2026 Volume 51 Issue 9 Pages 517-525
Reliable early prognostic tools for acute poisoning remain limited. This study examined whether four first-measured routine biomarkers — glucose, lactate, bicarbonate, and white blood cell count (WBC) — carry prognostic signal for 30-day mortality in emergency department (ED) patients with acute poisoning, using a pre-processed, de-identified electronic health records dataset derived from MIMIC-IV (Medical Information Mart for Intensive Care, version 3.1; a publicly available critical care database from Beth Israel Deaconess Medical Center, Boston, USA, hosted on PhysioNet; n=12,039 ED visits; 59 deaths; 30-day mortality 0.49%). Due to extreme non-random missingness across all four biomarkers (lactate 95.2%; bicarbonate/glucose/WBC 88–89%), multivariable modelling was restricted to a complete-case subgroup (n=503; deaths=19). This subgroup was markedly non-representative: intensive care unit (ICU) admission rate 66.4% versus 2.1% in excluded visits (standardised mean difference [SMD]=1.844), reflecting selective laboratory ordering in clinically severe patients. In univariable analyses across available cases, all four biomarkers were significantly associated with 30-day mortality. Within the complete-case subgroup, glucose alone demonstrated good discrimination (area under the receiver operating characteristic curve [AUROC]=0.812; 95% confidence interval [CI] 0.721–0.885), and glucose and WBC were independently associated with mortality in multivariable analysis (adjusted odds ratios [aOR] 1.05 and 1.13, respectively). The calibration slope of 0.646 for the primary four-biomarker model indicated systematic overestimation of absolute risk. These findings are strictly exploratory. Critical limitations — extreme non-random missingness, ICU-enriched complete-case subgroup, absent age and vital signs, low event count (n=19), and absence of external validation — preclude clinical application. This work defines data and design requirements for a future adequately powered prospective prognostic study in acute poisoning.
Acute poisonings represent a persistent and substantial burden on emergency departments (EDs) worldwide, accounting for significant critical illness, mortality, and intensive care utilisation (Gummin et al., 2020; Gummin et al., 2023). Early identification of patients at highest risk of death is essential for optimal triage, monitoring intensity, toxicology consultation, antidote administration, and intensive care unit (ICU) allocation — all decisions that must frequently be made within the first hours of presentation, often before definitive toxicological information is available.
Existing severity classification tools for poisoning — principally the Poisoning Severity Score (PSS) — were developed primarily for case description and standardisation rather than outcome prediction, and have shown limited generalisation across heterogeneous toxicological presentations (Persson et al., 1998; Schwarz et al., 2017; Sam et al., 2009). More recent approaches, including the new Poisoning Mortality Score (new-PMS), nomogram-based models, and modified early warning score adaptations, have demonstrated promising discrimination, but most are restricted to specific agent classes, selected admission cohorts, or settings with comprehensive data (Han et al., 2021; Lee et al., 2024; Saeed and Elmorsy, 2023; Lionte et al., 2017; Martín-Rodríguez et al., 2021; Lee et al., 2022).
Objective biomarkers obtainable at first blood draw — serum lactate, bicarbonate, glucose, and white blood cell count (WBC) — may offer physiological surrogates of systemic toxicity severity that are available before agent identification. Prior work has linked elevated lactate to adverse outcomes in drug overdose and beta-blocker poisoning (Manini et al., 2010; Cheung et al., 2018; Mégarbane et al., 2010; Abdelhamid et al., 2024), and metabolic acidosis (reflected by bicarbonate), stress hyperglycaemia, and leukocytosis have been proposed as markers of systemic inflammatory and physiological stress response in poisoning (Martín-Rodríguez et al., 2021; Abdelhamid et al., 2024; Odakha et al., 2022; Sontakke and Kalantri, 2023; Fernando et al., 2020). Whether these signals survive mutual adjustment in a real-world, ICD-defined, heterogeneous ED cohort remains uncertain.
We present an exploratory analysis examining whether early routine biomarkers carry prognostic signal for 30-day mortality in a derived MIMIC-IV dataset of acute poisoning ED visits. Due to extreme non-random missingness in all biomarkers, multivariable analyses were confined to a complete-case subgroup that is substantially non-representative of the full cohort. This report should be interpreted as hypothesis-generating and data-requirement-defining rather than as a validated prognostic model.
