Skip to content
HEAL-SL

How HEAL-SL works

HEAL-SL is Sierra Leone's Sample Registration System (SL-SRS). It counts deaths in a fixed, nationally representative sample of households and assigns a cause to each one. The Ministry of Health and Sanitation launched it after the Bo Healthy City Study pilot, and national scale-up began on 1 September 2019. It is implemented by Njala University with the National Public Health Agency, with scientific support from the Centre for Global Health Research (University of Toronto) and the University of Oxford.

This page explains how the data are generated. Access, downloads, and papers can be found on separate pages, linked at the side.

The process in four steps

  1. Sample and enumerate

    Census enumeration areas are drawn from the census frame in all 16 districts. Teams visit every household in each selected enumeration area, record general information about every household member and take note of every death in the previous two years, and return to the same households and update the information in each successive round. Design and sampling

  2. The verbal autopsy: Interview a family member about each death

    With written consent, a trained surveyor interviews a family member about the death in the household using the WHO verbal autopsy instrument, capturing the information on a laptop that works offline. Instrument and fieldwork

  3. Cause of death, assigned by trained health professionals

    Two physicians review each record independently and assign a standard cause of death code (International Classification of Diseases, version 10 [ICD-10]). If they disagree on the cause of death, they reconcile (without disclosing the identity of the other physician; the system allows either of the physicians to accept the justifications of the other) and come to agreement. A senior physician adjudicates any remaining disagreements. The final cause is the ICD-10 underlying cause. How coding works

  4. Group, estimate, and publish

    Coded deaths are grouped into the WHO's related cause groups and scaled to the UN's death totals for the country. That gives yearly national estimates and district estimates for three-year periods, published as open data. From coded deaths to estimates

Design and sampling

  • 660+ census enumeration areas drawn from the Statistics Sierra Leone census frame, in all 16 districts and ~190 chiefdoms, rural and urban.
  • Areas are drawn from low, medium and high development strata (housing and female literacy), with equal households in each.
  • About 60,000 households and 340,000 people in all, about 5% of the population.
  • Every household in a selected area is surveyed, and households are never replaced.

The areas are grouped into the regions used in the data. The design gives separate urban and rural estimates for each district.

Sources: Carshon-Marsh et al., The Lancet Global Health, 2022 (678 areas and about 343,000 people in round 1); Causes of Death Report 2018-2023, p. 1.

District coverage

Each hexagon on the interactive map is a 10 km area with surveyed houses; shading indicates the district's share of the study deaths.

  • Survey houses (per 10 km area)
  • Share of study deaths (round 1)
  • Under 2%
  • 2–4%
  • 4–6%
  • 6–8%
  • 8–15%
  • 15% and over

Hexagons show surveyed houses per 10 km area (small areas left out). Shading shows each district's share of study deaths, not its death rate. Bo includes the Bo Healthy City Study (Carshon-Marsh et al., Lancet Global Health, 2022). Boundaries: HEAL-SL district GeoJSON.

Instrument and fieldwork

In the field

  • Community entry followed a fixed order: ethics approval from the Sierra Leone Ethics and Scientific Review Committee, then district health management teams, paramount chiefs, and community meetings, then the household.
  • Teams of four surveyors spent about four days in each area. They listed every house and household, recorded (updated) every member, and recorded every death in the past two years.
  • Household heads gave verbal consent for the listing. The family member interviewed about a death gives written consent.

The verbal autopsy

  • Surveyors use the WHO verbal autopsy instrument (2016/2022 version), with separate forms for neonates, children, and adults. They capture it on laptops that work offline and sync when there is a signal.
  • Key symptoms are covered first, then probes, then the order of events and the treatment history. Absence of symptoms is recorded too, to help rule out causes.
  • The surveyor writes a half-page narrative while part of the interview is audio-recorded.

In early rounds, verbal autopsies were done only for deaths under 70 years of age; deaths at 70 and over appear in the data from 2020 and fully from 2021. The adult instrument and the WHO form are in Documents.

Survey rounds

RoundFieldworkVerbal autopsies
12019 – Jan 20218,518
2Apr 2021 – Oct 20223,312
3Jan – Sep 20235,437
4Nov 2023 – Aug 20246,954
5Oct 2024 – Jul 20257,441
6Oct 2025 – ongoing6,500+

38,000+ verbal autopsies to date, including Round 6, which is still in the field.

Protocol and ethics approval

Ethics approval comes from the Sierra Leone Ethics and Scientific Review Committee, before any community is approached. Household heads give verbal consent for the listing, and the family member interviewed about a death gives written consent.

Cause of death, assigned by doctors

  • Records with too little information are set aside before coding.
  • Two physicians code each record independently and blind to each other's assigned cases, recording keywords, a quality grade, and a certainty grade.
  • If they disagree, each sees the other's reasoning anonymously and they reconcile. If they still disagree, a senior physician adjudicates.
  • The final cause is the ICD-10 underlying cause. The target is to complete from interview to final cause in under 15 days.
  • Since Round 4, AI models (GPT, InterVA-5, InSilicoVA) code the same records to assist physicians. They support physicians and never code alone.

Maternal deaths also go to an obstetric panel.

Source: Carshon-Marsh et al., The Lancet Global Health, 2022.

