Development and validation of an early prediction model for bee-sting-induced acute kidney injury using machine learning Authors: Yiqin Fang, Juan Zhang, Zhonghua Hu, Xiaoling Yu, Yiyan Xie, Yueming Liu, Huizhen Wu Published: 2026 ABSTRACT AbstractAcute kidney injury (AKI) is a severe and frequent complication following bee stings, with a reported incidence of 30–50%. Early identification is clinically challenging due to the delayed rise in serum creatinine. This retrospective study aimed to develop and validate an interpretable machine learning model for the early prediction of AKI in bee-sting patients. We included 305 patients admitted between January 2018 and September 2025, of whom 99 (32.5%) developed AKI according to Kidney Disease: Improving Global Outcomes (KDIGO) criteria. Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed for feature selection from 33 baseline variables, identifying eight key predictors, including lactate, creatinine, and dark urine. Six machine learning algorithms – logistic regression, random forest, support vector machine, neural network, extreme gradient boosting (XGBoost), and light gradient boosting machine – were developed and evaluated. The XGBoost model demonstrated superior performance, achieving an area under the receiver operating characteristic curve of 0.964 and an accuracy of 92.6% on the test set. SHapley Additive exPlanations analysis was applied to interpret the optimal model, highlighting creatinine, lactate dehydrogenase, lactate, and dark urine as the most influential predictors. A clinically applicable nomogram was constructed and showed good calibration and utility. In conclusion, this study successfully developed a high-performance and interpretable XGBoost model for early prediction of bee-sting-induced AKI. The validated nomogram provides an accessible bedside decision-support tool to facilitate timely risk stratification and intervention, potentially improving patient outcomes. Introduction Bee stings represent a significant public health issue globally, particularly in tropical and subtropical regions [1,2]. Bee venom comprises a complex mixture of bioactive substances capable of inducing pathophysiological changes ranging from localized allergic reactions to multi-organ dysfunction [3]. Among the affected organs, the kidneys are especially susceptible to venom-induced damage. Acute kidney injury (AKI) is one of the most severe complications associated with bee envenomation. Published reports indicate an AKI incidence rate of 30–50% among patients stung by bees [4], with AKI linked to prolonged hospitalization, higher treatment costs, and increased mortality [5]. Clinically, AKI caused by bee stings is primarily diagnosed using criteria established by the Kidney Disease: Improving Global Outcomes (KDIGO) organization, which considers elevated serum creatinine (Scr) levels and reduced urine output as core indicators [6]. However, Scr, a breakdown product of creatine phosphate in muscle, accumulates over time before rising in the blood. Relying on delayed Scr increases the risk that the window for intervention has already closed, thereby posing a substantial threat to patient outcomes [7]. Consequently, early identification and timely preventive management of at-risk patients is the optimal strategy. Developing early-warning methods to detect AKI at its initial or preclinical stages is thus crucial. Furthermore, traditional statistical methods, such as logistic regression (LR), are insufficient for capturing the complex, nonlinear interactions among various risk factors [8]. In recent years, machine-learning methods have shown exceptional predictive capabilities, enabling the automatic identification of complex disease patterns within large clinical datasets. These techniques support early warning systems and prognostic predictions that often surpass traditional methods [9,10]. For example, Hu et al. developed models to predict early mortality among ICU patients with AKI [11]. Despite their strong performance, the ‘black-box’ nature of machine-learning algorithms limits their clinical adoption. To address this, explainable artificial intelligence techniques, such as SHapley Additive exPlanations (SHAP), have been introduced to elucidate the decision-making process of these models, thereby enhancing physicians’ trust in their predictions [12]. Research on predictive models for bee-sting-induced AKI remains scarce. Wu et al. constructed a nomogram using LASSO regression to identify high-risk individuals [13]. However, they did not systematically compare predictive performance across different