SitRep-NLP: An Open-Source Pipeline for Automated Multi-Agency Situation Report Synthesis Aligned with NIMS ICS-209 and FEMA Community Lifelines
American Journal of Health and Medical Sciences · 01.05.2026 · 16 (05) · 01-22
DOI: 10.63125/q2p5q504 ↗Kurz gefasst
Fachartikel in der Zeitschrift American Journal of Health and Medical Sciences, online erschienen am 01.05.2026. Verfasst von Razia Sultana.
Abstract (Originalsprache)
Emergency operations centres (EOCs) compile the ICS-209 incident status summary and the status of FEMA's Community Lifelines by hand from many agencies' text. We present SitRep-NLP, an open-source pipeline that does this with a 1.5-billion-parameter open-weight language model (Qwen2.5-1.5B-Instruct, 4-bit) running on a 2015 desktop processor, so no data leave the EOC and there is no per-use cost. It was evaluated on public data for Hurricane Harvey (2017) and three later storms in Southeast Texas, with Jefferson County as case study. Six tasks were scored against human labels (HumAID), FEMA support-function codes, rule-based parses of fixed-format National Weather Service products, and draft labels released for verification. The model read storm position, intensity and motion correctly from 94% of 124 hurricane local statements and transcribed storm-report casualty counts reliably, but gave the issuance time correctly in 2% of statements and all nine storm-report fields in 31%. On message triage its macro-F1 was 0.32, against 0.65 for a logistic-regression classifier trained on labelled Harvey posts. It over-tagged lifelines (precision 0.19 against draft labels) and matched the coded lifeline of 68% of federal mission assignments, no better than keywords. With the model writing all of it, a daily county ICS-209 draft took 2.4 minutes of machine time on average (8.6 minutes including extraction). Free-text drafts looped or lost sections. Grammar-constrained drafts were complete, but 8 of 10 contained unsupported statements (25% of statements). A hybrid variant gave the model only the two narrative blocks, with rules, a supervised classifier and cited evidence doing the rest. It cut unsupported statements to 10%, all in the model's blocks, listed requests that the model-only drafts had missed, and needed 37 seconds per report instead of 8.6 minutes; 54% of its cited lifeline evidence was relevant. A small local model is useful for reading semi-structured text, not for judgement or unsupervised drafting; rule parsers, supervised classifiers and analyst review must carry the rest. Time saved against manual compilation was not measured; a measurement protocol is given.
Zitieren
APA
Sultana, R. (2026). SitRep-NLP: An Open-Source Pipeline for Automated Multi-Agency Situation Report Synthesis Aligned with NIMS ICS-209 and FEMA Community Lifelines. American Journal of Health and Medical Sciences, 16(05), 01-22. https://doi.org/10.63125/q2p5q504
Vancouver
Sultana R. SitRep-NLP: An Open-Source Pipeline for Automated Multi-Agency Situation Report Synthesis Aligned with NIMS ICS-209 and FEMA Community Lifelines. American Journal of Health and Medical Sciences. 2026;16(05):01-22. doi:10.63125/q2p5q504
BibTeX
@article{Sultana2026,
author = {Sultana, Razia},
title = {SitRep-NLP: An Open-Source Pipeline for Automated Multi-Agency Situation Report Synthesis Aligned with NIMS ICS-209 and FEMA Community Lifelines},
journal = {American Journal of Health and Medical Sciences},
year = {2026},
volume = {16},
number = {05},
pages = {01--22},
doi = {10.63125/q2p5q504}
}
EUVIMED