URGENTIAPARSE: Parsing Emergency & Crisis Data

Updated 19 July 2026

URGENTIAPARSE is an integrated framework that processes mixed structured and unstructured data to parse urgency-bearing messages in both emergency medicine and disaster informatics.
It leverages transformer embeddings (FlauBERT) combined with XGBoost and SHAP explainability to generate robust triage predictions and rapid classification of crisis alerts.
The framework outperformed comparative models in clinical triage (composite score 2.514) and achieved high accuracy in disaster notice classification, demonstrating its practical impact.

URGENTIAPARSE denotes an emergency-information parsing framework with two distinct usages in the cited literature. In emergency medicine, it is the LLM-based system developed to predict adult emergency department triage levels and immediate disposition from triage narratives and structured clinical variables, benchmarked against consensus adjudication on the FRENCH scale at Roger Salengro Hospital, CHU Lille (Lansiaux et al., 1 Jul 2025). In disaster informatics, the same name is used in integrated system designs for rapid detection, classification, geolocation, and redistribution of urgent notices and crisis messages, especially hurricane evacuation notices, by combining transformer classification, spatial targeting, urgency detection, and real-time stream parsing (Zhao et al., 2024).

1. Scope, nomenclature, and problem domain

In the emergency department study, URGENTIAPARSE is one of three artificial intelligence models compared for triage prediction: TRIAGEMASTER (NLP), URGENTIAPARSE (LLM), and EMERGINET (JEPA). The motivating problem is persistent undertriage and overtriage in the setting of increasing patient influx and staffing constraints. The clinical objective is concordance with a gold-standard adjudication using the national FRENCH triage scale, with secondary evaluation on GEMSA coding, which reflects immediate emergency department outcome and consumption of medical resources (Lansiaux et al., 1 Jul 2025).

In the disaster-notice setting, the problem formulation is different but structurally related. Evacuation notices are issued by numerous local authorities across multiple states, undergo frequent updates, and appear in heterogeneous online formats including press releases, alerts, webpages, and social media posts. The operational objective is rapid detection, classification into actionable categories such as mandatory or voluntary evacuation, and redistribution through a Web GIS for situational awareness and archiving (Zhao et al., 2024).

A common feature across these usages is the dependence on mixed structured and unstructured inputs under real-time or near-real-time constraints. This suggests a shared emphasis on parsing urgency-bearing language embedded in noisy operational environments, although the validated, explicitly described URGENTIAPARSE implementation is the emergency department triage model (Lansiaux et al., 1 Jul 2025).

2. Clinical formulation in emergency department triage

The emergency department study was conducted in the adult Emergency Department of Roger Salengro Hospital, CHU Lille, France, over seven consecutive months from June 1 to December 31, 2024. It analyzed a prospectively recruited cohort of adult patients aged at least 18 years. The full emergency department volume was 73,236 visits; 681 triage interviews were recorded; 24 were excluded, including 6 due to missing data; and 657 patients were included. The mean age was 43, and 49.77% were women (Lansiaux et al., 1 Jul 2025).

The dataset integrated structured variables and unstructured triage narratives. Structured features came from ResUrgences® version 2024.1.148 and included age, sex, systolic and diastolic blood pressure, heart rate, temperature, pain E.V.A., oxygen saturation, oxygen flow, nurse-provided triage classifications, and timestamps. Unstructured data consisted of audio recordings of the intake interview, manually transcribed verbatim by medical students and reviewed by medical annotators. Text and structured data were synchronized through unique identifiers (Lansiaux et al., 1 Jul 2025).

Gold-standard labeling used the FRENCH scale adjudicated by two senior emergency physicians and two triage nurses, all blinded to model outputs and initial nurse triage; disagreements were reconciled by consensus. The dataset included FRENCH categories 1, 2, 3A, 3B, 4, and 5, but 3A was absent in the included sample. The primary endpoint was concordance between model-predicted FRENCH level and the gold-standard consensus. Secondary analyses targeted GEMSA prediction and robustness across input modalities, specifically structured nurse-recorded history versus raw transcript (Lansiaux et al., 1 Jul 2025).