machine-learning algorithms, clarify the internal workings of their model, or assess its clinical benefits. This study aims to develop and validate a clinical early-warning system for AKI following bee stings through a multi-level framework that integrates clinical-course analysis with predictive modeling. Initially, we systematically characterize the onset, progression, and prognostic features of AKI following bee stings. Next, multiple machine-learning algorithms are developed and compared to identify the optimal prediction model. Traditional statistical methods, combined with interpretable artificial intelligence techniques, are then used to identify key predictive factors and quantify their contributions. Finally, a simplified nomogram based on these core factors is developed, and its clinical net benefit is evaluated using decision-curve analysis (DCA). This approach provides clinicians with a bedside decision-support tool characterized by high accuracy, interpretability, and practicality. Materials and methods Participants and selection criteria This retrospective study was conducted at the First People’s Hospital of Chun’an County. Patients admitted to the emergency department (ED) or intensive care unit (ICU) for bee-sting injuries between January 2018 and September 2025 were screened. Eligibility criteria included: (1) age ≥18 years; (2) documented history of a bee sting; (3) admission to the ED or ICU within 72 h after the sting; and (4) availability of complete clinical data. Exclusion criteria were: (1) baseline chronic kidney disease (CKD) (eGFR <60 mL/min/1.73 m2) [14]; (2) renal replacement therapy (RRT) prior to admission; (3) other known causes of AKI (e.g., septic shock, nephrotoxic drug exposure); or (4) missing critical variables. The study protocol was approved by the Ethics Committee of the First People’s Hospital of Chun’an County (Approval No. 2025-04-12-21). Due to the retrospective design, informed consent was waived. This study was conducted in accordance with the Declaration of Helsinki. Data collection and variable definitions Clinical data were obtained from the hospital’s electronic medical record system. A total of 33 baseline variables were collected, including demographic characteristics (age, biologic sex), medical history (hypertension, diabetes mellitus, CKD, allergy history), bee sting-related details (number of stings, time from sting to hospital admission), vital signs upon admission (systolic and diastolic blood pressure, heart rate, body temperature), and clinical symptoms (dark urine, nausea and vomiting, altered consciousness). Admission laboratory tests included routine blood parameters (white blood cell (WBC) count, neutrophil percentage, hemoglobin, platelet count), liver function markers (alanine aminotransferase, aspartate aminotransferase, total bilirubin), renal function indicators (Scr, blood urea nitrogen (BUN)), electrolytes (potassium, sodium), myocardial enzymes (creatine kinase, lactate dehydrogenase (LDH)), coagulation tests (prothrombin time, activated partial thromboplastin time, D-dimer), and additional biomarkers (C-reactive protein, lactate). All tests were conducted within 24 h of ED or hospital admission using standardized methods and instruments in the hospital’s clinical laboratory. Subsequent treatments (e.g., glucocorticoid administration, blood purification therapy) and clinical outcomes (e.g., ICU admission status, RRT requirement, and hospital outcomes) were also recorded and analyzed. Diagnosis and staging were based on the 2012 KDIGO Clinical Practice Guidelines. AKI was defined by meeting at least one of the following criteria: (1) increase in Scr ≥26.5 μmol/L within 48 h; (2) ≥1.5-fold increase in Scr from baseline within seven days; or (3) urine output <0.5 mL/(kg·h) for ≥6 h. AKI severity (stages 1–3) was classified according to KDIGO criteria. Baseline Scr was defined as the most recent value obtained within 3 months before admission. If unavailable, the lowest Scr measured 48–72 h after admission was used as a substitute. Chronic kidney disease was diagnosed within a 3-month follow-up after discharge, following KDIGO guidelines for CKD evaluation and management. Criteria included persistent estimated glomerular filtration rate (eGFR) <60 mL/(min·1.73 m2) for ≥3 months or the presence of other renal injury markers (e.g., proteinuria). Renal and clinical outcomes were defined as follows: for AKI patients, complete renal recovery meant Scr returned to ≤ baseline/reference level without RRT. Partial recovery was defined as a ≥50% reduction from peak in-hospital