Abstract

Objective: To systematically map and synthesize the scientific evidence on the quantitative methodologies used for the identification, cleaning and redistribution of junk codes in population-based mortality registries in low- and middle-income countries (LMICs), evaluating their technical applicability for the parameterization of microsimulation models in health.

Inclusion criteria: This review will consider studies utilizing population-based mortality records or vital statistics from low- and middle-income countries (Population). Eligible literature must describe or apply traditional statistical methods, multiple imputation, Bayesian models, record linkage, or machine learning algorithms to correct ill-defined causes of death and “garbage codes” (Concept). The context includes health information systems characterized by scarce or heterogeneous data, specifically where the feasibility of disaggregating results to the individual level to parameterize stochastic microsimulation models is analyzed (Context).

Methods: This scoping review will follow the JBI methodology and the PRISMA Extension for Scoping Reviews (PRISMA-ScR). Comprehensive searches will be performed across PubMed, Scopus, Web of Science, LILACS, and SciELO, supplemented by grey literature from the WHO, PAHO, and IHME. To ensure reproducibility, the workflow will be executed within the R statistical environment. We will utilize litsearchr for search syntax optimization, synthesisr for deduplication, and revtools for blinded screening supported by latent Dirichlet allocation topic clustering. Discrepancies will be resolved by discussion or a third reviewer. Following paired data extraction, findings will be synthesized narratively and visualized using ggplot2 matrix heatmaps and igraph flow networks.

Keywords: health information systems; causes of death; vital statistics; data quality; machine learning algorithms; proportional redistribution; Computer simulation.

1. INTRODUCTION

Mortality statistics derived from Civil Registration and Vital Statistics (CRVS) systems serve as the bedrock of global epidemiological surveillance and public health planning1. However, their utility is frequently undermined by inconsistencies in the medical certification of causes of death.

A critical challenge is the assignment of deaths to “garbage codes”, defined in the Global Burden of Disease (GBD) literature as International Classification of Diseases (ICD) codes that represent ambiguous, non-specific, or biologically implausible underlying causes2. These include ill-defined symptoms (e.g., R-codes), intermediate conditions (e.g., heart failure, septicemia), or unspecified tumor sites3,4.

To standardize the assessment of these data quality issues, garbage codes are systematically classified into four hierarchical levels based on their degree of specificity and their potential to distort public health inferences. Level 1 codes are the least specific, encompassing broad categories that can hinder identification across major disease groups, while Levels 2 through 4 represent increasing, yet still insufficient, degrees of specificity—ranging from intermediate or immediate causes of death to unspecified sites within larger disease groupings. By categorizing these unusable codes, researchers can better prioritize redistribution efforts, focusing on the most damaging “major” garbage codes (Levels 1 and 2) that pose the greatest risk to accurate health policy and planning4.

The prevalence of “garbage codes” within official mortality databases severely obscures population health trends, underestimating the true burden of critical pathologies—such as ischemic heart disease and specific malignancies5,6. To mitigate these systematic biases, a diverse array of correction and redistribution methodologies has evolved over the last two decades. These approaches range from traditional proportional redistribution stratified by age and sex to multiple imputation and regression modeling. More recently, sophisticated methods have emerged, involving record linkage with clinical datasets and advanced machine learning algorithms7–9.

Despite the proliferation of “garbage code” redistribution algorithms spearheaded by international bodies like the WHO and IHME, the extant literature remains fragmented. Current research is predominantly localized, focusing on specific causes or geographical regions in isolation. Consequently, systematic mapping is lacking to consolidate the methodological diversity, underlying mathematical assumptions, and data structures of these approaches. This review aims to bridge this gap by synthesizing the existing evidence and evaluating the technical transferability of these methodologies to microsimulation models.

