This was a single-centre observational hospital cohort study, assembled by retrospective analysis of the register of consecutive admissions to the Department of Neurosurgery of Bouaké Teaching Hospital (Côte d’Ivoire) between January 2025 and June 2026, and reported in accordance with the STROBE recommendations [9]. The department receives patients with traumatic brain injury referred by peripheral health facilities as well as by the hospital emergency department; the most severe cases are subsequently transferred to the general intensive care unit, and deaths occurring there were counted, vital status being established by cross-checking the registers of both departments. Patients managed exclusively in the emergency department or in intensive care, without neurosurgical consultation, fell outside the scope of data collection and their number could not be established.

All consecutive patients admitted for traumatic brain injury during the study period were included, irrespective of age, provided they had undergone at least one brain computed tomography (CT) scan and had a usable medical record. The analytical cohort comprised 772 patients. Admissions that did not meet these criteria were not logged separately in the register: the number of patients excluded because computed tomography was unavailable or the medical record was unusable, and the reasons in individual cases, could not therefore be established. Recruitment was accordingly consecutive among eligible patients rather than among all neurosurgical admissions. Mortality analyses covered the whole cohort; functional outcome analyses covered the 703 patients for whom the Glasgow Outcome Scale score was documented (69 lost to follow-up). The study flow is summarised in Figure 1.

Figure 1. Study flow diagram. All 772 patients in the analytical cohort contributed to the mortality analysis; the functional outcome analysis covered the 703 patients for whom the Glasgow Outcome Scale score was documented.

The variables collected covered six domains: sociodemographic characteristics; circumstances of injury (mechanism, type of road user, helmet use, initial loss of consciousness, time to and mode of admission, transport); clinical and haemodynamic data; CT findings; therapeutic management; and outcome measures.

Initial neurological severity was defined according to the Glasgow Coma Scale [10]: mild TBI (13 - 15), moderate (9 - 12) and severe (8 or below). The score analysed was the first Glasgow Coma Scale score recorded in the medical file on admission to the neurosurgical pathway. The register did not systematically document whether sedation, intubation or resuscitation had preceded this assessment in the referring facility or in the emergency department, and no imputation rule was applied to intubated patients, whose recorded admission score was used as documented; this limitation is discussed in Section 4.3. Hypotension was defined as a systolic blood pressure below 90 mmHg. Time to admission was the number of days elapsed between injury and admission to neurosurgery, transfers were included: it measures access to the specialised pathway and not access to first aid. Mass effect was defined as sulcal effacement or ventricular or cisternal compression, midline shift as displacement of midline structures (severity threshold of 10 mm or more), diffuse axonal injury when the admission CT showed punctate or petechial haemorrhages at the grey-white matter junction, in the corpus callosum or in the brainstem, or diffuse cerebral swelling out of proportion to focal lesions, in a compatible clinical context, and polytrauma as the presence of at least one extracranial lesion threatening life or function. A surgical indication was the operative decision recorded in the medical file before any procedure, distinct from the procedure actually documented in the operating register. The CT findings considered were those of the first scan available on admission to the neurosurgical pathway; the interval between injury and CT acquisition was not recorded as a separate variable and is bounded by the time to admission. Lesion coding was based on the radiology reports and on the imaging findings recorded in the medical file, harmonised into homogeneous categories during the quality-control step described in Section 2.3, with systematic re-reading of the images by the investigators.

The primary outcome was all-cause in-hospital mortality; late functional outcome, assessed using the Glasgow Outcome Scale [11]—a score of 1 to 3 defining an unfavourable outcome—was a secondary outcome. It was collected at outpatient review (n = 386), by telephone interview (n = 177) or from the medical file (n = 104), with the 36 deaths having occurred during the hospital stay; median follow-up was 3.2 months (interquartile range 2.1 - 5.4).

Before any analysis and independently of it, the database underwent systematic quality control covering the coding of binary variables, plausibility bounds for haemodynamic variables, recoding of diagnostic labels into homogeneous lesion categories, and agreement between the Glasgow Coma Scale score and the severity classification, with no discrepancy being observed. A data dictionary and a versioned change log accompany the database.

Completeness of admission variables was 100%. The only notable missing data concerned functional outcome (69 patients lost to follow-up, 9.4% of survivors), and helmet use being undocumented in only 0.7% of motorised two- or three-wheeler users. Midline shift measurement was missing in 62 of the 230 patients with mass effect (27%): this absence is not random and was addressed in sensitivity analyses. The analysis was conducted on complete cases, without imputation. Baseline characteristics of patients lost to follow-up did not differ from those of patients followed up (all p-values 0.19 or above).

