], making it misleading to assume that when the physical force causing the TBI is weak, the consequences will be less severe [8]. Widely used decision rules such as the Canadian CT Head Rule (CCHR) [9] and New Orleans Criteria [10] have been proposed to guide imaging in mild cases, while prognostic models like CRASH [11] focused on mortality and unfavourable outcomes but not specifically on acute complications. It is important to note that, despite the various published data, the incidence of TBI is likely higher, given that underestimation of mild cases is a common trend in most studies [12]. Recent research indicates that between 50 and 60 million new cases of TBI are recorded worldwide annually, with a gradual increase in prevalence (8.4% between 1990 and 2016); 80% of TBI cases occur in developing countries [13].

Furthermore, the mechanisms of TBI vary greatly by age group: in people under 14 and over 65 years old, falls are the leading cause, while in the 16 to 65-year age group, traffic accidents are predominant [10]. There are also notable differences between patients under 75 years compared to those over 75, with the latter showing high rates of re-hospitalization and low rates of rehabilitation after TBI compared to younger patients [14].

Although diagnosis has improved, TBIs remain underdiagnosed. This issue is particularly notable in mild TBI, where there is often insufficient objective evidence of brain injury. Often, TBI is just the tip of the iceberg for underlying cardiac, respiratory, and endocrine pathologies. In recent years, significant progress has been made in the research of blood biomarkers, which has improved the detection and clinical description of patients with possible brain injuries [15].

Advances in our understanding of biomarkers associated with TBI could facilitate a more accurate description of its severity, improve our understanding of stratification both within the injury process and the recovery process, and contribute to the development of quantifiable measures that reflect brain injury reversal and recovery after trauma. Biological markers indicating brain damage commonly consist of proteins released by neuronal cells or that experience an increase in their levels, indicating a pathological change. Among these brain markers, S100 and NSE are some of the most researched; however, glial fibrillary acidic protein (GFAP) and ubiquitin carboxy-terminal hydrolase L1 (UCH-L1) are gaining increasing relevance [16].

Biomarkers represent a promising tool for detecting and assessing the severity of TBI. However, due to the complexity and heterogeneity of TBI, biomarkers for TBI are under continuous study and development. More research is needed to validate and define existing biomarkers, as well as to identify new biomarkers with greater sensitivity and specificity. Advances in this field have the potential to significantly improve medical care and outcomes for patients with TBI. Our research team found that TBI patients with complications (traumatic, delirium, heart and respiratory) presented significantly higher S100 values (p < 0.05) [17]. S100B has been incorporated into Scandinavian guidelines to reduce unnecessary CT scans in mild TBI [18] and GFAP/UCH-L1 have received FDA approval for triage purposes [19], although their role in predicting acute post-discharge complications remains under investigation.

Most of the patients with a TBI does not experience clinical complications after a 24 h hospitalization in observation. Computed tomography (CT) has become the imaging modality of choice for the initial evaluation of TBI, due to its speed, availability, and diagnostic accuracy. However, its widespread use has raised concerns, particularly in cases of mild TBI, where only 7–10% of patients present with significant intracranial findings [20] and fewer than 1% require neurosurgical intervention [21]. This is compounded by the risks associated with radiation exposure [22], increased healthcare costs, and the overburdening of emergency departments. Considering that a normal cranial CT does not preclude the possibility of subsequent complications, we propose this study to assess whether it is possible to develop a predictive score for complications in mild TBI that does not rely on cranial CT imaging. The Goliat score is specifically designed for patients with non-severe TBI, defined as a Glasgow Coma Scale (GCS) score of 9–15. The score is intended for use after the initial ED, to support decision-making regarding discharge versus extended observation. Its primary aim is to predict complications that may develop within the first 48 h after injury, particularly those occurring beyond the initial 24-hour observation period, when patients are often discharged. This makes it particularly relevant for non-severe TBI patients, where the decision to discharge is clinically challenging. Our group has previously reported the epidemiology and biomarker profiles of this prospective cohort [17], providing the foundation for the current secondary analysis focused on model development and internal validation.

This study is a secondary analysis of a previously published prospective cohort conducted at the Emergency Department of Arnau de Vilanova University Hospital, Lleida, Spain [17]. The original cohort included adult patients with TBI of any severity. For the present analysis, we selected only those with non-severe TBI to develop and internally validate the Goliat score, a prediction model for acute complications. This specific clinical focus and model design address a decision point not covered by existing tools, such as the Canadian CT Head Rule, New Orleans Criteria, or CRASH model, which are oriented toward imaging decisions or long-term prognosis.

Methods

This is a secondary analysis of data from a prospective cohort study previously published by our group [1, 17]. The original cohort was designed to investigate prognostic factors in patients with TBI of any severity presenting to the Emergency Department of Arnau de Vilanova University Hospital between 1 June 2019 and 31 December 2020. For the current study, we restricted the analysis to patients with non-severe TBI (Glasgow Coma Scale 9–15) without immediate indications for hospital admission, with the aim of developing and internally validating the Goliat score. The study follows the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines. Sample size was calculated and explained in previous articles(1).

A patient flow diagram (Supplementary File1) shows the number of patients screened, included, excluded (with reasons), and analysed in the final model.

Eligibility criteria applied

This study included only adult patients (age ≥ 18 years) with a TBI showing a GCS in 9–15, thus excluding the severe TBI .

Patients with a cranial or a facial contusion or who had suffered a TBI in the last 6 months, together with those who did not give or were unable to give their signed informed consent for inclusion were excluded.

Variables

The available patients’ characteristics have been already described in previous publications [1, 14]. In summary, data on socio-demographical and clinical variables at baseline as well as CT scan and blood test results including the assessment of 4 promising biomarkers (S100, NSE, UCHL1 and GFAP) were collected.

