Intracranial suppurative infections—brain abscess, subdural or extradural empyema, ventriculitis, scalp suppuration with intracranial extension—are among the few neurosurgical emergencies whose prognosis has been transformed by a single technological innovation. Since computed tomography came into general use, mortality has fallen from over 40% to under 10% in contemporary high-income series [2]-[5].

Disorders of the nervous system are, moreover, the leading global cause of disability-adjusted life years, with a burden concentrated in low- and middle-income countries [6]. This transformation has not been universal. In sub-Saharan Africa the disease remains common, is seen late, and carries higher case fatality and greater functional morbidity [7]-[11]. Its determinants are documented: prolonged time to consultation, unequal access to cross-sectional imaging, inconsistent microbiological documentation and limited critical-care capacity [7][9][12][13].

The threat to life does not arise from the infection as such. It arises from three mechanical processes: mass effect from the collection and the surrounding oedema, hydrocephalus from obstruction of cerebrospinal fluid pathways, and intraventricular rupture [3][14]-[16]. All three are identifiable on unenhanced computed tomography. The information that determines how urgent the pathway is, therefore, is precisely the information that remains available where technical resources are limited.

The Lancet Commission on Global Surgery estimated that five billion people lack access to safe, affordable and timely surgery [17]. Deaths avertable by surgery and anaesthesia now exceed the combined total attributable to tuberculosis, HIV and malaria [18]. Applied to neurosurgery, this analysis reveals an annual shortfall of several million essential operations, concentrated in sub-Saharan Africa and South Asia [19][20]. Resolution WHA68.15 placed essential surgery within universal health coverage, and national surgical, obstetric and anaesthesia plans are its operational instrument [13][21].

Within this framework, mortality attributable to poor-quality health systems exceeds that attributable to lack of access alone [22]. Intracranial suppurative infections illustrate this mechanism: the required procedure is technically simple, survival depends on how early it is performed, and the obstacle is organisational rather than technical.

In practical terms, the neurosurgeon on call in a teaching hospital of inland Côte d’Ivoire is not trying to estimate a three-month probability of death. He has to solve a scheduling problem: should this patient occupy the only theatre available tonight? Should one of the few close-monitoring beds be kept for him? Can he be placed on a general ward on antibiotics, with reassessment at twelve hours? This decision is taken several times a week, often by a junior operator, and almost always without any formal support.

Published scores do not answer this question. They were developed in populations where resources are not the binding constraint, which shifts the informative content of the variables; they predict endpoints—mortality, late Glasgow Outcome Scale—that do not coincide with the action to be decided; and they frequently include variables unavailable in an emergency in the settings concerned (magnetic resonance imaging, bacteriological documentation, specialised markers). Triage systems validated in Africa, such as the Cape Triage Score and the paediatric triage promoted by the World Health Organization, have demonstrated their value for in-hospital mortality but remain general-purpose and lack neurosurgical granularity [23]-[25]. To our knowledge, no clinical prediction rule for the triage of intracranial suppurative infections has been developed and internally validated in an African population.

Our hypothesis was that clinical and computed tomography variables available before any therapeutic decision are sufficient to identify, at presentation, the patients whose admission will consume heavy resources or will follow an unfavourable course.

The primary objective was to develop and internally validate an integer triage score, calculable without a calculator, from the variables available at presentation alone.

There were three secondary objectives: to evaluate this score simultaneously against four pre-specified triage targets—in-hospital death, poor functional outcome, surgery within the first 24 hours and use of critical-care therapies; to compare it with a pre-specified comparator, the Glasgow Coma Scale treated as a continuous variable; and to delimit its domain of validity explicitly by reporting in full the targets for which prediction fails.

This is a development and internal validation study of a clinical prediction rule (TRIPOD type 1b: development with internal validation by resampling, without an external dataset) [26][27]. Reporting follows the TRIPOD checklist together with the TRIPOD + AI extension for the penalised learning components [28] (Supplementary material S2 and S3); an abridged compliance summary is given in Supplementary Table S7-4. The statistical protocol, including the targets, the candidate pool, the stability threshold, the point-conversion unit and the two pre-specified negative analyses, was fixed before any performance figure was examined and is provided in full (S1).

The cohort comprises all admissions for imaging-documented intracranial suppurative infection managed in the neurosurgery department of Bouaké Teaching Hospital between January 2016 and January 2026. The analytical database holds 150 admissions corresponding to 148 patients; two patients had two distinct episodes, kept as separate admissions, a choice discussed in the limitations section. One hundred and twenty-seven admissions come from a retrospective case-note review (2016-2024) and 23 from structured prospective collection (2025-2026). The flow diagram is shown in Figure 1.

Figure 1. Study flow diagram (TRIPOD format). Admissions included, sources of data collection, candidate variables retained and discarded, multiple imputation, selection and internal validation.

Patients of any age with a brain abscess, a subdural or extradural empyema, a ventriculitis or a scalp suppuration with intracranial repercussion, documented by computed tomography or magnetic resonance imaging, were included. No exclusion criterion was applied a priori, in order to preserve the representativeness of the real patient flow—a necessary condition for the validity of a triage instrument, which must apply to every patient who presents and not to a selected subgroup.

Four targets were pre-specified, each corresponding to a distinct pathway decision:

1) In-hospitaldeath, recorded at the end of the admission;

2) Poorfunctionaloutcome, defined by a Glasgow Outcome Scale score ≤ 3 at the last documented assessment of the index admission, in practice at discharge [29];

3) Earlysurgery, defined by an evacuation procedure performed within 24 hours of admission (time to surgery ≤ 1 day);

4) Critical-careuse, defined by the administration, at any time during the index admission, of osmotherapy and/or the insertion of an external ventricular drain.

The fourth definition calls for an explicit methodological comment. The database holds no “admission to intensive care” variable, the department having had no dedicated neurosurgical intensive care unit over the whole period. We therefore took as a proxy the use of therapies that, in this setting, mandate critical-care-level monitoring. The proxy is imperfect: it captures therapeutic intensity rather than an administrative destination, and it is liable to be influenced by prescribing habits. It is reported as such, and every result relating to it must be read with that reservation. None of the three components of the composite was time-anchored to surgery: death, poor functional outcome and critical-care use were ascertained over the whole index admission, whether they arose before or after any operative procedure, whereas early surgery—the only time-defined target—was a procedure performed within 24 hours of admission.

The primary development endpoint is a pragmatic composite, the critical trajectory, defined by the occurrence of at least one of the following: death, Glasgow Outcome Scale ≤ 3 or critical-care use. Four arguments support this choice; they are set out in full, with the corresponding methodological references, in the pre-specified statistical protocol (S1).

First,thecompositematchesthedecisionitinforms. An endpoint should reflect the action it is meant to guide [30][31]. At the moment of triage, the clinician wants to know whether the patient belongs to the category for which a priority pathway is justified. From the standpoint of scheduling, the three components are manifestations of a single underlying state. A composite endpoint is legitimate when its components share a common causal mechanism and a common decisional implication, a condition met here [32].

Second,eachtargettakenaloneisunsuitablefordevelopment. Mortality yielded only 16 events, far below recommended thresholds whichever rule is applied [33]-[35]. The Glasgow Outcome Scale is almost entirely explained by the Glasgow Coma Scale, which would have produced an instrument redundant with a variable already measured universally. Critical-care use is a process variable, sensitive to the prescribing habits of the department.

Third,earlysurgerywasdeliberatelyexcludedfromthecomposite. It is a therapeutic decision and not an event undergone by the patient. Including it would have introduced circularity between the predictor and the predicted action. It is retained as a secondary evaluation target, in order to document this limitation.

Fourth,thedecompositionisreportedinfull. The main criticism levelled at composite endpoints is that they mask the heterogeneity of their components [30][32]. Performance of the score is therefore reported for each component separately, including where it is nil.