Categorical variables are expressed as counts and percentages, quantitative variables as mean ± standard deviation or as median with range or interquartile range. Bivariate comparisons used the Fisher exact test for all cross-tabulations, irrespective of expected cell counts. Confidence intervals for proportions were estimated by the Wald method.

Variables associated with each outcome at p below 0.20 in bivariate analysis, together with those judged clinically relevant, were entered into multivariable logistic regression models. An explicit causal framework guided the adjustment strategy (Figure 2): exposures present on admission were distinguished from downstream markers reflecting clinical management, such as intubation or surgical procedure, in order to avoid any adjustment for mediators. Helmet use and time to admission were treated as distal exposures. Two outcomes were modelled: in-hospital mortality and unfavourable functional outcome, all predictors being measured on admission. The mortality model was fitted on the 772 patients of the analytical cohort (36 events) and included the Glasgow Coma Scale score as a continuous variable, diffuse axonal injury, hypotension, anisocoria and mass effect. The functional outcome model was fitted on the 703 patients with a documented Glasgow Outcome Scale score (39 events) and included a

Figure 2. Causal framework guiding the adjustment strategy. Distal exposures and severity measured on admission are entered into the models; markers arising downstream of clinical management, such as intubation or surgical procedure, are excluded in order to avoid any adjustment for mediators.

Glasgow Coma Scale score of 8 or below, diffuse axonal injury, hypotension, mass effect and anisocoria. In accordance with this framework, helmet use and time to admission were treated as distal exposures whose effect on outcome is expected to operate through severity on admission, including the Glasgow Coma Scale score: they were therefore assessed in bivariate analysis without adjustment for admission severity and were not entered into the adjusted models, in which the score would have acted as a mediator of their effect. The Glasgow Coma Scale score was modelled as a continuous variable, dichotomisation at 8 being retained for sensitivity analysis. Because the small number of events and the quasi-separation observed for some variables made conventional logistic regression unstable, penalized regression using the Firth method [12] was applied, with confidence intervals estimated by profile penalized likelihood and p-values by likelihood ratio test [13]. Discrimination was quantified by the area under the ROC curve, corrected for optimism by bootstrapping (2000 replications). The significance threshold was 0.05 and analyses were performed with Python 3.12.

The cohort consisted mainly of young adults, with more than three-quarters of patients being under 45 years of age. Mean age was 33.9 ± 17.9 years (median 30 years; range 1 - 90 years) and the male-to-female ratio was 5.2. Slightly more than half of the patients (50.4%) were referred from another facility, and non-medicalised transport was the main mode of conveyance (60.1%) (Table 1).

Road-traffic crashes accounted for 89.4% of mechanisms (Table 2). Among users of motorised two- or three-wheeled vehicles for whom helmet use was documented, 5.4% (95% CI: 3.6 - 7.2) were wearing one. Median time to admission was two days, exceeding 48 hours in 34.8% of cases.

Table 1. Sociodemographic characteristics (n = 772).

The category “over 60 years” comprises 82 patients; the category “60 years and over” used in the bivariate analyses comprises 94, twelve patients having turned 60.

Table 2. Circumstances of injury (n = 772).

Helmet use is reported for the 611 users of motorised two- or three-wheeled vehicles for whom this item was documented.

Most patients presented with mild injury, with severe forms accounting for only 7.9% (95% CI: 6.0 - 9.8) of the cohort (Table 3). Signs of neurological severity on admission remained infrequent: anisocoria 10.5%, intubation 14.4% and motor deficit 16.6%. Haemodynamic status was broadly stable (mean systolic pressure 130 mmHg, heart rate 85/min, pulse oximetry 97%) and systemic secondary brain insults were uncommon: hypotension in 8 patients (1.0%), desaturation in 9 (1.2%) and hyperglycaemia in 22 (2.8%).

Table 3. Clinical presentation on admission and comorbidities (n = 772).

Cerebral contusions, isolated or associated with an intraparenchymal haematoma, were the most frequent CT findings (43.5%), ahead of mass effect (29.8%), subarachnoid haemorrhage (15.0%), anterior skull base fracture (14.8%) and acute subdural haematoma (14.2%) (Table 4, Figure 3). A midline shift of 10 mm or more was present in 10.8% of cases. Among associated injuries, wounds and facial fractures predominated, polytrauma being recorded in 8.2% of patients.

Table 4. Computed tomography findings and associated injuries (n = 772).

