Context-specific predictors of mortality and the performance of paediatric sepsis scores in Tanzanian children: a secondary analysis of a prospective cohort study

Abstract

Introduction

Sepsis is a significant cause of global child mortality, disproportionately affecting resource-limited settings (RLS). Although clinical scores have been proposed in RLS, few have been externally validated. This study aimed to identify risk factors for mortality among Tanzanian children with sepsis and test three paediatric clinical scores in an RLS.

Methods

We conducted a secondary analysis of a prospective cohort study of 755 children (28 days to 14 years) with sepsis in the Emergency Department at Muhimbili National Hospital, Tanzania (July 2022–November 2024). Demographic, clinical and outcome data were collected. We performed multivariable logistic regression to identify independent predictors of in-hospital mortality. We then tested three clinical scores, Lambaréné Organ Dysfunction Score (LODS), Blantyre quick Sequential Organ Failure Assessment (BqSOFA) and Phoenix Sepsis Score (PSS), using sensitivity, specificity and area under the receiver operating characteristic curve (AUC).

Results

In-hospital mortality was 20% (n=148); independent predictors included age, malnutrition, HIV status, referral status and delayed presentation (p<0.001). All three scores demonstrated strong discrimination for mortality (LODS AUC 0.79, 95% CI 0.75 to 0.83; BqSOFA AUC 0.75, 95% CI 0.71 to 0.79; PSS AUC 0.75, 95% CI 0.70 to 0.79) and differences between scores were not statistically significant (p=0.066). LODS showed high sensitivity (93%) and negative predictive value (96%). BqSOFA and PSS demonstrated high specificity (93% and 91%) and positive predictive values (51% and 56%), but low sensitivity (28% and 44%).

Conclusions

Multiple clinical and sociodemographic factors were independently associated with mortality in children with sepsis. LODS demonstrated high sensitivity, while BqSOFA and PSS demonstrated high specificity; the scores’ discriminative ability for mortality was not statistically significant. The identification of simple risk factors and moderately performing clinical scores provides a valuable opportunity to develop context-specific risk stratification tools for paediatric sepsis in RLS.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Paediatric sepsis significantly impacts children in resource-limited settings. While paediatric clinical scores have been shown to guide patient risk stratification and significantly improve outcomes in high-resource countries, they often require laboratory or monitoring data that are unavailable in resource-limited settings. Several clinical scores have been developed for resource-limited settings, but comparative external validation is limited.

WHAT THIS STUDY ADDS

• We provide an external validation of three commonly used paediatric clinical scores and demonstrate significant predictors of mortality in a cohort of Tanzanian children with sepsis.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Our results may guide the development of future paediatric sepsis clinical scores, which should adapt existing tools to include simple, context-relevant clinical and system factors (malnutrition, referral status, delayed presentation) rather than creating wholly new scores.

Introduction

Sepsis is defined as life-threatening organ dysfunction caused by a dysregulated immune response related to infection.¹ In 2021, it was estimated that three million sepsis-related deaths occurred globally in the paediatric population.² Paediatric sepsis remains a leading cause of mortality worldwide, with a disproportionate number of both cases and deaths in sub-Saharan Africa and South Asia. In these resource-limited settings (RLS), late presentation, delayed recognition, high prevalence of malnutrition and scant critical care resources, including workforce shortages, all contribute to sepsis-related mortality.^3–6

Timely recognition and treatment of sepsis are crucial for improving sepsis outcomes.⁵ In poorly resourced health systems, clinical scores can be used to triage, guide clinical management and improve resource allocation. These clinical scores need to be context-specific and should rely on readily available clinical data rather than on laboratory support, which is difficult or impossible to use in RLS.^7–10 Thus, identifying readily available demographic and clinical risk factors for mortality relevant to RLS and validating clinical scores for these settings may help guide triage, early intervention and the deployment of resources to those most likely to benefit. Several clinical scores have been developed for children in RLS, but have not been widely validated.¹¹

The Lambaréné Organ Dysfunction Score (LODS) is a clinical score initially developed to predict early death in African children with malaria and subsequently validated to predict mortality in children with severe systemic infections in RLS.^{12 13} The Sequential Organ Failure Assessment (SOFA) score was proposed in 2016 to assess the degree of organ dysfunction in adults with sepsis. It was then modified to an age-adjusted quick SOFA (qSOFA) score, which has predictive value for all-cause mortality in critically ill paediatric patients with sepsis in HRS.^14–16 The Blantyre qSOFA (BqSOFA) score is an adaptation of qSOFA, which eliminates the need for invasive monitoring or laboratory testing and has been shown to predict mortality in children with sepsis.¹⁷ While LODS and BqSOFA were developed to predict mortality in children with sepsis in RLS, they have not been externally validated, which limits their broad implementation.^{11 16 18}