The reported metrics comprised micro- and macro-F1, Weighted Cohen’s Kappa, Spearman correlation, Mean Absolute Error, Root Mean Square Error, exact agreement, near agreement within $\pm 1$ class, and a composite z-score. The composite score was defined as

$\mathrm{Score} = Z(-\mathrm{MAE}) + Z(-\mathrm{RMSE}) + Z(\mathrm{Kappa}) + Z(\mathrm{Spearman}).$

The exact normalization constants for the $Z$ operator and the exact weighting scheme for weighted kappa were not fully specified in the text (Lansiaux et al., 1 Jul 2025).

3. Architecture and training of the LLM-based triage model

URGENTIAPARSE is described conceptually as an embedding-and-boosting pipeline rather than a chat-style instruction-following system. Its LLM component is FlauBERT, used to produce contextual embeddings of free-text chief complaints and triage interviews in French. These contextual embeddings are concatenated with structured clinical variables and passed to an XGBoost classifier. Feature importance is interrogated with SHAP values, allowing attribution of predictive influence to words or tokens and to structured variables (Lansiaux et al., 1 Jul 2025).

This design contrasts with the two comparison systems. TRIAGEMASTER uses Doc2Vec paragraph embeddings concatenated with structured features and a feedforward neural network classifier with ReLU activations and dropout/L2 regularization. EMERGINET uses a Joint Embedding Predictive Architecture that maps inputs and targets into a shared latent space via dual encoders, with text processed by FlauBERT plus LSTM, and optimizes a contrastive energy objective regularized by VICReg (Lansiaux et al., 1 Jul 2025).

Training used a stratified split preserving class balance by FRENCH level, with 80% of data for training and 20% for validation, corresponding to $n=526$ and $n=131$ . Hyperparameters were tuned by grid search. Training proceeded for 100 iterations, and learning-curve diagnostics were recorded. No explicit probability calibration, thresholding, or post-processing was reported for URGENTIAPARSE. The paper notes that TRIAGEMASTER’s probability thresholds were set at maximum F1 score, but no analogous thresholding was described for URGENTIAPARSE. Tokenization or normalization details for FlauBERT, the exact FlauBERT variant, parameter count, XGBoost hyperparameters, computational hardware, and calibration details were not reported (Lansiaux et al., 1 Jul 2025).

The paper attributes URGENTIAPARSE’s superior performance to LLM-based “abstraction of patient representations,” meaning the ability to encode nuanced semantics and clinical context from unstructured narratives into latent features that align more closely with triage categories and downstream outcomes. The absence of prompt design is consequential: URGENTIAPARSE is not presented as a conversational LLM, but as a hybrid representation-learning and gradient-boosting architecture (Lansiaux et al., 1 Jul 2025).

4. Comparative performance in FRENCH and GEMSA prediction

The primary analysis found URGENTIAPARSE to be the top-performing system for FRENCH triage prediction, with a composite score of 2.514. This exceeded EMERGINET at 0.438, TRIAGEMASTER at $-3.511$ , and initial nurse triage at $-4.343$ (Lansiaux et al., 1 Jul 2025).

System	Composite score	Selected FRENCH metrics
URGENTIAPARSE	2.514	MAE 0.228; RMSE 0.790; Kappa 0.800; Spearman 0.802; F1-micro 0.900; F1-macro 0.894; exact 0.900; near 0.928
EMERGINET	0.438	MAE 0.401; RMSE 0.979; Kappa 0.560; Spearman 0.602; F1-micro 0.731; F1-macro 0.747; exact 0.820; near 0.860
TRIAGEMASTER	-3.511	MAE 0.637; RMSE 1.180; Kappa 0.370; Spearman 0.005; F1-micro 0.618; F1-macro 0.613; exact 0.554; near 0.696
Nurse triage	-4.343	MAE 1.393; RMSE 1.834; Kappa 0.080; Spearman 0.024; F1-micro 0.303; F1-macro 0.275; exact 0.303; near 0.498

Confusion-matrix analyses showed URGENTIAPARSE and EMERGINET slightly overestimated lower triage levels, whereas TRIAGEMASTER tended to underestimate and nurse triage displayed variable dispersion. Bland–Altman plots showed URGENTIAPARSE and EMERGINET with low, centered errors; TRIAGEMASTER had positive bias; nurse triage was widely dispersed. Class-wise F1 was reported as consistently strong for URGENTIAPARSE and EMERGINET, with per-class F1 between 0.8 and 1.0 (Lansiaux et al., 1 Jul 2025).