The following operational definitions were pre-specified:

• Polytrauma: arrival following activation of the regional time-dependent polytrauma code (Catalonia), as recorded in the EHR trauma-code field (high-energy mechanism and/or multi-regional injury).

• Fluctuating GCS: any ≥ 2-point change between the initial (prehospital/first ED) and ED-arrival scores or a ≥ 2-point fluctuation during the first 6 h of ED observation, not attributable to sedation.

• Pupil abnormality: non-isocoria (anisocoria > 1 mm) and/or sluggish/non-reactive pupil on standardized neuro checks.

• Baseline factors: Charlson Comorbidity Index (continuous), prior anticoagulant treatment (binary), platelet count (continuous), systolic blood pressure (continuous), age (years), sex.

Functional forms followed linearity-in-the-logit checks: age modeled with a knot at 84 years; S100 and platelets entered log-transformed, with S100 modeled using a spline with a knot at log(0.08). Two interactions were retained a priori and by LRT: sex × log(S100) and hypertension history × SBP.

Biomarkers

Blood samples for biomarker analysis (S100B, neuron-specific enolase [NSE], glial fibrillary acidic protein [GFAP], and ubiquitin carboxy-terminal hydrolase L1 [UCH-L1]) were collected at 6 h post-trauma, together with routine laboratory tests (complete blood count, coagulation). In clinically stable patients, these 6-hour samples were available before the hospitalisation decision. A second biomarker sample was obtained at 12 h. Serum and plasma were processed following a standardised protocol and stored at − 80 °C in the institutional biobank until analysis. S100B and NSE were quantified using a sandwich-type electrochemiluminescence immunoassay (ECLIA, Roche Diagnostics, Mannheim, Germany) at the hospital’s clinical laboratory. GFAP and UCH-L1 were measured using a commercial sandwich enzyme-linked immunosorbent assay (ELISA, DuoSet, R&D Systems) on automated Triturus equipment (Grifols^®).

CT results

Pathological findings in CT scan included atrophy, fracture (occipital and cranial rock excluded), haematoma (epicranial or soft tissues), chronic cerebral infarct, chronic injuries or neoplasia). According to the institutional protocol for non-severe TBI management at our hospital, the first cranial CT scan is performed approximately 6 h after injury in clinically stable patients without neurological deterioration. This timing is intended to reduce unnecessary acute imaging while allowing for the detection of delayed intracranial complications. For patients presenting with abnormal initial neurological status or showing rapid deterioration, CT imaging is performed immediately after ED assessment. All predictor variables used for model development were collected during the initial ED evaluation and were available for all patients within the first two hours after presentation. In our institution, ED discharges routinely include strict at-home follow-up at ~ 24 h; in practice, events that influence disposition (safe discharge vs. observation/admission) typically declare within 1–2 days. We therefore pre-specify an acute (initial 48 h) window to guide ED/short-stay decisions while avoiding harms from unnecessarily prolonged observation in frail older adults.

Complications: main outcome definition

The primary outcome was the occurrence of acute complications within 48 h of ED arrival. Acute complications were defined as any of the following [17]:

• Neurological: post-traumatic amnesia, new headache or dizziness requiring medical intervention, acute change in mental status (as per the 4AT scale), seizures.

• Cardiac: acute heart failure, angina, or myocardial infarction.

• Respiratory: acute respiratory failure, pulmonary embolism.

• Death from any cause.

At 7 days after the index ED visit, outcomes were ascertained using two sources: (1) a structured telephone follow-up conducted by a trained study fellow using a predefined checklist of complications and return-visit triggers; and (2) a systematic abstraction of the electronic health record (EHR) to capture ED revisits, hospital admissions, imaging results, ICU transfers, and in-hospital events within the 7 days window. Outcome assessors were independent of clinical care and, given the retrospective internal-validation design, were unaware of the study hypotheses and model predictions at the time of outcome collection. Disagreements were resolved by consensus.

Predictor assessment and modelling strategy

Candidate predictors were selected based on clinical plausibility and previous literature. Continuous variables were initially examined for their relationship with the logit of the outcome using scatterplots and locally weighted smoothing. Non-linear relationships were modelled using natural cubic splines, and inflection points were identified for age and S100B levels (in logarithmic scale). Interaction terms were tested a priori for sex × S100B and hypertension × systolic blood pressure (SBP). Predictors with a significant contribution according to the likelihood ratio test (p < 0.05) were retained in the final multivariable logistic regression model. The Goliat score is intended for application after the initial ED assessment, with the purpose of predicting complications within the first 48 h after injury, especially those arising after the typical 24-hour observation period.

Intended use and clinical pathway (operationalization at ~ 6 h)

“At approximately 6 hours after index TBI, the Goliat score informs disposition over an acute (initial 48 h) decision horizon:

• Low risk → Discharge without CT, with standard 24-h at-home follow-up and return precautions, given a very low probability of acute complications within 48 h.

• Intermediate risk → Obtain CT. If CT is normal, consider early discharge with hospital-at-home support or close outpatient follow-up; if abnormal but not severe, tailor observation.

• High risk → Obtain CT and avoid discharge: ≥24 h observation in ED/short-stay; admit if CT is abnormal or clinical course warrants.

• This pathway clarifies that the model aims at safe disposition, not routine CT reduction.”

Internal validation and performance metrics

Internal validation was performed using leave-one-out cross-validation (LOOCV) to estimate unbiased prediction error. Model performance was evaluated in terms of:

• Discrimination: area under the receiver operating characteristic curve (AUC) with 95% confidence interval.

• Calibration: Hosmer–Lemeshow goodness-of-fit test, calibration plot.

• Overall performance: Brier score.

• Diagnostic accuracy: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (LR+), and negative likelihood ratio (LR–) for predefined cut-offs (Youden Index, F1 score, and high-sensitivity threshold).