This knowledge gap becomes critical when addressing the transition from population-level mortality records to stochastic microsimulation models. Unlike aggregated demographic models that rely on macroscopic trends, microsimulation platforms operate at the individual level. They simulate single life trajectories, chronic disease progression, and the dynamic impact of public health interventions. To remain predictive, these frameworks require parameterization with highly disaggregated baseline mortality risks. Consequently, “garbage codes” create a fundamental barrier to valid modeling by obscuring the individual-level precision necessary for simulation. When input data used to calibrate microsimulation models incorporate “garbage codes”, the models inevitably propagate and exacerbate systematic biases in individual-level projections.

Currently, no evidence synthesis exists on how to adapt or transfer redistribution methodologies—originally designed for aggregated mortality data—to generate granular datasets suitable for stochastic microsimulation. Consequently, this scoping review will map the current state-of-the-art, evaluate the technical feasibility of these approaches within simulation environments, and provide a framework for future public health mathematical modeling.

This scoping review aims to systematically map and synthesize evidence on “garbage code” identification and redistribution methodologies in low- and middle-income country mortality registries. Concurrently, it will evaluate their technical characteristics and applicability for calibrating and parameterizing health microsimulation models.

1.1 Research Questions

  1. What analytical methodologies (statistical, probabilistic, deterministic, linkage-based or machine learning) have been described in the literature for the correction and redistribution of “garbage codes” in mortality records?
  2. How are they classified and what types of “garbage codes” (poorly defined causes of ICD chapters, intermediate or terminal causes) are the main objective of each identified methodology?
  3. What is the documented level of applicability or compatibility of these correction methods to structure databases aimed at the development, calibration and validation of microsimulation models in health?
  4. What are the main methodological challenges, limitations and advantages reported when transferring or adapting analytical correction algorithms from the population (macro) level to the individual level (micro)?

2. ELIGIBILITY CRITERIA

2.1 Population

This component defines eligible data records, populations, and geographic or demographic coverages.

  • Inclusion Criteria:
    • Data Type: Studies utilizing population-based mortality databases, vital statistics systems, official national or subnational health statistics registers, or civil registration and vital statistics (CRVS) systems.
    • Geography and Income Level: Research focusing on populations within low- and middle-income countries (LMICs) as defined by historical or current World Bank classifications (e.g., regions within Latin America and the Caribbean, Sub-Saharan Africa, and South Asia).
    • Technical Exception: Global or multinational studies (e.g., Global Burden of Disease [GBD] analyses), provided they independently disaggregate methodologies and data applicable to LMIC geographies.
    • Demography: Records tracking general human mortality without restrictions on age groups (including neonatal, infant, adult, or geriatric mortality) or sex and gender.
  • Exclusion Criteria:
    • Clinical studies restricted to hyper-specific hospital specimens or controlled trials that do not utilize—or aim to correct—population-based records or broader epidemiological information systems.

2.2 Concept

This component delineates the core methodologies, algorithms, and analytical tools evaluated in this review.

  • Inclusion Criteria:
    • Studies describing, evaluating, validating, or implementing quantitative methods to identify, clean, reclassify, or redistribute “garbage codes” (levels 1 to 4) or ill-defined causes of death (as classified under ICD-10/ICD-11 Chapter XVIII, or ICD-9 equivalents).
    • Methodologies within the following analytical categories are eligible:
      • Traditional statistics and mathematical demography: Proportional redistribution (simple or stratified by age and sex), fractional algorithms, and indirect demographic adjustment methods3.
      • Advanced statistical modeling: Multiple imputation of missing or non-specific data, spatial or spatiotemporal regression modeling, and empirical or hierarchical Bayesian approaches (e.g., CODEm or CoDcorrect analytical algorithms)7–9.
      • Data integration (record linkage): Deterministic or probabilistic algorithms linking mortality records with complementary databases (e.g., hospital discharge records, automated verbal autopsies, or epidemiological surveillance systems)10,11.
      • Artificial intelligence and machine learning: Data-driven predictive architectures (e.g., decision trees, gradient boosting, random forests, or other data-mining frameworks) applied to classify or predict underlying causes of death masked by non-specific codes.
  • Exclusion Criteria:
    • Studies analyzing mortality data quality in a purely descriptive manner (e.g., reporting isolated completeness percentages or garbage code prevalences) without implementing or evaluating an active mathematical or algorithmic correction method.
    • Research focused solely on correcting numerical under-registration (omission of death capture), unless the methodology explicitly integrates cause-of-death redistribution.