creatinine, remaining above baseline/reference but no longer requiring RRT. No recovery was defined as failure to meet partial-recovery criteria or continued RRT dependence at discharge. Patients with complete or partial recovery were classified as ‘improved’ at discharge. Those without recovery were monitored post-discharge to determine whether persistent impairment met CKD criteria. Other outcomes (e.g., ICU admission, death) were recorded based on standard clinical criteria. Dark urine was defined as visibly dark, reddish-brown, or tea-colored urine, indicating rhabdomyolysis or severe hemolysis with significant myoglobin or hemoglobin release into the urine. Statistical analysis All analyses were performed using Python 3.9 and associated scientific computing libraries. Continuous variables are presented as mean ± SD or median (interquartile range (IQR)) according to distribution assessed by the Shapiro–Wilk test. Categorical variables are expressed as frequency and percentage. Between-group comparisons of normally distributed continuous variables were performed using independent-samples t-tests. Non-normally distributed data were compared using the Mann–Whitney U-test. Categorical variables were compared using the χ2 test or Fisher’s exact test. All statistical tests were two-sided, with p < 0.05 considered statistically significant. Model construction, validation, and explanation Feature screening and data preprocessing Variables with missing values were handled by complete-case analysis; lactate had the highest missing rate (10.16%), and all other variables had missing rates below 2% (Supplementary Table 1). After excluding missing values, 267 patients were included in the modeling dataset. Initially, univariate analysis (p < 0.05) was performed on 33 clinical variables for preliminary screening. Least Absolute Shrinkage and Selection Operator (LASSO) regression was then used for feature selection, with the optimal regularization parameter (λ) identified via 10-fold cross-validation. Continuous variables underwent Z-score standardization, and categorical variables were one-hot encoded. To evaluate model performance and prevent data leakage, the dataset was randomly split into training and test sets (7:3 ratio). All preprocessing parameters were fitted on the training set and subsequently applied to the test set. Feature selection (LASSO) and hyperparameter tuning were performed exclusively within the training set. LASSO regression identified eight features with non-zero coefficients: lactate, Scr, WBC count, dark urine, number of stings, LDH, BUN, and nausea and vomiting. To quantify the incremental value of Scr and address concerns regarding diagnostic leakage, we performed a sensitivity analysis by removing Scr from the feature set and re-training the primary models. The results are presented in Supplementary Table 2. Machine learning model development and hyperparameter tuning Using the key feature subset selected by LASSO regression, we systematically developed and compared six algorithms: LR, random forest (RF), support vector machine (SVM), neural network (NN), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). To handle class imbalance and optimize performance, class weights were assigned during training. Hyperparameters for each model were tuned using fivefold cross-validation with grid search, maximizing the area under the receiver operating characteristic curve (AUC) as the primary objective. The final hyperparameters were: LR (C = 10); RF (n_estimators = 100, max_depth = 5, min_samples_split = 2); SVM (C = 1, gamma = auto); XGBoost (n_estimators = 100, max_depth = 3, learning_rate = 0.05, subsample = 0.8, colsample_bytree = 0.8); LightGBM (n_estimators = 100, max_depth = 5, learning_rate = 0.05, num_leaves = 15, subsample = 0.8); NN (hidden_layer_sizes = (64), alpha = 0.01). Model evaluation, explanation, and clinical translation Model performance was assessed on the independent test set using accuracy, precision, recall, F1-score, specificity, and AUC with 95% confidence intervals (1,000 bootstrap replicates). To assess optimism and overfitting, 500 bootstrap resamples were performed on the training set to estimate the optimism-corrected AUC for the best-performing model. Variance inflation factor (VIF) was calculated to evaluate multicollinearity among selected features (Supplementary Table 3). To enhance transparency and clinical credibility, the best-performing model was interpreted with SHAP, quantifying each variable’s contribution to the final prediction [12]. Decision curve analysis (DCA) was performed to evaluate the clinical