Time origin and predictor windows

“The time origin (T0) is ED presentation. The model is intended for use at ~ 6 hours after T0 (the decision time). Predictors are restricted to data available by ~ 6 h:

• Baseline: age, sex, Charlson index, prior anticoagulant treatment.

• Early ED (T0–T6 h): polytrauma-code activation, pupil findings, systolic blood pressure, platelet count, serum S100 (sampled < 6 h), and fluctuating GCS (any ≥ 2-point change from prehospital/first ED score to ED-arrival or during T0–T6 h, excluding sedation).

• The outcome window is the acute (initial 48 h) period after injury. For decision support at ~ 6 h, primary analyses consider events occurring after the decision time (≈ 6–48 h). No post-6 h information is used as a predictor.”

Assumed causal pathway and leakage control

“We posit a pathway whereby host factors (age, comorbidity, anticoagulation, platelets) and injury-severity signals (polytrauma context, pupil findings), together with early physiology/biomarker (SBP, S100 < 6 h), influence the risk of 48 h complications. To avoid information leakage, we excluded variables measured after ~ 6 h and performed a sensitivity analysis excluding patients with events before ~ 6 h; results were unchanged.

Statistical analysis

The quartiles (or the median and the 25th and 75th percentiles) were used to summarize quantitative variables, whereas the absolute and relative percentages were applied to qualitative variables. Patients with clinical complications at 48 h were compared with those without them using the Mann–Whitney U-test for the quantitative variables and the Pearson’s chi-squared test for the qualitative variables. The unadjusted associated odds ratio (OR) was also estimated using the midp method.

A multivariable logistic regression model for the development of clinical complications in the first 48 h was adjusted. The proposed model included inflection points for the predictors with non-linear associations with the logit of complications assessed by natural splines. The final model included significant predictors according to the likelihood ratio test. The unbiased prediction error was estimated from leave-one-out cross-validation. The model calibration and discrimination abilities were assessed by the Hosmer-Lemeshow test and the area under the ROC curve and illustrated graphically. The overall model performance was measured by the Brier score. These measures of model calibration, discrimination and overall performance were also assessed in a subgroup analysis by age group ( > = 65 vs. youger) and sex.

The selection of a cut-off point for patients in high vs. non-high risk of clinical complications used two methods: the Youden Index (to maximize the sum of sensitivity and specificity) and the F1 score method (to maximize the harmonic mean between sensitivity and the proportion of true positives among positive results). An additional cut-off point with high sensitivity is proposed to identify those patients with “very low risk”.

The statistical analysis was performed in R [23] and a significance level of 0.05 was applied.

Ethics statement

All clinical information was obtained during the emergency department visit for the TBI episode. A patient follow-up to determine patient survival was obtained using SAP software. The information was collected by the research team. Patient results and information were managed according to the recommendations of our ethics committee.

The study was approved by the CEIC of the Hospital Universitario Arnau de Vilanova de Lleida (CEIC-1952). All patients included in the study, or their legal representatives, consented to participate in the study prior to initiating the study in the emergency department. The processing, communication, and transfer of the personal data of all participating subjects complied with the provisions of Spanish Organic Law 3/2018, on the Protection of Personal Data and Guarantee of Digital Rights (LOPD-GDD 3/2018) and Regulation 2016/679 (EU) of the European Parliament and of the Council of Europe of 27 April 2016.

Results

The Goliat score is applied approximately 6 h after ED arrival for minor/moderate TBI to estimate the individual probability of any clinically relevant complication within 48 h (the acute/initial 48 h window) and to guide disposition (safe discharge, short observation, or admission).

Patient charateristics

A total of 525 patients with mild TBI were included in the analysis. Median age was 84.0 years [IQR 73.0–88.0], 84.6% were 65 years or older, and 53.3% were men. At 48 h, 155 patients (29.5%) experienced at least one acute complication. Table 1 summarises the distribution of acute complications according to clinical and risk factors with their associated p-values. Odds ratios (OR) with 95% confidence intervals (CI) and are provided consistently for each variable.

Table 1.

Distribution of acute complications by clinical risk factors and calculation of unadjusted odds ratios