2.3 Context

This component defines the data environments and final fields of application toward which the evidence synthesis is oriented.

  • Inclusion Criteria:
    • Data-Scarce or Heterogeneous Settings: Research situated in contexts with weak or fragmented health information systems characterized by low medical certification coverage, high regional heterogeneity, or small-area spatiotemporal misalignments.
    • Microsimulation Applicability: Studies evaluating or discussing the impact of correction techniques on individual-level microdata disaggregation, the generation of consistent mortality risk matrices, individual trajectory estimations, or parameterizing inputs for health microsimulation platforms.
    • Emergent Methodological Scope: Studies originally operating at a macro or aggregate level, provided they detail the mathematical assumptions or individual-record classification algorithms necessary to calibrate and parameterize stochastic microsimulations or agent-based models.
  • Exclusion Criteria:
    • Methodologies operating exclusively at an irreducible macro-population level whose designs preclude individual-level micro-disaggregation, thereby lacking utility for agent-based or individual-level simulations.

2.4 Sources of Evidence

  • Publication Types: Peer-reviewed original research articles published in indexed journals, methodological book chapters, and authoritative institutional gray literature (e.g., technical reports from the WHO, PAHO, or IHME).
  • Timeframe: Material published from 1996 onwards, capturing the global introduction and widespread adoption of the Tenth Revision of the International Classification of Diseases (ICD-10), which consolidated the term “garbage codes.”
  • Languages: Studies published in English, Spanish, and Portuguese, reflecting the substantial volume of methodological literature generated in Latin America (particularly Brazil).

3. METHODS

3.1 Search Strategy

The search strategy will locate published academic literature and online grey literature. The search will focus on evidence published from 1996 onwards, corresponding to the global implementation and widespread adoption of the International Classification of Diseases, Tenth Revision (ICD-10)—a timeframe during which the term “garbage codes” became conceptually consolidated.

  • Academic Literature: To identify academic literature, a three-step search strategy will be developed in collaboration with a medical librarian:
    1. Initial Limited Search: An initial search of MEDLINE (Ovid) will be undertaken to analyze text words in titles and abstracts, Medical Subject Headings (MeSH), and author-supplied keywords from relevant articles.
    2. Full Search Expansion: A comprehensive search strategy will be developed using these identified terms, then translated and adapted for Scopus, Web of Science, LILACS, and SciELO. The complete search syntax for MEDLINE (Ovid) is presented in Appendix I.
    3. Reference Hand-Searching: The reference lists of all selected articles and relevant systematic or scoping reviews will be hand-searched to identify primary sources missed by the electronic search.
  • Grey Literature: To locate technical reports and eligible grey literature, a targeted list of national and international health organization websites will be compiled following the two-step Google search methodology.
    1. Preliminary Term Identification: A preliminary Google search will be conducted to identify key text words, accounting for highly variable terminology such as “unspecific causes” or “ambiguous causes.”
    2. Structured Execution: Ten unique structured Google searches will be executed using distinct keyword combinations representing the population, concept, and context (PCC) of interest.

The first 100 results of each structured search will be screened. Identified organizational websites will be manually evaluated for technical reports. Finally, vital registration epidemiologists, public health officials, and key collaborators with expertise in health information systems within LMICs will be consulted to identify any additional missing sources.