Twenty-four candidates were pre-specified under a strict availability rule: only variables obtainable at presentation without magnetic resonance imaging, without microbiology and without specialised assays were retained—that is, from the history, the neurological examination, an unenhanced brain computed tomography scan and a full blood count. These candidates covered four domains: patient background (age ≥ 65 years, age < 5 years, immunosuppression, postoperative origin, referred patient, rural residence, symptom-to-admission interval ≥ 21 days); neurology (Glasgow Coma Scale ≤ 8, Glasgow Coma Scale 9 - 12, impaired consciousness, raised intracranial pressure syndrome, meningeal syndrome, seizures, motor deficit, anisocoria, scalp swelling/cellulitis); imaging (subdural empyema, ventriculitis, multiple collections, maximum size ≥ 30 mm, mass effect, midline shift ≥ 5 mm, hydrocephalus); and laboratory tests (leucocytosis). Immunosuppression was defined operationally as human immunodeficiency virus infection, prolonged corticosteroid therapy, active malignancy or any other cause of immunosuppression documented in the notes; diabetes and sickle cell disease did not fall within this definition. The three items whose bedside assessment is least standardised were, in addition, defined operationally, and these definitions were reproduced verbatim on the score card and in the electronic calculator (S10): impaired consciousness denoted any impairment of the level of consciousness recorded at presentation, whatever its depth; raised intracranial pressure syndrome, the presence of headache, vomiting, papilloedema or a combination of these signs; and mass effect on unenhanced computed tomography, the displacement of adjacent structures, the effacement of the cortical sulci or compression of the ventricular system. Scalp swelling or cellulitis denoted inflammatory involvement of the soft tissues of the scalp overlying the suppuration, and hydrocephalus on computed tomography an active ventricular dilatation. The full dictionary of definitions and coding rules is given in S4.

C-reactive protein was excluded from the candidate pool despite its apparent clinical relevance: it was missing in 38% of admissions and is not routinely available as an emergency test in a substantial proportion of the settings for which this score is intended. A triage score with an item measurable only once in every two patients is not a triage score.

Dichotomising continuous variables (Glasgow Coma Scale, size, midline shift) is methodologically unfavourable and entails a recognised loss of information [36]. It was nevertheless adopted in a pre-specified manner for the construction of the point scale, the aim being an instrument calculable mentally in the admissions room. The cost of this choice is measured explicitly in the study by the comparison with the Glasgow Coma Scale treated as a continuous variable, and it constitutes one of the two negative findings reported.

Missing data concerned mainly midline shift (18.0%), white cell count (8.0%) and maximum size of the collection (4.0%). Multiple imputation by chained equations was performed (ten imputed datasets, posterior draw, twenty iterations) [37]-[39], the derived binary variables being recomputed after imputation and then dichotomised. All selection and estimation steps were conducted across the ten datasets, with pooling according to Rubin’s rules [37]. A complete-case sensitivity analysis is provided in S6.

Selection relied on L1-penalised logistic regression (LASSO) [40], the penalty parameter being chosen by five-fold cross-validation under the one-standard-error rule, which deliberately favours parsimony over apparent performance. The procedure was repeated on each of the ten imputed datasets, and only variables selected in at least 60% of the datasets were retained—a stability criterion inspired by stability selection approaches [41], intended to prevent a variable from owing its presence in the score to a particular imputation draw.

The final model was estimated by Firth penalised logistic regression [42][43], chosen because quasi-complete separation was expected on rare variables (scalp swelling was associated with a critical trajectory in all 12 patients concerned, a configuration that makes ordinary maximum likelihood unusable). Coefficients and standard errors were pooled according to Rubin’s rules.

The Firth coefficients were divided according to a pre-specified conversion rule—division by the smallest retained coefficient, whose estimated value here is 0.642—then rounded to the nearest integer, with a cap of 4 points per item. This cap, decided before any performance figure was examined, limits the influence of estimates arising from quasi-separation, whose confidence interval is by construction very wide. The total score ranges from 0 to 14 points.

The score-to-probability mapping was obtained by a Firth regression of the composite endpoint on the score. The model is therefore fully specified by the following three elements, reproduced here so that it can be implemented without recourse to the authors (full detail in S5):

1)Scorecalculation.

Score = 4 × (scalp swelling or cellulitis) + 3 × (mass effect on computed tomography) + 2 × (impaired consciousness) + 2 × (immunosuppression) + 2 × (hydrocephalus) + 1 × (raised intracranial pressure syndrome)

each item scoring 1 if present and 0 otherwise.

2)Transformationintoaprobability.

logit(p) = −2.213 + 0.681 × Score

that is

p = 1/[1 + e^−(−2.213 + 0.681 × Score)]

3)Underlyingmultivariablemodel (provided to allow recalibration or reweighting at external validation, and not for routine calculation):

logit(p) = −2.096 + 3.270 × scalp + 1.795 × mass effect + 1.463 × consciousness + 1.399 × immunosuppression + 1.279 × hydrocephalus + 0.642 × raised intracranial pressure

Step-by-step worked examples are given in Supplementary TableS7-1.

Internal validation followed Harrell’s method [44][45]: 1000 bootstrap resamples with replacement, each replaying the entire construction procedure—fixed imputation, LASSO selection under the one-standard-error rule, Firth estimation, conversion to points, score-to-probability recalibration—the score thus rebuilt being then applied to the original sample. Imputation was carried out once, before the bootstrap, and the ten imputed datasets were then held fixed within the resampling loop rather than regenerated inside each resample; because re-imputing within every resample would propagate imputation uncertainty more completely, holding the imputations fixed is a pragmatic simplification that may render the optimism-corrected intervals slightly narrower than a fully nested imputation-in-bootstrap procedure would yield. Given the low overall fraction of missing data—at most 18.0% for a single variable and 8.0% or less for the others—this effect is expected to be small; it is recorded among the limitations. Optimism was estimated as the mean difference between performance in the resample and performance in the original sample, and subtracted from apparent performance. This approach is more conservative than cross-validation applied after selection. It is also the only one that accounts for the optimism induced by selection itself, in line with recent recommendations on the evaluation of prediction models [46][47].

Corrected performance measures comprised, in line with current recommendations for prognostic models [48]-[50]:

Discrimination, by the area under the ROC curve;Weakcalibration, by the calibration slope and calibration-in-the-large (intercept estimated with the slope fixed at 1);Moderatecalibration, by a smoothed calibration curve obtained by locally weighted logistic regression (tricube kernel), together with the E_mean, E_90 and E_max indices, defined as the mean, the 90th centile and the maximum of the absolute difference between predicted probability and smoothed observed probability;Overallperformance, by the Brier score and the scaled Brier score (1 − Brier/Brier_max).

Calibration was also tested formally by the Spiegelhalter test and by the Hosmer-Lemeshow test with ten groups, and displayed by deciles and quintiles of predicted risk.

The full bootstrap distribution of the corrected area under the curve and the optimism-corrected calibration curve, with its 95% confidence envelope, are reported in the Results and in Supplementary.

The Glasgow Coma Scale treated as a continuous variable was pre-specified as the comparator for each target [51][52]. This choice is deliberately demanding: it poses the question that any new score must answer—does it add anything to the clinical variable the clinician already measures? The areas under the curve of the score and of the comparator were compared for each target by the DeLong test for correlated ROC curves [53], with estimation of the difference in area under the curve and its 95% confidence interval.

Clinical utility was assessed by decision curve analysis [54][55], comparing the net benefit of the score, over a range of probability thresholds from 0.05 to 0.95, with the two reference strategies: directing every patient to the priority pathway, or directing none. Net benefit was converted into a directly interpretable measure, the net reduction in unnecessary priority referrals per 100 patients, obtained by dividing the difference in net benefit by the threshold odds. The range of clinically plausible thresholds (0.15 - 0.60) was discussed a priori with the clinicians of the department and corresponds to the range within which a neurosurgeon on call would accept to commit a scarce resource [56]. A clinical impact curve completes this analysis: for 1000 admissions and at each threshold, it displays the number of patients the score would classify as high risk and, among them, the number who would actually experience the event.