	Overall	No	Yes	p-value	OR
TBI severity
Mild	505 (96.2%)	360 (71.3%)	145 (28.7%)	0,072
Moderate	20 (3.81%)	10 (50.0%)	10 (50.0%)		2.48 [0.99;6.23]
Unfrequent high risk events
Fluctuating CGS
Yes	6 (1.14%)	2 (33.3%)	4 (66.7%)		4.69 [0.85;38.3]
No	519 (98.9%)	368 (70.9%)	151 (29.1%)	0,066
Pupil alteration
Yes	2 (0.38%)	1 (50.0%)	1 (50.0%)		2.39 [0.06;93.6]
No	523 (99.6%)	369 (70.6%)	154 (29.4%)	0,504
TBI code activation
Yes	36 (6.86%)	19 (52.8%)	17 (47.2%)		2.27 [1.13;4.53]
No	489 (93.1%)	351 (71.8%)	138 (28.2%)	0,026
Reversed anticoagulation
Yes	14 (2.67%)	8 (57.1%)	6 (42.9%)		1.83 [0.58;5.46]
No	511 (97.3%)	362 (70.8%)	149 (29.2%)	0,372
Polytraumatism
Yes	30 (5.71%)	18 (60.0%)	12 (40.0%)		1.65 [0.75;3.49]
No	495 (94.3%)	352 (71.1%)	143 (28.9%)	0,276
Any unfrequent high risk event
Yes	55 (10.5%)	29 (52.7%)	26 (47.3%)		2.37 [1.33;4.19]
No	470 (89.5%)	341 (72.6%)	129 (27.4%)	0,004
Clinical Variables
Age (mean) [IQR]	84.0 [73.0;88.0]	83.0 [72.2;87.8]	84.0 [75.0;89.0]	0,059	1.01 [1.00;1.02]
Sex: Woman	245 (46.7%)	169 (69.0%)	76 (31.0%)	0,544
Man	280 (53.3%)	201 (71.8%)	79 (28.2%)		0.87 [0.60;1.27]
Charlson Comorbidity index	2.00 [1.00;3.00]	2.00 [1.00;3.00]	2.00 [1.00;3.00]	0,966	1.01 [0.90;1.13]
0	75 (14.3%)	52 (69.3%)	23 (30.7%)	0,908
1_2	256 (48.8%)	180 (70.3%)	76 (29.7%)		0.95 [0.55;1.69]
3_4	146 (27.8%)	102 (69.9%)	44 (30.1%)		0.97 [0.53;1.81]
5+	48 (9.14%)	36 (75.0%)	12 (25.0%)		0.76 [0.32;1.71]
Dementia
Yes	99 (18.9%)	69 (69.7%)	30 (30.3%)		1.05 [0.64;1.68]
No	426 (81.1%)	301 (70.7%)	125 (29.3%)	0,947
Type II Diabetes*
Yes	169 (32.2%)	117 (69.2%)	52 (30.8%)		1.09 [0.73;1.62]
No	356 (67.8%)	253 (71.1%)	103 (28.9%)	0,742
Hypertension*
Yes	418 (79.6%)	298 (71.3%)	120 (28.7%)		0.83 [0.53;1.32]
No	107 (20.4%)	72 (67.3%)	35 (32.7%)	0,49
Dyslipidemia*.
Yes	212 (40.4%)	149 (70.3%)	63 (29.7%)		1.02 [0.69;1.49]
No	313 (59.6%)	221 (70.6%)	92 (29.4%)	1
Smoker
Yes	18 (3.43%)	9 (50.0%)	9 (50.0%)		2.47 [0.93;6.53]
No	507 (96.6%)	361 (71.2%)	146 (28.8%)	0,094
Alcohol Consumer
Yes	16 (3.05%)	11 (68.8%)	5 (31.2%)		1.11 [0.34;3.14]
No	509 (97.0%)	359 (70.5%)	150 (29.5%)	1
Drug Abuse
Yes	11 (2.10%)	9 (81.8%)	2 (18.2%)		0.56 [0.08;2.23]
No	514 (97.9%)	361 (70.2%)	153 (29.8%)	0,52
Anticoagulant treatment
Yes	214 (40.8%)	157 (73.4%)	57 (26.6%)		0.79 [0.53;1.16]
No	311 (59.2%)	213 (68.5%)	98 (31.5%)	0,269
Antiplatelet treatment
Yes	189 (36.0%)	127 (67.2%)	62 (32.8%)		1.28 [0.86;1.88]
No	336 (64.0%)	243 (72.3%)	93 (27.7%)	0,256
TBI VARIABLES
Place of TBI: At home	326 (62.1%)	227 (69.6%)	99 (30.4%)	0,202
Outdoors	142 (27.0%)	107 (75.4%)	35 (24.6%)		0.75 [0.48;1.17]
Institutionalized	57 (10.9%)	36 (63.2%)	21 (36.8%)		1.34 [0.73;2.40]
CGS level: 15	490 (93.3%)	349 (71.2%)	141 (28.8%)	0,319
14	21 (4.00%)	13 (61.9%)	8 (38.1%)		1.53 [0.59;3.76]
10–13	14 (2.67%)	8 (57.1%)	6 (42.9%)		1.87 [0.59;5.56]
CGS	15.0 [15.0;15.0]	15.0 [15.0;15.0]	15.0 [15.0;15.0]	0,157	0.81 [0.59;1.11]
SBP	144 [129;165]	145 [130;167]	142 [125;162]	0,188	1.00 [0.99;1.00]
DBP	77.0 [67.0;87.0]	77.0 [68.0;87.0]	77.0 [66.5;86.0]	0,467	0.99 [0.98;1.01]
HR	74.0 [65.0;86.0]	74.0 [66.0;86.0]	73.0 [65.0;83.5]	0,288	0.99 [0.98;1.01]
Temperature	36.0 [35.5;36.3]	36.0 [35.5;36.3]	36.0 [35.5;36.3]	0,305	0.86 [0.63;1.17]
Hemoglobin	12.9 [11.7;14.1]	12.9 [11.7;14.2]	12.8 [11.6;14.1]	0,32	0.95 [0.86;1.05]
Platelets (100 units)	2.03 [1.59;2.45]	2.06 [1.66;2.48]	1.92 [1.48;2.32]	0,01	0.75 [0.58;0.97]
INR	1.13 [1.03;1.69]	1.13 [1.04;1.79]	1.13 [1.02;1.56]	0,299	0.89 [0.72;1.10]
S100	0.13 [0.08;0.25]	0.12 [0.07;0.23]	0.16 [0.09;0.30]	0,002	1.27 [0.85;1.92]
UCHL	0.00 [0.00;66.7]	0.00 [0.00;64.7]	0.00 [0.00;72.9]	0,483	1.00 [1.00;1.00]
NSE	14.8 [11.1;22.4]	14.8 [11.2;22.0]	14.6 [11.1;25.0]	0,892	1.00 [1.00;1.01]
GFAP	0.00 [0.00;0.30]	0.00 [0.00;0.30]	0.00 [0.00;0.00]	0,274	1.01 [0.94;1.08]
Pathological CT scan
Yes	67 (12.8%)	39 (58.2%)	28 (41.8%)		1.87 [1.09;3.17]
No	458 (87.2%)	331 (72.3%)	127 (27.7%)	0,027

*In clinical history or receiving treatment for. GCS: Glasgow Coma Score; SBP: Systolic Blood Pressure; DBP: Dyastolic Blood Pressure; HR: Heart Rate

