Study design and setting

This was a prospective cohort study that was conducted among children aged 1–59 months admitted to the general paediatric wards at Muhimbili National Hospital. Muhimbili National Hospital is the national referral hospital and a teaching hospital for Muhimbili University of Health and Allied Sciences and has a bed capacity of 2000 beds. At the time of data collection, the Department of paediatrics had specialized units that included neurology, nephrology, haematology, endocrinology, oncology, gastroenterology, and infectious and malnutrition ward. The general paediatric wards admit children aged 1 month to 14 years with an average daily admission of 3 to 15 children in all units per day. Each unit is staffed with intern doctors, medical officers, residents, and specialists.

Study population

Children aged 1–59 months admitted to general paediatric wards at MNH were eligible for this study. Children admitted with surgical diagnosis/trauma died within 30 min of admission and those whose parents/guardians refused consent were excluded.

Sample size

The sample size for this study was calculated from OpenEpi using Fleiss Methods for Rates and proportions [9]. The death rate due to respiratory tract infection of 2% reported by Ezeonwu et al. [10] was used for estimating sample size, making use of a two-sided confidence level (1-alpha) of 95, power of 80%, ratio of sample size; unexposed/exposed of 4, percentage of exposed population with an outcome (patient who died with respiratory tract infection) of 2%, percentage of unexposed population with an outcome (patient who died with no respiratory tract infection) of 7% and odds ratio of 0.27. Respiratory tract infection was selected because it was presumed to be the leading cause of mortality among children. The exposed and unexposed sample sizes were estimated to be 185 and 737 respectively, making a total sample size of 922. The total sample size was adjusted with an assumption of a 5% non-response rate making the minimum sample size to be 975.

Sampling procedure

We conveniently enrolled all participants meeting inclusion criteria consecutively until the sample size was reached. Information about the study was provided to parents/guardians and informed written consent was sought before recruitment.

Data collection

The principal investigator and two research assistants, who were intern doctors trained in the protocol, collected data for this study using a structured questionnaire starting on 2nd October 2017 and finishing on 13th April 2018. Information collected was filled in the four sections of the questionnaire which included socio-demographic characteristics, clinical presentation, laboratory investigations, and causes of death. Participants were recruited within 24 h of admission. Data was collected from parents/caregivers and participants' case notes. Participants were then followed up every two days until discharge or death, confirmed diagnoses and laboratory findings were recorded on follow-up.

Study variables Dependent variables (outcome)

Death or survival was the primary outcome measured as a binary variable where death = 1, and discharge = 0. Cause of death was defined as the primary underlying illness; for a patient with multiple diagnoses both the underlying cause of death, the immediate cause of death, and co-morbid conditions were recorded. The immediate cause of death was reported as the immediate cause of death as recorded in the patient file after a mortality audit, which is done by more than one clinician in the respective unit. Causes of death were discussed in a review clinical presentation, and any supportive laboratory findings if any as well as circumstances around that death are taken into consideration when reporting the cause of death. Finally, causes of death were documented in the file after a consensus was reached. Mortality audits were done weekly or once in two weeks depending on the unit. Comorbid conditions are those related to the cause of death or complication of underlying disease, for example, malnutrition in a patient with cerebral palsy or congenital heart disease, anaemia, or diarrhoea.

Independent variable

Social demographics: Age was recorded in terms of months as a continuous variable. Sex and primary caregiver were recorded as nominal variables. The primary caregiver, occupation, level of education, and wealth quintiles were recorded as ordinal variables. Health insurance and vaccination status were recorded as a dichotomous categorical variable.

The wealth index was generated using 11 items on asset ownership, having a bank account, housing material used to construct roofs, walls, and floors, and the main source of drinking water and cooking fuel. All the items were converted into dichotomous variables 1 = wealth and 0 = poor. (eg housing material floor; made of mud or sand = 0, cement/tile/polished wood = 1). The first factor was then used to construct a household wealth index that was summarized as quintiles – highest, rich, middle, poor, and lowest quintiles.

Clinical features: cough, fever, convulsion, vomiting, diarrhoea, inability to feed, reduced urine output, lethargy, unconscious and respiratory distress was recorded as a dichotomous variable.

Biochemical markers relevant to the diagnosis recorded included white blood cell count, haemoglobin level, platelet counts, C – reactive protein, electrolyte, serum creatinine recorded as a continuous variable, and blood culture as a nominal variable.

Nutrition status: assessment was done on admission; weight was measured using a weighing scale (Seca beam balance) calibrated before measurement and weight recorded to the nearest 10gm. The length was taken for children who were not able to stand using the length board while the child was lying supine on a flat surface with the head touching the top board and the soles of the feet touching the footboard. Children who were able to stand had their height measured using a height board while the child was standing and the head touching the top board. Weight for height was calculated and interpreted using WHO standard growth chart Z scores where moderate wasting was defined as weight for height between (-2SD to -3SD) and severe wasting (< -3SD).

Human Immunodeficiency Virus infection (HIV)status was extracted from the patient file. Every child's HIV status was checked on admission as per local paediatric protocol. HIV exposure was defined as children below 18 months born with HIV positive mother but had a DNA PCR negative test at 6 weeks of life and the child was still breastfeeding. HIV infection was defined as children confirmed HIV positive by DNA PCR if they were below 18 months. Children above 18 months of HIV infection were diagnosed using two rapid tests; Bioline then confirmed by Unigold test, manufacture of these tests are Standard Diagnostic, Inc., 65, Borahagal-ro, Giheung-gu, Yongin-si, Gyeonggi-do, Republic of Korea.

Respiratory distress: These are children presented with higher respiratory rate than normal for age, lower chest wall in drawing, nasal flaring, or grunting.

Anaemia: This study categorizes anaemia into very severe anaemia – Hb < 5 g/dl, severe anaemia – Hb 5.0–6.9 g/dl, moderate anaemia Hb 7.0–9.0 g/dl, mild anaemia- Hb 9.1–11.0 g/dl and normal is > 11.0 g/dl.

Fever was defined as a temperature above 37.5 °C.

Admission diagnosis: This was a confirmed diagnosis after being reviewed by the team of the unit either resident or specialist or both with laboratory and radiological confirmation as per protocol. The provisional diagnosis was used if a child died before the diagnosis was confirmed.

Data management and analysis

Questionnaires were manually checked for completion daily by Principal Investigator and were entered into the computer system using Statistical Package for Social Sciences version 16 (SPSS Inc., Chicago, IL, USA). Data was cleaned before the analysis and all analyses were done using SPSS.

The mortality rate was calculated as the cumulative proportion of the total number of deaths divided by the total number of admissions during the study period. Kaplan–Meier curves are presented showing the overall mortality rate and by select characteristics. Mean ± SD and median (IQR) were calculated for continuous normally distributed and skewed variables, respectively. The wealth index was constructed after component factor analysis (data reduction procedure).

Bivariate analysis was done to determine the association between patient characteristics and mortality. Statistical significance was assessed using chi-square tests and Fisher's exact test for categorical variables. Differences between continuous (normally distributed) variables were analyzed using an independent sample t-test. A variable was considered statistically significant if the p-value was < 0.05.

Factors associated with mortality in bivariate analysis with a p-value < 0.1 were entered in the multivariable logistic regression model to identify and quantify predictors of deaths while controlling for potential confounders. Crude and adjusted odds ratio with 95% confidence intervals were calculated and factors with P-value < 0.05 were considered significant after analysis.