Midline shift could not be measured in 62 patients, all of whom had mass effect. The low frequency of skull fractures other than depressed fractures and anterior skull base fractures reflect under-documentation of simple linear fractures.

Figure 3. Frequency of computed tomography findings (n = 772). Categories are not mutually exclusive: a single patient may present with several lesions.

A surgical indication, in the sense of the operative decision recorded before any procedure, was retained in 72.9% of patients; a procedure was documented in the operating register in 47.8%. Evacuation or drainage of a haematoma was the most frequent procedure (18.0%), ahead of decompressive craniectomy (8.2%), external ventricular drainage (5.2%) and craniotomy (4.3%). Antibiotic therapy was administered to 75.5% of patients (Table 5).

In-hospital mortality was 4.7% (36 deaths; 95% CI: 3.2 - 6.2) and 78% of deaths occurred within the first forty-eight hours. Functional outcome, documented in 703 patients, was favourable in 94.5% of cases (95% CI: 92.8 - 96.1); an unfavourable outcome was observed in 39 patients (5.5%), of whom 36 died. Specialist opinions were requested mainly from oral and maxillofacial surgery (n = 135), trauma and orthopaedics (n = 76) and ophthalmology (n = 51) (Table 6).

In bivariate analysis, markers of neurological severity and space-occupying or diffuse lesions were most strongly associated with mortality: a Glasgow Coma Scale score of 8 or below (26.2% versus 2.8%), intubation (18.9% versus 2.3%), anisocoria (13.6% versus 3.6%), diffuse axonal injury (12.5% versus 3.9%), a midline shift of 10 mm or more (10.8% versus 3.9%) and hypotension (50.0% versus 4.2%). Conversely, age of 60 years or over, helmet non-use and a time to admission beyond 48 hours were not associated with mortality (Table 7). Among the 611 users for whom helmet use was documented, mortality was 5.0% in those without a helmet and 9.1% in those with one (3 deaths out of 33; p = 0.246): this apparent reversal, non-significant and based on three events, permits no interpretation.

Table 5. Therapeutic management (n = 772).

Procedures are not mutually exclusive. The surgical indication refers to the operative decision recorded in the medical file; the documented procedure is the one recorded in the operating register.

Table 6. In-hospital and functional outcome.

¹Percentage among the 703 patients whose functional outcome was documented; ²Percentage among the 736 survivors.

Table 7. Factors associated with in-hospital mortality, bivariate analysis (36 deaths).

Fisher exact test. The helmet row is restricted to the 611 documented users.

In multivariable analysis, depth of coma was the only independent determinant of in-hospital mortality: each Glasgow point lost multiplied the odds of death by 1.56 (95% CI: 1.31 - 1.86; p < 0.001). No other admission variable reached the significance threshold, including diffuse axonal injury (adjusted odds ratio 1.83; 95% CI: 0.67 - 4.77; p = 0.23) (Table 8, Figure 4). Anisocoria, hypotension and mass effect, strongly associated in crude analysis, lost their significance after adjustment, their information being largely captured by overall neurological severity. Unfavourable functional outcome shared these determinants, with an independent association for a Glasgow Coma Scale score of 8 or below (odds ratio 7.89; 95% CI: 3.22 - 18.85) and for diffuse axonal injury (3.05; 95% CI: 1.14 - 7.61). Discrimination of the mortality model, corrected for optimism by bootstrapping, was 0.79. In sensitivity analysis, dichotomisation of the Glasgow Coma Scale score at 8 degraded discrimination, the corrected area under the curve falling to 0.74, and made diffuse axonal injury appear as an independent determinant (odds ratio 3.59; 95% CI: 1.34 - 9.03): part of the effect usually attributed to it therefore reflects the gradient in depth of coma. Restriction to patients aged 16 years and over left the estimates unchanged.

The profile observed is consistent with that of West and Central African series: male predominance, mean age around the third decade and predominance of road-traffic crashes. In-hospital mortality of 4.7% is of the same order as that reported in N’Djamena in a series of 1022 head-injured patients of all grades (5.3%) [15]; it is probably underestimated by the non-inclusion of prehospital deaths. The concentration of deaths within the first forty-eight hours sheds light on two findings of this work. It explains first why time to admission is not associated with mortality: in patients who have survived until specialised care, it is initial lesion severity and the hours following arrival, not the time elapsed beforehand, that determine the outcome. It then places the lever for improvement upstream and at the very start of the hospital stay rather than in prolonging inpatient monitoring.