From coded deaths to estimates

Coded deaths are grouped into the WHO's cause groups, a standard classification based on the WHO Global Health Estimates.

The sample-weighted share of each cause is applied to the UN Population Division's death totals for Sierra Leone, partitioned across the four regions using the study's death proportions, the 2015 census and the 2019 DHS, and the regional totals are summed back to the national UN total; confidence intervals use the delta method.

The Level One tables show, for each cause (with its category), sex, age group and region, the estimated national deaths and the 2020 UN population (thousands) used for the estimates.

Sources: Carshon-Marsh et al., The Lancet Global Health, 2022; the Level One table columns.

The data this produces

  • From enumeration, household and member characteristics: age, sex, education, occupation, and more.
  • Every death, with the full verbal autopsy, narrative, each physician's code and the final ICD-10 cause.
  • In participating rounds and areas, health check-up measurements: blood pressure, heart rate, chronic disease history, tobacco and alcohol use.

These are linked through a HEAL-SL identifier. Actual names never leave the internal database.

What information is openly available and what requires a special request is on the data access plan. Level One and Level Two are open; the Level Two table gives region (shown as province in the Level Two table) and a data dictionary describes its columns.

Quality checks

  • About 10% of households in each area are re-surveyed by a different staff member.
  • Every house is GPS-geocoded as a check against fabricated interviews.
  • Reviewers score every surveyor's narratives and audio each week and return low-scored ones with the reason.
  • Unannounced supervision; performance reviews each round underpin surveyor contract renewal.
  • The eVA application validates answers as they are entered.
  • Low-quality narratives fell from 24.7% (Round 1) to 20.9% (Round 2) and 7.4% (Round 3) following rigorous QA measures.

Limitations

  • A verbal autopsy records what the family reports about the symptoms and circumstances of a death. It is not a medical record or a laboratory result. HEAL-SL worked with CHAMPS on biological confirmation (minimally invasive tissue sampling) of child and adult deaths, starting in Bo district.
  • Verbal autopsy identifies some causes more reliably than others. For malaria, see the programme's review (Carshon-Marsh et al., Malaria Journal, 2024).
  • HEAL-SL is a sample survey of about 5% of the population. Its figures are estimates, not a count of every death in the country, and less common causes carry more uncertainty.

Updates and versions

The open Level Two file holds 31,110 deaths with a final cause assigned by physicians from the completed rounds (updated 20 Jan 2026); deaths coded from the ongoing Round 6, once it is complete, are added at the next refresh.

Open files are replaced in place at the same address, so a file you downloaded earlier may differ from the current one. Each data page shows its file's date, and the citation it gives carries the year of the latest update and the day you accessed it.

Use and cite

Levels One and Two are published under CC BY 4.0: use and share them, and credit HEAL-SL. Level Three is covered by its signed agreements.

To cite the method, cite Carshon-Marsh et al., The Lancet Global Health, 2022.

Each data page carries the citation for its data set, with the year of its latest update: Level One, Level Two.

Documents

  • Neonate eVA Instrument - 2024 version

    HEAL-SL3 Nov 2026
  • Child eVA Instrument - 2024 version

    HEAL-SL3 Nov 2026
  • Adult eVA Instrument - 2024 version

    HEAL-SL3 Nov 2026
  • Causes of Death Report 2018-2023

    HEAL-SL4 Dec 2023

    HEAL-SL Launch: Causes of Death Report 2018-2023

Publications on the method

  • Patterns and trends of maternal mortality in Sierra Leone: nationally representative mortality survey, 2018-2024

    Meh et al.1 Oct 2026

    Accepted for publication in the Bulletin of the World Health Organization (October 2026).

  • Computer assisted verbal autopsy: comparing large language models to physicians for assigning causes to 6939 deaths in Sierra Leone from 2019–2022

    Wen et al.24 Dec 2025

    BMC Medicine 2026;24:49. Compares large language models with physicians for assigning causes to 6,939 deaths in Sierra Leone, 2019–2022.

  • Estimating annual deaths from stroke in Freetown Sierra Leone

    Youkee et al.5 Aug 2025

    Estimating annual deaths from stroke in Freetown Sierra Leone: a comparative analysis of a hospital-based stroke register and a population-based verbal autopsy study

  • Educational Attainment and Child and Adult Mortality in Sierra Leone

    Wu et al.8 Jul 2025

    Educational Attainment and Child and Adult Mortality in Sierra Leone: A Nationally Representative Mortality Study, 2019-2024

  • Child, Maternal, and Adult Mortality in Sierra Leone: Nationally Representative Mortality Survey 2018-2020

    Carshon-Marsh et al.25 Nov 2021

    The Lancet Global Health

Terms

Verbal autopsy
A structured interview with the family about the symptoms and circumstances of a death.
eVA
Electronic verbal autopsy: the same interview on an electronic form that works offline.
Enumeration area
A census enumeration area, an urban census block or a village: the unit HEAL-SL samples.
Survey round
One visit to every enumeration area: the baseline, then one each year.
Underlying cause
The disease or injury that started the chain of events leading to death; the cause the ICD-10 code records.
ICD-10
The WHO's International Classification of Diseases, tenth edition. Each death gets one ICD-10 code.