Age and Early Death: What Three Years of Stroke Admissions Reveal in Luanda

At a public tertiary hospital on the eastern edge of Luanda, nearly one in six patients admitted with a stroke never leaves alive. A new retrospective cohort study from Hospital do Prenda, published in Frontiers in Neurology, has combed through three years of routine admission records to ask a deceptively simple question: which characteristics, available at the bedside, travel with in-hospital death? The answer, drawn from 479 hospitalisation episodes between January 2023 and December 2025, is stark. Age stands out as the only admission-available variable strongly and consistently associated with dying, and the pattern is not the smooth staircase clinicians might expect. Mortality climbed from 6.7 percent among patients aged 45 to 54 to 28.0 percent among those aged 75 and older, yet it was also strikingly elevated, at 17.2 percent, among patients younger than 45, a group in whom the excess deaths were concentrated almost entirely in haemorrhagic stroke.

The headline number, 75 deaths among 479 episodes, corresponds to an in-hospital mortality of 15.7 percent, with a 95 percent confidence interval of 12.7 to 19.2 percent. But the timing of those deaths may be the most consequential finding for anyone thinking about stroke systems in sub-Saharan Africa. Nearly half of the deaths, 44.0 percent, occurred within the first 24 hours of admission. By day three, 58.7 percent of the eventual deaths had happened, and by day seven the figure reached 85.3 percent. Median length of stay was six days among survivors against three days among those who died, a difference the authors explicitly frame as a consequence of early mortality rather than a predictive exposure. When patients are dying this fast, the window in which hospital-based interventions could plausibly alter the trajectory is measured in hours, not days.

That compressed timeline mirrors what earlier Angolan studies have documented about how patients reach care. A cross-sectional study from another Luanda hospital reported a median onset-to-door time of 14 hours, with only 10.8 percent of patients arriving within three hours of symptom onset, well inside the 4.5-hour thrombolysis window. The multicentre INTERSTROKE Angola case-control study found that only 27.4 percent of stroke cases achieved a favourable Glasgow Outcome Scale score of 5 at discharge. Hospital do Prenda, which receives neurovascular referrals from surrounding municipalities but has no dedicated stroke unit and no dedicated neuro-intensive care unit, sits at the end of that delayed chain. When most patients arrive beyond the point where reperfusion therapy is possible and critical-care escalation capacity is limited, deaths concentrate in the first days of admission, exactly the pattern the registry reveals.

The statistical architecture of the study is careful to a degree that rewards attention. The researchers fitted three exploratory multivariable logistic regression models, adjusting for age, sex, admission year, stroke subtype, recorded comorbidities and head-CT status. With only 75 deaths distributed across multiple covariates, the events-per-variable ratio sat at roughly seven, and the Hosmer-Lemeshow test flagged poor calibration in the smaller complete-case models. The authors therefore present their adjusted estimates as exploratory associations rather than independent prognostic effects, a distinction that matters for how the findings should be used. In the model using age bands on the full cohort, which achieved a c-statistic of 0.649 and acceptable calibration, the gradient was pronounced: compared with the 45-to-54 reference band, the adjusted odds ratio for death was 2.97 for patients under 45, 2.90 for ages 55 to 64, 3.59 for ages 65 to 74, and 5.92 for those 75 and older.

The non-monotonic shape of that gradient, with mortality dipping in middle age and rising again in the youngest band, prompted a subgroup analysis that produced one of the study’s most intriguing signals. Among patients younger than 45, 22.2 percent of haemorrhagic stroke episodes ended in death, against 10.8 percent of ischaemic episodes. A likelihood-ratio test of the age-by-subtype interaction returned a p-value of 0.0009, suggesting the age pattern genuinely differs between the two major stroke mechanisms. The authors offer two non-exclusive explanations. Younger Angolan stroke patients may present with more severe syndromes, carrying a heavier haemorrhagic load and more malignant middle-cerebral-artery events, consistent with observations from the earlier Luanda cross-sectional study. Alternatively, the effect could be a small-denominator artefact, in which a handful of very early deaths in a modestly sized group inflates the proportion. The registry, lacking admission severity scores, cannot adjudicate between these readings.