The Phoenix Sepsis Score (PSS) was developed in 2024 to improve the diagnosis of paediatric sepsis and septic shock by identifying organ dysfunction, which is associated with a significantly increased risk of mortality.¹⁹ The PSS included data from three RLS sites with electronic medical records and, as a result, does not represent most RLS cases, where laboratory support and advanced care and monitoring are rarely available. Thus, we test the utility of LODS, BqSOFA and PSS for the early identification of high-risk children with sepsis in RLS.

To identify the children at highest risk for mortality, this study aimed to identify demographic and clinical factors associated with in-hospital mortality among paediatric patients with sepsis at Muhimbili National Hospital (MNH) in Dar es Salaam, Tanzania, and to evaluate the predictive performance of three clinical scores (LODS, BqSOFA and PSS) for predicting child sepsis mortality via secondary analysis. Together, these aims support efforts to improve context-specific risk stratification, triage and clinical decision-making in RLS.

Materials and methods

Study design and setting

This is a secondary analysis of a prospective cohort study of paediatric patients with sepsis presenting to the Emergency Medicine Department (EMD) at MNH in Dar es Salaam, Tanzania. Data were collected from 26 July 2022 to 26 November 2024. MNH is a national referral hospital and the country’s largest tertiary care centre, serving as a key academic and clinical training site affiliated with the Muhimbili University of Health and Allied Sciences (MUHAS). The hospital receives referrals from across the country, making it a critical entry point for acutely ill children from both urban and rural settings.

Study population

Children aged 28 days to 14 years who presented to the MNH EMD with sepsis were consecutively screened. Sepsis was defined as fever (axillary temperature >38°C or history of fever) accompanied by one or more of the following WHO warning signs: difficulty breathing, altered mental status, impaired perfusion or at risk for dehydration.²⁰ Difficulty breathing was defined as tachypnoea for age, increased work of breathing or hypoxia (peripheral oxygen saturation (SpO₂)>90%). Altered mental status was defined as a score of ‘Verbal’ or lower on the Alert-Verbal-Pain-Unresponsive scale or a report of seizure within 24 hours in a child without known epilepsy. Impaired perfusion was defined as the presence of at least one of the following: hypotension for age, cold extremities or severe tachycardia (online supplemental table 1).^21–25 Risk of dehydration was determined based on caregiver report of loose stools, emesis, reduced urine output or inability to drink or breastfeed.^{26 27} Exclusion criteria included patients with acute trauma or burn injuries, cardiac arrest on arrival, weight <4 kg, non-infectious fever (eg, rheumatologic aetiology), known epilepsy with isolated convulsion, caregivers unable to communicate in Kiswahili or English and those whose guardians were not available or declined to participate. Informed consent was obtained from the parent or guardian before patient enrolment in the study.

Data collection

Trained research personnel collected demographic data, vital signs, comorbidities (including HIV, malaria, tuberculosis and malnutrition), socioeconomic indicators (eg, maternal education), clinical interventions and the patients’ course of illness from initial triage to hospital discharge or death. Delayed presentation to care was defined as a fever lasting more than 2 days before presentation, regardless of prior treatment. HIV positivity was defined as a known history of HIV or a positive screening rapid diagnostic test (RDT) (HIV-1/2 Alere) with confirmatory testing. Malaria infection was diagnosed by RDT (SD Bioline pan-Pf). Data were obtained from direct observation, parental inquiry and chart review. All patients were followed throughout their hospital stay or until death. Relevant laboratory values were obtained from point-of-care testing or hospital laboratory records. Data were entered into REDCap, an electronic data capture tool hosted at MNH, and de-identified before analysis.²⁸

Outcome

The primary outcome was in-hospital mortality, defined as death occurring any time from EMD presentation through the hospital admission.

Clinical scores

We tested three clinical scores via secondary analysis: LODS, BqSOFA and PSS.

