Methods
Design and preregistration
This is a descriptive, denominator-type census. Two denominators are fixed first — a 14-area publication pool and, inside it, a seven-cause burden pool — and every quantity is a composition within them. The protocol, disease pool and analysis layers were fixed in a signed work report and preregistered on the Open Science Framework before record-level production began (registration DOI 10.17605/OSF.IO/NUC4B). One amendment is documented: the Institute for Health Metrics and Evaluation retired GBD 2021, so Amendment v1.1 substituted GBD 2023 (burden years 1995–2023). The study measures the configuration of research allocation only; it makes no clinical-value judgement and no claim about how much research any disease deserves.
Publication census (L1)
Publications were counted in Europe PMC under one frozen title/abstract query per area, per year, 1995–2024 (420 area–year cells), and aggregated into six five-year periods. The L1 counts use the Europe PMC hitCount basis (392,458 records). Record-level retrieval for the geographic layer delivered 392,456 records; two parts each contained one "phantom" record — counted in the hitCount but never delivered (D-1: oral cancer 2010–2014, 4,806 vs 4,805 retrievable; D-2: obstructive sleep apnoea 2015–2019, 11,604 vs 11,603). Each was accepted only after three scripted checks passed (unique identifiers; stable hitCount on re-query; identical identifier sets from an independent second retrieval channel) and both are itemised in a frozen discrepancy record shipped with the archive.
Country parsing (L3 production)
A first-author affiliation was present for 356,332 records (90.80%). Affiliation strings were parsed to a country or region by a four-layer dictionary procedure (institution and city dictionaries with disambiguation rules; no imputation), succeeding for 340,582 records (95.58% of those with an affiliation) across 159 countries or regions. A 90-record manual audit found no errors. Author country is a property of the publishing institution, not of the study population.
Burden data (L2 inputs)
GBD 2023 DALYs (primary) and prevalence (sensitivity calibre), all ages, both sexes, global aggregate plus 204 countries and territories (205 locations), annually 1995–2023, for cause ids 329, 674, 423, 444, 447, 450 and 480. Country DALY sums reproduce 99.83–99.96% of the corresponding global totals. The DALY Number metric is primary; the all-ages rate in the downloads is not age-standardised and is not used. Last-period pairing: burden periods match publication periods for the first five periods; in the last period the burden side covers 2020–2023 against publications 2020–2024. This one-year asymmetry is disclosed wherever last-period burden-relative quantities appear, and a sensitivity analysis restricts last-period publications to 2020–2023.
Attention–burden ratio (L2)
For period t, ABR = (a cause's share of the seven-cause publication pool) ÷ (its share of the seven-cause DALY pool). ABR 1 is proportionality within the pool; log2 ABR is reported, so +1 is a twofold publication excess over burden share and −1 a twofold deficit. Both shares are compositions inside the same seven-cause pool: the ABR is a within-specialty relative measure and supports no comparison outside the pool.
Geographic layer (L3 analysis)
Double-centroid distance. The burden centroid is the DALY-weighted mean centre of country centres (frozen snapshot of the mledoze/countries dataset); the research centroid is the same construction weighted by parsed first-author records. The primary estimator is the arithmetic mean centre: the weighted arithmetic mean of latitudes combined with the circular (unit-vector) mean of longitudes; distances are haversine great-circle distances (R = 6,371.0088 km). The estimator was chosen after a documented diagnostic episode: the chordal (unit-vector) mean, initially frozen, placed several research centroids at implausible Arctic latitudes (oral cancer 1995–99 at 82.0°N) because first-author distributions ring the Pacific and their horizontal components cancel; the spherical geometric median avoided the pole but degenerated for nasopharyngeal cancer, both centroids snapping to the same country centre and the distance collapsing to 0.0–0.3 km for three consecutive periods. All three estimators are frozen side by side for all 42 cause–period cells in the estimator-sensitivity table shipped with the archive; all three agree on the direction of every first-to-last distance change.
Author shares. The LMIC share is the proportion of parsed records whose first author is affiliated in a low- or middle-income country (World Bank low-, lower-middle- and upper-middle-income groups combined), with a decomposition excluding China. Income groups are the World Bank classification as frozen on 12 November 2026, held fixed across all periods — a disclosed simplification, since countries' classifications changed over 30 years. For burden-tier causes, the share of first authors in that cause's ten highest-DALY countries is also reported.
Decoupling typology (L4)
Each burden-tier cause is classified from three signals comparing the first and last periods: S, the relative change in its share of the 14-area pool; B, the relative change in its share of the seven-cause DALY pool; and A, the change in log2 ABR. The signed protocol fixed the type names and definitions — resolution-dilution, technology-ignited, burden-tracking, constant — and anything matching no rule is reported as unclassified, never forced. The protocol did not fix numeric thresholds; these were operationalised at the typology stage, before the classification was run (main ±25% for S and B, ±0.5 for A; sensitivity sets ±15%/±0.3 and ±35%/±0.7 are reported with every assignment). The typology is therefore a descriptive classification under pre-stated definitions, not preregistered inference. Two reading rules apply: B is a share, so "burden falls" means relative dilution inside the pool only (annualised absolute DALYs rose for all seven causes, +3.9% to +101.3%); and type names are shape labels, not mechanism claims.
Sensitivity programme and verdicts
- Lag window. Pairing last-period publications with burden 0, 5 or 10 years earlier changed no cause's log2 ABR sign (7/7 identical across lags); drift is at most 0.136 per five-year step and 0.271 over the full span.
- Publication window. Restricting last-period publications to 2020–2023 moved each last-period log2 ABR by less than 0.1, with no sign change.
- Burden calibre. The prevalence calibre agrees with the primary calibre on five of seven last-period signs and five of seven trend directions. The sign exceptions are laryngeal cancer and otitis media (hearing loss is 92.76% of the seven-cause prevalence pool, which compresses the other shares); the trend exceptions are other pharyngeal and thyroid cancer, both near-flat in both calibres. The two headline level claims — hearing loss deeply below parity and thyroid cancer at the largest excess — hold in both calibres.
- Terminology drift. For each area, one backup query was specified and frozen in writing before any backup count was retrieved. The two queries agree on share direction for 12 of the 15 rows. Cochlear implants and tinnitus differ only across the ±5% direction band. CRS/polyps reverses outright (−26.6% primary vs +251.0% backup): the backup query uses only modern terms that entered routine use after the 2003–2004 consensus terminology, so it is structurally blind to the sinusitis-era literature. The reversal is read as terminology migration, and under the protocol's two-query rule the CRS share decline is withdrawn as a headline conclusion.
The lag-window and terminology analyses were completed on 12 November 2026, after the primary layers were frozen; the backup queries were fixed before their counts existed. They are reported as completions of the protocol's sensitivity programme at the analysis stage, not as elements of the original preregistration timestamp.
Use of large language models
Large language models were used as assistive tools in this project: for development and debugging of analysis code, for orchestration of the analysis pipeline, and for assistance in drafting the manuscript. All data were produced by deterministic scripts from the public sources named above; every number reported here was verified against the frozen result files; no data, result or figure was generated by a language model.