Equally instructive is what the study did not find. Sex showed no association with mortality, with 15.7 percent among women and 15.2 percent among men. Stroke subtype overall, comparing haemorrhagic with ischaemic cases, returned a null result, with crude mortality of 14.5 percent versus 16.0 percent. Recorded hypertension and recorded diabetes were likewise unassociated with death. But the authors are emphatic that these nulls cannot be read as evidence of absence. The comorbidity data came from a single free-text line on the discharge summary, extracted using predefined string rules rather than natural language processing, and the field was non-empty in only 33.0 percent of episodes. Recorded hypertension reached just 19.8 percent of the cohort, against roughly 88 percent case prevalence measured by INTERSTROKE Angola. In other words, approximately four of five hypertensive patients in this catchment carry no recorded comorbidity in the registry, and such misclassification biases observed associations toward the null.

Two ascertainment-related signals reinforce that caution. Episodes with any comorbidity recorded had higher mortality than those without, 20.9 percent versus 13.1 percent, a pattern the authors interpret as the field being completed more often for patients who were more thoroughly investigated or more severely ill. Episodes without a head CT, 4.2 percent of the cohort, had double the mortality of those scanned, 30.0 percent versus 15.0 percent, consistent with imaging being withheld from the most severe or most rapidly deteriorating patients. Both mechanisms also help explain the surprising haemorrhagic-versus-ischaemic null: the unclassified category, with 24.0 percent mortality, almost certainly absorbs the most severe haemorrhagic cases who die before reaching CT, simultaneously raising the unclassified death rate and stripping the highest-mortality events out of the haemorrhagic denominator. What looks like biology in a crude table can, on inspection, be a story about documentation.

The transparency of the analysis is itself notable. The registry was assembled for clinical administration, not research, and captures only administrative fields, discharge diagnosis and a one-line comorbidity note. It does not record the NIH Stroke Scale, the Glasgow Coma Scale on admission, pre-stroke functional status on the modified Rankin Scale, admission glucose, ICU admission, mechanical ventilation or thrombolysis use. The authors treat these absences as explicit boundaries of inference, repeatedly stating that no variable’s independent prognostic effect can be estimated while holding admission severity and acute treatment constant. They also published the anonymised episode-level dataset and the complete Python analysis code as supplementary material, so that every figure and odds ratio in the paper can be independently reproduced, and they report the study against the STROBE checklist for observational research. A data-quality audit identified and explicitly handled one duplicate episode, two episodes with comorbidity text entered in the outcome field and three date anomalies, none of which changed the primary results.

Why should readers far from Luanda care? Because the prognostic scores most clinicians reach for, including iScore, PLAN and the Weimar NIHSS-derived model, were built on North American, European and Australasian cohorts whose case-mix and care pathways bear little resemblance to a hospital where patients arrive a median of 14 hours after onset, nearly half of strokes are haemorrhagic, and no stroke unit exists. The Global Burden of Disease 2019 study placed stroke at roughly 12 percent of all deaths worldwide, with most of the absolute increase since 1990 occurring in low- and middle-income countries, and forecasts expect the gap to widen through 2050. Studies like this one are the raw material for building locally valid prediction and, more importantly, for making the structural case that the authors close on: routine documentation of admission severity, structured coding of comorbidities and a national Angolan stroke registry harmonised with its sub-Saharan neighbours are overdue. Until then, age at the door, and the clock that started long before it, remain the strongest signals of who will not survive the admission.

Subject of Research: In-hospital mortality after stroke at a tertiary hospital in Angola

Article Title: Factors associated with in-hospital mortality after stroke at a tertiary hospital in Angola

Article References: Neto, S., Cuenda, B., Major, M. N., Kapele, J., & Pinto, C. (2026). Factors associated with in-hospital mortality after stroke at a tertiary hospital in Angola. Frontiers in Neurology, 17, Article 1944436. https://doi.org/10.3389/fneur.2026.1944436

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1944436

Keywords: stroke, in-hospital mortality, Angola, Luanda, sub-Saharan Africa, haemorrhagic stroke, ischaemic stroke, hypertension, registry study, retrospective cohort, thrombolysis window, prognostic factors