net benefit across decision thresholds [15]. Calibration curves assessed the agreement between predicted probabilities and observed risk. Finally, a visual nomogram based on independent predictors identified by multivariable LR was constructed as a bedside risk assessment tool for routine clinical use. Results Patient selection and baseline characteristics After applying the inclusion and exclusion criteria (Figure 1), a total of 305 bee-sting patients were included in this study, of whom 99 (32.5%) developed AKI. Among them, 47 (47.5%) were stage 1, seven (7.1%) stage 2, and 45 (45.5%) stage 3. There were no significant differences between the two groups in terms of age, gender, or underlying conditions. However, the AKI group had a significantly higher number of bee stings (p < 0.001) compared to the non-AKI group. Additionally, the mean systolic blood pressure was significantly higher in the AKI group (p = 0.01) and dark urine (p < 0.01) were higher in the AKI group at the time of presentation. Laboratory test results also revealed several indicators that were inconsistent between the two groups. Patient selection flow diagram depicting the inclusion of bee-sting patients in the analysis. Clinical course and outcomes of AKI Among the 99 patients who developed AKI, the severity of the disease showed a polarized distribution: 47 (47.5%) were classified as stage 1, seven (7.1%) as stage 2, and 45 (45.5%) as stage 3 (Figure 2(A)). The median onset times differed across stages (Figure 2(B)). For stages 1 and 2, median onset occurred on day 1 post-sting, with 76.6% (36/47) and 100% (7/7) diagnosed within 24 h, respectively. In contrast, stage 3 AKI had a median onset on day 2, with 55.6% (25/45) diagnosed at or after this point. Clinical characteristics and treatment analysis of AKI patients: (A) distribution of AKI stages; (B) onset-time distribution across AKI stages; (C) comparison of clinical outcomes among AKI stages; (D) treatment modalities according to AKI stages. AKI severity was correlated with clinical outcomes (Table 1; Figure 2(C)). ICU admission rates increased from 21.3% at stage 1 to 95.6% at stage 3, while mortality and CKD progression also rose with increasing severity of AKI [16]. In stage 2, one patient (14.3%) died from multi-organ dysfunction within 4 h of admission. Regarding treatment and renal recovery (Figure 2(D)), the proportion receiving aggressive interventions (e.g., RRT) increased with AKI stage. Conversely, the clinical improvement rate decreased progressively: stage 1: 95.7%, stage 2: 85.7%, and stage 3: 44.4%. Table 1.Comparison of clinical outcomes among patients with different stages of acute kidney injury.Clinical outcomesNon-AKI (n = 206)AKI stage 1 (n = 47)AKI stage 2 (n = 7)AKI stage 3 (n = 45)ICU admission, n (%)12 (5.8)10 (21.3)3 (42.9)43 (95.6)Mortality, n (%)0 (0.0)0 (0.0)1 (14.3)8 (17.8)Progression to CKD, n (%)0 (0.0)2 (4.2)0 (0.0)17 (37.8)Improvement, n (%)206 (100)45 (95.7)6 (85.7)20 (44.4)AKI: acute kidney injury; CKD: chronic kidney disease; ICU: intensive care unit.Data are presented as number (%). Comparison of clinical outcomes among patients with different stages of acute kidney injury. AKI: acute kidney injury; CKD: chronic kidney disease; ICU: intensive care unit. Data are presented as number (%). Model construction and evaluation Selection of predictive factors Eighteen potential predictive variables were initially identified through univariate analysis (Table 2). LASSO regression was then employed for further feature selection [17], with the optimal regularization parameter (λ = 0.042) identified using 10-fold cross-validation (Supplementary Fig. S1). Eight features with non-zero coefficients were selected: lactate (LAC), Scr, WBC count, dark urine, number of stings, LDH, and BUN and nausea and vomiting (Supplementary Fig. S1A–D). VIF values for all selected features were <5 (Supplementary Table 3), indicating no severe multicollinearity. Table 2.Comparison of baseline characteristics between AKI and non-AKI groups.VariablesAKI group (n = 99)Non-AKI group (n = 206)p ValueAge (years)63.00 (57.00, 70.00)63.00 (57.00, 68.00)0.473Sex, %Female61/99 (61.6%)119/206 (57.8%)0.606Male38/99 (38.4%)87/206 (42.2%)0.606Sting-count20.00 (10.00, 30.00)10.00 (5.75, 10.00)<0.001Time from sting to admission (hours)6.00 (4.00, 11.75)6.00 (5.00, 8.00)0.965Systolic blood pressure (mmHg)153.28 ± 29.28144.80 ± 23.770.013Diastolic blood pressure (mmHg)84.47 ± 14.4482.05 ± 12.680.156Heart rate (beats/min)79.00 (70.00, 91.50)78.00 (70.00, 88.00)0.308Body temperature (°C)36.90 (36.70, 37.20)36.90 (36.60, 37.20)0.485Dark