In the secondary GEMSA analysis, URGENTIAPARSE again led, with MAE 0.082, RMSE 0.402, weighted kappa 0.863, Spearman 0.864, F1-micro 0.957, F1-macro 0.628, exact agreement 0.957, near agreement 0.960, and composite score 2.382. EMERGINET and TRIAGEMASTER numerics for GEMSA were not reported in the table, but narrative and visual summaries placed EMERGINET close behind URGENTIAPARSE and TRIAGEMASTER as notably weaker (Lansiaux et al., 1 Jul 2025).

Robustness analyses favored structured inputs over raw transcripts across models, especially for URGENTIAPARSE, which achieved Weighted Kappa and Spearman greater than 0.8 on structured inputs. EMERGINET showed moderate alignment on transcripts, with Kappa 0.43 and Spearman 0.56. TRIAGEMASTER was poor on structured data, with near-zero Kappa and Spearman 0.16, and worse on transcripts. Learning-curve diagnostics for URGENTIAPARSE showed training accuracy at 1.0 and validation accuracy plateauing around 0.5, with training log loss near zero and validation loss remaining high, indicating overfitting in that setup, even though final validation metrics for FRENCH and GEMSA were strong (Lansiaux et al., 1 Jul 2025).

5. URGENTIAPARSE in disaster and crisis informatics

In integrated summaries of disaster-response research, URGENTIAPARSE is used as the name of a real-time urgent notice parsing, classification, and redistribution system. For hurricane evacuation notices, the proposed pipeline begins with spatially targeted source discovery using National Weather Service APIs to identify counties under hurricane or tropical storm alerts, retrieve FIPS codes, and match them to official county and state sources. Collection uses twitter-scraper, facebook-scraper, and BeautifulSoup, with manual collection from government websites and news outlets for earlier periods. The classification layer uses BERT base and RoBERTa base encoders with WordPiece tokenization, a maximum sequence length of 512 tokens, AdamW at learning rate $5.0\times10^{-6}$ , batch size 4, cross-entropy loss, early stopping, and 10-fold cross-validation (Zhao et al., 2024).

The reported evacuation-notice corpus contained 1,305 texts collected from 218 counties across 19 states since 2001, with 1,227 labeled items. The labeled set included 730 evacuation notices, subdivided into 489 mandatory and 241 voluntary notices, plus 497 non-evacuation texts. In binary classification, RoBERTa achieved weighted F1 of approximately 0.918 and accuracy of approximately 0.919. In three-class classification, RoBERTa achieved accuracy of approximately 0.882 and weighted F1 of approximately 0.882, while mandatory-evacuation recall reached approximately 0.961 for BERT and approximately 0.937 for RoBERTa. The paper highlights high recall for mandatory notices as an operational requirement to minimize missed critical orders (Zhao et al., 2024).

A second component relevant to broader URGENTIAPARSE designs is real-time location extraction from social media. SAVITR performs unsupervised extraction of locations from non-geotagged tweets using hashtag decompounding, spaCy POS tagging and dependency parsing, heuristic candidate extraction, and gazetteer verification against GeoNames or OpenStreetMap. On 101 annotated tweets, its GeoLoc method achieved precision 0.7485, recall 0.8389, and F1 0.7911, while processing the batch in about 1.1901 seconds, roughly 12 ms per tweet. In a larger emergency corpus, 317,567 tweets were collected, 239,276 distinct English tweets remained after de-duplication and language filtering, and 68,793 tweets, about 26.15% of the 239,276, were tagged with inferred locations using GeoNames (Dutt et al., 2018).

