Summary
Tigecycline (TGC) is widely used to treat severe multidrug-resistant infections but can cause drug-induced liver injury. Existing prediction studies frequently lack external validation and offer limited clinical interpretability. We performed a multicenter retrospective cohort study of 1,946 adult inpatients who received TGC at three tertiary hospitals in Jiangsu, China. Liver injury was defined by biochemical criteria, and causality was assessed with a structured clinical method. After variable selection and model comparison, seven routinely available predictors were retained: TGC duration, concomitant hepatotoxic medications, neurological disease as the principal admission diagnosis, trauma as the principal admission diagnosis, baseline hepatic impairment, alcohol drinking history, and hemodialysis. eXtreme Gradient Boosting (XGBoost) showed the best overall performance and maintained high sensitivity in external validation. We implemented the final explainable model as a web-based calculator to support early screening, monitoring prioritization, and clinical reassessment.
Keywords: tigecycline, drug-induced liver injury, RUCAM, machine learning, prediction model, SHAP, multicenter study
Graphical abstract
Highlights
-
•
Seven clinical predictors estimate tigecycline-related liver injury risk
-
•
Multicenter validation supports high-sensitivity early screening
-
•
Treatment duration is the strongest contributor to predicted risk
-
•
A web calculator supports individualized monitoring decisions
Pharmaceutical science; Biochemistry; Artificial intelligence
Introduction
Antimicrobial resistance is a major global public health challenge of the 21st century and has profoundly influenced routine antimicrobial prescribing.1 Tigecycline (TGC) is a broad-spectrum glycylcycline with potent activity against Gram-positive and Gram-negative bacteria, atypical pathogens, and multidrug-resistant organisms.2 In adults, TGC is approved as monotherapy for three indications: complicated skin and skin-structure infections (excluding diabetic foot infections), complicated intra-abdominal infections, and community-acquired bacterial pneumonia. The standard regimen is a 100 mg loading dose followed by 50 mg every 12 h administered intravenously (i.v.). In real-world practice, however, TGC is frequently used off-label for other indications and at alternative doses; this occurs particularly for ventilator-associated pneumonia, hospital-acquired pneumonia, and bloodstream infections caused by carbapenem-resistant organisms.3 Several factors may explain this pattern, including concerns about suboptimal TGC concentrations in serum and epithelial lining fluid,4 the severity of antimicrobial resistance, and limited availability of alternative agents.5 Despite widespread use, particularly off-label, real-world safety data on TGC remain limited, and extensive off-label prescribing has raised concerns that adverse drug reactions may be more frequent and severe than indicated by the product label.6 Recent clinical studies have reported an association between TGC therapy and drug-induced liver injury (DILI).7,8,9,10 Mild cases may resolve after prompt withdrawal of the offending agent, while severe cases can progress to acute liver failure or death.11 Therefore, determining the incidence, clinical outcomes, and risk factors for TGC-associated DILI in routine practice is essential to support safer use of TGC.
Machine learning (ML) methods can integrate demographic, clinical, and laboratory data to improve DILI prediction and to support early identification of high-risk patients and individualized risk assessment.12 However, many ML models are structurally complex and therefore difficult to interpret directly, so they are often regarded as black boxes.13 Most reports of TGC-associated DILI are limited to case reports or small retrospective series. Although several predictive models have been proposed, some studies did not assess causality,10 which introduces a high risk of bias. Other investigations relied only on descriptive analyses or conventional regression methods,7,8 which are interpretable and robust but may be less flexible for capturing complex nonlinear relationships. Shapley additive explanations (SHAP) provide a framework for interpreting ML models by quantifying and visualizing each predictor’s contribution to model outputs at both the population and patient levels.14 Therefore, externally validated and interpretable models for TGC-associated DILI remain needed.
In this study, we conducted a multicenter retrospective cohort study of hospitalized adults who received TGC at three tertiary hospitals. Our primary objectives were to develop an explainable ML model to predict the risk of TGC-associated DILI and to assess its performance in internal and external validation cohorts using routinely collected clinical data. We used SHAP to quantify each predictor’s contribution and to provide explanations at the individual-patient level. This study also aimed to deliver an interpretable tool for early risk stratification of TGC-associated DILI and to support safer TGC use in routine clinical practice.