This table summarizes the presence of clinical complications at 48 h according to selected variables, including fluctuating GCS, TBI severity, pupil alteration, TBI code activation, need for anticoagulant reversal, and polytraumatism. Unadjusted odds ratios (OR) and p-values are provided for each comparison

Univariable analysis

When we examined the data in univariable analyses, several factors emerged as significant. The higher unadjusted estimated odds ratio of 48 h clinical complications was observed for clinically meaningful characteristics such as fluctuating GCS, moderate TBI severity, pupil alteration, TBI code activation, anticoagulant reversion or polytraumatism. Since all of these characteristics were observed in a low number of patients each, they were summarized in a single variable identifying the occurrence of any of them. Thus, complications at the first 48 h were experienced by a 47.3% of patients with any of these evenss compared with the 27.4% of patients without them, with a significant unadjusted OR of 2.37 and 95% CI in [1.33, 4.19]. Complications at the first 48 h were experienced by a 41.8% of patients with pathological CT findings vs. 27.7% without them (p-value = 0.027), with a significant unadjusted OR of 1.87 and 95% CI in [1.09, 3.17], Other factors significantly associated with acute complications in univariable analysis included lower platelet count, with p-value = 0.010 and OR 0.75 per 100 units change, 95% CI 0.58–0.97, as well as higher S100B levels (in log scale) at 6 h post-trauma, with, p = 0.045 and OR 1.42, 95% CI 1.01–2.01.

Complications at the first 48 h were not significantly associated with sex (31.0% in women vs. 28.2% in men, p-value = 0.54), anticoagulant treatment (26.6% with vs. 31.5% without it, p-value = 0.27), or antiplatelet treatment (32.8% with vs. 27.7% without it (p-value = 0.26). Neither were Age, Charlson comorbidity index, or hypertension. No significant associations were observed for NSE, GFAP, or UCH-L1 at 6 h, nor for changes in biomarker levels between 6 and 12 h.

Multivariable model

We assessed the linearity of the relationship of the potential predictors with the logit of complications at 48 h. Changes in trend or inflection points were detected for S100 levels in logarithmic scale and age at values log(0.08) and 84, respectively. Interactions were also identified (Fig. 1), one between sex and S100 levels and another one between clinical history of hypertension and systolic blood pressure (SBP). Finally, the platelet levels exhibited a linear relationship in logarithmic scale.

Fig. 1.

Linearity and interactions between predictors and the logit of 48-hour complications. This figure illustrates the relationship between selected continuous predictors (S100, age, platelet levels) and the logit of acute complications, including detected inflection points and identified interactions between sex and S100, and between history of hypertension and systolic blood pressure

The proposed multivariable logistic regression model included these estimated inflection points and interactions as well as platelet levels in logarithmic scale, previous anticoagulant treatment and the occurrence of either fluctuating GCS, moderate TBI severity, pupil alteration, TBI code activation, anticoagulant reversion or polytraumatism. All predictors showed significant contribution to the multivariable model according to the LRT. Pathological CT findings did not show a significant contribution to the multivariate model predicting complications, with a LRT p-value of 0.100. Neither did other risk factors such as smoking or antiplatelet therapy.

Final model structure and key predictors

The final multivariable logistic regression model included age, sex, TBI severity (minor vs. moderate; fluctuating GCS; pupil alteration), TBI code activation / polytrauma, previous anticoagulant treatment (and reversal), history of hypertension, systolic blood pressure (SBP), platelet count, Charlson Comorbidity Index, and serum S100 measured < 6 h after injury. Based on pre-specified functional assessments, S100 and platelet count entered the model on the logarithmic scale. Age showed a non-linear relationship with the logit of 48-hour complications with an inflection at 84 years, and S100 exhibited a change in slope around log(S100) = 0.08 (Fig. 1). We therefore modeled age using a two-piece linear spline with a knot at 84, and S100 using a two-piece linear spline on the log scale with a knot at log(0.08).

Two clinically plausible interactions improved model fit and were retained: sex × log(S100) and history of hypertension × SBP (Fig. 1). The sex × log(S100) interaction indicates a steeper risk gradient with increasing S100 in men than in women, whereas the hypertension × SBP interaction indicates that lower SBP is more strongly associated with complications among patients with hypertension history than among those without. All retained predictors (including interaction terms and spline components) contributed significantly to the model by likelihood-ratio testing.

In terms of effect direction, higher age (particularly ≥ 84), moderate TBI vs. minor, fluctuating GCS, pupil alteration, TBI code activation/polytrauma, prior anticoagulant treatment, higher Charlson index, lower platelet count (log scale), and higher S100 (log scale) were each associated with increased 48-hour complication risk, while higher SBP was protective—especially in those without prior hypertension.

Model performance

The proposed model showed no significant lack of Calibration, with a Hosmer–Lemeshow p-value of 0.18 and a calibration plot (Fig. 2a). The discrimination ability measured by the area under the ROC curve reached 0.72 with 95%CI in [0.68, 0.77], as shown in Fig. 2b. The overall performance of the predictive model as measured by the Brier score was 0.195.

Fig. 2.

a. Calibration plot of the Goliat score using leave-one-out cross-validation. The figure shows the observed versus predicted probability of 48-hour complications. The Hosmer–Lemeshow test indicates no significant lack of fit (p = 0.18). b. Receiver Operating Characteristic (ROC) curve of the Goliat score. Displays the discrimination ability of the model with an area under the curve (AUC) of 0.72 [95% CI: 0.68–0.77]

Both the Youden and the F1 score index provided a unique cut-off point with 59.5% specificity and 76.1% sensitivity (Table 2).

Table 2.

