Prediction of severe caffeine intoxication: a bedside clinical model with implications for extracorporeal therapy.

INTRODUCTION: Severe caffeine intoxication can cause life-threatening complications, yet serum caffeine measurements are rarely available at presentation. We aimed to develop a simple clinical prediction model using readily available clinical parameters to predict severe caffeine intoxication. METHODS: This retrospective study involved patients with acute caffeine intoxication admitted between April 2016 and March 2022. Data on clinical variables at presentation were collected. Severe intoxication was defined as serum caffeine concentration ≥ 80 mg/L (412 μmol/L). Candidate predictors included heart rate and serum potassium and bicarbonate levels. A ridge logistic regression model was developed and evaluated using the area under the receiver operating characteristic curve, calibration plots, and Hosmer-Lemeshow test. Internal validation was performed using bootstrap resampling and leave-one-out cross-validation. Conventional logistic regression was performed as a sensitivity analysis. Clinical utility was assessed using decision curve analysis. RESULTS: Of 30 patients included, 13 (43%) had serum caffeine concentration ≥ 80 mg/L (412 μmol/L). Patients with severe intoxication had higher ingested doses, shorter time to presentation, higher heart rates and respiratory rates, lower bicarbonate and potassium levels, and more frequent use of hemodialysis and activated charcoal. The ridge model retained heart rate and bicarbonate and potassium levels as predictors. Internal validation demonstrated excellent discrimination and good calibration. Decision curve analysis indicated net clinical benefit across a range of threshold probabilities. Sensitivity analysis using conventional logistic regression revealed consistent results, with heart rate remaining a significant predictor. DISCUSSION: The selected predictors are biologically plausible and reflect key pathophysiological features of severe caffeine intoxication. Internal validation demonstrated excellent discrimination and calibration, supporting the robustness of the model. CONCLUSION: We developed a practical bedside prediction model for caffeine intoxication using heart rate and bicarbonate and potassium levels. The model demonstrated excellent discrimination and calibration. It may support early risk stratification and guide intensive monitoring or extracorporeal therapy.