Lambaréné Organ Dysfunction Score

LODS was calculated using published and validated criteria.¹² LODS includes three clinical variables: altered mental status, prostration and deep breathing (box 1).¹² A score greater than zero was considered high risk, based on initial studies showing an 85% sensitivity in predicting death.¹² LODS was initially proposed as a mortality prediction tool for children with malaria, but it has since been validated for predicting mortality in febrile illness.^{12 13}

Box 1: Clinical scores

Lambaréné Organ Dysfunction Score¹²

This score assigns one point for each item:

• Prostration: defined as ≥1 of the following, depending on age: inability to breastfeed, sit, stand or walk unsupported.

• Coma: Blantyre Coma Score ≤2.

• Deep breathing (difficulty breathing, retractions, head bobbing).

Blantyre quick Sequential Organ Failure Assessment¹⁷

This score assigns one point for each item:

• Abnormal respiratory rate (90th percentile of age-specific thresholds).^{17 29}

• Altered mental status.

• Delayed capillary refill.

• Pallor.

Phoenix Sepsis Score¹⁹

This score assigns points across four organ dysfunction domains:

Respiratory dysfunction

• PaO₂/FiO₂ or SpO₂/FiO₂ ratio below defined thresholds.*

• Oxygen supplementation.

• Non-invasive positive pressure ventilation.

• Invasive mechanical ventilation.

Cardiovascular dysfunction

• Use of vasoactive or inotropic support.

• Elevated lactate level.

• Age-adjusted hypotension.

Coagulation dysfunction

• Thrombocytopenia.

• Elevated INR.**

• Elevated D-dimer.**

• Low fibrinogen.**

Neurologic dysfunction

• Glasgow Coma Scale ≤10.

• Abnormal pupillary exam.**

*FiO2, fraction of inspired oxygen; PaO2, arterial oxygen partial pressures; PaO2/FiO2, ratio of arterial oxygen partial pressure to fraction of inspired oxygen; SpO2, peripheral oxygen saturation; SpO2/FiO2, ratio of peripheral oxygen saturation to fraction of inspired oxygen.

**Not used to calculate the final score.

Blantyre quick Sequential Organ Failure Assessment

The BqSOFA score is an adaptation of the qSOFA score designed to identify critically ill children in RLS. It assigns one point for each component (Box 1).¹⁷ We considered scores of 2 or greater to be high risk, in accordance with recommendations from the initial validation study.¹⁷

Phoenix Sepsis Criteria Score

The composite PSS includes four organ dysfunction domains: respiratory, cardiovascular, coagulation and neurologic (Box 1).¹⁹ Consistent with the original PSS publication, missing components were coded as normal.¹⁹ A PSS of two or greater was considered high risk, in accordance with guideline recommendations.¹⁹ Because international normalised ratio, D-dimer and fibrinogen were not routinely measured locally, we excluded them from the coagulation domain, leaving one coagulation component. Fixed pupils were not routinely documented and were therefore excluded from the neurologic domain, such that no patients could receive a neurologic score of 2. Given limitations in data availability, we additionally performed a sensitivity analysis in which the PSS was recalculated for patients who had all of the locally available PSS data recorded. While the initial PSS was calculated within the first 24 hours of hospital care, some variables used to calculate PSS (eg, platelet count) fell outside this window.

Statistical analysis

We used descriptive statistics to summarise the baseline characteristics of the cohort. We presented continuous variables as means (SD) or medians (IQRs) and categorical variables as counts and percentages.

We conducted bivariate analyses to assess for associations between patient characteristics and mortality using χ² tests for categorical variables and t-tests or Wilcoxon rank-sum tests for continuous variables, as appropriate. We determined covariates a priori based on our prior work^{29 30} and literature review^{3 31} and included age, sex, nutritional status, comorbidities (including HIV and malaria), vaccination status, maternal educational level, referral status and the duration of illness prior to presentation.^{13 21 22 29 32} Variables significantly associated with in-hospital mortality were included in multivariable regression models to identify independent predictors of mortality. Multicollinearity among candidate predictors was assessed using variance inflation factors (VIFs), and model fit was evaluated using the Hosmer-Lemeshow goodness-of-fit test.

