📄 medicine

Early risk stratification for in-hospital mortality after acute diquat poisoning: a retrospective cohort study with external validation

This retrospective cohort study developed and externally validated a four-variable model (HCO₃⁻, lactate, ingested volume, and white blood cell count) that demonstrates high discrimination for predicting in-hospital mortality after acute diquat poisoning, though calibration drift across settings necessitates local recalibration before clinical implementation.

Original authors: Meiwen Xie, Yifan Ye, Yuqiang Lin, Ziyan Wang, Zhiqian Yang, Zhi Wang, Bao Wang, Yuquan Chen

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Meiwen Xie, Yifan Ye, Yuqiang Lin, Ziyan Wang, Zhiqian Yang, Zhi Wang, Bao Wang, Yuquan Chen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a person swallows a toxic chemical, the body's clock starts ticking with terrifying speed. The substance does not just sit in the stomach; it races through the blood, attacking organs and disrupting the delicate chemical balance that keeps cells alive. In the case of a chemical called diquat, this process can lead to total organ failure within hours, and doctors currently have no specific antidote to stop it. Because there is no magic pill to reverse the damage, the most critical tool a medical team has is time. They must decide immediately which patients need the most intense care, which ones might need to be moved to a specialized center, and which ones are likely to survive. The challenge is that early symptoms can be misleading; a patient might look relatively stable at first while their internal systems are already collapsing. To make these life-or-death decisions, doctors need a way to see the future of the illness based on the limited information available the moment a patient walks through the emergency room door.

A team of researchers set out to build a simple, reliable guide for this exact moment. They looked back at the records of 229 patients who had been admitted to a hospital in Guangzhou after swallowing diquat. Their goal was to find a small group of facts that could predict who would die in the hospital and who would go home. They did not want a complicated system that required rare, expensive tests or hours of waiting for results. Instead, they hunted for data that is already collected for almost every patient: how much poison they likely swallowed, and a few standard blood tests that measure how their body is handling stress, acid, and inflammation. By analyzing these records, the researchers identified four specific pieces of information that, when combined, created a clear picture of the risk. These were the estimated amount of liquid ingested, the level of a chemical called bicarbonate which indicates how acidic the blood is, the level of lactate which signals how well the body is getting oxygen, and the white blood cell count which shows how much stress the immune system is under.

Using these four factors, the team created a mathematical model that could sort patients into high-risk and low-risk groups with remarkable accuracy. When they tested this model on their own group of patients, it correctly distinguished between those who survived and those who did not in nearly every case. The model was so precise that it could separate the two groups almost perfectly, much like a sieve that lets only the smallest grains fall through while holding back everything else. The researchers then took this same model, without changing a single number or rule, and tested it on a completely different group of 201 patients from other hospitals. The results were striking. The model still worked very well at ranking patients by risk; it knew who was in danger and who was safe. However, while it was excellent at sorting patients, it was not perfect at guessing the exact percentage chance of death for any single individual in the new group. In this second group, the model tended to underestimate the danger, predicting that fewer people would die than actually did.

This difference is a crucial detail for doctors to understand. The model acts like a highly skilled navigator who can tell you which path leads to a cliff and which leads to safety, but who might slightly misjudge the distance to the edge. The researchers found that the model's ability to rank patients remained strong, but the specific numbers it gave for the chance of death needed adjustment when used in a different hospital setting. This happens because different hospitals treat different types of patients, or because the severity of the poisonings varies from place to place. The study also looked at the journey of the patients who survived, using a special method to account for the fact that some patients died before they could be discharged. They found that the same factors that predicted death also predicted how long it took for a patient to recover enough to leave the hospital. Higher levels of acid in the blood and larger amounts of swallowed poison were linked to a slower path to recovery.

The final conclusion is one of cautious optimism. The researchers have developed a tool that is simple, fast, and uses information that is already available in almost any emergency room. It can help medical teams quickly identify the patients who are in the most danger and need immediate, aggressive support. However, the study makes it clear that this tool is not yet ready to be used as a final verdict in every hospital without a local check. Before a doctor uses the specific numbers from this model to make a decision, the hospital must first test the model on its own patients to ensure the predictions match their reality. The model is a powerful map for the early hours of a poisoning crisis, but like any map, it must be calibrated to the specific terrain it is being used on. Until that local calibration is done and the tool is tested in future real-world situations, it serves best as a guide to flag the most critical cases rather than a final calculator of fate.