3.2 Data Management and Search Automation

To ensure the methodological reproducibility and transparency required by the PRISMA-ScR guidelines, the information retrieval process will be systematized within the R statistical environment (v4.5 or higher). The litsearchr package will be employed for the algorithmic optimization of Boolean syntax via term co-occurrence networks. Indexed databases will be interrogated programmatically through Application APIs using the rentrez and rscopus libraries. Data consolidation and duplicate removal will be managed via the synthesisr package using fuzzy string-matching models, supplemented by manual verification of borderline cases. Finally, title and abstract screening will be conducted through the interactive graphical interface of the revtools package, supported by Latent Dirichlet Allocation (LDA) algorithms for topic modeling to optimize evidence selection in accordance with the PCC framework.

3.3 Data Extraction and Charting

To ensure integrity, traceability, and reproducibility, we will conduct data extraction programmatically within the R statistical environment (v4.5 or higher). This automated approach mitigates the transcription errors inherent in manual processes and ensures a standardized workflow. We have designed an electronic data charting form to capture information across six dimensions: study metadata, geographic and health system context, mortality database characteristics, “garbage code” typology, correction methodologies (including algorithmic and software ecosystems), and microsimulation applicability.

This form will be piloted with a random sample of three included studies to refine consistency and ensure variable granularity. Two reviewers will perform the extraction independently, resolving discrepancies through consensus. The resulting data will be consolidated into a relational structure (e.g., tibble or data.frame) to facilitate transparent downstream analysis. The full extraction schema (exported as JSON/CSV) and corresponding R scripts will be made available via the Open Science Framework (OSF) upon study completion. For a detailed breakdown of the variables and the classification framework utilized, refer to the Data Charting Instrument (Appendix II).

3.4 Data Analysis, Synthesis, and Presentation

This scoping review adheres to the Joanna Briggs Institute (JBI) methodology and the PRISMA Extension for Scoping Reviews (PRISMA-ScR). Given the methodological scope, we will prioritize descriptive mapping, algorithmic categorization, and theoretical evaluation of data transfer flows from aggregate population (macro) levels to individual (micro) levels. We will conduct all data processing and visualization programmatically within the R statistical environment (v4.5 or higher).

  • Descriptive Quantitative Analysis: We will perform descriptive statistical analysis on the variables extracted into our R matrix. Absolute frequencies and percentages will characterize the literature based on:
    • Chronological distribution: Temporal trends from 1996 to the present.
    • Geographic reach: Studies by World Bank income classification and region.
    • Taxonomic frameworks: Criteria used to define “garbage codes” (e.g., GBD vs. traditional WHO typologies).
    • Analytical techniques: Frequencies of techniques such as proportional redistribution, multiple imputation, Bayesian modeling, record linkage, and machine learning.
    • Cause categories: Most frequently corrected ICD-10/ICD-11 chapters (e.g., cardiovascular, neoplasms, or external causes).
  • Thematic Narrative Synthesis: We will accompany the analyzed data with an explanatory narrative addressing the research questions derived from the PCC framework. The narrative will be structured around:
    • The status of “garbage codes” in LMICs: A critical description of the magnitude and typology of ill-defined mortality causes in data-scarce settings.
    • Methodological diversity: A systematic review of mathematical assumptions, technical advantages, and data infrastructure requirements for each algorithmic family.
    • Transferability to microsimulation: A conceptual analysis evaluating whether existing methodologies preserve microdata validity and addressing analytical gaps in calibrating agent-based stochastic models.
  • Mapping and Visualization: To visualize the structure and knowledge gaps of the field, we will generate advanced graphics using ggplot2, igraph12,13, and tidyr:
    • Structured Evidence Tables: Master tables will detail the methodological characteristics of each study, cross-referencing correction methods with their minimum input data requirements.
    • Two-Dimensional Heatmaps: We will generate matrix plots using ggplot2 to map analytical correction methods against specific ICD Chapters. Color intensity will reflect literature density, facilitating the identification of consolidated methodological combinations versus underexplored areas.
    • Flow Network Diagrams: Using igraph, we will construct network graphs where nodes represent correction methods and links map their relationship with individual-level variables. Nodes will be categorized by output granularity (macro-aggregate vs. micro-individual), visually illustrating the technical pathway required to connect traditional vital statistics with microsimulation parameterization.