To validate the clinical scores, we evaluated the predictive performance of LODS, BqSOFA and PSS for in-hospital mortality. Associations between each score (and its individual components) and mortality were assessed using unadjusted logistic regression models. All score performance metrics were derived from unadjusted analyses. We calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and the area under the receiver operating characteristic (ROC) curve (AUC) with 95% CIs for each score. Primary analyses used previously published score thresholds (LODS >0, BqSOFA ≥2, PSS ≥2).^{12 17 19} We additionally evaluated locally optimised thresholds derived using Youden’s index, which identified the threshold that maximised the balance between sensitivity and specificity. We conducted a sensitivity analysis restricted to children with complete data for all PSS components (n=403). As an additional sensitivity analysis, we excluded children enrolled solely on the basis of risk of dehydration (without cardiovascular, neurologic or respiratory dysfunction, as defined by inclusion criteria) to evaluate score performance among a higher-acuity subgroup, without restricting the score components themselves. Differences in AUCs between scores were formally compared using DeLong’s test for correlated ROC curves. The proportion of participants experiencing mortality was stratified by increasing score categories to assess dose-response relationships.

We additionally generated forest plots depicting unadjusted ORs for each clinical sign included in the scores, as well as for the total scores. A p value <0.05 was considered statistically significant for all tests. Analyses were conducted using Stata V.18 (StataCorp).

Reporting guidelines

We followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines (online supplemental file 2).³³

Patient and public involvement

Neither patients nor the public were involved in the design, conduct, recruitment or reporting of this study.

Results

We screened 15 280 patients, enrolled 756 and followed 755 until hospital discharge or death (figure 1). We observed an in-hospital mortality of 20% (n=148). Table 1 presents the baseline characteristics of patients included in our study, along with comparisons by outcome. The median age was 1.4 years (IQR 0.7–3.5), 56% (n=424) were male and 57% (n=432) had delayed presentation to care (table 1). Among study participants, a total of 64% (n=481) had a high-risk LODS score, 11% (n=81) had a high-risk BqSOFA score and 15% (n=117) had a high-risk PSS score.

Figure 1. Flow chart depicting participant enrolment and clinical outcomes for Tanzanian children with sepsis.

Table 1. Patient characteristics and prehospital illness course in Tanzanian children with sepsis by survival status.

Patient characteristics	Total n (%) (n=755)	Survivors n (%) (n=607)	Non-survivors n (%) (n=148)	P value
Age				<0.001
0–12 months	272 (36)	203 (33)	69 (47)
1–2 years	181 (24)	140 (23)	41 (28)
2–5 years	174 (23)	157 (26)	17 (12)
>5 years	128 (17)	107 (18)	21 (14)
Male sex	424 (56)	340 (56)	84 (57)	0.87
Nutritional status
Moderate acute malnutrition	114 (15.1)	84 (14.0)	30 (20.8)	<0.001
Severe acute malnutrition	109 (14.4)	71 (11.0)	38 (26.4)	<0.001
Comorbidities
HIV	12 (2)	5 (1)	7 (5)	<0.001
Malaria	47 (6)	36 (6)	11 (7)	0.44
Fully vaccinated	738 (98)	593 (98)	145 (98)	0.78
Maternal education level				0.005
No school or primary school	276 (37)	207 (34)	69 (47)
Secondary or postsecondary school	479 (63)	400 (66)	79 (53)
Referred from outside	411 (54)	280 (46)	131 (89)	<0.001
Referral location				0.051
From clinic	8 (1)	8 (3)	0 (0)
From hospital	403 (53)	272 (97)	131 (100)
Duration of illness				<0.001
Fever ≤2 days	323 (43)	284 (47)	39 (26)
Fever >2 days	432 (57)	323 (53)	109 (74)
Antibiotics before arrival	393 (52)	284 (47)	109 (74)	<0.001

In unadjusted analyses of the individual components of the clinical scores, the three components with the strongest association with mortality were abnormal respiratory rate (OR 16.83, 95% CI 1.87 to 151.74), deep breathing (OR 9.04, 95% CI 6.01 to 13.58) and altered mental status (OR 8.23, 95% CI 5.54 to 12.21) (figure 2).

Figure 2. Forest plots of unadjusted ORs for individual clinical signs and overall scores within each clinical score. ORs and 95% CIs were derived from logistic regression models with in-hospital mortality as the dependent variable. Individual signs contributing to LODS, BqSOFA and PSS are shown alongside the unadjusted ORs for the total score, illustrating the relative strength of each component in predicting mortality. The abnormal respiratory rate estimate should be interpreted with caution, due to the wide confidence interval and the small number of events (n=5). BqSOFA, Blantyre quick Sequential Organ Failure Assessment; LODS, Lambaréné Organ Dysfunction Score; PSS, Phoenix Sepsis Score.

