1 PhD Program in Translational Specialistic Medicine ‘G.B.
Morgagni’, Department of Cardiac, Thoracic, Vascular Sciences and Public
Health, University of Padua, Padua, Italy.
2 Faculty of Health Sciences / Research Institute,
Universidad Dr. José Matías Delgado, Antiguo Cuscatlán, El
Salvador.
3 National Institute of Health of El Salvador (INS), San
Salvador, El Salvador.
* Corresponding author: Giovanni F. Guevara
(giovannifrancisco.guevaravasquez@phd.unipd.it).
Abstract
Objective: To map quantitative methodologies for identifying, cleaning, and redistributing mortality “garbage codes”, evaluate their feasibility for parameterizing individual-level health microsimulations, and formulate a translational framework for mathematical modeling.
Introduction: Garbage codes represent ambiguous, intermediate, or biologically implausible ICD causes of death. Most redistribution algorithms operate on aggregated data under missing-at-random assumptions. How these macro-level corrections transfer to calibrate stochastic, individual-level microsimulations and agent-based models remains unmapped.
Inclusion criteria:
Population: Studies using population-based mortality records (civil registration and vital statistics); hospital and clinical registries are excluded.
Concept: Quantitative algorithms (proportional redistribution, multiple causes of death, Bayesian modeling, record linkage, matching, or machine learning) that reclassify garbage codes. Purely descriptive assessments and under-registration adjustments are excluded.
Context: Any mortality reporting setting globally, focusing on analytical feasibility for disaggregation or transferability to parameterize individual-level disease simulation models. Peer-reviewed and institutional grey literature in English, Spanish, or Portuguese are eligible.
Methods: Following JBI scoping review methodology and PRISMA-ScR (registered in OSF), six bibliographic databases (PubMed/MEDLINE, Embase, Scopus, Web of Science, LILACS and SciELO) and grey literature will be searched. Using R-based tools (revtools), two reviewers will independently screen records and extract data across six dimensions: study metadata, health system context, mortality data characteristics, garbage code typology, correction methodology, and microsimulation applicability. Findings will be synthesized through thematic narrative synthesis, evidence tables, method-by-ICD heatmaps, and conceptual flow networks linking vital statistics to simulation parameterization.
Keywords: health information systems; causes of death; vital statistics; International Classification of Diseases; data accuracy; machine learning; algorithms; computer simulation.
Mortality statistics derived from Civil Registration and Vital Statistics (CRVS) systems serve as the bedrock of global epidemiological surveillance, health technology assessment, and health policy planning1–3. However, despite decades of global advocacy, the analytical utility and international comparability of these databases are frequently undermined by persistent deficiencies in the medical certification of causes of death4,5. A critical challenge is the assignment of deaths to “garbage codes” (GCs)—a nosological concept introduced by Murray and Lopez in the 1990 Global Burden of Disease (GBD) study6 and subsequently expanded by Naghavi et al.7,8. These codes encompass International Classification of Diseases (ICD-9, ICD-10, or ICD-11) categories that represent ambiguous conditions, non-specific signs and symptoms (e.g., Chapter XVIII of ICD-10/ICD-11), intermediate mechanisms of dying (e.g., septicemia, cardiac arrest, heart failure), or unspecified anatomical sites within broader disease classes7,9. The World Health Organization (WHO) explicitly establishes that such uninformative diagnoses cannot and should not be selected as the underlying cause of death (UCoD)10,11.
To standardize the empirical assessment of data quality, garbage codes are systematically classified into four hierarchical severity levels based on their degree of specificity and potential to misguide public health policy8,9,12. Level 1 codes are the least specific, encompassing broad diagnostic categories (e.g., severe sepsis or ill-defined symptoms) that could plausibly be redistributed across the three top-level GBD groups: communicable, maternal, neonatal, and nutritional disorders; non-communicable diseases; or injuries. Levels 2 through 4 represent increasing, yet still insufficient, degrees of diagnostic detail—ranging from intermediate causes confined to a single broad group (Level 2) to unspecified sites within organ-specific families (Level 3) and unclassified subtypes within a single clinical condition, such as unspecified stroke (Level 4)12,13. Prioritizing redistribution algorithms targeting “major” garbage codes (Levels 1 and 2) is essential, as these misclassifications introduce the greatest distortion into national epidemiological profiles8,9.