Results
Patient characteristics
This multicenter retrospective study included 1,946 patients who received TGC, including 1,640 patients from The First Affiliated Hospital with Nanjing Medical University in the model development cohort and 306 patients from Suqian First Hospital and The Third Affiliated Hospital of Soochow University in the independent external validation cohort. Within the development cohort, 1,312 patients were assigned to the training set and 328 patients to the internal validation set. The study design and patient flow are shown in Figure 1. Using the updated Roussel Uclaf Causality Assessment Method (RUCAM), we evaluated the causal relationship between TGC exposure and liver injury and classified TGC-treated patients into the TGC-associated DILI group (RUCAM score ≥6) and the non-DILI control group (RUCAM score <6), as shown in Figure S1. Baseline characteristics of the development cohort are summarized in Table 1, and the comparison between the development and external validation cohorts is shown in Table S6. Overall, 145 of 1,640 patients (8.84%) in the development cohort and 38 of 306 patients (12.42%) in the external validation cohort were diagnosed with TGC-associated DILI. Compared with the development cohort, patients in the external validation cohort were older, more frequently male, more often admitted to the intensive care unit (ICU), and differed in TGC duration and dosing patterns.
Figure 1.
Flowchart of the study design
LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; PR, precision-recall; AUC, area under the receiver operating characteristic curve; DCA, decision curve analysis; SHAP, Shapley additive explanations.
Table 1.
Baseline demographic and clinical characteristics of the development cohort according to drug-induced liver injury status
| Variables | Total (n = 1640) | DILI (n = 145) | Non-DILI (n = 1495) | P |
|---|---|---|---|---|
| Male, n (%) | 1081 (65.91) | 102 (70.34) | 979 (65.48) | 0.238 |
| Age (years) | 59.00 (49.00, 71.00) | 60.00 (51.00, 70.00) | 59.00 (49.00, 71.00) | 0.752 |
| Height (cm) | 168.00 (160.00, 172.00) | 168.00 (161.25, 171.00) | 168.00 (160.00, 173.00) | 0.863 |
| Weight (kg) | 65.00 (56.52, 73.30) | 66.00 (60.00, 75.00) | 65.00 (56.00, 73.10) | 0.073 |
| BMI(kg/mˆ2) | 23.41 (20.76, 25.95) | 23.88 (21.30, 26.49) | 23.24 (20.72, 25.95) | 0.048 |
| ICU admission during hospitalization, n (%) | 769 (46.89) | 93 (64.14) | 676 (45.22) | <0.001 |
| Principal admission diagnosis, n (%) | ||||
| Respiratory disease | 267 (16.28) | 22 (15.17) | 245 (16.39) | 0.705 |
| Cardiovascular disease | 203 (12.38) | 23 (15.86) | 180 (12.04) | 0.182 |
| Gastrointestinal disease | 187 (11.40) | 22 (15.17) | 165 (11.04) | 0.135 |
| Hematological disease | 646 (39.39) | 41 (28.28) | 605 (40.47) | 0.004 |
| Neurological disease | 173 (10.55) | 4 (2.76) | 169 (11.30) | 0.001 |
| Trauma | 111 (6.77) | 28 (19.31) | 83 (5.55) | <0.001 |
| Alcohol drinking history, n (%) | 363 (22.13) | 47 (32.41) | 316 (21.14) | 0.002 |
| Hypertension, n (%) | 555 (33.84) | 58 (40.00) | 497 (33.24) | 0.101 |