In unadjusted logistic regression analyses, each 1-point increase in total LODS score was associated with higher odds of in-hospital mortality (OR 3.35, 95% CI 2.69 to 4.18). Similarly, each 1-point increase in BqSOFA score (OR 3.47, 95% CI 2.69 to 4.48) and PSS (OR 2.12, 95% CI 1.83 to 2.46) was associated with increased odds of mortality.

In a multivariable analysis of baseline clinical and sociodemographic factors, younger age, malnutrition, HIV infection, referral from another health facility and longer duration of illness were independently associated with in-hospital mortality (p<0.001) (online supplemental table 3). In a sensitivity analysis excluding HIV-positive participants (n=12), the direction and magnitude of the associations between mortality and the remaining predictors were similar to those observed in the primary multivariable model (online supplemental table 4). In a separate combined adjusted model including both baseline predictors and individual score components, age 2–5 years was associated with lower odds of mortality, while moderate-severe malnutrition, severe acute malnutrition and referral from another health facility remained associated with higher odds of death. Among the individual score components, altered mental status, deep breathing, abnormal respiratory rate and the respiratory and cardiovascular domains of PSS were independently associated with mortality after adjustment for baseline clinical factors (online supplemental table 5). There was no evidence of multicollinearity among predictors (all VIFs ≤1.32). Both the primary multivariable model and the combined adjusted model demonstrated adequate fit by the Hosmer-Lemeshow goodness-of-fit test (χ²=8.14, p=0.42 and χ²=5.87, p=0.66, respectively).

The prognostic performance of the three clinical scores is shown in table 2 and figure 3. LODS demonstrated an AUC of 0.79 (95% CI 0.75 to 0.83), BqSOFA demonstrated an AUC of 0.75 (95% CI 0.71 to 0.79) and PSS demonstrated an AUC of 0.75 (95% CI 0.70 to 0.79); these differences were not statistically significant (p=0.066). At the prespecified cut-offs, LODS yielded high sensitivity (92.6%, 95% CI 87.1% to 96.2%) and NPV (96.0, 95% CI 92.9 to 98.0) compared with BqSOFA (27.7%, 95% CI 20.7% to 35.7% and 84.1%, 95% CI 81.1% to 86.8%, respectively) and PSS (43.9%, 95% CI 35.8% to 52.3% and 87.0%, 95% CI 84.1% to 89.5%, respectively) (table 2). LODS had low specificity (43.3%, 95% CI 39.3% to 47.4%) and low PPV (28.5%, 95% CI 24.5% to 32.7%) compared with BqSOFA (93.4%, 95% CI 91.1% to 95.3% and 50.6%, 95% CI 39.3% to 61.9%, respectively) and PSS (91.4%, 95% CI 88.9% to 93.5% and 55.6%, 95% CI 46.1% to 64.7%, respectively). In the sensitivity analysis restricted to participants with complete locally available PSS data (n=403), the AUC was 0.72 (95% CI 0.66 to 0.78) (online supplemental table 6).

Table 2. Test characteristics of Lambaréné Organ Dysfunction Score (LODS), Blantyre quick Sequential Organ Failure Assessment (BqSOFA) and Phoenix Sepsis Score (PSS) for predicting in-hospital mortality.

Score	AUC* (95% CI)	Sensitivity % (95% CI)	Specificity % (95% CI)	PPV % (95% CI)	NPV % (95% CI)
LODS >0	0.791 (0.753 to 0.829)	92.6 (87.1 to 96.2)	43.3 (39.3 to 47.4)	28.5 (24.5 to 32.7)	96.0 (92.9 to 98.0)
BqSOFA ≥2	0.752 (0.711 to 0.793)	27.7 (20.7 to 35.7)	93.4 (91.1 to 95.3)	50.6 (39.3 to 61.9)	84.1 (81.1 to 86.8)
PSS† ≥2	0.747 (0.702 to 0.791)	43.9 (35.8 to 52.3)	91.4 (88.9 to 93.5)	55.6 (46.1 to 64.7)	87.0 (84.1 to 89.5)

LODS >0 showed very high sensitivity and negative predictive value, supporting its utility as a screening tool to identify children at low risk of in-hospital mortality. In contrast, BqSOFA ≥2 and PSS ≥2 demonstrated higher specificity and positive predictive value, suggesting better performance for identifying children at the highest risk of death.

^*AUC was calculated using the clinical scoring tools as continuous variables.

^†Missing data considered normal in score calculation.

AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value.

Figure 3. Receiver operating characteristic (ROC) curves for LODS, BqSOFA and PSS in predicting in-hospital mortality among paediatric patients with sepsis. The AUC for each scoring tool is shown, indicating moderate to good discriminatory performance for all.