Three sensitivity analyses were pre-specified: restriction to complete cases; restriction to the first admission of each patient; and exclusion of patients operated on within the first 24 hours. This last analysis tests the robustness of the score against the residual circularity linked to the surgical decision: if performance is maintained in the subgroup not operated on immediately, the score does not reduce to a reproduction of the operative indication.

Patients and the public were involved neither in the design, nor in the conduct, nor in the choice of endpoints, nor in the reporting of this research, nor in its dissemination. This absence of involvement is an acknowledged limitation. A prospective impact study, such as the one proposed in the discussion, will however have to involve patients, families and admissions staff in defining the endpoints and the decision thresholds, since the relative value placed on an unnecessary priority referral and on an unanticipated critical trajectory is a trade-off that is not purely clinical.

Analyses were conducted in Python 3.12 (numpy, pandas, scikit-learn [57], scipy), with an in-house implementation of Firth regression by Newton-Raphson iterations with the score modified by the Jeffreys term, and an in-house implementation of the DeLong test by the method of structural components. The full code, allowing exact reproduction of every figure reported, is provided as supplementary material (S9).

The flow of admissions is shown in Figure 1 and the characteristics of the 150 admissions in Table 1. Median age was 26 years (IQR 14 - 50), with a male predominance (69.3%) and an almost equal distribution between rural and urban residence. One patient in two had been referred from another facility.

Table 1. Characteristics of the 150 admissions at presentation, on imaging and during management.

IQR: interquartile range; GCS: Glasgow Coma Scale; GOS: Glasgow Outcome Scale; CT: computed tomography; MRI: magnetic resonance imaging. Missing data: midline shift 27/150 (18.0%), white cell count 12/150 (8.0%), maximum size of the collection 6/150 (4.0%), GOS 8/150 (5.3%); all other candidate predictors were complete. Candidate prevalences not listed above were: age ≥ 65 years 11/150, age < 5 years 8/150, symptom-to-admission interval ≥ 21 days 75/150, meningeal syndrome 30/150, ventriculitis 18/150, maximum size ≥ 30 mm 114/150, midline shift ≥ 5 mm 17/150 and leucocytosis 138/150 (TRIPOD item 13c; full dictionary in S4). Documented immunosuppression: human immunodeficiency virus infection, prolonged corticosteroid therapy, active malignancy or any other cause of immunosuppression documented in the notes; diabetes and sickle cell disease did not fall within this definition. So-called pure forms exclude associated locations; the candidate variable “subdural empyema” used in Table 2 andTable 3 denotes any subdural involvement, isolated or associated (27/150).

The median interval between symptom onset and admission reached 19 days (IQR 15 - 25). Median GCS at presentation was 13 (IQR 10 - 15), with 12.7% of patients at GCS ≤ 8. The six items of the future score were present in 71.3% (raised intracranial pressure), 48.7% (impaired consciousness), 73.3% (mass effect), 10.7% (hydrocephalus), 8.7% (immunosuppression) and 8.0% (scalp swelling) of admissions.

Imaging rested on computed tomography alone in 67.3% of cases. The median maximum size of the collections was 36 mm (IQR 30 - 44). Microbiological documentation was attempted in 77.3% of patients, with a positive culture in 43.1% of the samples obtained; sterile cultures accounted for 44.0% of the cohort. A combined medical and surgical strategy was adopted in 83.3% of patients, with a median time to surgery of 2 days (IQR 1 - 3).

The characteristics of admissions with and without a critical trajectory are compared in Supplementary TableS7-5. The two groups differed neither in age, nor in sex, nor in place of residence, nor in the symptom-to-admission interval. Differences concerned exclusively the neurological and radiological variables recorded at presentation, which supports the choice of the candidate pool.

A critical trajectory was observed in 97 admissions (64.7%). The individual targets were: in-hospital death, 16 (10.7%); poor functional outcome (GOS ≤ 3), 57 of 142 assessable (40.1%); surgery within 24 hours, 41 (27.3%); critical-care use, 65 (43.3%). The three components of the composite overlapped substantially: of the 97 critical trajectories, 2 were defined by death alone, 24 by poor functional outcome alone and 35 by critical-care use alone, whereas 6 combined death and poor functional outcome, 3 death and critical-care use, 22 poor functional outcome and critical-care use, and 5 all three components; death and poor functional outcome co-occurred in 11 admissions. The number of events per candidate variable was therefore favourable for the composite (97 events for 6 final parameters, about 16 events per variable) but plainly insufficient for death taken alone (16 events), which is a major determinant of the results set out below [33][35].

Univariable associations with the critical trajectory are given in Table 2. The strongest associations concerned mass effect (79.1% versus 25.0%; OR 10.82 (95% CI 4.69 - 24.96); p < 0.001), raised intracranial pressure syndrome (75.7% versus 37.2%; OR 5.13 (2.42 - 10.87); p < 0.001), impaired consciousness (82.2% versus 48.1%; OR 4.84 (2.31 - 10.13); p < 0.001), size ≥ 30 mm (OR 3.52 (1.63 - 7.60); p = 0.001), GCS ≤ 8 (OR 8.14 (1.49 - 44.59); p = 0.004) and midline shift ≥ 5 mm (OR 7.09 (1.29 - 39.06); p = 0.006). Scalp swelling showed a perfect association (12/12 patients concerned). Postoperative origin was associated with a lower risk (44.4% versus 69.1%; OR 0.36 (0.16 - 0.84); p = 0.025), a finding consistent with earlier diagnosis in patients already admitted and under follow-up—a pattern also described in series of postoperative neurosurgical infections [58][59].

Table 2. Univariable associations with the critical trajectory (main candidates).

OR: odds ratio estimated with a continuity correction (0.5); p: Fisher’s exact test. Univariable associations were estimated on the ten imputed datasets, so denominators are out of 150 for every candidate, including midline shift and maximum size. Continuous variables (Mann-Whitney test): median GCS 12 (IQR 9 - 14) in patients with a critical trajectory versus 14 (13 - 15) without (p < 0.001); maximum size 39.5 mm (33.0 - 47.2) versus 31.0 mm (23.0 - 36.2) (p < 0.001); midline shift 3.1 mm (1.3 - 4.7) versus 0.0 mm (0.0 - 1.1) (p < 0.001).

LASSO selection repeated across the ten imputed datasets retained six variables above the 60% stability threshold (Table 3, Supplementary FigureS8-1): impaired consciousness, raised intracranial pressure syndrome, mass effect and scalp swelling (stability 100%), immunosuppression and hydrocephalus (80%). No other variable exceeded 20% stability across imputed datasets.

Table 3. Selection stability of the predictors (Endpoint: Critical trajectory).

The median number of variables selected per bootstrap resample was 11 (out of 24 candidates), the final selection resting on the stability threshold and not on penalisation alone.

The selection frequency across the 1000 bootstrap resamples, reported here for the first time, confirms this hierarchy while showing the true uncertainty of the procedure: mass effect 99.6%, impaired consciousness 94.6%, scalp swelling 87.6%, raised intracranial pressure syndrome 75.1%, hydrocephalus 75.0%, immunosuppression 64.0%. The six items retained are therefore also the six variables most frequently selected under resampling, with a ten-point margin over the seventh candidate (midline shift ≥ 5 mm, 54.0%). The median number of variables selected per resample was 11, which indicates that penalisation alone would have produced a markedly less parsimonious model and underlines the role of the stability threshold.

Notably, the dichotomised GCS was not retained: in this dataset, clinically recorded impaired consciousness captures most of the information carried by the GCS for this composite, while being more robust to variation in scoring between operators—inter-observer variability being well documented for the GCS, including among trained assessors [51].