urine36/99 (36.4%)6/206 (2.9%)<0.001Nausea and vomiting21/99 (21.2%)21/206 (10.2%)0.015Altered consciousness8/99 (8.1%)8/206 (3.9%)0.206Comorbidity, % Chronic kidney disease34/99 (34.3%)57/206 (27.7%)0.290Diabetes7/99 (7.1%)11/206 (5.3%)0.733Hypertension3/99 (3.0%)1/206 (0.5%)0.102Allergy history0/99 (0.0%)3/206 (1.5%)0.554Laboratory findings WBC, ×109/L15.40 (11.44, 22.10)10.40 (7.43, 14.20)<0.001NEUT%84.80 (74.30, 90.70)84.05 (76.33, 90.20)0.759Hb, g/L133.57 ± 24.35136.15 ± 19.740.359PLT, ×109/L204.00 (134.00, 252.00)172.00 (127.00, 209.50)0.007CRP, mg/L0.73 (0.50, 3.96)0.58 (0.49, 2.25)0.227LAC, mmol/L2.45 (1.50, 3.85)1.50 (1.20, 2.10)<0.001ALT, U/L44.00 (26.00, 120.50)25.00 (16.25, 34.00)<0.001AST, U/L95.00 (38.50, 595.00)37.00 (27.25, 52.00)<0.001TBil, μmol/L38.40 (22.90, 61.85)24.10 (16.80, 35.00)<0.001CK, U/L480.00 (187.00, 1,890.00)198.50 (123.00, 404.50)<0.001LDH, U/L516.00 (273.00, 1,659.50)272.50 (225.25, 334.00)<0.001Scr, μmol/L98.00 (72.00, 119.50)67.00 (58.00, 77.50)<0.001BUN, mmol/L8.59 (7.06, 10.52)6.51 (5.30, 7.92)<0.001K+, mmol/L3.97 (3.39, 4.29)3.72 (3.47, 3.95)0.005Na+, mmol/L139.00 (137.00, 141.00)140.00 (138.00, 141.00)0.133PT, s12.05 (11.30, 13.30)11.60 (11.00, 12.30)<0.001APTT(s)42.00 (28.95, 71.45)29.10 (25.70, 44.73)<0.001D-dimer, mg/L1.02 (0.56, 3.26)0.56 (0.30, 0.99)<0.001WBC: white blood cell count; NEUT: neutrophil percentage; Hb: hemoglobin; PLT: platelet count; CRP: C-reactive protein; LAC: lactate; ALT: alanine aminotransferase; AST: aspartate aminotransferase; TBil: total bilirubin; CK: creatine kinase; LDH: lactate dehydrogenase; Scr: serum creatinine; BUN: blood urea nitrogen; PT: prothrombin time; APTT: activated partial thromboplastin time; D-dimer: D-dimer.Normally distributed variables are presented as mean ± standard deviation; non-normally distributed variables are presented as median (P25, P75); categorical variables are presented as number (%). Comparison of baseline characteristics between AKI and non-AKI groups. WBC: white blood cell count; NEUT: neutrophil percentage; Hb: hemoglobin; PLT: platelet count; CRP: C-reactive protein; LAC: lactate; ALT: alanine aminotransferase; AST: aspartate aminotransferase; TBil: total bilirubin; CK: creatine kinase; LDH: lactate dehydrogenase; Scr: serum creatinine; BUN: blood urea nitrogen; PT: prothrombin time; APTT: activated partial thromboplastin time; D-dimer: D-dimer. Normally distributed variables are presented as mean ± standard deviation; non-normally distributed variables are presented as median (P25, P75); categorical variables are presented as number (%). Machine learning model construction and performance comparison Using these eight selected features, six algorithms were trained and compared. XGBoost achieved the best performance on the independent test set, with an AUC of 0.964, accuracy of 92.6%, and an F1-score of 89.3% (Table 3(A,B); Figure 3). The confusion matrix confirmed high true-positive and true-negative rates (Supplementary Fig. S2). After 500 bootstrap iterations, the optimism-corrected AUC for XGBoost was 0.943, with an average optimism of 0.021, indicating limited overfitting. Receiver operating characteristic curves for the six machine learning models. Table 3.Comparison of complete classification metrics between training sets and test sets for six machine learning models.ModelAUCAccuracyPrecisionRecallSpecificityF1-scoreConfusion matrix (TP/FN/FP/TN) (A) Classification metrics of training set Logistic regression0.9630.8980.8950.7970.9510.84351/13/6/117 Random forest0.9980.9791.0000.9381.0000.96860/4/0/123 SVM0.9820.9470.9660.8750.9840.91856/8/2/121 XGBoost0.9960.9790.9690.9690.9840.96962/2/2/121 LightGBM1.0001.0001.0001.0001.0001.00064/0/0/123 Neural network0.9920.9680.9530.9530.9760.95361/3/3/120 ModelAUCAUC 95%CIAccuracyPrecisionRecallSpecificityF1-scoreConfusion matrix(B) Classification metrics of test setLogistic regression0.953(0.890–0.995)0.9010.8570.8570.9250.85724/4/4/49Random forest0.960(0.900–0.9970)0.9380.9260.8930.9620.90925/3/2/51SVM0.943(0.877–0.990)0.8770.8460.7860.9250.81522/6/4/49XGBoost 0.964 (0.906–0.997)0.9260.8930.8930.9430.89325/3/3/50LightGBM0.949(0.882–0.993)0.9260.8670.9290.9250.89726/2/4/49Neural network0.939(0.870–0.990)0.9010.8850.8210.9430.85223/5/3/50SVM: support vector machine.The AUC 95%CI was calculated using 1,000 bootstrap iterations (test set only); XGBoost achieved the highest AUC on the test set (0.964, bold row); LightGBM exhibited an AUC of 1.000 on the training set, indicating complete overfitting; model performance metrics varied between training and test sets, derived from computations using the