The prevalence of garbage codes in official registries severely obscures population health trends, leading to substantial underestimations of critical chronic conditions, most notably ischemic heart disease14 and specific malignancies4,15,16. To mitigate these systematic biases, a wide spectrum of analytical correction frameworks has emerged over the past two decades. These methodologies range from Ledermann-type subnational regression modeling17 to traditional age- and sex-stratified proportional reallocation9 to multi-cause regression frameworks utilizing multiple causes of death (MCoD) data18,19. More recently, advanced data integration frameworks have incorporated deterministic or probabilistic record linkage with clinical and hospital discharge databases20,21, as well as non-parametric approaches such as Coarsened Exact Matching (CEM) and supervised machine learning architectures22–25. Prior to algorithmic correction, baseline data reliability is increasingly evaluated using standardized metrics such as the Vital Statistics Performance Index (VSPI) and automated audit tools like ANACONDA12,26.
Despite the proliferation of these algorithmic frameworks spearheaded by international bodies9,27, the extant literature remains fragmented. Most empirical applications focus narrowly on isolated disease entities or localized municipal jurisdictions22,25,28. Consequently, there is an absence of systematic mapping that consolidates the underlying mathematical assumptions, required data architectures, and trade-offs among competing algorithmic families27,29.
Crucially, this methodological gap becomes a major barrier when transitioning from aggregate demographic estimates to stochastic, individual-level computational models30. While macro-level demographic adjustments frequently assume that diagnostic misclassification occurs at random across populations, real-world vital registries demonstrate that garbage codes are concentrated disproportionately in vulnerable populations, elderly cohorts, and socioeconomically deprived settings12,31. Stochastic microsimulations and agent-based models require highly disaggregated baseline transition probability matrices. Directly applying unadjusted macro-level proportions into individual-level simulation architectures risks compounding classification errors and propagating structural biases into simulated life trajectories. A rigorous evidence synthesis is therefore urgently required to assess whether and how macro-level correction algorithms can be adapted or transferred into micro-level simulation environments.
A preliminary search of MEDLINE, the Cochrane Database of Systematic Reviews, PROSPERO, Open Science Framework (OSF) Registries, and JBI Evidence Synthesis was conducted, and no currently underway or existing systematic or scoping reviews on this specific topic were identified.
Therefore, this scoping review aims to systematically map and synthesize the evidence on quantitative methodologies used to identify, clean, and redistribute garbage codes in population-based mortality records, evaluate their analytical feasibility and transferability for parameterizing individual-level health microsimulations and agent-based models, and establish a translational conceptual framework to guide future public health mathematical modeling.
To achieve the overarching objective of this scoping review, the inquiry is guided by one primary review question formulated through the Population, Concept, and Context (PCC) framework:
What quantitative methodologies are documented in the literature to identify, clean, and redistribute garbage codes in population-based mortality registries, and what is their analytical feasibility and transferability for parameterizing individual-level health microsimulations and agent-based models?
To comprehensively operationalize this overarching inquiry, the review will address the following four specific sub-questions:
This scoping review will consider studies utilizing population-based mortality datasets, vital statistics systems, official national or subnational health statistics registries, or Civil Registration and Vital Statistics (CRVS) databases. The target population encompasses all deceased human individuals documented within these registries, without restrictions regarding age groups (including neonatal, infant, child, adult, and geriatric mortality) or sex and gender.
The core concept of this review encompasses quantitative, analytical, and computational methodologies designed to identify, clean, reclassify, or redistribute “garbage codes” or ill-defined causes of death in mortality registries. “Garbage codes” are defined in accordance with the World Health Organization (WHO) and Global Burden of Disease (GBD) frameworks as International Classification of Diseases (ICD-9, ICD-10, or ICD-11) codes representing ambiguous conditions, non-specific symptoms and signs (such as Chapter XVIII of ICD-10/ICD-11), intermediate or immediate mechanisms of dying (e.g., cardiac arrest, heart failure, septicemia), or unspecified sites within broad disease categories.
Eligible methodologies will include, but are not limited to:
Traditional demographic and statistical techniques: Proportional redistribution (simple, fixed, or stratified by age, sex, and region), fractional reallocation algorithms, and indirect demographic adjustments.
Advanced statistical and multi-cause modeling: Multiple imputation algorithms, spatial and spatiotemporal regression frameworks, multinomial logistic models, and empirical or hierarchical Bayesian algorithms (e.g., Cause of Death Ensemble model [CODEm] or CoDCorrect).
Record linkage approaches: Deterministic or probabilistic algorithms linking mortality microdata with complementary health information systems (such as hospital discharge registries, disease surveillance networks, forensic reports, or automated verbal autopsies).
Machine learning and data mining: Supervised algorithms (decision trees, Random Forest, Gradient Boosting, artificial neural networks) and non-parametric approaches (e.g., Coarsened Exact Matching [CEM]) applied to predict underlying causes of death.
Exclusion criteria:
Purely descriptive evaluations of mortality data quality (such as studies calculating isolated completeness percentages or reporting garbage code frequencies) that do not evaluate, propose, or implement an active quantitative algorithm for correction or redistribution.