The pooled Firth model and the corresponding point scale are given in Table 4, which reports for each item, on a single line, the coefficient β, its pooled standard error, the odds ratio and its confidence interval, the p value and the points assigned. Adjusted odds ratios, with profile penalised-likelihood confidence intervals and penalised likelihood-ratio test p values, were 26.32 (95% CI 1.6 to >1000; p = 0.010) for scalp swelling, 6.02 (2.1 - 18.0; p < 0.001) for mass effect, 4.32 (1.8 - 10.5; p < 0.001) for impaired consciousness, 4.05 (0.7 - 25.0; p = 0.12) for immunosuppression, 3.59 (0.6 - 23.0; p = 0.17) for hydrocephalus and 1.90 (0.7 - 5.3; p = 0.21) for raised intracranial pressure syndrome. Once the quasi-complete separation is handled by penalised likelihood, the confidence interval for scalp swelling no longer includes 1 and this item reaches significance.

Table 4. Final firth model (Rubin pooling) and point scale of the TRIAGE-SIC score.

Confidence intervals are profile penalised-likelihood intervals and p values are from the penalised likelihood-ratio test (Firth method [27]); under the quasi-complete separation of scalp swelling (12/12) the Wald statistic is invalid and was therefore not used. Points = β/0.642, rounded to the nearest integer, capped at 4. Standard error pooled according to Rubin’s rules (within variance + between variance × [1 + 1/m]). Score-to-probability mapping: logit(p) = −2.213 + 0.681 × score. Full equation in Supplementary Table S7-1.

Three of the six items—immunosuppression, hydrocephalus and raised intracranial pressure syndrome—are not significant taken in isolation on the basis of the penalised likelihood-ratio test; they were retained because selection rested on the stability of the penalised contribution and not on a significance threshold—selection by p value being explicitly discouraged in the development of prognostic models [44][45]—and because removing them degraded calibration in the upper bands.

The resulting TRIAGE-SIC point scale (0 - 14 points) assigns 4 points to scalp swelling/cellulitis, 3 points to mass effect, 2 points to impaired consciousness, to immunosuppression and to hydrocephalus, and 1 point to raised intracranial pressure syndrome. The corresponding nomogram is presented in Figure 2, the paper version of the score in Table 5, and the full model equation in Supplementary TableS7-1. The distribution of the score in the cohort is shown in Supplementary FigureS8-2: it covers the whole 0 - 14 range, with a main mode at 6 points (n = 44), a secondary mode at 4 points (n = 32) and a distinct group of 20 patients at 0 points. The extreme values are sparsely populated (12 points, n = 2; 14 points, n = 1), which justifies not interpreting predicted probabilities individually beyond 10 points.

Table 5. TRIAGE-SIC score—paper version (for display in the admissions room).

This score predicts neither death nor individual functional outcome. For functional prognosis, use the Glasgow Coma Scale. The estimated probabilities are the apparent probabilities of the derivation cohort; as the optimism-corrected calibration slope is 0.640, they are slightly too widely dispersed at both ends of the scale and call for local recalibration before any numerical use. Operationaldefinitionsoftheitems. Scalp swelling or cellulitis: inflammatory involvement of the soft tissues of the scalp overlying the suppuration. Mass effect: displacement of adjacent structures, effacement of the cortical sulci or ventricular compression on unenhanced computed tomography. Impaired consciousness: any impairment of the level of consciousness at presentation, whatever its depth. Hydrocephalus: active ventricular dilatation on computed tomography. Raised intracranial pressure syndrome: headache, vomiting, papilloedema or a combination of these signs.

Figure 2. TRIAGE-SIC nomogram. Points assigned to each item, points scale, total score and corresponding probability of a critical trajectory. The associated logistic equation is restated below the probability axis: p = 1/[1 + e^−(−2.213 + 0.681 × score)].

For the critical trajectory, the apparent area under the curve was 0.858. The bootstrap encompassing selection estimated an optimism of 0.070, giving a corrected area under the curve of 0.788 (95% CI 0.726 - 0.852). The full bootstrap distribution of the corrected area under the curve is shown in Supplementary: it is approximately symmetrical and centred on 0.788; the apparent value of 0.858 lies beyond the upper bound of the confidence interval of the corrected area under the curve, which gives the measure of the optimism that a naive validation would have let through.

The calibration indices, reported in full in Table 6 and in Supplementary TableS7-6, were as follows. Corrected calibration-in-the-large was 0.011 (95% CI −0.548 to 0.604). The corrected calibration slope was 0.640 (0.367 - 0.977). The corrected Brier score was 0.185 for a maximum Brier score of 0.229. The corrected scaled Brier score was 0.196 (95% CI −0.010 to 0.372); it was corrected for optimism as a quantity in its own right and is therefore not exactly the transformation of the two preceding values. This interval contains zero: at this sample size, and unlike discrimination and net benefit, overall performance is not distinguishable from that of an uninformative strategy. The corrected moderate calibration indices were E_mean 0.058, E_90 0.099 and E_max 0.131.

Table 6. Internal validation by bootstrap (1000 resamples of the entire procedure)—endpoint: critical trajectory.

Maximum Brier score (Brier score of the uninformative strategy): 0.229. Scaled Brier score = 1 − Brier/Brier_max. E_mean, E_90 and E_max: mean, 90th centile and maximum of the absolute difference between predicted probability and observed probability smoothed by local logistic regression. Optimism is the mean difference between performance in the resample and performance of the resampled model applied to the original cohort; for indices where a lower value is better (Brier, E), a negative optimism means that the corrected value is higher than the apparent value. The apparent calibration slope equals 1.000 by construction, the score-to-probability relation being estimated on the very data used to evaluate it; it therefore carries no information, and the corrected value of 0.640 is determined entirely by the estimation of optimism. The lower bounds of the intervals for E_90 and E_max are truncated at 0.000: the corresponding bootstrap percentiles were negative, which is impossible for absolute differences.

The formal calibration tests applied to the apparent model showed no significant departure from perfect calibration: Spiegelhalter test z = −0.094 (p = 0.925), Hosmer-Lemeshow test C = 5.13 on 8 degrees of freedom (p = 0.744). The latter must be read with caution: the score takes only fifteen distinct values and twenty admissions share a score of 0, so that the deciles separate patients with identical predicted probabilities; it is reported as secondary to the calibration curve. Calibration by deciles of predicted risk is shown in Figure 3(b) and detailed in Supplementary TableS7-6: the observed proportion of events rose from 20.0% in the first decile (mean predicted probability 0.099) to 100% in the tenth (0.977), and no decile departed from the diagonal beyond its confidence interval.

Figure 3. Calibration of the score on the composite endpoint. (a) Calibration curve smoothed by local logistic regression: apparent curve (solid line), optimism-corrected curve (dashed), 95% confidence envelope and line of identity. (b) Calibration by deciles of predicted risk, with Wilson confidence intervals and the results of the Spiegelhalter and Hosmer-Lemeshow tests. The lower panels show the distribution of predicted probabilities according to occurrence of the event.

The apparent calibration curve and the optimism-corrected curve, with the line of identity and the 95% confidence envelope, are shown in Figure 3, together with the histogram of predicted probabilities by status. Correction for optimism shifts the curve upwards in the lower half of the range, which means that the score, transposed to an unobserved population, would slightly underestimate the risk of low-scoring patients—the clinically preferable direction of error for a triage instrument, whose most costly error is underestimation of high risk.

Applied to the individual targets (Table 7, Figure 4), the score retained useful discrimination for critical-care use (AUC 0.787; 95% CI 0.714 - 0.851), moderate discrimination for poor functional outcome (0.711; 0.628 - 0.791), weak discrimination for early surgery (0.588; 0.494 - 0.683) and none for death (0.472; 0.336 - 0.608).

Figure 4. Discrimination of the score by triage target, compared with the continuous Glasgow Coma Scale. (a) Critical trajectory (composite); (b) critical-care use; (c) poor functional outcome (GOS ≤ 3); (d) in-hospital death. Each panel reports the difference in area under the curve and the p value of the DeLong test for correlated ROC curves.