This is a retrospective cohort study with pre-specified exploratory multivariable modelling. Reporting follows TRIPOD and TRIPOD+AI guidelines (Collins et al., 2015; Collins et al., 2024); methodological bias was assessed with reference to PROBAST (Wolff et al., 2019). All modelling results are presented as exploratory and hypothesis-generating; this analysis does not constitute a validated clinical prediction model.
Data source and its limitations relative to full MIMIC-IVThe dataset is a pre-processed, de-identified tabular extract derived from MIMIC-IV (Medical Information Mart for Intensive Care, version 3.1; PhysioNet), a freely accessible electronic health records (EHR) database of Beth Israel Deaconess Medical Center, Boston, USA (Johnson et al., 2023). Each row represents a unique ED visit identified by stay_id. An important methodological caveat: this pre-processed extract is missing variables present in full MIMIC-IV source tables. Specifically: (a) age was entirely absent (100% missing), whereas the patients table in MIMIC-IV contains anchor_age; (b) triage vital signs were absent, whereas the MIMIC-IV-ED edstays and triage tables contain heart rate, blood pressure, oxygen saturation, and temperature at ED triage; and (c) out-of-hospital death records could not be verified from this extract, whereas the admissions table in MIMIC-IV includes date of death sourced from Massachusetts state mortality records. These are limitations of the derived dataset, not of MIMIC-IV itself.
Cohort definitionAcute poisoning visits were identified using ICD-10-CM codes T36–T65 (poisoning by drugs, medicaments, and biological substances) applied to primary or secondary diagnosis fields. This code range encompasses a heterogeneous spectrum of toxicological exposures including pharmaceutical drugs, alcohol, pesticides, corrosives, and industrial chemicals. The specific distribution of agent classes could not be characterised from the derived dataset (ICD codes were not preserved in the extract), which limits comparability with agent-specific published models. Each row was treated as an independent ED visit. The dataset contained 12,039 total visits from 6,687 unique subjects; 968 subjects (14.5%) had two or more visits, accounting for 6,320 visits (52.5% of the total).
Primary and secondary outcomesThe primary outcome was 30-day all-cause mortality (death_30d), defined as any recorded death within 30 days of ED presentation based on hospital administrative records. Out-of-hospital deaths and follow-up completeness could not be independently verified. Secondary outcomes were ICU admission and model discrimination.
ICU admission derivationIn the derived dataset, hospital admission identifier (hadm_id) was null for 5,772 visits (47.9%), representing structurally ED-only visits (patients discharged home, eloped, or who left without being seen or against medical advice). For these visits, ICU admission is structurally inapplicable and was recoded as 0, not treated as missing. For the 6,267 visits with hadm_id present, ICU admission was directly available. After recoding, 574 of 12,039 visits (4.8%) resulted in ICU admission.
Missing data and complete-case subgroupMissingness in laboratory variables was extreme and strongly consistent with data missing not at random (MNAR): lactate was measured in 63.1% of ICU visits versus 3.7% of non-ICU visits (p<0.001). Lactate was missing in 95.2% of all visits; glucose, creatinine, and bicarbonate in 88.3–88.4%; and WBC in 89.0%. Multiple imputation was not feasible at >95% missingness. All multivariable analyses were restricted to the 503 visits with all four biomarkers present (deaths=19). The included-versus-excluded comparison used standardised mean differences (SMD); SMD >0.1 was considered a meaningful imbalance.
Statistical analysisContinuous variables are summarised as median [interquartile range]; categorical variables as count (percentage). Non-normality was confirmed by Shapiro–Wilk test (Shapiro–Wilk p<0.001 for all biomarkers in the complete-case subgroup). Non-parametric tests were used: Mann–Whitney U test for continuous variables and chi-squared or Fisher exact test for categorical variables.
'First' biomarker values refer to the earliest laboratory measurement recorded within each individual ED stay (stay_id). When multiple measurements of a given biomarker were available within a single visit, only the earliest recorded value was used. This definition is applied consistently across all four biomarkers (lactate, glucose, bicarbonate, WBC). The repeat-visit sensitivity analysis (first-visit-only; n=343; deaths=17) addresses the separate question of between-visit clustering.