Studies focusing exclusively on correcting numerical under-registration (the missing volume of unregistered deaths), unless the analytical framework integrates cause-of-death redistribution packages.
Research employing machine learning or statistical modeling exclusively for patient-level clinical prognosis, individual risk prediction, or survival forecasting without the explicit aim of reallocating or redistributing non-specific mortality codes.
To ensure an unbiased mapping of the global methodological landscape, this review encompasses any mortality reporting setting globally, across both high-income countries (HICs) and low- and middle-income countries (LMICs). The operational context considers health information systems characterized by varying levels of data completeness, diagnostic capacity, and medical certification accuracy.
The analytical focus of this context centers on assessing the feasibility, technical assumptions, and transferability of these redistribution methodologies to inform and parameterize individual-level health microsimulations and stochastic agent-based models.
Inclusion criteria:
Studies situated in any civil registration, vital statistics, or sample mortality surveillance system, ranging from well-resourced CRVS infrastructures to fragmented or data-scarce settings.
Studies describing, developing, or evaluating redistribution algorithms at either the aggregate population level (macro) or the individual record level (micro), provided they present sufficient detail regarding mathematical assumptions, classification rules, or data structures to evaluate their analytical transferability toward individual-level disease modeling platforms.
Exclusion criteria:
This scoping review will consider a broad range of quantitative and analytical study designs. Eligible designs include experimental and quasi-experimental studies (before-and-after studies evaluating training or certification interventions), analytical observational studies (prospective and retrospective cohort studies, analytical cross-sectional designs, and secondary analyses of population registries), and descriptive observational studies focusing on methodological validation. In addition, existing systematic and scoping reviews evaluating cause-of-death data quality and redistribution methods will be considered to identify relevant primary literature through reference chaining. Methodological treatises, academic dissertations, and peer-reviewed technical reports will also be eligible. Purely qualitative studies (e.g., phenomenological or ethnographic inquiries) and narrative text or opinion papers lacking mathematical, statistical, or algorithmic content will be excluded.
Authoritative institutional grey literature—including technical guidelines, working papers, and official reports from international and regional health bodies (the World Health Organization [WHO], the Pan American Health Organization [PAHO], and the Institute for Health Metrics and Evaluation [IHME])—will be considered for inclusion.
Timeframe: Literature published from inception to the present, ensuring the capture of foundational demographic adjustments (such as subnational regressions and classical GBD algorithms) alongside modern machine learning developments.
Languages: Peer-reviewed articles and institutional reports published in English, Spanish, and Portuguese will be included, reflecting the substantial methodological literature generated in Latin America (notably from Brazil’s mortality surveillance initiatives).
The proposed scoping review will be conducted in accordance with the JBI methodology for scoping reviews32. The protocol was registered a priori on the Open Science Framework osf.io/374tm.
The search strategy will aim to locate both published and unpublished (grey) literature across the entire available historical continuum, capturing foundational demographic adjustments alongside contemporary machine learning developments.
A three-step search strategy will be utilized: 1. Initial search: An initial limited search of MEDLINE (via PubMed) was undertaken to analyze text words in titles, abstracts, Medical Subject Headings (MeSH), and author-supplied keywords of relevant articles. 2. Full search strategy: A comprehensive search strategy was developed using these identified terms and index descriptors, then adapted for each database’s syntax and thesaurus (see Appendix I for the full PubMed/MEDLINE search strategy). Electronic databases to be searched from inception to the present include: PubMed/MEDLINE (National Library of Medicine), Embase (Elsevier), Scopus (Elsevier), Web of Science Core Collection (Clarivate), LILACS (BIREME/PAHO/WHO), and SciELO (FAPESP/BIREME). Studies published in English, Spanish, and Portuguese will be included. 3. Reference list screening: The reference lists of all included evidence sources and relevant methodological reviews will be screened for additional primary studies.
Unpublished and institutional grey literature will be searched through targeted exploration of key health authorities—including the World Health Organization (WHO), Pan American Health Organization (PAHO), Institute for Health Metrics and Evaluation (IHME), as well as official repositories of national ministries of health, national statistical institutes (NSIs), and specialized repositories (CADTH Grey Matters, BASE, and OpenGrey). Open web searches via Google Scholar will be systematized using Publish or Perish software. In adherence to PRISMA-S guidelines, search strings, platforms, dates, and hit counts will be logged in an audit tracking table.
Following the search, all identified citations will be collated and uploaded into the R statistical computing environment (version 4.5.1; R Core Team, Vienna, Austria). Deduplication will be conducted programmatically using the synthesisr package33 via fuzzy string-matching algorithms, supplemented by manual verification of edge cases.