Technical Summary: Early Risk Stratification for In-Hospital Mortality After Acute Diquat Poisoning

Problem Statement
Acute diquat poisoning is a rapidly deteriorating condition with no specific antidote, often leading to multiorgan failure and death within hours. Effective early risk stratification is critical for determining monitoring intensity, transfer decisions, and organ-support planning. However, existing prognostic models vary widely in their prediction timing, outcome definitions, and reliance on specialized measurements (e.g., plasma diquat concentrations) that are often unavailable in routine emergency settings. Furthermore, many previous models lack rigorous external validation, particularly regarding calibration—the accuracy of absolute risk estimates in new populations. There is a need for a parsimonious model based on routinely available admission data that has been validated for both discrimination and calibration in independent cohorts.

Methodology
This study employed a retrospective cohort design with external validation, adhering to the TRIPOD and STROBE guidelines.

Cohorts:
- Development Cohort: 229 patients admitted to Guangzhou Twelfth People's Hospital (2018–2026) within 24 hours of ingestion. Patients with co-poisoning, severe underlying chronic diseases, or uncertain diagnoses were excluded.
- External Validation Cohort: An independent, publicly available multicenter dataset of 201 patients with the same inclusion criteria and outcome (in-hospital death).
Data Handling: Missing predictor data were managed using Multiple Imputation by Chained Equations (MICE) with 20 imputed datasets.
Variable Selection: Thirteen candidate predictors (demographics, clinical symptoms, and early laboratory values) were assessed for redundancy and collinearity. Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression was applied to each imputed dataset using 10-fold cross-validation and the one-standard-error rule ( $\lambda_{1SE}$ ) for variable selection.
Model Development: Selected variables were entered into a multivariable logistic regression model. Coefficients were pooled using Rubin's rules.
Validation Strategy:
- Internal Validation: Conditional bootstrap resampling (1,000 resamples) with the final predictor set fixed to estimate optimism-corrected performance.
- External Validation: The model was applied to the independent cohort with all intercepts and coefficients fixed (no re-estimation). Performance was assessed via discrimination (AUC), calibration (slope, intercept, O:E ratio), overall error (Brier score), and decision-curve analysis (DCA).
Secondary Analysis: A Fine-Gray competing-risks model was performed to evaluate the cumulative incidence of discharge alive, treating in-hospital death as a competing event.

Key Contributions

Parsimonious Model: The study developed a four-variable model using only routinely available admission data: Bicarbonate (HCO₃⁻), Lactate, Estimated Ingested Volume, and White Blood Cell (WBC) Count.
Rigorous Validation Protocol: The study emphasized the distinction between discrimination and calibration. It validated the model in an independent multicenter cohort without retraining, providing a realistic assessment of transportability.
Competing-Risks Perspective: Beyond mortality prediction, the study utilized a Fine-Gray model to characterize the association of early factors with the competing outcome of discharge alive, offering a more nuanced view of hospital course.

Results

Development Cohort: 63 patients (27.5%) died. The final model included the four variables listed above.
- Internal Performance: The optimism-corrected Area Under the Curve (AUC) was 0.965. The Brier score was 0.068, and the calibration slope was 0.903, indicating high discrimination but mild overfitting.
External Validation: 100 patients (49.8%) died.
- Discrimination: The model maintained high discrimination with an AUC of 0.931 (95% CI: 0.896–0.967).
- Calibration: The model exhibited calibration drift. The Observed-to-Expected (O:E) ratio was 1.124, indicating an underestimation of mortality risk. The calibration slope was 0.639, suggesting the model produced overly extreme predictions (probabilities were too close to 0 or 1) in the external population.
- Re-calibration: Exploratory updates to the intercept and slope improved the Brier score, but the authors note these require validation in a new cohort.
Competing-Risks Analysis: Greater ingested volume, higher lactate, higher creatinine, and lower bicarbonate were associated with a lower subdistribution hazard of discharge alive (sHR < 1).

Significance and Claims
The paper claims that the primary contribution of this work is a parsimonious, early-risk stratification tool that retains high discrimination across different clinical settings. The model successfully identifies patients at relatively high versus low risk of in-hospital mortality using only standard admission data.

However, the authors are explicit about the model's limitations regarding clinical implementation:

Calibration Drift: While the model ranks risk well, the absolute probability estimates are not directly transportable across centers without local recalibration due to differences in case mix, referral patterns, and treatment strategies.
Implementation Requirements: The authors state that the model should not be used for direct clinical decision-making across centers without prior local validation and recalibration.
Future Steps: Prospective multicenter validation and clinical impact evaluation are required before routine implementation.

In summary, the study provides a robust, data-driven framework for early risk stratification in diquat poisoning but cautions that its utility as a treatment decision rule is currently limited by calibration issues that must be addressed locally.

Technical Summary: Early Risk Stratification for In-Hospital Mortality After Acute Diquat Poisoning

More like this