real XGBoost and LightGBM libraries; confusion matrix format: TP/FN/FP/TN; training set size n = 187 (AKI = 64), test set size n = 81 (AKI = 28). Comparison of complete classification metrics between training sets and test sets for six machine learning models. SVM: support vector machine. The AUC 95%CI was calculated using 1,000 bootstrap iterations (test set only); XGBoost achieved the highest AUC on the test set (0.964, bold row); LightGBM exhibited an AUC of 1.000 on the training set, indicating complete overfitting; model performance metrics varied between training and test sets, derived from computations using the real XGBoost and LightGBM libraries; confusion matrix format: TP/FN/FP/TN; training set size n = 187 (AKI = 64), test set size n = 81 (AKI = 28). Clinical utility validation Decision-curve analysis demonstrated that the XGBoost model provided a greater net clinical benefit across a wide range of decision thresholds (Supplementary Fig. S3A). Calibration curves showed good agreement between predicted probabilities and observed risks (Supplementary Fig. S3B). For bedside application, four independent predictors identified by multivariable LR (Table 4) were incorporated into a nomogram (Figure 4). In the multivariable model, the events-per-variable (EPV) was 8 (64 events with eight predictors), below the recommended value of 10, which was noted as a limitation. Nomogram for predicting AKI. Table 4.Multivariable logistic regression analysis of predictors for AKI.VariablesCoefficient βOR (95%CI)p ValueEPVSerum creatinine7.7632,352.586 (110.002–50,313.972)<0.001EPV = 8.00Lactate1.3453.838 (1.737–8.480)<0.001 Dark urine1.1383.120 (1.547–6.293)0.001 Lactate dehydrogenase2.67714.536 (2.120–99.660)0.006 BUN1.0242.784 (0.668–11.606)0.160 Sting-count0.1601.174 (0.633–2.175)0.611 WBC0.0891.093 (0.556–2.148)0.796 nausea-vomiting0.0651.068 (0.610–1.867)0.819 Intercept0.485–– EPV: event per variable; the training set contained 64 events with eight predictors included, resulting in an EPV of 8.00 (below the recommended value of 10). The wide confidence interval for serum creatinine reflects coefficient instability and should be interpreted in conjunction with sensitivity analysis. All variables were standardized.Note: All continuous variables were standardized (Z-score) before entry into the multivariable logistic regression. Therefore, the reported odds ratios (ORs) indicate the change in risk per one standard deviation increase of the respective predictor. The original clinical units are: serum creatinine (μmol/L), lactate (mmol/L), lactate dehydrogenase (U/L), blood urea nitrogen (mmol/L), and white blood cell count (×10⁹/L). Multivariable logistic regression analysis of predictors for AKI. EPV: event per variable; the training set contained 64 events with eight predictors included, resulting in an EPV of 8.00 (below the recommended value of 10). The wide confidence interval for serum creatinine reflects coefficient instability and should be interpreted in conjunction with sensitivity analysis. All variables were standardized. Note: All continuous variables were standardized (Z-score) before entry into the multivariable logistic regression. Therefore, the reported odds ratios (ORs) indicate the change in risk per one standard deviation increase of the respective predictor. The original clinical units are: serum creatinine (μmol/L), lactate (mmol/L), lactate dehydrogenase (U/L), blood urea nitrogen (mmol/L), and white blood cell count (×10⁹/L). Subgroup analysis by AKI stage In the test set, XGBoost achieved a recall of 0.818 for stage 1, 1.000 for stage 2, and 0.846 for stage 3, with median predicted probabilities increasing from 0.866 (stage 1) to 0.979 (stage 3). Specificity among non-AKI patients was 0.944, indicating stable predictive performance across different AKI severities (Supplementary Table 5). Feature importance analysis To elucidate the prediction mechanism of the best-performing XGBoost model, SHAP analysis was performed. Serum creatinine, LDH, lactate, and dark urine emerged as key drivers of AKI risk (Supplementary Fig. S4). Serum creatinine exhibited the highest SHAP value, indicating the greatest influence on model predictions. Dependency plots further revealed the quantitative relationship between feature values and predicted risk. Elevated Scr, lactate, LDH, and the presence of dark urine markedly increased the probability of AKI. Discussion Main findings In this study of 305 bee-sting patients, we identified two key features of bee-sting-induced AKI: rapid progression with poor prognosis, and the feasibility of early prediction