Structuring evidence in air pollution and cognitive impairment research using a role-analysis framework built on LLM-based entity extraction

Rapid growth in biomedical literature has increased the need for auditable methods that convert article-level text into structured evidence data. Air pollution is increasingly framed as a modifiable risk factor for cognitive decline and dementia, but publication growth alone does not indicate a structurally diverse evidence base. We aimed to develop and apply a human-validated large language model (LLM)-assisted role-analysis framework based on structured entity extraction to characterize recent air pollution–cognition evidence. Operationally, role-analysis retains pollutant and cognition-related entities as multi-label fields while assigning a single predominant study-level analytical role to each domain. We conducted an abstract-level evidence-structure analysis of PubMed-indexed human original studies published from 1 January 2020 through 22 March 2026. PubMed formed the primary analytic corpus, and Web of Science was used as a supplementary coverage check. A predefined coding schema was used to extract pollutants, cognitive outcomes, study designs, statistical methods, populations, data sources, geographic settings, and analytical roles from titles and abstracts. LLM-assisted screening, extraction, and role coding were constrained by field definitions, conservative coding rules, and consistency checks. Human validation evaluated screening, extraction, and role-coding outputs. Of the 1,061 PubMed records that entered the four-domain eligibility screening, 406 eligible studies were included. In the screening-validation sample ( n = 106), precision was 1.000, recall was 0.881, accuracy was 0.953, F1 score was 0.937, and Cohen’s κ was 0.899. In the structured-extraction and role-coding validation sample ( n = 81), field-level agreement ranged from 0.765 to 1.000; agreement was 0.926 for pollutant roles and 0.901 for cognition roles. The structured evidence dataset showed concentration around particulate matter and gaseous pollutants, older-adult populations, and cognitive outcomes including global cognitive performance, domain-specific cognition, and dementia-related outcomes. Direct pollutant exposure–cognitive outcome analyses dominated the role structure, whereas mediation, effect modification, and alternative role configurations were uncommon. Evidence sparsity reflected overlapping limitations in pollutant characterization, cognitive domain coverage, analytical role diversity, population representation, geographic distribution, and data-source dependence. The post-2020 air pollution–cognition literature is volume-rich but structurally uneven. A human-validated LLM-assisted role-analysis framework can convert biomedical abstracts into auditable evidence data that preserve multi-label study entities together with study-level analytical-role summaries and help distinguish growth in study volume from structural diversity of the evidence base.