When locally optimised cut-offs derived using Youden’s index were applied, test performance shifted, as demonstrated in table 3. For LODS, increasing the threshold from >0 to ≥2 reduced sensitivity from 92.6% to 65.5% while increasing specificity from 43.3% to 81.2%. Similarly, lowering the BqSOFA threshold from ≥2 to ≥1 increased sensitivity from 27.7% to 77.0%, while reducing specificity from 93.4% to 70.3%. For PSS, lowering the threshold from ≥2 to ≥1 increased sensitivity from 43.9% to 67.6% while reducing specificity from 91.4% to 74.6% (table 3).

Table 3. Performance of published and locally optimised cutoffs for evaluated clinical scores.

Score	Cutoff type	Threshold	Sensitivity % (95% CI)	Specificity % (95% CI)	PPV % (95% CI)	NPV % (95% CI)
LODS	Published	>0	92.6 (87.1 to 96.2)	43.3 (39.3 to 47.4)	28.5 (24.5 to 32.7)	96.0 (92.9 to 98.0)
	Youden-optimal	≥2	65.5 (57.3 to 73.2)	81.2 (77.9 to 84.3)	46.0 (39.1 to 52.9)	90.6 (87.9 to 92.9)
BqSOFA	Published	≥2	27.7 (20.7 to 35.7)	93.4 (91.1 to 95.3)	50.6 (39.3 to 61.9)	84.1 (81.1 to 86.8)
	Youden-optimal	≥1	77.0 (69.4 to 83.5)	70.3 (66.5 to 74.0)	38.8 (33.2 to 44.6)	92.6 (89.8 to 94.8)
PSS	Published	≥2	43.9 (35.8 to 52.3)	91.4 (88.9 to 93.5)	55.6 (46.1 to 64.7)	87.0 (84.1 to 89.5)
	Youden-optimal	≥1	67.6 (59.4 to 75.0)	74.6 (71.0 to 78.0)	39.4 (33.3 to 45.7)	90.4 (87.5 to 92.9)

BqSOFA, Blantyre quick Sequential Organ Failure Assessment; LODS, Lambaréné Organ Dysfunction Score; NPV, negative predictive value; PPV, positive predictive value; PSS, Phoenix Sepsis Score.

The availability of score components by survival status varied across the three clinical scores (online supplemental table 7). There were no missing data in the variables comprising the LODS and BqSOFA scores, resulting in 100% of participants having fully calculable scores. In contrast, PSS exhibited substantial missingness for certain variables, most notably mean arterial pressure (MAP), which was missing in 39% (n=292) of participants overall and disproportionately missing among survivors compared with non-survivors (42% (n=256) vs 24% (n=36), respectively, p<0.001). Other PSS components had minimal missingness, including lactate (9%, n=68), platelet count (5%, n=36) and GCS score (1%, n=5). Overall, only 53% of patients had available data for all components of the PSS, with fully calculable scores more common among survivors than non-survivors (66% (n=98) vs 50% (n=305), p<0.001). In an analysis evaluating the timing of PSS components, 8.1% (n=61) of participants had at least one score component obtained outside the evaluation window proxy, most commonly CBC (n=35) and coagulation (n=24) measurements. Restricting the analysis to participants without evidence of out-of-window PSS components based on available timing data (n=694) yielded results similar to those of the primary analysis, with an AUC of 0.763 (95% CI 0.72 to 0.81), sensitivity of 45.0%, and specificity of 92.6%.

Finally, in a sensitivity analysis excluding children enrolled solely for risk of dehydration (n=452), AUCs for all three scores were numerically similar to the full cohort, with overlapping CIs (LODS 0.72, 95% CI 0.66 to 0.78; BqSOFA 0.70, 95% CI 0.65 to 0.74; PSS 0.71, 95% CI 0.66 to 0.76), and differences among scores remained non-significant (p=0.50) (online supplemental table 8). LODS retained high sensitivity in this subgroup (94.2%, 95% CI 88.8% to 97.4%), though its specificity was further reduced (25.4%, 95% CI 20.7% to 30.6%) relative to the full cohort (43.3%, 95% CI 39.3% to 47.4%), consistent with the higher overall mortality rate in this higher-acuity subgroup (30.3% vs 19.6%). PSS specificity (83.5%, 95% CI 78.9% to 87.4%) and NPV (78.5%, 95% CI 73.7% to 82.8%) were lower than in the full cohort (91.4% and 87.0%, respectively), with non-overlapping CIs; BqSOFA showed a similar, though not statistically distinguishable, reduction in specificity (88.3% vs 93.4%).