Urgency detection in short crisis messages provides a further layer. A low-supervision ensemble combines crisis-local fastText embeddings, Wikipedia fastText embeddings, and 11 manual features consisting of ten urgency-related stems plus a digit-present feature. Three linear regression classifiers are trained and combined by a weighted sum in the low-supervision setting or by averaging in the transfer setting. On Nepal earthquake data, the ensemble achieved accuracy 69.25%, recall 70.50%, and F1 69.44%; on Kerala floods, it achieved accuracy 68.50%, recall 52.76%, and F1 51.62%. Transfer experiments showed that simple up-sampling of scarce target labels and averaging across component models could outperform several baselines on Macedonia and cross-crisis transfer tasks (Kejriwal et al., 2019).

For streaming text segmentation and parsing, USTEP contributes an online evolving-tree method originally developed for log parsing. Its per-message complexity is $O(|t_i|)$ and expected $O(1)$ when mean message length is stable. Across 10 labeled datasets, it achieved average parsing accuracy 0.937, exceeding Drain at 0.903, LenMa at 0.855, Spell at 0.810, and SHISO at 0.707. On large datasets such as HDFS-2, OpenStack-2, and Internal-1, USTEP and Drain exhibited constant processing-time curves, and both completed Internal-1 in roughly 5 hours, whereas Spell, LenMa, and SHISO did not complete within a week on Internal-1 (Vervaet et al., 2023).

Taken together, these results define a broader URGENTIAPARSE design space in which source targeting, text classification, urgency estimation, location resolution, and online parsing are treated as interoperable components. This is a systems-level interpretation derived from the integrated summaries rather than a single end-to-end implementation reported in one paper.

6. Limitations, ethics, and future directions

The emergency department URGENTIAPARSE study has several explicit limitations. It is single-center, adult-only, retrospective, and covers a very small proportion of total emergency department encounters: 657 included patients out of 73,236 visits, or 0.897%. Generalizability is limited by the academic tertiary setting and by local protocols. Transcript quality varied, fewer nurses were represented in the audio corpus, one recorder lost data, and data and full code are not available because of patient data constraints, with only gold-standard construction code accessible. External validation is therefore necessary (Lansiaux et al., 1 Jul 2025).

Model-level risks in the triage setting include overfitting and potential LLM hallucinations. The authors recommend regularization, data augmentation, simplification, and TimeLLM-inspired reprogramming as possible remedies. Ethical and regulatory management included CESREES approval and adherence to CNIL MR-004. The deployment concept is decision support at intake rather than autonomous triage replacement, with SHAP-derived explanations used to maintain transparency (Lansiaux et al., 1 Jul 2025).

The disaster-response usage also has explicit constraints. Some jurisdictions do not use social media; government websites can become unstable during crises; PDFs and scanned notices add extraction difficulty; and 3–4% of actual evacuation notices may be missed under the reported models. Multilingual and code-switched content were not covered in the evacuation-notice paper, although integrated system guidance recommends bilingual support for English and Spanish in U.S. coastal regions (Zhao et al., 2024).

For location extraction, SAVITR’s English-only evaluation, ambiguity from homonyms, hashtag segmentation errors, GeoNames’ limited granularity, and OpenStreetMap’s latency and false-positive profile remain important trade-offs (Dutt et al., 2018). For urgency detection, domain shift across crises, sparse early labels, and multilingual degradation are central issues, motivating active learning, target up-sampling, and thresholding that favors recall when false negatives are considered more harmful than false alarms (Kejriwal et al., 2019). For stream parsing, USTEP does not implement pruning or merging, and its future directions include distribution across machines and alternative split criteria (Vervaet et al., 2023).

Across these literatures, the principal future directions are multicenter validation, real-time deployment studies, stronger regularization against overfitting, richer interpretability, rigorous bias audits, automated extraction of who-when-where attributes, and resilient operational deployment on scalable infrastructure. In the clinical setting, the central next step is prospective workflow integration. In the disaster-informatics setting, the central next step is robust end-to-end linkage between ingestion, classification, geospatial indexing, and public or institutional redistribution.