using initial clinical indicators. First, we observed a high AKI incidence (32.5%) and a distinct polarized pattern of rapid progression. Nearly, half (45.5%) of AKI cases were classified as stage 3, with very few at stage 2 (7.1%). This distribution suggests a pattern of rapid progression in a substantial subset of patients, rather than a mechanistic inference of ‘bypassing’ intermediate stages. Stage 3 AKI was associated with very high rates of ICU admission (95.6%), mortality (17.8%), and CKD progression (37.8%). Second, AKI severity was inversely correlated with clinical improvement (stage 1: 95.7%; stage 3: 44.4%). This finding aligns with previous reports indicating that AKI severity is a critical predictor of poor renal recovery [18]. These results underscore the importance of early identification of high-risk bee-sting patients. To address this need, we developed and validated a machine-learning prediction model using admission indicators. The XGBoost model exhibited excellent performance (AUC = 0.964 on test set, optimism-corrected AUC = 0.943) and provided interpretability via SHAP analysis. Serum creatinine, dark urine, lactate, and LDH emerged as the four critical predictors, creating a precise and explainable tool for clinical early warning. Comparison with previous studies The prediction model developed in this study significantly advances previous research on bee-sting-related AKI prediction. First, compared to earlier single-center studies (e.g., Wu et al. [13], n = 214), this investigation included a larger patient cohort (n = 305), which improved statistical power and the generalizability of identified predictors. Second, while prior models (e.g., Wu et al. [13] and Tang et al. [19]) focused mainly on risk identification upon admission, this study provided a detailed longitudinal analysis of AKI progression, including onset timing, stage distribution, and dynamic clinical outcomes. Thus, our results offer more comprehensive insights into the natural history of bee-sting-induced AKI. Methodologically, earlier studies predominantly relied on conventional LR or limited machine-learning methods. For instance, Wu et al. [13] created a nomogram using LASSO regression and multivariable analysis, reporting an AUC of 0.757. In contrast, this study systematically compared six different machine-learning algorithms and found that the XGBoost model had superior predictive performance (AUC = 0.964), setting a new performance benchmark for clinical prediction models in this area [20]. Moreover, previous models primarily emphasized discriminative performance but offered limited clinical utility validation. This study utilized DCA and calibration plots to quantify clinical benefit and predictive accuracy [21], ensuring that the model is both statistically robust and clinically applicable. Additionally, SHAP-value analysis enhanced model interpretability, bridging complex machine-learning results and bedside decision-making – a feature rarely demonstrated in earlier studies on bee-sting AKI prediction. In summary, by incorporating a larger patient sample, comprehensive disease trajectory analysis, systematic algorithm comparison, and rigorous clinical validation, this study significantly advances AKI prediction research. Our results provide a reliable and practical tool for early risk stratification in bee-sting injury patients. Clinical and pathophysiological significance of the predictors Each of the four independent predictors identified in this study has a well-defined pathophysiological basis and clear clinical relevance. Serum creatinine, the principal biomarker of renal function, indicates early renal stress or injury, even before meeting KDIGO criteria, reflecting a subclinical stage of renal impairment [22]. Dark urine is a classic manifestation of hemoglobinuria resulting from hemolysis or rhabdomyolysis [7,23]. It signals the massive release of myoglobin or hemoglobin, which can block renal tubules and cause oxidative damage, making its presence an important clinical warning sign. Elevated lactate indicates inadequate tissue perfusion and a shift toward anaerobic metabolism [24]. In bee-sting victims, elevated lactate may reflect anaphylactic shock, hemolysis, or circulatory collapse [25]. Additionally, the systemic inflammatory response induced by bee venom is itself a known driver of AKI [26]. Serial lactate measurements thus provide valuable information regarding the severity of systemic shock. Lactate dehydrogenase, a nonspecific marker of cellular injury [27], rises during hemolysis, muscle injury, or hepatic damage, indicating