Performance metrics for different risk thresholds of the goliat score

	Threshold	Sensitivity	Specificity	Youden index	F1 score	PPV	NPV	PLR	NLR	Accuracy
Youden index maximization	0,2675	0,761	0,5946	0,3559	0,5579	0,4403	0,856	1,8778	0,4015	0,6438
F1 score maximization	0,2675	0,761	0,5946	0,3559	0,5579	0,4403	0,856	1,8778	0,4015	0,6438
Higher sensitivity	0,1631	0,936	0,2892	0,2247	0,5151	0,3554	0,9145	1,3161	0,2231	0,48

PPV and NPV refer to the positive and negative predictive values assuming that the prevalence of complications at 48 h in the population of study is the one observed, 29.5%

PLR and NLR are the positive and negative likelihood ratios

Diagnostic accuracy indicators (sensitivity, specificity, positive and negative predictive values, likelihood ratios, and accuracy) are presented for three different thresholds: one maximizing the Youden index, another maximizing the F1 score, and a third prioritizing high sensitivity for very low-risk identification

The performance of the proposed risk classification, or Goliat score, in relation with complications is described in Table 3. The group of patients with complications were classified 118 (76.1%) in high risk. The group of patients without complications were classified 220 (59.5%) in low risk. The coverage of the estimated high-risk class in each type of clinical complication goes from a minimum 54.5% of the patients suffering from cardiac complications to the maximum 84.2% of the patients exhibiting an acute change in mental state. Looking at the distribution of patients classified at high risk by the number of complications at 48 h, they are 76.8% of the patients with 2 or more complications and 75.0% of the patients with one complication. The 3 patients dying during hospitalization were all classified at high estimated risk by the model. In the study population, the distribution of patients according to pathological findings, Goliat predicted risk, and 48-hour complications is shown in Table 4. Among patients without pathological findings, those in the very low–risk group (0–0.163) represented 105 cases (25.9% without and 5.81% with complications), the low–risk group (0.163–0.267) comprised 126 cases (27.6% without and 15.5% with complications), and the high–risk group (0.267–1) included 227 cases (35.9% without and 60.6% with complications). Among patients with pathological findings, the very low–risk group included 12 cases (2.97% without and 0.65% with complications), the low–risk group comprised 14 cases (2.97% without and 1.94% with complications), and the high–risk group contained 41 cases (4.59% without and 15.5% with complications). Notably, in none of the patients with pathological CT findings was there a change in clinical management: discharge destination was home or admission for causes unrelated to the head injury, suggesting that pathological CT findings do not necessarily lead to changes in care and could, in some cases, be safely avoided.

Table 3.

Distribution of complications by goliat score risk group

	Very low risk (0,0.163]	Low risk (0.163,0.267]	High risk (0.267,1]	Total
Any 48 h complication: No	107 (28.9%)	113 (30.5%)	150 (40.5%)	370 (71.5%)
Yes	10 ( 6.5%)	27 (17.4%)	118 (76.1%)	155 (29.5%)
Drowsiness	1 (2.78%)	6 (16.7%)	29 (80.6%)	36 (6.86%)
Post-traumatic amnesia	3 (17.6%)	1 (5.88%)	13 (76.5%)	17 (3.24%)
Headache and dizziness	6 (5.77%)	17 (16.3%)	81 (77.9%)	104 (19.8%)
Acute change in the mental state	0 (0.00%)	6 (15.8%)	32 (84.2%)	38 (7.24%)
Easily distracted or difficulty following conversations	0 (0.00%)	6 (22.2%)	21 (77.8%)	27 (5.14%)
Expression of incoherent ideas or conversations	0 (0.00%)	4 (21.1%)	15 (78.9%)	19 (3.62%)
Altered states of consciousness	0 (0.00%)	5 (22.7%)	17 (77.3%)	22 (4.19%)
Respiratory complications	2 (18.2%)	3 (27.3%)	6 (54.5%)	11 (2.10%)
Cardiac complications	1 (14.3%)	2 (28.6%)	4 (57.1%)	7 (1.33%)
Death	0 ( 0.0%)	0 ( 0.0%)	3 ( 100%)	3 (0.57%)
Pathological TC scan	12 (17.9%)	14 (20.9%)	41 (61.2%)	67 (12.8%)

The table displays the percentage of patients with and without clinical complications within 48 h according to three risk categories: very low, low, and high. It also details the distribution of specific complications—such as drowsiness, post-traumatic amnesia, headache, and dizziness—across risk groups

Table 4.

Distribution of patients by pathological CT findings, goliat predicted risk, and occurrence of 48-hour complications

Pathological findings	Goliat predicted risk	48 h complications	No 48 h complication	Total
No	Very low risk (0,0.163]	9 (5.81%)	96 (25.9%)	105 (20.0%)
No	Low risk (0.163,0.267]	24 (15.5%)	102 (27.6%)	126 (24.0%)
No	High risk (0.267,1]	94 (60.6%)	133 (35.9%)	227 (43.2%)
Yes	Very low risk (0,0.163]	1 (0.65%)	11 (2.97%)	12 (2.29%)
Yes	Low risk (0.163,0.267]	3 (1.94%)	11 (2.97%)	14 (2.67%)
Yes	High risk (0.267,1]	24 (15.5%)	17 (4.59%)	41 (7.81%)

Percentages refer to the proportion of each subgroup within the total sample. No pathological CT finding led to a change in clinical management; all patients were discharged home or admitted for causes unrelated to the head injury

When stratified by age, the model maintained good performance in both groups. In patients aged < 65 years (n = 81), calibration was acceptable (Hosmer–Lemeshow p = 0.059) with a Brier score of 0.172 and an AUC of 0.70 (95% CI 0.56–0.84). In patients aged ≥ 65 years (n = 444), calibration remained good (p = 0.247), the Brier score was 0.183, and discrimination was slightly higher, with an AUC of 0.73 (95% CI 0.68–0.78).

