Predictive Factors of Severe Outcome in Glufosinate Ammonium Poisoning: A Retrospective Study in Vietnam.

BackgroundGlufosinate ammonium (GLA) is a widely used herbicide in rural Vietnam, often associated with severe toxicity. Given the limited diagnostic resources at primary care levels, identifying simple and effective predictors of severity is critical for early triage and intervention.ObjectiveTo identify clinical and biochemical predictors of severe outcomes in patients with glufosinate ammonium poisoning using multivariable logistic regression.MethodsWe conducted a retrospective study of 83 patients with confirmed glufosinate ammonium poisoning admitted to our poison control center at Bach Mai hospital between March 2023 and October 2024. Demographic, clinical, and laboratory data were collected. Severe outcomes were defined as the need for mechanical ventilation, inotropic support, Glasgow Coma Score ≤8, or in-hospital mortality. Logistic regression was used to identify significant predictors and evaluate model performance.ResultsThe mean age was 48.6 ± 17.3 years; 62.7% were male; 88% of cases resulted from intentional ingestion; six patients died (7.2% mortality). Significant predictors of severe outcomes included glufosinate ammonium ingestion >100 ml, (Odds Ratio (OR) 4.5, p < 0.001), time to hospital >6 h (OR 3.2, p = 0.004), pH < 7.35 (OR 5.7, p < 0.001), lactate >4.0 mmol/L (OR 6.2, p < 0.001), creatinine >110 µmol/L (OR 4.3, p = 0.005), and NH3(ammonia) > 100 µmol/L (OR 5.4, p = 0.002). The final logistic regression model including pH < 7.35, time to hospital >6 h, NH3 > 100 µmol/L, and ingested volume >100 mL demonstrated good discrimination (AUC = 0.84, 95% CI 0.74-0.93; p < 0.001) and acceptable calibration by the Hosmer-Lemeshow test (p = 0.47).ConclusionReadily available clinical and laboratory indicators, such as pH < 7.35, delayed hospital arrival, high ammonia levels, and large ingested volume, were strong predictors of severe outcomes in glufosinate ammonium poisoning. The simple logistic model (AUC = 0.84) may support early triage in low-resource settings, but prospective validation is needed.