The earlier series from the same institution, covering the general intensive care unit between 2013 and 2014, reported a mortality of 66% among 92 patients with a Glasgow Coma Scale score below 14, with a mean time to admission of three hours [8]. The two cohorts describe two complementary segments of the same pathway: intensive care recruits its most severe end and the most direct admissions, neurosurgery the whole lesion spectrum, mostly after transfer. Both the difference in mortality and the difference in time to admission reflect this difference in recruitment setting rather than a change in practice, and serve as a reminder that any comparison between African hospital series presupposes knowing which segment of the pathway is being observed. Unlike cohorts from high-income countries, where falls in older people take an increasing share and prehospital care is medicalised and rapid, our series was characterised by the youth of the victims, the overwhelming share of two-wheelers and largely non-medicalised transport [16][17].

One of the main findings of this study is the very low rate of helmet use (5.4%) among documented users of motorised two- or three-wheeled vehicles. This level is markedly lower than that reported among motorcycle crash victims admitted to the emergency unit of Mulago Hospital in Kampala, where 19.2% of users were wearing a helmet [7], placing Bouaké among the most exposed settings documented in sub-Saharan Africa. Measured here on good-quality data, this figure constitutes a reference value for monitoring prevention policies. That helmet non-use is not associated with in-hospital mortality is expected: the helmet acts upstream, on the occurrence and initial severity of injury, rather than on lethality once the patient is hospitalised. The extent of non-use nonetheless identifies generalised helmet wearing as the most accessible primary prevention lever [18]. The prolonged times to admission further illustrate the obstacles in the prehospital pathway and constitute a priority target for organisational improvement, through regulation of referrals, medicalisation of transport and shortening of delays in access to computed tomography and to the operating theatre.

Clinically, the variables associated with outcome—Glasgow Coma Scale score, pupillary status, blood pressure, CT findings—are simple to collect from admission, without sophisticated equipment beyond the initial CT scan, and are consistent with the Brain Trauma Foundation guidelines [19]. Each Glasgow point lost is accompanied by an increase of about 56% in the odds of death, which justifies considering the score in its granularity rather than the threshold of 8 alone. Diffuse axonal injury should draw attention to neurological risk even in the absence of an evacuable surgical lesion, without however constituting a prognostic factor independent of the level of consciousness.

The strengths of this work lie in the consecutive recruitment of all eligible patients in the neurosurgical pathway over the study period, in the ascertainment of deaths from two sources, and in the very low proportion of missing data on admission variables, the result of a documented preliminary audit supported by a data dictionary and a change log. To these are added the sample size of 772 patients, among the largest in West African series, the joint assessment of clinical, haemodynamic and radiological factors, and the use of penalized regression suited to rare events.

Several limitations must be taken into account. The number of events remains moderate (36 deaths), which widens confidence intervals and calls for caution in interpreting rare variables such as hypotension. The single-centre design limits external validity. The mortality of 4.7% is that of the neurosurgical pathway and does not represent hospital lethality of traumatic brain injury at Bouaké Teaching Hospital: patients who died before a neurosurgical opinion was obtained—at the scene, during transport, in the emergency department or in intensive care—do not appear in the cohort and their number could not be established. This uncounted group probably leads to an underestimation of overall lethality and to an attenuation of the effect of determinants acting before specialised admission. In addition, the admission Glasgow Coma Scale score was abstracted retrospectively: whether sedation or intubation had preceded its recording could not be verified for every patient, and scores measured after intubation may underestimate the true level of consciousness in a minority of cases.

Functional outcome relied on the Glasgow Outcome Scale, assessed at variable intervals, with 9.4% of patients lost to follow-up; the absence of a standardised reassessment schedule may have introduced heterogeneity. The unfavourable outcome measure is moreover made up of deaths in more than nine tenths of cases (36 of 39 events): the two analyses therefore do not constitute two independent assessments. Finally, the cohort included 86 children aged 15 years or under (11.1%), in whom the properties of the Glasgow scales differ from those observed in adults, without the numbers allowing a separate analysis.

The bivariate association between midline shift and mortality does not withstand the extreme assumption regarding the 62 missing measurements and should not be over-interpreted. The analysis of helmet use rests on a denominator of 611 users, of whom only 33 were wearing a helmet: comparisons in this subgroup lack power and the observed difference in mortality, in an unexpected direction, must be regarded as a sampling artefact. The low frequency reported for skull fractures other than depressed fractures and anterior skull base fractures (3.1%) probably reflects under-documentation of simple linear fractures in CT reports. Residual confounding related to unmeasured variables, inherent to the observational design, cannot be excluded. Lastly, the absence of intracranial pressure monitoring and of detailed prehospital data limits the assessment of secondary brain insults.