Univariable logistic regression was performed on available-case data. The primary multivariable model was a ridge-penalised logistic regression (L2 regularisation, C=0.1, class-weight balancing) including: first lactate (per 1 mmol/L), first bicarbonate decrease (per 1 mmol/L), first glucose (per 10 mg/dL), and first WBC (per 1 ×103/µL). Ridge penalisation was selected to reduce overfitting given the low event-per-variable ratio (~4.75). Point estimates (adjusted odds ratios [aORs]) from a standard unpenalised logistic regression model are reported for interpretability, with bootstrap 95% confidence intervals (CIs; 2,000 resamples). ICU admission was examined only as a secondary sensitivity variable. A glucose-only single-predictor model was evaluated as a parsimony benchmark.
Model performance was evaluated by 5-fold stratified cross-validation with bootstrap 95% CIs. Metrics reported: area under the receiver operating characteristic curve (AUROC), precision–recall AUC (PR-AUC), Brier score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at the Youden-optimal threshold. The Brier score measures the mean squared difference between predicted probabilities and observed binary outcomes (range 0–1; lower = better). It is prevalence-dependent: in the complete-case subgroup (event rate 3.78%), a null model predicting zero probability for all observations achieves a Brier score of 0.038. Brier scores above this null benchmark indicate worse-than-null probability estimation despite potentially good discrimination. Calibration was assessed by regressing observed outcomes on logit-transformed cross-validated predictions. A gradient boosting machine (GBM; scikit-learn GradientBoostingClassifier, n_estimators=100, max_depth=3, learning_rate=0.1) was evaluated as an exploratory secondary analysis (Naemi et al., 2021; Kareemi et al., 2021; Tschoellitsch et al., 2023). A first-visit-only sensitivity analysis (n=343, deaths=17) was performed to assess potential repeat-visit clustering effects.
As an additional sensitivity analysis addressing class imbalance, SMOTE (Synthetic Minority Oversampling Technique) was applied within each training fold of the 5-fold cross-validation (never to the validation fold, to avoid data leakage). Within each training fold, the minority class was oversampled to achieve a 1:1 class ratio. The SMOTE-augmented logistic regression model was compared with the original class-weighted ridge-penalised model on all performance metrics.
All analyses used Python 3.10 (pandas, SciPy, scikit-learn 1.3, imbalanced-learn 0.11). A two-sided p<0.05 was considered statistically significant for pre-specified tests; no correction for multiple comparisons was applied given the exploratory nature of the analysis.
A total of 12,039 acute poisoning ED visits were identified from 6,687 unique subjects; 59 visits (0.49%) resulted in 30-day mortality (Table 1). Age was entirely absent from the derived dataset. Most patients arrived by ambulance (73.6%) and were discharged home (74.9%). Non-surviving patients had higher lactate, creatinine, glucose, and WBC, and lower bicarbonate. ICU admission was substantially more frequent among non-survivors (33.9% vs 4.7%; p<0.001).
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t1.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
The 503 complete-case visits differed substantially from the 11,536 excluded visits (Table S1). ICU SMD was 1.844, confirming extreme ICU enrichment in the complete-case group (66.4% vs 2.1%). Mortality SMD was 0.243 (3.8% vs 0.35%). These differences confirm that all biomarker model results are applicable only to this highly selected, high-acuity subgroup.
Univariable analysisResults are presented in Table 2. Lactate (OR 1.37; 1.22–1.54; p<0.001), bicarbonate decrease (OR 1.25; 1.16–1.35; p<0.001), glucose per 10 mg/dL (OR 1.06; 1.03–1.09; p<0.001), and WBC (OR 1.20; 1.13–1.28; p<0.001) were each significantly associated with 30-day mortality. ICU admission showed the strongest univariable association (OR 9.75; 5.25–18.10).
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t2.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
In the primary model (n=503; events=19; EPV~4.75), glucose (aOR 1.05; 95% CI 1.00–1.10; p=0.035) and WBC (aOR 1.13; 95% CI 1.04–1.23; p=0.004) were independently associated with mortality (Table 3A). Lactate showed a borderline association (aOR 1.17; 95% CI 0.99–1.38; p=0.066). The secondary sensitivity model including ICU admission yielded an extreme aOR for ICU (89.92; 95% CI 28.39–>550,000), confirming that five-predictor modelling with 19 events is statistically intractable (Table 3B).