Table 7. Performance of the TRIAGE-SIC score by triage target, and comparison with the continuous GCS (DeLong test).

*Apparent AUC; the value corrected for optimism by 1000 bootstrap resamples encompassing selection is 0.788 (0.726 - 0.852). ΔAUC = AUC of the score−AUC of the continuous GCS. A positive value favours the score. The 95% confidence intervals of the AUC of the continuous GCS are given in the table. For critical-care use the AUC of the GCS is exactly 0.500 (95% CI 0.406 - 0.594), computed on all 150 admissions with the conventional orientation of the scale; the interval includes 0.5, confirming the absence of any discrimination and supporting one of the two negative findings.

Direct comparison with the pre-specified comparator by the DeLong test gives these figures their meaning. For the two resource-allocation targets, TRIAGE-SIC is significantly superior to the continuous Glasgow Coma Scale: critical trajectory, 0.858 versus 0.723, ΔAUC + 0.135 (95% CI 0.050 - 0.221; p = 0.002); critical-care use, 0.787 versus 0.500, ΔAUC + 0.287 (0.203 - 0.372; p < 0.001). This second result deserves emphasis: the Glasgow Coma Scale, taken alone, does not discriminate at all for the use of critical-care therapies in this cohort (AUC 0.500), whereas the triage score reaches 0.787. The information that determines the need for osmotherapy or ventricular drainage is therefore essentially morphological—mass effect, hydrocephalus—and not related to the depth of coma.

Conversely, for the other three targets the comparator is significantly superior to the score: poor functional outcome, ΔAUC −0.209 (−0.289 to −0.130; p < 0.001); death, ΔAUC −0.182 (−0.284 to −0.080; p < 0.001); early surgery, ΔAUC −0.097 (−0.187 to −0.006; p = 0.037). The two instruments therefore do not answer the same question, and this divergence is not a sampling artefact but a statistically established result across all five comparisons.

Two pre-specified analyses produced negative results, reported here on the same footing as the positive ones.

3.8.1. Negative Finding No. 1: In-Hospital Mortality Is Not Predictable from Variables Available at Presentation Alone

The selection procedure applied specifically to death retained no variable. The three most frequently selected candidates—age ≥ 65 years, GCS ≤ 8 and subdural empyema—reached only 40% stability, below the pre-specified threshold of 60%. The continuous Glasgow Coma Scale discriminated death only weakly (AUC 0.654) and the TRIAGE-SIC score not at all (0.472; 95% CI 0.336 - 0.608, an interval containing 0.500).

3.8.2. Negative Finding No. 2: For Functional Outcome, the Glasgow Coma Scale Outperforms the Point Score

The pre-specified comparator reached an area under the curve of 0.920 for GOS ≤ 3, against 0.711 for the point score, a difference of 0.209 that is significant on the DeLong test (p < 0.001). A model specific to this target, selected by the same procedure (GCS ≤ 8, GCS 9 - 12, impaired consciousness, subdural empyema, midline shift ≥ 5 mm), reached 0.889, still below the continuous GCS. Conversely, for the composite endpoint the score outperformed the continuous GCS (0.858 versus 0.723; p = 0.002). This dissociation can be read directly in the distribution by band (Table 8(a)): the proportion of poor functional outcomes peaks at 66.0% in the 5 - 6 band and then falls back to 50.0% in the ≥7 band, and mortality, also non-monotonic, falls from 14.5% to 4.5% between these same two bands. The score ranks resource needs, not individual neurological severity.

Table 8. Observed risk by score band, operating characteristics of the thresholds and likelihood ratios. (a) Observed risk by band, (b) Operating characteristics of the thresholds (target: critical trajectory; prevalence 64.7%).

Sn: sensitivity; Sp: specificity; PPV/NPV: positive and negative predictive values; LR +/LR−: positive and negative likelihood ratios. Post-test probabilities computed from the observed prevalence of 64.7%. At the ≥ 7 threshold, specificity reaches 100% and LR + is not estimable (no false positives); this threshold is therefore not proposed as a decision rule.

Four bands were defined a priori on the distribution of points (Table 8, Figure 5). The observed risk of a critical trajectory was 11.5% (3/26; Wilson 95% CI 4.0 - 29.0) for 0 - 1 point, 55.3% (26/47; 41.2 - 68.6) for 2 - 4 points, 83.6% (46/55; 71.7 - 91.1) for 5 - 6 points and 100% (22/22; 85.1 - 100) for ≥7 points. Critical-care use followed the same gradient: 0%, 40.4%, 49.1% and 86.4%. No patient in the lowest band required osmotherapy or an external ventricular drain.

Figure 5. Observed risk of a critical trajectory by score band. Observed proportions with Wilson 95% confidence intervals, numbers per band and a graded colour code corresponding to the four priority levels.

At the operating thresholds, a score ≥ 2 identified the critical trajectory with a sensitivity of 96.9% and a negative predictive value of 88.5%, for a negative likelihood ratio of 0.07 (95% CI 0.02 - 0.23)—a configuration suited to rule-out use, that is, to allow a non-priority referral: a patient scoring below 2 sees his probability of a critical trajectory fall from 64.7% (prevalence) to 11.5%. A score ≥ 5 gave a sensitivity of 70.1%, a specificity of 83.0%, a positive predictive value of 88.3% and a positive likelihood ratio of 4.13 (2.24 - 7.59), raising the post-test probability to 88.3%—a configuration suited to rule-in use, that is, to trigger the priority pathway. A score ≥ 6 (LR + 6.67; 2.85 - 15.57) raises the post-test probability to 92.4%.

These two thresholds correspond to the two distinct triage decisions and are proposed jointly rather than as a single cut-off, a choice consistent with the logic of an instrument whose rule-out use and rule-in use do not carry the same consequence when they err.

Decision curve analysis (Figure 6(a)) showed a net benefit superior to both reference strategies across the whole range of clinically plausible thresholds. Expressed as a net reduction in unnecessary priority referrals per 100 patients admitted, the gain was 7 at the 0.20 threshold, 15 at the 0.50 threshold and 19 at the 0.65 threshold. At this last threshold, the “refer everyone” strategy became harmful (negative net benefit) whereas the score retained a net benefit of 0.342.

The clinical impact curve (Figure 6(b)) expresses the same result in absolute numbers. For 1000 admissions and at the 0.20 threshold, the score would classify 827 patients as high risk, of whom 627 would actually experience a critical trajectory. At the 0.50 threshold, these figures become 727 and 587 respectively.

Figure 6. Clinical utility. (a) Clinical decision curve: net benefit of the score compared with the “refer everyone” and “refer no one” strategies. The shaded area delimits the range of clinically plausible thresholds for a triage decision (0.15 - 0.60); the reference threshold of 0.20 is indicated. (b) Clinical impact curve: for 1000 admissions and at each threshold, the number of patients the score would classify as high risk and, among them, the number who would actually experience a critical trajectory.

Excluding the 41 patients operated on within the first 24 hours left 109 admissions and 65 events. Discrimination of the score there was 0.852, against 0.858 in the whole cohort. Calibration-in-the-large was 0.002 and the calibration slope 1.000; both values result from re-estimation of the score-to-probability relation in the 109 admissions and therefore do not constitute a test of the transportability of the derivation equation: the calibration that would have been obtained by applying that equation to this subgroup without re-estimation was not computed. The gradient between bands was preserved: 13.0%, 52.8%, 79.4% and 100% for 0 - 1, 2 - 4, 5 - 6 and ≥7 points. A LASSO selection conducted de novo in this subgroup retained seven variables, including five of the six items of the final score; immunosuppression was no longer retained there, whereas GCS ≤ 8 and midline shift ≥ 5 mm were.

The complete-case analysis and the analysis restricted to each patient’s first admission did not change the variables selected (S6).