3.5 Data Representation and Synthesis

Upon completing the data extraction process, we will synthesize the retrieved information using both qualitative and quantitative approaches. We will visualize the results by programming graphical solutions in R with the ggplot2 and igraph packages to generate the following: * Heatmaps: These will cross-reference analytical methods (e.g., AI-based models, Bayesian inference, proportional redistribution) with the medical causes or ICD codes where they have been successfully applied in LMIC settings. * Co-occurrence Network Diagrams: These will map the methodological transitions from macro-level algorithms to individual-level parameters, illustrating their scalability for microsimulation models. * Structured Evidence Tables: These will hierarchically consolidate the advantages, limitations, computational requirements, and theoretical assumptions associated with each methodological approach.

Conflicts of Interest

The authors declare no competing interests. This scoping review was conducted for academic and methodological purposes only, with no commercial, financial, or institutional support that could be construed as a potential conflict of interest.

References

1.
AbouZahr C, Savigny D de, Mikkelsen L, et al. Civil registration and vital statistics: Progress in the data revolution for counting and accountability. Lancet. 2015;386:1373–85.
2.
Murray C, Lopez A. The global burden of disease: A comprehensive assessment of mortality and disability from diseases, injuries and risk factors in 1990 and projected to 2020. Boston: Harvard School of Public Health; 1996.
3.
Naghavi M, Makela S, Foreman K, et al. Algorithms for enhancing public health utility of national causes-of-death data. Popul Health Metr. 2010;8(1):9.
4.
Naghavi M, Richards N, Chowdhury H, et al. Improving the quality of cause of death data for public health policy: Are all “garbage” codes equally problematic? BMC Med. 2020;18(1):55.
5.
Hart J et al. Improving medical certification of cause of death: Effective strategies and approaches based on experiences from the data for health initiative. BMC Public Health. 2020;20(1):1083.
6.
Miki J et al. Saving lives through certifying deaths: Assessing the impact of two interventions to improve cause of death data in peru. BMC Public Health. 2018;18(1):1329.
7.
Foreman K, Naghavi M, Ezzati M. Improving the usefulness of US mortality data: New methods for reclassification of underlying cause of death. Popul Health Metr. 2016;14:14.
8.
Grigoriev P, Bonnet F, Perdrix E. Method for redistributing ill-defined causes of death. Popul Health Metr. 2024;22(1):1–14.
9.
Johnson S et al. Public health utility of cause of death data: Applying empirical algorithms to improve data quality. BMC Med Inform Decis Mak. 2021;21(1):175.
10.
Bierrenbach A et al. Redistribution of heart failure deaths using two methods: Linkage of hospital records with death certificates. Rev Bras Epidemiol. 2019;22:e190006.
11.
França E et al. Ill-defined causes of death in brazil: A redistribution method based on the investigation of such causes. Rev Saúde Pública. 2014;48(4):671–81.
12.
Wickham H. ggplot2: Elegant graphics for data analysis. New York: Springer-Verlag; 2016.
13.
Csardi G, Nepusz T. The igraph software package for complex network research. InterJournal, Complex Systems. 2006;1695:1–9.

Appendices

Appendix I: Search strategy

Database: MEDLINE (via Ovid)
Search conducted on: July 2026
Planned Limits: * Date restrictions: from 1996 to present to capture historical development of algorithms like GBD. * Language restrictions: None at search level (any required filters based on reviewer capacities will be applied during the programmatic screening stage). * Document types: All source types (including technical notes, methodology papers, and electronic articles).