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t3A.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t3B.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
All results are presented in Table 4 and Fig. 1; SMOTE sensitivity analysis results are presented in Table 5. Glucose alone (AUROC 0.812; 95% CI 0.721–0.885) performed comparably to the four-biomarker ridge-penalised model (AUROC 0.795; 95% CI 0.645–0.916; PR-AUC 0.342; Brier 0.134), indicating that additional predictor complexity does not improve discrimination in this data-sparse setting. The ridge model Brier score of 0.134 and the glucose-alone Brier score of 0.187 both exceed the null Brier score of 0.038, confirming that neither model improves upon a null predictor in terms of absolute probability estimation. The calibration slope of 0.646 for the primary four-biomarker model reveals significant systematic overestimation of predicted probabilities. At the Youden-optimal threshold, the ridge model achieved PPV 0.131 and NPV 0.987. The GBM achieved AUROC 0.790 (95% CI 0.645–0.907) with a calibration slope of 0.398, indicating poor probability estimation. The first-visit-only sensitivity analysis (AUROC 0.810; n=343; deaths=17) was consistent with the primary result.
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t4.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
ROC curves for 30-day mortality prediction in the complete-case subgroup (n=503; events=19). Curves represent out-of-fold predictions from 5-fold stratified cross-validation. Three models are shown: ridge-penalised logistic regression (primary model; AUROC=0.801), glucose-alone single-predictor model (AUROC=0.812), and standard logistic regression (AUROC=0.794). Dashed diagonal = chance performance (AUROC=0.50). AUROC values with bootstrap 95% confidence intervals are reported in Table 4. Metrics were recomputed in a separate cross-validation run; the minor differences from Table 4 reflect fold-assignment variability.
;%0A%09%09%09newWindow.document.open();%0A%09%09%09newWindow.document.write('<img src=%22./Graphics/51_517-t5.jpg%22>');%0A%09%09%09newWindow.document.close();%0A%09%09)
This exploratory analysis examined prognostic signals of early routine biomarkers for 30-day mortality in a derived MIMIC-IV dataset of acute poisoning ED visits. Three principal findings emerge. First, glucose, WBC, lactate, and bicarbonate each showed univariable associations with mortality across available-case analyses. Second, within the complete-case subgroup, glucose alone demonstrated discrimination comparable to the four-biomarker multivariable model (AUROC 0.812 vs 0.795), though the multivariable model had better precision-recall performance. Third, and most critically, the complete-case subgroup is so profoundly non-representative — ICU admission 66.4% vs 2.1% in excluded visits (SMD 1.844) — that no findings can be generalised to the broader acute poisoning ED population.
The finding that glucose alone approaches multivariable model discrimination warrants discussion. In an ICU-enriched subgroup, elevated glucose may reflect several overlapping mechanisms: catecholamine-mediated physiological stress response (Dungan et al., 2009; Marik and Bellomo, 2013), pre-existing diabetes mellitus (which could not be ascertained from the derived dataset), glucose-containing resuscitation fluids, dextrose administration as antidote therapy, or direct toxic metabolic effects of specific agents. The relative contribution of these mechanisms cannot be distinguished from the available data. Regardless of mechanism, glucose appears to capture a severity signal in this ICU-enriched subgroup that is not substantially augmented by adding lactate, bicarbonate, and WBC — likely because all four biomarkers co-vary with illness severity in this highly selected population. The multivariable model did outperform glucose alone on PR-AUC (0.342 vs 0.147) and Brier score (0.134 vs 0.187), suggesting that additional biomarkers contribute to risk probability estimation even where they do not substantially improve rank discrimination. This mechanistic uncertainty further underscores the exploratory nature of these findings and the need for prospective data collection including comorbidity status, medication history, and agent identification.
The calibration slope of 0.646 for the primary four-biomarker ridge-penalised model is a clinically important finding. A slope below 1.0 indicates systematic overestimation of absolute risk — the model assigns higher predicted probabilities than are observed. This is expected in sparse data with only 19 events and EPV~4.75. This finding further precludes direct clinical application of model-derived probabilities.