Finally, a post-hoc sensitivity analysis, requested to gauge the dependence of the score on its treatment-dependent component, restricted the development endpoint to death or poor functional outcome (GOS ≤ 3), excluding critical-care use. Sixty-two of 150 admissions (41.3%) met this restricted endpoint. Repeating the imputation and LASSO selection procedure retained four variables above the 60% stability threshold: GCS ≤ 8, GCS 9 - 12, impaired consciousness and mass effect. The resulting Firth model had an apparent AUC of 0.831; bootstrap correction for optimism yielded an AUC of 0.817 (bootstrap 95% CI approximately 0.763 - 0.892), with an optimism-corrected calibration slope of 0.99. Removing the process-dependent critical-care component therefore reduced the number of selected variables and shifted selection towards neurological severity, while preserving useful discrimination.

Four supports are provided: a paper version fitting on half a page, with colour coding of the bands and the suggested management (Table 5); the nomogram, intended for teaching and for reading the relative weight of each item (Figure 2); the clinical algorithm (Figure 7); and a standalone calculator contained in a single HTML file, running offline on a mobile phone with no data transmission, which returns the total, the estimated probability, the risk band and an explicit reminder that the score predicts neither death nor individual functional outcome (S10). None of these supports requires a network connection or a licence.

We have developed and internally validated a triage score for intracranial suppurative infections in an Ivorian cohort of 150 admissions. Six variables available at presentation, without magnetic resonance imaging or microbiology, stratify the risk of an unfavourable in-hospital course with an optimism-corrected area under the curve of 0.788, and critical-care use with an area under the curve of 0.787—in both cases significantly better than the continuous Glasgow Coma Scale. The score predicts neither death nor individual functional outcome. These results matter because they provide, for the first time in this setting, a reproducible instrument for a decision that is currently taken without any formal support.

One point of vocabulary conditions the reading of everything that follows. TRIAGE-SIC is not a prognostic score: it is a resource-allocation instrument intended to structure the pathway decision in a resource-limited setting. It supports clinical judgement and does not replace it.

Figure 7. Clinical algorithm for use of the TRIAGE-SIC score. Sequence of steps from admission to referral, with the three decision bands, the corresponding observed risks, the reassessment loop at the twelfth hour and a reminder of the limits of use of the instrument.

4.2.1. Rationale for the Methodological Choices

Four methodological choices structure this work and deserve to be made explicit, because they determine the credibility of the model more than its apparent performance does.

LASSO was chosen because stepwise selection based on p values produces, at this sample size, unstable and overfitted models [44][45]. L1 penalisation performs selection and shrinkage of the coefficients simultaneously; the one-standard-error rule favours parsimony over apparent performance; and the 60% stability threshold applied across the ten imputed datasets ensures that a variable does not owe its inclusion to a particular draw [41].

Firth regression answers a specific difficulty: scalp swelling was associated with a critical trajectory in all 12 patients concerned, a quasi-complete separation configuration that makes ordinary maximum likelihood diverge. Penalisation by the Jeffreys term makes the estimate finite and corrects the second-order bias of classical estimators, a bias that is substantial when events are rare [42][43].

The bootstrap encompassing selection is more conservative than cross-validation applied after selection. Replaying the entire procedure on each resample is the only way to quantify the optimism induced by selection itself [44]-[46]; a validation applied to the final model alone would have underestimated it.

Rubin pooling combines the estimates from the ten imputed datasets while incorporating imputation uncertainty into the variance [37][39]. Analysing a single dataset would have produced spuriously narrow intervals; analysing complete cases would have discarded 18% of admissions because of midline shift.

4.2.2. Interpretation of Calibration

The corrected calibration slope, below 1, reflects moderate residual overfitting. This is the expected behaviour of a first derivation cohort. It implies that extreme predictions are too widely dispersed: high risks are slightly overestimated and low risks slightly underestimated. The practical consequence is not a revision of the coefficients but recalibration whenever the score is transposed [50][60].

Calibration-in-the-large close to zero indicates that the model neither systematically overestimates nor systematically underestimates the mean probability of the event. The Spiegelhalter test and the Hosmer-Lemeshow test show no significant departure from perfect calibration, and calibration by deciles remains within the expected confidence intervals across the whole range of risk.

4.2.3. Discrimination, Prevalence and Transportability

An area under the curve approaching 0.80 is usually regarded as good discrimination. It must nevertheless be recalled that, for an implementation decision, calibration and clinical utility are more informative than discrimination alone [50][55][61]. That is why we report all three.

The prevalence of a critical trajectory reaches 64.7% in our cohort, a high value reflecting the recruitment of a referral centre receiving late-presenting patients. The positive and negative predictive values reported depend directly on this prevalence and are not transposable as they stand. Likelihood ratios, by contrast, are independent of prevalence and are the quantity to be carried across to an external validation. That is why they are reported with their confidence intervals, and why their application to any local prevalence is facilitated by the Fagan nomogram (Supplementary FigureS8-3).

Three levels of transportability must be distinguished. Statistical transportability concerns recalibration of the model: it is almost always necessary and technically simple, the intercept being the first parameter to readjust. Clinical transportability concerns the preservation of discrimination in a population whose case-mix differs. Organisational transportability, finally, concerns adaptation to hospitals whose capacities differ: a centre with an identified intensive care unit does not share our operational definition of “critical care”. All three must be assessed separately at external validation [60][62][63].

4.3.1. Plausibility of the Point Scale

A score is adopted only if its content is recognisable to the clinician who has to use it. The six items retained by a wholly data-driven procedure describe the three mechanisms of decompensation set out in the introduction, which amounts to a form of construct validation.

Masseffect (3 points) reflects the first mechanism. It aggregates the volume of the collection and the surrounding oedema, whose contribution often exceeds that of the collection itself [14][15], and proved more informative than size alone: two 30 mm collections do not have the same consequence depending on their location. Hydrocephalus (2 points) represents the second mechanism; it predicts the critical trajectory beyond the mere need for drainage, which suggests that it marks a more extensive involvement of cerebrospinal fluid pathways.

Impairedconsciousness (2 points) reflects decompensation, whatever its cause. That the procedure retained it in preference to the dichotomised GCS deserves comment. The GCS remains the reference instrument for quantifying the depth of coma [52], but its inter-observer reliability is imperfect, particularly for the verbal component and in the aphasic, intubated or non-French-speaking patient [51]. The observation “this patient is not normally alert” is cruder but more reproducible. For an instrument intended for the least experienced operator in the chain, this robustness has a value of its own—at the cost, as our results show, of a loss of prognostic information.

Immunosuppression (2 points) modulates the response to infection, delays encapsulation and broadens the microbiological spectrum [16][64][65]; in a region where the prevalence of human immunodeficiency virus infection remains substantial, it is a demographic determinant of the patient flow rather than a rare variable. The definition adopted is deliberately restrictive: diabetes and sickle cell disease are not included, and the item concerned 8.7% of admissions. An external validation that included them would mechanically raise the prevalence of the item and shift the calibration of the score; the definition must therefore be taken over unchanged. Raisedintracranialpressuresyndrome (1 point) makes a weak but stable contribution, its modest weighting being consistent with its prevalence of 71.3%: a sign present in three patients out of four necessarily discriminates little.

4.3.2. The “Scalp Swelling” Item

All twelve patients concerned had a severe in-hospital course. This complete separation produces a very high but very imprecise estimate, which Firth penalisation renders estimable without rendering it reliable [42][43]. Its weighting was capped at 4 points.

Clinically, the sign is coherent. Scalp suppuration with cellulitis in a patient with an intracranial suppurative infection generally reflects either an infection spreading through tissue planes by contiguity, with intercurrent osteitis, or a surgical site infection after craniotomy [58][59]. In both cases, management requires wide debridement, often removal of a bone flap or of implanted material, and prolonged antibiotic therapy. The sign therefore marks surgical complexity rather than neurological severity—precisely the information a triage instrument needs.