By sex, performance was consistent. In women (n = 245), calibration was acceptable (p = 0.150) with a Brier score of 0.188 and an AUC of 0.71 (95% CI 0.65–0.78). In men (n = 280), calibration was good (p = 0.363), with the lowest Brier score among subgroups (0.176) and the highest discrimination (AUC 0.73, 95% CI 0.66–0.79). Figure 3 illustrates the decision curve analysis showing the added value of GOLIAT in comparison with either treating everyone, treating no one, treating depending on a pathological CT scan, treating depending on S100B value.

Fig. 3.

Decision curve analysis showing the added value of GOLIAT in comparison with either treating everyone, treating no one, treating depending on a pathological CT scan, treating depending on S100B value

Across all subgroups, there was no significant lack of calibration, and the direction of predictor effects remained consistent, suggesting that the Goliat score maintains robustness regardless of age group or sex.

Discussion

This study fills a critical gap in the management of non-severe TBI by proposing a clinical score that predicts acute complications, rather than relying solely on imaging findings. We have developed the Goliat score, that can identify the 76.1% of patients without a severe TBI but who will develop clinical complications in the first 48 h after a TBI. The score classification at non-high risk shows a 59.5% of specificity for those patients who will not suffer from clinical complications, who represent a 71.5% of the non-severe TBI patients. We have observed that having pathological CT findings in the early hours does not significantly contribute to the proposed predictive model for acute complications. Of the total, there were four patients (2,56%) with pathological CT findings who were not captured by the model (classified as low or very low risk) and they had clinical complications. Despite having pathological CT findings, these patients were discharged home or admitted for other reasons, as the CT finding did not influence prognosis, diagnostic management or the severity of complications. If the CT had not been performed, clinical practice would not have changed, which we consider an important observation. Considering the impact of radiation exposure and the need for observation in the emergency department, our proposed score could help avoid performing cranial CT in a subset of patients for whom the scan offers limited clinical value. Clinically, the score’s purpose is disposition rather than routine CT reduction. At ~ 6 h, low-risk patients can be discharged safely (with 24-h home checks), intermediate-risk patients undergo CT and may be discharged early if CT is normal with structured follow-up (including hospital-at-home), and high-risk patients receive CT and remain under observation ≥ 24 h or are admitted. This risk-stratified clinical pathway is designed to reduce unnecessary ED observation in older adults, while concentrating monitoring where it is most beneficial.

To make the score easy to remember, we have thought of the biblical character Goliath. Goliath died of a severe head trauma at the hands of little David. We believe that in this way we try to humanize a score that we think could have a long way to go given the increase of patients with TBI that we attend every day in the Emergency Departments.

Nowadays, there is no clinical score predictive of complications in patients with non-severe TBI for adults. In 2021, in Australia the emergency services developed PREDICT [24] score for the management of head injuries. However, it was just for paediatric population. There are some clinical guidelines and management recommendations, but not a prognostic score. In 2022, a study from Sweden [25] shown the most important predictors of complications in elderly TBI Patients: the need for neurosurgical intervention, cardiac risk, and measures of injury severity. There are also new resources based on Artificial Intelligence using machine learning to manage TBI patients after Emergency Department discharge [26]. However, no score has been developed to predict complications in non-severe TBI patients. In fact, efforts are already underway in Spain to implement the use of biomarkers as an alternative to cranial CT; however, these approaches are not yet fully developed [27]. To our knowledge, this is the first predictive score specifically developed for adult patients with non-severe TBI to estimate the risk of clinical complications within the first 48 h.

That is why we believe it is very important to have one, given the increase in the incidence of these cases and their high comorbidity. In fact, in the latest guidelines for the management of TBI [28], it is already proposed that these patients can be observed in separate areas of the emergency room. Several existing clinical decision tools address head injury management, but they differ in scope from the Goliat score. The Canadian CT Head Rule (CCHR) and the New Orleans Criteria [29] are designed to identify patients who require initial CT imaging, not to predict post-assessment complications. The CRASH model [30] focuses on predicting mortality and unfavourable outcomes in moderate-to-severe TBI, but does not target short-term complication risk in non-severe cases. Biomarker-based strategies, such as the Scandinavian guidelines incorporating S100B to reduce unnecessary CTs, have demonstrated value in selected populations, yet are not designed to predict delayed complications after an initial normal CT.

The Goliat score differs in that it specifically addresses non-severe TBI in adults without immediate indications for admission, aiming to predict complications within the first 48 h, including those arising after the typical 24-hour observation period. By integrating clinical variables, physiological parameters, and biomarker data, the score provides a multimodal risk estimate rather than relying on a single domain. This approach may be particularly relevant for older adults, who have higher complication risk and atypical presentations. While the Goliat score has shown promising internal validation, prospective external validation will be essential to confirm its performance and generalisability before clinical adoption. One of the novel aspects of the Goliat score is that it does not include cranial CT findings among its predictors. In our analysis, having a CT with pathological results was not a significant independent predictor of acute complications, consistent with previous observations that complications can develop even when CT is normal. This omission could have important clinical implications, potentially reducing unnecessary radiation exposure in older adults—who represent the majority of non-severe TBI patients—while still maintaining predictive accuracy.

Given that complications in patients with non-severe TBI are relatively low, as we have seen in our previous study [14], we believe that this score may help patients to return safely to their homes earlier. The Goliat score may support more personalized decision-making, potentially reducing unnecessary imaging and overcrowding in emergency departments, especially among older adults. And this can be very important in emergency departments that are becoming more and more overcrowded, mainly by our elderly. In fact, a recent study [31] showed that observation maintained in an Emergency Care Unit was associated with an increased risk of delirium and provided little clinical benefit.