Line Search Query (Terms, Truncations, and Syntax) Conceptual Block / Target
#1 exp Mortality/ or exp Cause of Death/ PCC: Population / Focus (Core mortality registry filters)
#2 (mortality data* or death certificate* or vital statistic* or vital registration or CRVS).tw,kf,ot. Keywords for vital statistics and national death registry systems
#3 1 or 2 Result: Core Mortality Data
#4 (garbage code* or ill-defined or misclassified or misclassification* or undefined cause* or ill-defined cause* or unspecific cause* or vague code* or redistribution algorithm*).tw,kf,ot. PCC: Concept (Block A) - Specific typology of Data Errors (“Garbage Codes”)
#5 (data cleaning or data correction or diagnostic accuracy or underlying cause of death).tw,kf,ot. Data quality attributes specific to mortality databases
#6 4 or 5 Result: Garbage Code Concepts
#7 exp Algorithms/ or exp Models, Statistical/ or exp Computer Simulation/ MeSH Indexing for Analytical Methods
#8 (demographic redistribution or multiple imputation or MICE or Bayesian model* or Markov chain* or machine learning or random forest* or neural network* or artificial intelligence or algorithmic correction or fractional assignment).tw,kf,ot. PCC: Concept (Block B) - Specific statistical and computational correction frameworks
#9 (microsimulation* or micro-simulation* or agent-based model* or stochastic simulation* or individual-level dynamic* or synthetic population*).tw,kf,ot. PCC: Context / Transferability - Target micro-level application architectures
#10 7 or 8 or 9 Result: Methodological Ecosystem
#11 exp “Developing Countries”/ MeSH Indexing for LMICs
#12 (low income country or low-and-middle income countr* or LMIC or LMICs or developing nation* or resource-constrained setting* or transitional econom*).tw,kf,ot. Geographic Context Keywords (Based on World Bank specifications)
#13 (Africa* or Asia* or South America* or Central America* or Latin America*).tw,kf,ot. Broad regional geographic filters
#14 11 or 12 or 13 Result: Developing Context (LMIC)
#15 3 and 6 and 10 and 14 COMBINED STRATEGY (Boolean Intersection)

Appendix II: Data extraction instrument

Block Variable Description / Purpose Suggested Format / Values
I. Identification & Metadata Study ID Unique identifier assigned by R workflow Numeric (e.g., 001)
Bibliographic Data Author, year, title, and journal/institution Text
Evidence Type Original article, technical report, or method book Dropdown
II. Geographic & Health Context Geographic Scope Country, region, or sub-national area Text
Economic Classification World Bank income level Dropdown (Low, Lower-Middle, Upper-Middle)
CRVS System Status Description of Civil Registration and Vital Statistics status Text
III. Mortality Database Data Periodicity Years covered by mortality records Text / Numeric
Data Volume Sample size (number of deaths analyzed) Numeric
Primary Source Origin of data (e.g., forensic, hospital, surveys) Text
ICD Framework Version of the ICD utilized (ICD-9, ICD-10, ICD-11) Dropdown
IV. Junk Code Typology Definition Framework Criteria used to identify “junk” (e.g., GBD list) Text
Evaluated ICD Codes Specific codes being targeted for correction Text
Baseline Magnitude Percentage/volume of junk codes in original records Percentage (%)
V. Correction Methodology Central Technique Algorithmic classification Dropdown (e.g., MICE, Bayesian, ML)
Theoretical Assumptions Underlying mathematical principles/assumptions Text
Software Ecosystem Tools used (R, SAS, Stata, etc.) Text
Uncertainty/Bias Report Reported confidence intervals or limitations Yes / No / Partial
VI. Microsimulation Applicability TRL (Technology Readiness Level) Maturity of the development method (1-4) Scale (1-4)
Output Format Aggregate vs. individual-level data output Dropdown
Demographic Granularity Consistency at sub-group/small-area levels Scale (1-5)
Parametrization Potential Capability to feed agent-based/microsimulation models Scale (1-5)
Reviewer Notes Professional notes on applicability and transferability Free text