Following a pilot test on a random sample of 50 records to calibrate inter-rater consistency, titles and abstracts will be screened independently in duplicate by two reviewers against the eligibility criteria using the revtools package (version 0.4.1; Westgate), supported by Latent Dirichlet Allocation (LDA) topic modeling strictly as an unsupervised decision-support tool.
Potentially relevant sources will be retrieved in full text and evaluated in detail against the inclusion criteria by two independent reviewers. Reasons for exclusion of full-text sources will be recorded hierarchically (see Appendix IV for the dual-independent screening protocol, inter-rater calibration criteria, and exclusion decision tree) and reported in the scoping review. Any disagreements at any stage of the selection process will be resolved through discussion or by arbitration with a third reviewer. The final search will be re-run immediately prior to final manuscript preparation. Results of the search and study inclusion process will be reported in full in the final review and presented in a PRISMA-ScR flow diagram.
Data will be extracted from included sources by two independent reviewers using a customized data extraction instrument developed in R (see Appendix V). The instrument captures variables across six core dimensions: (I) Study Identification and Metadata, (II) Geographic and Health System Context, (III) Mortality Database Characteristics, (IV) Garbage Code Typology and Nosology, (V) Correction and Redistribution Methodology, and (VI) Microsimulation Applicability and Transferability.
The draft extraction tool will be piloted with a random sample of three included studies to ensure clarity and variable granularity; any resulting modifications will be detailed in the final review. Discrepancies between reviewers will be resolved through consensus or a third reviewer. If appropriate, authors of papers will be contacted to request missing data or technical clarifications where required.
In accordance with JBI scoping review methodology, formal critical appraisal of individual sources of evidence will not be conducted, as the objective is to map and synthesize the spectrum of existing quantitative methodologies rather than to evaluate study-level risk of bias.
The charted data will be synthesized and presented to directly address the primary review objective and four sub-questions, utilizing R for all data manipulation and visualization.
Ethical approval is not required for this scoping review protocol, as the methodology involves the collection, synthesis, and secondary analysis of publicly available literature, published studies, and open-source documentation, with no direct contact with human participants or identifiable private data. The findings will be disseminated through peer-reviewed publication and academic presentations
All datasets generated or analyzed during this study are included in this published protocol and its supplementary information files. The complete R scripts, customized search syntaxes, and electronic data charting instruments are publicly accessible via the Open Science Framework (OSF) repository osf.io/374tm. Any additional customized scripts, codebooks, or intermediate extraction matrices are available from the corresponding author upon reasonable request.
The authors declare that they have no competing interests. This scoping review was conducted solely for academic, educational, and methodological research purposes, with no commercial, financial, corporate, or industrial support that could be construed as a potential conflict of interest.
The authors expresses sincere gratitude to Dr. Giulia Lorenzoni and Prof. Dario Gregori (Unit of Biostatistics, Epidemiology and Public Health, Department of Cardiac, Thoracic, Vascular Sciences and Public Health, University of Padova, Italy) for their intellectual guidance, mentorship, and invaluable methodological supervision throughout the conceptualization and development of this doctoral scoping review protocol. Gratitude is also extended to Dr. Pablo Salazar (School of Medicine, Universidad Dr. José Matías Delgado, San Salvador, El Salvador) for his for his dedicated advisory support, insightful technical discussions, and critical contributions to contextualizing national mortality registries and health information.
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The study was conducted entirely through the academic time, institutional affiliation resources, and professional dedication of the authors.
During the preparation of this scoping review protocol, the authors used Gemini (Google) exclusively as a writing and language-editing support tool to assist in translating, refining academic English syntax, improving textual conciseness to adhere to word-count constraints, and formatting document sections according to JBI guidelines. The authors did not use this tool to conceive the research questions, design the methodology, select inclusion criteria, or extract analytical parameters. Following its use, the authors thoroughly reviewed, edited, and verified all outputs against the primary literature and assume full responsibility for the intellectual integrity, factual accuracy, and content of the final manuscript.
All underlying search syntaxes, programmatic R pipelines, PRISMA reporting checklists, screening protocols, and data extraction codebooks are permanently archived and openly available in the Open Science Framework (OSF) project repository: - Appendix I: Electronic Search Strategy (PubMed / MEDLINE API Syntax). - Appendix II: PRISMA-S Literature Search Checklist and Audit Table. - Appendix III: PRISMA-ScR Reporting Checklist. - Appendix IV: Dual-Independent Screening Protocol, Inter-Rater Calibration, and Exclusion Taxonomy. - Appendix V: Standardized Data Extraction Instrument, Codebook, and Microsimulation Readiness Level (MRL) Criteria.