Discussion

In this cohort of paediatric patients with sepsis in Tanzania, malnutrition, delayed presentation to care, referral status and HIV positivity were the most strongly associated with mortality, and LODS, BqSOFA and PSS demonstrated comparable discrimination, although their operating characteristics differed substantially.

Our findings emphasise the vulnerability of young and malnourished children, those with comorbidities, and those who have a delayed presentation to tertiary care. Notably, referral from an outside facility was among the strongest predictors of mortality (OR 7.23), suggesting mortality risk in this population may be due to delayed recognition, transfer to advanced facilities and resuscitation.⁴ Current sepsis screening tools, developed primarily in HRS, often overlook the mortality predictors and potential health system barriers we identified, underscoring the need to adapt prognostic tools to incorporate contextually relevant factors in RLS.

Among the evaluated clinical scores, LODS demonstrated high sensitivity and trended toward higher accuracy, but differences among the scores were not statistically significant. We demonstrate that LODS is a strong initial screening tool for paediatric sepsis mortality, with high sensitivity and NPV. However, its low specificity may limit its use as a standalone prognostic tool, as a large proportion of children would be classified as high-risk, potentially overwhelming limited resources and reducing the effectiveness of triage systems intended to prioritise the most critically ill. Importantly, despite differences in baseline mortality between our cohort and the original LODS derivation cohort (4% vs 20%, respectively),¹² discrimination for mortality was preserved, suggesting its utility as a clinical screening tool across varied settings. In contrast to the valuable initial screening benefit of LODS, BqSOFA and PSS demonstrated substantially higher specificity and may help identify a smaller subset of children at particularly high risk of death.

BqSOFA demonstrated high specificity and PPV but markedly lower sensitivity than in the initial development cohort.¹⁷ This pattern suggests that, while a positive score identified children at extremely high risk of mortality, the majority of deaths occurred among children who did not meet BqSOFA criteria. Similarly, PSS demonstrated moderate discrimination and high specificity but lower sensitivity. When validating PSS, we found a higher PPV than reported in the original validation study, reflecting the substantially higher baseline mortality in our cohort.¹⁹ These findings indicate that both BqSOFA and PSS perform better as late markers of severe illness than as effective screening tools. This is consistent with the original designation of PSS as a tool to identify life-threatening organ dysfunction, rather than as a screening tool.¹⁹

Although PSS demonstrated moderate discrimination, its validation was limited by substantial missingness of component variables, particularly MAP. Importantly, missingness was differential, with MAP missing more often among survivors than among non-survivors. This may reflect the allocation of limited resources and a more focused approach to children with more severe illness. An additional challenge concerns missing PSS components. In the primary analysis, missing components were coded as normal in accordance with the published PSS methodology, as the original derivation setting assumed the absence of a value reflected a normal value.¹⁹ However, interpreting missing data in RLS is often more challenging than in HRS, as it may also result from the limited availability of monitoring equipment or laboratory resources. Consequently, coding missing variables as normal may underestimate organ dysfunction in some children and may have contributed to the high specificity observed in our cohort. Future efforts to adapt organ dysfunction scores for RLS should explicitly address this challenge. Finally, in a complete-case sensitivity analysis of participants with all locally available PSS values, the AUC for PSS decreased slightly (0.72 vs 0.75), a relevant finding in RLS, where invasive medical tests and resources are limited and may distort performance estimates for scores that rely on routine or invasive laboratory measurements or resources.

The choice of scoring threshold has important implications for the implementation of triage tools in RLS. At the published LODS threshold (>0), sensitivity was very high (92.6%), but specificity was low (43.3%), resulting in more than half of surviving children being classified as high risk for mortality. While this approach minimises missed mortality, it may also lead to over-triage and increased use of limited monitoring and treatment resources in RLS. In contrast, a locally optimised threshold (≥2) improved specificity (81.2%) and PPV (46.0%), while reducing sensitivity (65.5%). Similar trade-offs were observed for BqSOFA and PSS (table 3). These findings suggest updating existing scores may represent a practical strategy for balancing early identification of high-risk children with efficient resource utilisation, although prospective implementation studies are needed before recommending alternative thresholds.