In our study we did not detect differences with other biomarkers that did seem to be related to TBI prognosis, such as GFAP or UCHL1 [32]. Probably in non-severe TBI, these biomarkers do not help to predict complications in a short period of time, since in most of our sample there is no expression. And this is very important, because in our case the observation time in the emergency department crutial. In our model, among the biomarker variables considered, only S100B levels at 6 h post-trauma showed a statistically significant association with acute complications. Although this association was modest, it contributed to the multivariable logistic regression model underlying the Goliat score. The absence of predictive value for NSE, GFAP, and UCH-L1 in this cohort highlights the importance of evaluating biomarkers within the context of a broader clinical prediction tool, rather than as isolated tests.

The Goliat score integrates S100B with other clinical and physiological variables to estimate risk, ensuring that the final prediction is not dependent on a single laboratory measurement. For older adults, in whom non-severe TBI presentations can be subtle and complication risk higher, this multimodal approach may improve decision-making around safe discharge versus extended observation. The current study provides internal validation of the score in a real-world ED population; external, prospective validation will determine whether the inclusion of S100B and other predictors enhances its generalisability and clinical utility.

This study has several limitations that should be considered when interpreting the results. First, it was conducted at a single center, which may limit the generalizability of the findings to other healthcare settings with different patient populations, clinical practices, or resource availability. Second, although the cohort was prospectively collected, the analysis was retrospective in nature, which may introduce selection or information biases. Third, while the internal cross-validation showed satisfactory performance, the model lacks external validation, and its applicability in broader or more diverse populations remains to be confirmed.

Additionally, the incidence of complications may be influenced by local clinical protocols and thresholds for intervention or observation, which may vary across institutions. The relatively low number of patients exhibiting certain key predictors—such as fluctuating GCS or pupil alterations—required these variables to be grouped, potentially reducing the granularity of the model. Moreover, although several biomarkers were assessed, their utility in predicting acute complications in non-severe TBI may have been limited by timing of sample collection or insufficient sensitivity in cases without visible intracranial injury. This retrospective analysis did not systematically record the precise time from ED admission to hospitalisation decision, nor the exact timing and number of lesion progressions detected on second CT scans. While our institutional protocol provides a structured framework for these decisions, we acknowledge that actual recorded times could provide additional insights into the model’s applicability in real-time ED settings. We plan to capture these metrics prospectively in the forthcoming multicentre external validation study.

Finally, while the proposed score could assist in decision-making regarding cranial CT and observation, any clinical implementation should be accompanied by well-defined follow-up protocols to ensure patient safety.

The Goliat score shows a good internal cross-validation performance but it lacks of an external validation to assess its replicability in other settings. That is why after the publication of this article, we are going to perform the external validation of the score and check its performance and functionality. Our subgroup analyses showed that the Goliat score maintained acceptable calibration and moderate discrimination across age group and sex categories, with no significant loss of performance in older adults—a population that represents the majority of non-severe TBI cases in emergency departments. Discrimination was slightly higher in patients aged ≥ 65 years (AUC 0.73) compared to those < 65 years (AUC 0.70), and performance was consistent in both women and men. The stability of performance in older adults is particularly relevant, given their higher baseline risk of complications and the potential benefit of avoiding unnecessary CT scans in this group.

Conclusions

The Goliat score is a novel clinical prediction tool developed and internally validated to estimate the risk of acute complications in adult patients with non-severe TBI. In this study, it demonstrated good internal validity and discriminatory capacity, identifying a substantial proportion of at-risk patients. Importantly, pathological CT findings were not an independent predictor in our model, meaning that the Goliat score does not benefit from CT results to stratify risk. This feature could have significant implications for clinical practice, potentially reducing unnecessary cranial CT scans—and thus radiation exposure—particularly in older adults, who represent the majority of mTBI presentations in the emergency department. This study’s particularity lies in its endpoint, timing, and spectrum: it estimates 48-hour complication risk—not CT positivity—at ~ 6 h post-injury, in a geriatric-enriched minor/moderate TBI population where non-neurological complications (e.g., delirium, cardiorespiratory events) are common and decision-shaping.

By integrating easily obtainable clinical and laboratory variables, the score may help inform decisions on observation and discharge, with potential to optimise resource use in overcrowded emergency departments. These findings represent a first step in the evaluation of the Goliat score. External, prospective validation in larger and more diverse populations is essential before any clinical implementation. Beyond clinical risk stratification, safe discharge decisions—particularly when considering home or telemedicine follow-up—require confirmation of adequate communication capacity, availability of transportation for urgent return, and the presence of a caregiver or appropriate home support. This is especially relevant in older adults, for whom social vulnerability may increase the risk of adverse outcomes after discharge. The Goliat score should therefore be applied only in conjunction with a structured assessment of these non-clinical factors and within the context of validated clinical protocols.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

FP, MM, and OY conceived the study and designed the project. OY and AV obtained research funding. FP supervised data collection. AV, MB and ILV undertook recruitment of patients and managed the data, including quality control. MM and ILV provided statistical advice on study design and analyzed the data; MB chaired the data oversight committee. ILV and OY drafted the manuscript, and all authors contributed substantially to its revision. OY takes responsibility for the paper as a whole.

Funding

The projecte was funded by Mutua Madrileña Foundation.

Data availability

All the data is available upon request to the authors.

Declarations

Ethical approval

The study was approved by the CEIC of the Hospital Universitario Arnau de Vilanova de Lleida (CEIC-1952).

Human ethics and consent to participate

The study was approved by the Ethics Committee of Hospital Universitario Arnau de Vilanova de Lleida, reference number [CEIC-1952]. Written informed consent was obtained from all participants (or their legal representatives when applicable) prior to inclusion. The study complies with the Declaration of Helsinki and relevant national regulations (Organic Law 3/2018 and EU Regulation 2016/679).

Competing interests

The authors declare no competing interests.