This finding should be regarded as a hypothesis to be confirmed and is the priority target of external validation. The decision rule is stated in advance: if the association is not reproduced in an independent cohort, the item will have to be removed and the scale recalibrated over 10 points.

4.3.3. Domain of Validity

The scope of the instrument deserves to be delimited precisely. A poorly framed triage score is quickly read as an individual prognosis, and it is by this mechanism that a methodologically sound instrument can cause clinical harm.

The score does not predict death. An area under the curve of 0.472 indicates a complete absence of discrimination. With 16 deaths, the available power is far below that required to identify a stable combination of predictors [33]-[35]: the result therefore does not mean that no variable recorded at presentation is associated with death, but that no stable combination is identifiable at this sample size. This distinction matters, because failing to make it probably contributes to the publication of fragile mortality scores that subsequently fail at external validation. An explicit restriction of use follows: TRIAGE-SIC should not be presented to a patient or to a family as an estimate of the risk of death.

A hypothesis also follows. In a cohort where death occurs once therapeutic resources have already been committed, mortality probably depends more on events occurring after triage—the actual time to theatre, the real availability of monitoring, antibiotic stock-outs, secondary complications—than on the state at presentation. This reasoning is consistent with the analysis according to which mortality attributable to the quality of health systems exceeds that attributable to lack of access [22].

The score does not replace the GCS for functional prognosis, and this gap is statistically established. The practical conclusion is simple: to discuss functional outcome with a family, the clinician uses the GCS; to organise the pathway, he uses the score.

Lastly, the score predicts early surgery poorly. This result was foreseeable: time to surgery is a decision variable, determined as much by the state of the patient as by the availability of the theatre and of the anaesthetist. A score that predicted current time to surgery perfectly would in fact be a poor triage score, since it would merely reproduce existing decisions, biases included.

The sensitivity analysis excluding patients operated on immediately supports this reasoning. In this subgroup, where the immediate surgical decision can no longer contribute to the endpoint, discrimination of the score remains 0.852 and calibration remains excellent. The score therefore does not reduce to a reproduction of the operative indication. The disappearance of immunosuppression in favour of GCS ≤ 8 and midline shift in a selection conducted de novo in this subgroup does, however, recall the instability of selection expected at this sample size, and underlines that the point scale must not be read as identifying the “true” causal factors.

4.3.4. Clinical Relevance and Use

TRIAGE-SIC is designed to support immediate triage decisions in resource-limited hospitals. It does not replace clinical judgement: it provides a standardised estimate of the risk of an unfavourable course, allowing a more rational allocation of neurosurgical resources and explicit communication between neurosurgeons, emergency physicians, anaesthetists and intensive care physicians.

Use proceeds in five steps: perform the unenhanced computed tomography scan, tick the six items, add up the total, identify the band, apply the suggested management (Figure 7, Table 5). Three clinical vignettes illustrate application of the score in Table 9.

Table 9. Three clinical vignettes illustrating use of the score.

These vignettes are constructed for teaching purposes and do not correspond to real patients. They illustrate the mechanics of the point scale, not a therapeutic recommendation: the suggested management remains subordinate to clinical judgement.

Three uses are conceivable without waiting for external validation, because they carry no risk for the patient: documenting the score on admission as a variable describing severity; using it as a formalised argument when negotiating an operating slot; and using it in teaching, including with the doctors of the peripheral facilities that refer the patients.

The expected organisational consequences—better theatre scheduling, anticipation of monitoring beds, fewer unjustified transfers—were not measured here and constitute the endpoints of a future impact study [66].

Three publications from 2015 structured the field of global surgery: the EssentialSurgery volume of the Disease Control Priorities [67], the Lancet Commission on Global Surgery [17][68] and Resolution WHA68.15 [13]. They put an end to the view of surgery as the “neglected stepchild of global health” [69]. Applied to neurosurgery, this analysis quantified a major shortfall in essential operations [12][19][20][70], whose translation into public policy runs through national surgical, obstetric and anaesthesia plans [21][71]—in which neurosurgery remains under-represented [72]. In this respect, intracranial suppurative infections are a sentinel condition: preventable upstream, curable downstream, and dependent on nothing but the speed of access to a technically simple procedure.

The score fits into this framework along three lines.

Resourceallocation. Decision curve analysis shows that the advantage of the score over a policy of systematic referral grows with the scarcity of the resource. The instrument is therefore most useful where it is most needed. In a department sharing a theatre with all the surgical specialties, it provides an objectifiable justification for the priority requested—an argument that, in practice, carries more weight than a subjective appraisal. Beyond the bedside, describing a cohort by its distribution of scores rather than by its mortality alone allows the resource burden to be compared between centres independently of their capacities, and hence a centre receiving more severely ill patients to be distinguished from a centre obtaining poorer results at equal severity. This descriptive function would give national plans a specifically neurosurgical indicator, where the indicators of the Lancet Commission remain general-purpose [17][21].

Equityofaccess. Triage that is not formalised is not neutral triage. In the absence of an explicit criterion, priority is set by factors that are not clinical: time of arrival, insistence of the accompanying relative, familiarity of the referring doctor with the team, ability to pay up front. This last point is not anecdotal in a system where out-of-pocket payment remains dominant and where catastrophic health expenditure is one of the indicators of the Lancet Commission [17][68]. An explicit score, displayed and documented, makes the decision auditable and shifts the discussion towards clinical criteria. This argument of distributive justice and decisional transparency is independent of the statistical performance of the instrument.

Essentialneurosurgery. The availability constraint was imposed before the analysis rather than observed afterwards. The six items require a neurological examination and an unenhanced computed tomography scan; none requires magnetic resonance imaging, microbiology, a biomarker, a computer algorithm or a network connection. This parsimony is a design specification, aligned with the definition of essential neurosurgery [20][70]. As the six items belong to the common clinical vocabulary of the subregion, a West African multicentre external validation is technically feasible from data already collected.

European series are the methodological reference in this field. The meta-analysis by Brouwer and colleagues places mortality at around 10% in series published after 2000, with a favourable outcome in about two thirds of survivors [2]. The prognostic determinants identified differ, however, from one cohort to another. The British series identifies impaired consciousness on admission [59]; the prognostic analysis of the Danish national cohort—whose incidence and mortality were described separately [73]—retains intraventricular rupture, immunosuppression, age over 65 years and a collection diameter over 30 mm, without impaired consciousness appearing among them [74]. Asian and Turkish series place impaired consciousness first and add intraventricular rupture [75]-[78]. This heterogeneity deserves note: three of the four factors of the Danish series describe the patient’s background or the mechanical repercussion of the collection rather than the depth of coma, which chimes with the dissociation we observe between resource needs and individual neurological prognosis. North American series historically established the transformative effect of computed tomography and remain the reference for the choice of therapeutic modality [4][5][79].

African series are the most relevant comparator. They report mortality of between 10% and 25% [7]-[11][80]. The South African series of Nathoo and colleagues, the largest published in the computed tomography era on the continent, shows that preoperative neurological deterioration governs outcome more than the organism or the location does [11][80]. Series from French-speaking West Africa describe a younger population, a predominance of otogenic and post-traumatic forms, and decisive delays in access to imaging [9][10]; Ghanaian and Nigerian work stresses the frequency of presentation with established raised intracranial pressure [7].

Three observations emerge from this comparison. The clinical picture on admission is remarkably homogeneous from one African country to another, which argues in favour of the transportability of the score. Mortality, by contrast, varies by a factor of two and a half without initial severity fully explaining it, which points to a system effect. Finally, none of these series produced a decision instrument, which leaves each team to rebuild its prioritisation rules empirically; the positioning of TRIAGE-SIC relative to existing instruments is detailed in Supplementary Table S7-3.

The differences from high-income series concern delays and microbiology more than mortality. Our median interval of 19 days between first symptoms and admission is far longer than European medians, which are counted in days [73][74][81]. Symmetrically, our proportion of sterile cultures far exceeds that of series where microbiological yield reaches 70% when the sample precedes antibiotic therapy [64][65][78]. These two differences define the context of use of the score and empirically justify the availability rule imposed on the predictors.