Across all three clinical scores, respiratory and neurologic abnormalities emerged with the strongest associations with mortality. In the combined adjusted model including both baseline clinical characteristics and individual score components, measures of respiratory dysfunction and altered mental status remained independently associated with mortality, alongside key contextual risk factors such as malnutrition and referral status. These findings were consistent across the evaluated scoring systems, suggesting that respiratory compromise and neurologic impairment represent the most critical early clinical warning signs in paediatric sepsis in RLS. As these features are simple to assess and routinely evaluated in clinical care, they represent high-yield targets for incorporation into context-adapted clinical scores. This would be feasible for many clinical sites in RLS, as these are key warning signs recommended for routine assessment by the WHO.^{20 26 34} These findings suggest that entirely new clinical scores may not be necessary; rather, adapting existing scoring tools, such as LODS, to incorporate context-relevant risk factors may be a more pragmatic approach.

While this study had many strengths, we acknowledge several potential limitations. First, our study was conducted at a single tertiary referral centre, and our findings may not be generalisable to other healthcare settings; however, this limitation is offset by our large sample size and broad representation of paediatric patients, given that MNH is the largest referral hospital in Tanzania. Additionally, because our cohort was defined using a broad WHO warning sign-based sepsis definition, score performance may differ in populations using stricter sepsis criteria. Overlap between some study inclusion criteria (eg, difficulty breathing and altered mental status) and components of the LODS and BqSOFA scores also introduces the potential for incorporation bias, which may modestly inflate the observed performance of these scores. In a sensitivity analysis excluding children enrolled solely for dehydration, AUCs for all three scores were preserved, while PSS specificity and NPV were lower than in the full cohort; this pattern suggests the broad cohort definition may have contributed to the high specificity of PSS in the primary analysis, though discrimination overall was not affected. Second, although the underlying cohort was prospectively enrolled, validation of clinical scores was conducted as a secondary analysis and was not part of the original study objectives. Third, a strong limitation was the substantial missingness of laboratory and organ dysfunction components of the PSS. Because complete-case analysis included only 53% of patients, its performance estimates may be unstable and less generalisable. In addition, PSS was calculated using the published convention for missing values as normal. In our setting, missingness likely reflects a combination of resource limitations and clinician-directed monitoring practices, as evidenced by differential MAP missingness between survivors and non-survivors. Consequently, coding missing values as normal may underestimate organ dysfunction and artificially increase specificity. Additionally, some PSS laboratory components were drawn outside the initial 24-hour presentation window, as intended in the published score, potentially introducing time bias.¹⁹ We also did not prospectively power this study to detect small differences between scores. Fourth, referral bias may have influenced our outcomes, given that MNH is the national public referral hospital and has disproportionately high acuity. Finally, although we adjusted for several important confounders, unmeasured confounders and variability in clinician decision-making may have influenced treatment timing and outcomes. We were also unable to account for several treatment-related factors that may influence mortality, including timing of antibiotic administration, respiratory support, and critical care interventions. Because these variables were not consistently available for all participants, they were not included in the analysis.

Future research should focus on adapting existing clinical scores, such as LODS, to incorporate simple, context-relevant risk factors for paediatric sepsis. By incorporating easily attainable and locally relevant risk factors into early sepsis recognition and triage, either through modified scoring tools or adapted RLS-specific protocols, there is a significant opportunity to improve sepsis triage and reduce paediatric mortality. Previous studies have demonstrated that combining triage-focused interventions with quality improvement measures may reduce antibiotic use, hospital admissions and mortality.³⁵ That said, we recognise that clinical scores alone are not sufficient to improve outcomes. In addition to our previously proposed changes, it is crucial to improve access to high-quality care, stronger health systems and better implementation systems in RLS to impart lasting, systemic change.^{31–33 35}

Conclusions

Overall, our study identifies critical risk factors for in-hospital mortality in paediatric patients with sepsis and evaluates existing clinical scores in paediatric sepsis. These identified risk factors highlight the vulnerability of children in RLS and are not fully captured by existing sepsis clinical scoring tools. We demonstrate that the evaluated clinical scores perform well enough to support clinical implementation as screening and risk-stratification tools, particularly in RLS, where early identification is critical, but suggest further adaptations to incorporate the identified clinical risk factors. There is an urgent need for research in RLS to identify context-specific risk factors to inform prognostic tools. Strengthening sepsis mortality prediction in RLS will ultimately improve care and outcomes worldwide for children with sepsis.

Data availability statement

Data are available on reasonable request.