An AI-assisted map reveals uneven attention in disaster mental health research

After a disaster, researchers can ask many different questions: who develops post-traumatic stress, how communities recover, or whether substance use changes. But the resulting literature poses another question: how evenly is that attention distributed?

Soheil Shapouri and Saber Soleymani address this in an arXiv preprint using Claude Sonnet 4.5 to classify scientific abstracts. Their analysis describes the contents of a particular research collection, rather than estimating how common mental health problems are among disaster survivors. It also illustrates both the usefulness and the uncertainty of asking AI to map scientific work at scale. The supplied source is a preprint, not evidence of completed peer review.

From abstracts to a research map

The researchers searched PubMed in May 2026 using “disaster” AND “mental health,” retrieving 8,586 records. Removing records without abstracts left 7,859. After removing duplicate titles, duplicate identifiers and parsing errors, and restricting the collection to relevant research, 5,614 articles remained, spanning 1970–2026.

The model received each article’s title and abstract, together with predefined categories and coding rules. It extracted information including disaster type, mental health outcomes, participant country, sample size and primary analytical method. This goes beyond counting publication keywords, but it remains an analysis of abstracts—not a full-text assessment of thousands of studies.

For a quality check, an author independently coded 15 randomly selected articles. Across applicable variables, the model’s classifications matched that coding exactly 95.8% of the time. Where several categories were possible, agreement required the entire set to match.

That is encouraging, but 15 articles provide only a small check on a varied collection. The reported agreement does not establish equally reliable performance for every category, particularly rare ones.

Where attention concentrates

Earthquakes appeared in 22.35% of the analyzed articles, followed by epidemics or pandemics at 17.76%, hurricanes or storms at 12.13%, and floods at 8.23%. Extreme temperatures appeared in 1.82%. Some articles covered multiple disaster types.

Post-traumatic stress disorder, or PTSD, was the most frequently studied mental health outcome, appearing in 37.1% of articles. Depression appeared in 30.5%, and anxiety or stress in 24.8%. Substance use and suicidality each appeared in fewer than 5%, while aggression appeared in 0.7%. These categories overlap: one article could examine several outcomes.

Those numbers describe research attention, not the relative frequency or seriousness of these conditions. Fewer publications on an outcome do not, by themselves, establish how much additional research it needs.

Geographic representation was also concentrated. The United States, Japan and China were the most commonly identified participant countries. However, participant country was not applicable for about 20% of articles and unclear for roughly 8.5%, limiting how completely the map locates the populations studied.

Few computational methods—but what follows?

Classical statistics and hypothesis testing dominated the methodological classifications, appearing in 53.5% of articles. Machine learning appeared in 0.84%, causal inference in 0.78%, and natural language processing—computational analysis of language—in 0.43%.

Among articles with an available sample size, the reported median was 452, with the middle half ranging from 142 to 1,568. The authors argue that larger samples could support greater use of computational approaches. But their analysis does not test whether switching methods would improve predictions or care. Sample size alone does not demonstrate that a particular method is appropriate. The paper also notes that some exceptionally large counts reflect weighted population estimates or social media posts rather than straightforward participant totals.

A useful map with boundaries

PubMed emphasizes biomedical and life-science literature. Other databases could produce a different picture, and excluding records without abstracts creates another boundary. Abstracts may omit methods or study details, leaving the model unable to recover them. Because collection stopped in May, the 2026 publication count is incomplete.

The authors report that extraction took a few hours and cost under $200. That suggests a practical way to organize a large literature. The strongest conclusion, however, is narrower than a diagnosis of an entire field: this PubMed collection shows marked concentrations in research attention, offering leads for closer investigation rather than a definitive ranking of future priorities.