Several normative texts frame the management of intracranial suppurative infections: reference syntheses [3][16] and the European ESCMID guidelines for brain abscess [64], the Infectious Diseases Society of America having published a dedicated guideline only for healthcare-associated ventriculitis and meningitis [82]. All converge on a three-stage architecture: cross-sectional imaging without delay, drainage or excision of collections larger than 2.5 cm or responsible for mass effect, then prolonged antibiotic therapy guided by the intraoperative sample.

Three observations follow from comparing the score with this framework.

Our point scale is consistent with the recommended surgical threshold but expresses it differently. Our procedure retained only the mechanical criterion, size ≥ 30 mm failing to reach the stability threshold despite a strong univariable association. In a late-presenting population where median size already reaches 36 mm, almost every patient crosses the dimensional criterion, which thereby loses its discriminating power. A threshold derived from a population with short delays may thus fail in a population with long delays.

The guidelines then presuppose a resource that our setting does not guarantee. They state what should be done, not in what order to do it when two eligible patients present on the same night for a single theatre. TRIAGE-SIC therefore does not replace them: it operates in a decision space they do not cover.

The place of unenhanced computed tomography deserves, finally, to be made explicit. The guidelines favour diffusion-weighted magnetic resonance imaging for the differential diagnosis between abscess and necrotic tumour, an undisputed superiority [3][16]. But the triage question is not the differential diagnosis question: once suppuration has been established, the three pieces of information that determine urgency are accessible without contrast. Designing the score around this constraint is not a degraded compromise but the matching of the instrument to the decision it serves.

A more complex model—random forest, gradient boosting, neural network—could achieve better discrimination on these data. This comparison was conducted separately in the same cohort and is the subject of a distinct paper: at this sample size, algorithmic complexity does not improve prediction. We deliberately chose here an instrument calculable in the head. The systematic review by Warman and colleagues shows that most machine learning models published in neurosurgery are single-centre and that fewer than one in six undergo external validation [83]. An opaque model, dependent on computing infrastructure and hard to verify at the bedside, would meet the stated constraint less well than a six-item point scale calculable in the head. Transparency and interpretability are not concessions here: they are conditions for adoption and for external validation. Future work may compare TRIAGE-SIC with these approaches, provided that discrimination, calibration and clinical utility are assessed jointly.

This work brings together four features that, to our knowledge, are not found together in the literature of this field: a neurosurgical triage score developed in an African population; an endpoint centred on resource allocation rather than on survival; internal validation encompassing the entire construction procedure, selection included; and reporting compliant with TRIPOD and TRIPOD + AI, together with publication of the full equation and of the code.

Three methodological points may be added. The availability constraint on the predictors was imposed before the analysis, guaranteeing that the score can actually be used in the settings concerned. Internal validation replays the entire procedure, selection included, which avoids the substantial optimism introduced by a validation applied after selection [44]-[46]. Calibration is reported at every recommended level, including by formal tests and by deciles [48]-[50]. The negative analyses and the sensitivity analyses were pre-specified, and the former are reported on the same footing as the positive results. Reporting follows TRIPOD and TRIPOD + AI [26]-[28]. Finally, the protocol, the full code, the complete equation and the implementation tools are provided: the whole analytical pipeline—pre-processing, multiple imputation, selection, estimation, bootstrap validation and implementation of the score—is reproducible as it stands.

Samplesizeandpower. One hundred and fifty admissions is a modest sample. The composite endpoint produced 97 events, about 16 events per retained predictor, which exceeds contemporary minimum recommendations for penalised modelling [34][35][84]. The sample nevertheless remains insufficient for the individual targets, and plainly insufficient for death (16 events). External validation remains indispensable before any clinical implementation.

Singlecentreandabsenceofexternalvalidation. This is a single-centre, predominantly retrospective study. The performance reported, even corrected for optimism, will probably be lower in another centre.

Absenceofinternaltemporalvalidation. The cohort spans ten years, over which service provision and practice have changed. We did not carry out development on the earlier period followed by evaluation on the recent period, the sample size not allowing the cohort to be split without compromising the stability of selection. Performance on the most recent admissions, which correspond to the future context of use, therefore remains unknown.

Datadriftovertenyears. Beyond the absence of temporal validation, a gradual shift in the distribution of the predictors and in the predictor-endpoint relation is plausible over such a period. This phenomenon is the main mechanism of failure of prediction models deployed in clinical practice [85] and could not be quantified here.

Absenceofgeographicalvalidation. No evaluation in a separate centre. Bouaké Teaching Hospital receives a predominantly rural and referred population, with long delays. A centre in Abidjan, Dakar or Johannesburg would have a different profile, with a probably lower prevalence of the composite, which would mechanically shift the predictive values.

Selectionbias. The cohort consists of patients who reached the neurosurgery department of a teaching hospital. Patients who died before transfer, were directed elsewhere, or never obtained imaging do not appear in it. The score therefore describes risk conditional on admission to a referral centre, not the risk of the disease in the population.

Treatmentindicationbias. The most severely ill patients received the heaviest treatments, and critical-care use—a component of the primary endpoint—is partly a consequence of the medical decision rather than an independent event. This bias is intrinsic to a composite endpoint incorporating a process variable and cannot be neutralised in a retrospective dataset. A sensitivity analysis restricted to the more patient-centred endpoint of death or poor functional outcome, excluding critical-care use, preserved useful discrimination (optimism-corrected AUC 0.817), although selection then shifted towards GCS-defined neurological severity—which supports reading TRIAGE-SIC as a triage rather than a purely prognostic instrument.

Inter-observerreproducibilitynotassessed. The six items, particularly impaired consciousness and mass effect, rest on an appraisal whose agreement between assessors was not measured. A dedicated agreement study is a prerequisite for the impact study.

Proxyforthe“criticalcare”target. A therapeutic definition in the absence of an intensive care admission variable, influenced by prescribing habits and not transposable as it stands to a centre with an identified unit.

Fixedimputationwithinthebootstrap. The imputations were generated once and held fixed across bootstrap resamples rather than regenerated within each resample. This pragmatic choice, adopted because the fraction of missing data was low, may make the optimism-corrected confidence intervals marginally narrower than those a fully nested imputation-in-bootstrap procedure would yield; the reported intervals should be read with that caveat.

Otherlimitations. The “early surgery” target is a decision and not an event. Microbiological sampling was not performed in every patient and culture conditions varied. Two patients each contribute two admissions; an analysis restricted to the first admission did not change the variables selected. Finally, the composite endpoint must not be broken down into specific predictions of each of its components, as the failure to predict death demonstrates.

We propose a seven-step programme, the first three steps of which are feasible in the short term:

1) Ivorianexternalvalidation—an independent cohort from one or two national centres, with recalibration of the intercept.

2) WestAfricanvalidation—five to ten centres, with a sample size calculated a priori according to the methods dedicated to models with a binary endpoint rather than set for convenience [86]-[88]; as an indication, about 500 patients would allow the calibration slope to be estimated with useful precision. The evaluation will address discrimination, calibration, the recalibration required and clinical utility jointly.

3) ValidationinFrench-speakingandthensub-SaharanAfrica—geographical extension and assessment of organisational transportability.

4) Inter-observeragreementstudy of the six items, a prerequisite for any impact study.

5) Prospectiveimpactstudy, with the admission-to-theatre interval of high-risk patients as the primary endpoint.

6) Clusterrandomisedtrial comparing formalised triage with usual practice.

7) Comparisonwithmachinelearningmodels and digital deployment, once external validity has been established.

Prospective data collection including process variables will also have to test the hypothesis that mortality depends more on the system than on the state at presentation. TRIAGE-SIC should be regarded as version 1.0 of an instrument that is meant to evolve.