JMIR Publications

ABSTRACT

Background:

Acute poisoning accounts for an estimated 1-2% of emergency department (ED) visits worldwide, yet most low- and middle-income country hospitals still record triage in unstructured, free-text PDF forms with no toxicological fields. Automated case-finding from this text is technically feasible, but its real-world performance and the clinical phenotypes it reveals are not well characterized.

Objective:

To develop and evaluate an open-source natural language processing (NLP) pipeline for identifying acute poisoning cases from unstructured ED triage records, to quantify the pipeline's sensitivity and its principal failure mode, and to derive interpretable clinical phenotypes and a severity-screening rule from the identified cohort.

Methods:

We conducted a retrospective cohort study of all ED triage records (n=36,581) from a tertiary public hospital in Mexico City over 14 months (January 2025-February 2026). The CINDER pipeline (Python 3.11; PyMuPDF, pandas, regex; no machine learning and no proprietary software at the extraction stage) performed deduplication, text extraction, structured parsing, broad and strict keyword filtering, and dictionary-based semantic enrichment. A physician (blinded to pipeline output) reviewed 50 broad-filter records to estimate precision, and a second, unblinded audit of 100 broad-filter-negative records estimated the rate of undocumented ("clinically silent") poisonings. Toxic agents that did not map to the pipeline's eight predefined categories were adjudicated by a clinical toxicologist. The 195 confirmed cases were characterized descriptively, then 188 cases with complete National Early Warning Score 2 (NEWS-2) data were partitioned with k-means clustering (k=3) on standardized age, NEWS-2, heart rate, oxygen saturation, and polyexposure category count; a Random Forest model and recursive partitioning identified severity predictors and screening rules.

Results:

The pipeline reduced 36,581 records to 195 confirmed cases (yield 0.53%) with 100% precision among strict-filter cases (95% CI 98.1-100%). The audit of broad-filter-negative records found probable poisoning in 8 of 100 records without toxicological keywords, implying an estimated pipeline sensitivity near 9-10% and roughly 1,800 additional cases missed across the full non-filtered pool; this documentation gap, not the filtering logic, is the dominant limitation of keyword-based case-finding in this setting. Among confirmed cases, one-third (65/195, 33.3%) involved a toxic agent outside the eight predefined dictionary categories, confirmed by toxicologist review; this is markedly higher than dictionary coverage assumptions typically applied in keyword-based surveillance. Clustering identified three phenotypes: Young/Polypharmacy (n=66, 34%; NEWS-2 2.7±1.5), Adult/Moderate (n=97, 50%; NEWS-2 3.3±2.0), and Older/Severe (n=32, 16%; NEWS-2 6.0±2.8, lower SpO2 and higher respiratory rate, both p<0.001). Respiratory rate, not SpO2, was the strongest Random Forest predictor of NEWS-2≥5 (AUROC 0.868±0.048). A two-variable OR screening rule (respiratory rate>22.5 OR systolic blood pressure≤91.5 mmHg) outperformed clinical triage priority (AUROC 0.750 vs 0.620, p=0.020), but missed half of severe cases (30/60), a blind spot shared with clinical triage.

Conclusions:

An open-source NLP pipeline can identify acute poisoning cases from unstructured triage text with high precision, but its sensitivity is low and should not be presented as a validated case-finding tool without this caveat. The pipeline's principal contribution is diagnostic: it quantifies how much acute poisoning is lost to non-standardized documentation and to agents outside fixed keyword dictionaries. Within the cases it does identify, phenotyping and a simple screening rule offer a starting point for severity stratification, pending prospective, outcome-based validation.