The profound non-representativeness of the complete-case subgroup is a direct consequence of clinical ordering behaviour in the emergency department. In standard practice, laboratory investigations including lactate, bicarbonate, glucose, and WBC are not obtained universally; they are ordered selectively based on the clinician's bedside assessment of illness severity at triage and during initial evaluation. Patients who appear clinically well are typically discharged without a full laboratory panel, while those with haemodynamic instability, respiratory compromise, altered mental status, or clinical signs of severe toxidrome are more likely to receive comprehensive testing. This is confirmed by the observation that lactate was measured in 63.1% of ICU visits versus only 3.7% of non-ICU visits (p<0.001) — a pattern strongly consistent with data missing not at random (MNAR). The 503 complete-case visits therefore represent those patients who were sufficiently ill to trigger full laboratory assessment; they are not a random sample of the full 12,039-visit cohort. Any biomarker associations observed within this subgroup are conditional on a degree of illness severity substantially greater than the broader acute poisoning ED population, and generalisation beyond this high-acuity subgroup is not warranted.
The repeat-visit rate (52.5% of visits from subjects with ≥2 presentations) violates the independence assumption of standard logistic regression. However, the first-visit-only sensitivity analysis (AUROC 0.810; n=343; deaths=17) yielded consistent results, suggesting robustness to this clustering effect. Future analyses should apply clustered standard errors or mixed-effects models.
The heterogeneity of the ICD-10-CM T36–T65 cohort is a fundamental limitation. Pharmaceutical overdose, alcohol poisoning, organophosphate ingestion, and caustic injury have different pathophysiological mechanisms and mortality risk profiles. The absence of agent-specific data in the extract precludes stratified analyses and limits interpretability of biomarker associations.
The absence of age, vital signs, and Glasgow Coma Scale (GCS) represents the single largest structural gap relative to existing validated models such as new-PMS (Han et al., 2021). Age is among the strongest non-toxicological predictors of mortality across all ED presentations. These variables are available in full MIMIC-IV source tables; the derived dataset limitations are not intrinsic to MIMIC-IV itself. The principal recommendation arising from this work is that future analyses extract directly from MIMIC-IV source tables.
Strengths of this study include: real-world EHR data; a priori model specification; transparent characterisation of missingness with SMD-based comparison; correction of ICU derivation logic; comprehensive performance reporting across all models; explicit calibration assessment; and first-visit sensitivity analysis addressing repeat-visit clustering.
In conclusion, in a highly selected, ICU-enriched complete-case subgroup of 503 acute poisoning ED visits, first glucose and WBC were independently associated with 30-day mortality. Glucose alone demonstrated discrimination (AUROC 0.812) comparable to the four-biomarker multivariable model (AUROC 0.795), though the multivariable model showed superior precision-recall performance. Critical limitations — calibration slope of 0.646 for the primary four-biomarker ridge-penalised model indicating systematic overestimation of absolute risk, extreme non-random missingness (ICU SMD=1.844), absent age and vital signs, low event count (n=19), repeat-visit clustering, and absence of external validation — collectively preclude any clinical application. This work primarily defines data and design requirements for a future adequately powered, prospective prognostic study. Such a study should prospectively collect age, vital signs at triage, GCS, agent identification, and complete biomarker panels; power for ≥100 events; report calibration and decision curve analysis; and undergo external validation.
The authors thank the developers of MIMIC-IV and PhysioNet for making the dataset publicly available.
FundingNo specific funding was received for this study.
Conflict of interestThe authors declare no competing interests.
Data availabilityThe data used in this study were obtained from the MIMIC-IV database (version 3.1), a publicly available critical care database hosted on PhysioNet. The authors have completed the required training and obtained access to the database. The derived dataset and analysis code are available from the corresponding author upon reasonable request.
Author contributionsConceptualization: Ramazan Sami Aktaş
Data curation: Ramazan Sami Aktaş
Formal analysis: Ramazan Sami Aktaş, Mehmet Şirin Büyükkaya
Methodology: Mehmet Şirin Büyükkaya
Writing – original draft: Ramazan Sami Aktaş
Writing – review & editing: Ramazan Sami Aktaş, Mehmet Şirin Büyükkaya, Fatma Okucu, İdris Tüzün, Ömer Okucu
Supervision: Fatma Okucu
Ethical approval and consent to participateThis study used MIMIC-IV (version 3.1), a publicly available critical care database hosted on PhysioNet. The authors have completed the required training and obtained access to the database. The original data collection at Beth Israel Deaconess Medical Center received IRB approval with waiver of individual consent for research use. Secondary analysis of de-identified MIMIC-IV data under a valid Data Use Agreement does not constitute human subjects research requiring separate IRB approval.
Patient consent for publicationNot applicable.