multi-organ involvement. Although its individual predictive value is modest, a persistent upward trend often heralds clinical deterioration. Together, these four indicators represent early warning signals from distinct pathophysiological domains: renal impairment, hemolysis/rhabdomyolysis, tissue hypoxia, and systemic cellular injury, creating an integrated framework for AKI risk assessment following bee envenomation. Recent histopathological studies have further revealed that bee-sting-associated AKI involves diverse renal lesions, including acute tubular injury, interstitial nephritis, thrombotic microangiopathy, and rare glomerular pathologies such as collapsing glomerulopathy [28]. These findings confirm, at the tissue level, that bee venom induces renal injury via direct nephrotoxicity, immune-mediated damage, and ischemic insults. Such lesions likely account for the rapid clinical deterioration and poor outcomes observed in some patients, emphasizing the critical importance of early recognition. Strengths and limitations This study’s principal strength lies in its systematic integration of clinical-course analysis and advanced predictive modeling for bee-sting-induced AKI. We not only characterized the rapid progression and poor prognosis associated with this complication but also translated these insights into a high-performance, interpretable early-warning system. Methodologically, the study has several distinguishing features. It is the first to systematically compare six machine-learning algorithms and establish the superior predictive performance of XGBoost. By integrating LASSO regression, multivariable LR, and SHAP-based interpretability analysis, the study ensures robust feature selection and transparent model decisions. A staged modeling strategy was adopted to maintain predictive accuracy while identifying core predictors to construct a concise, user-friendly clinical nomogram. Additionally, the clinical value of the model was comprehensively evaluated using DCA and related metrics. Nevertheless, several limitations merit consideration. First, as a single-center retrospective study with a relatively modest sample size, particularly within the AKI subgroup (EPV = 8.0 for LR), the investigation is susceptible to selection bias and coefficient instability (e.g., wide confidence intervals for Scr). External validation using multicenter prospective cohorts is therefore necessary. Second, complete-case analysis was used to handle missing data. Lactate had the highest missing rate at 10.16%, which is a level commonly encountered in retrospective clinical studies. Although this approach may introduce selection bias, the sensitivity analysis that excluded Scr (Supplementary Table 2) demonstrated that the model maintained good predictive performance (AUC > 0.86). This supports the robustness of our findings despite the use of complete-case analysis. To address concerns regarding potential diagnostic leakage and the reliance on Scr, we performed a sensitivity analysis removing Scr from the model; the results (Supplementary Table 2) showed that the models maintained acceptable performance (AUC > 0.86), indicating that other predictors provide substantial independent value. However, the impact of different baseline definitions warrants further evaluation. Finally, the model was built solely on static admission parameters. Incorporating longitudinal time-series data collected during treatment may further enhance future predictive performance. Conclusions and future directions In summary, this study systematically characterizes the rapid progression and poor prognosis of AKI following bee envenomation. Additionally, it successfully develops a high-performance early prediction model. This model enables early identification of AKI in affected patients, with Scr, dark urine, lactate, and LDH identified as the strongest predictors. SHAP analysis enhances model interpretability, while DCA and the nomogram confirm clinical utility. Immediate next steps include three actions: First, the model should undergo external validation in larger, multicenter prospective cohorts. Second, emerging renal injury biomarkers such as NGAL and KIM-1 [29] should be incorporated to further refine risk stratification. Finally, following integration into hospital electronic medical record systems for real-time clinical decision support, prospective interventional trials are needed to confirm whether model use genuinely improves patient outcomes [30]. Only these trials will definitively establish the model’s clinical value. Supplementary Material Supplementary Material.docx