| Diabetes Mellitus, n (%) | 309 (18.84) | 29 (20.00) | 280 (18.73) | 0.709 |
| Hematologic Malignancy, n (%) | 544 (33.17) | 42 (28.97) | 502 (33.58) | 0.260 |
| Solid Tumor, n (%) | 179 (10.91) | 11 (7.59) | 168 (11.24) | 0.178 |
| CCI, M (Q1, Q3) | 5.00 (3.00, 6.00) | 5.00 (3.00, 6.00) | 5.00 (3.00, 6.00) | 0.871 |
| Bloodstream infection, n (%) | 301 (18.35) | 33 (22.76) | 268 (17.93) | 0.151 |
| Pulmonary infection, n (%) | 1214 (74.02) | 119 (82.07) | 1095 (73.24) | 0.021 |
| Abdominal infection, n (%) | 135 (8.23) | 19 (13.10) | 116 (7.76) | 0.025 |
| Thoracic infection, n (%) | 37 (2.26) | 5 (3.45) | 32 (2.14) | 0.472 |
| Central nervous system infection, n (%) | 26 (1.59) | 4 (2.76) | 22 (1.47) | 0.403 |
| Skin infection, n (%) | 142 (8.66) | 22 (15.17) | 120 (8.03) | 0.003 |
| Urinary tract infection, n (%) | 205 (12.50) | 14 (9.66) | 191 (12.78) | 0.278 |
| A. baumannii, n (%) | 603 (36.77) | 69 (47.59) | 534 (35.72) | 0.005 |
| K. pneumoniae, n (%) | 434 (26.46) | 44 (30.34) | 390 (26.09) | 0.267 |
| P. aeruginosa, n (%) | 245 (14.94) | 18 (12.41) | 227 (15.18) | 0.372 |
| S. aureus, n (%) | 119 (7.26) | 11 (7.59) | 108 (7.22) | 0.872 |
| Gram-negative, n (%) | 1022 (62.32) | 110 (75.86) | 912 (61.00) | <0.001 |
| Gram-positive, n (%) | 328 (20.00) | 23 (15.86) | 305 (20.40) | 0.192 |
| Fungi, n (%) | 297 (18.11) | 36 (24.83) | 261 (17.46) | 0.028 |
| Virus, n (%) | 149 (9.09) | 13 (8.97) | 136 (9.10) | 0.958 |
| Tracheal intubation, n (%) | 730 (44.51) | 90 (62.07) | 640 (42.81) | <0.001 |
| Urinary catheter, n (%) | 919 (56.04) | 101 (69.66) | 818 (54.72) | <0.001 |
| Hemodialysis, n (%) | 203 (12.38) | 37 (25.52) | 166 (11.10) | <0.001 |
| Arterial/venous catheter, n (%) | 1304 (79.51) | 128 (88.28) | 1176 (78.66) | 0.006 |
| Tigecycline duration (hours) | 214.00 (143.00, 320.00) | 264.00 (192.00, 384.00) | 203.00 (143.00, 312.00) | <0.001 |
| Loading dose, n (%) | 0.022 | |||
| 50 mg | 269 (16.40) | 19 (13.10) | 250 (16.72) | |
| 100 mg | 1318 (80.37) | 116 (80.00) | 1202 (80.40) | |
| 200 mg | 53 (3.23) | 10 (6.90) | 43 (2.88) | |
| Maintenance dose, n (%) | 0.004 | |||
| 50 mg | 634 (38.66) | 40 (27.59) | 594 (39.73) | |
| 100 mg | 1006 (61.34) | 105 (72.41) | 901 (60.27) | |
| Number of concomitant hepatotoxic medications | 1.00 (0.00, 2.00) | 2.00 (1.00, 3.00) | 1.00 (0.00, 2.00) | <0.001 |
| Baseline hepatic impairment, n (%) | 997 (60.79) | 117 (80.69) | 880 (58.86) | <0.001 |
| WBC (10ˆ9/L) | 7.78 (2.00, 12.59) | 8.27 (3.04, 12.77) | 7.75 (1.93, 12.57) | 0.275 |
| NE (10ˆ9/L) | 6.08 (1.21, 10.39) | 7.04 (2.30, 10.62) | 5.96 (1.13, 10.30) | 0.131 |
| NE (%) | 78.75 (60.45, 86.00) | 82.30 (69.60, 86.90) | 78.40 (59.44, 85.80) | 0.008 |
| LY (10ˆ9/L) | 0.68 (0.32, 1.13) | 0.65 (0.41, 1.00) | 0.68 (0.32, 1.14) | 0.963 |
| LY (%) | 10.80 (6.20, 22.21) | 9.20 (5.60, 17.65) | 10.90 (6.25, 22.70) | 0.058 |
| MO (10ˆ9/L) | 0.48 (0.13, 0.83) | 0.50 (0.23, 0.84) | 0.48 (0.12, 0.83) | 0.298 |
| MO (%) | 6.40 (3.80, 9.53) | 6.10 (4.40, 8.50) | 6.50 (3.80, 9.70) | 0.738 |
| BA (10ˆ9/L) | 0.02 (0.00, 0.03) | 0.02 (0.01, 0.04) | 0.02 (0.00, 0.03) | 0.082 |
| BA (%) | 0.20 (0.00, 0.40) | 0.20 (0.10, 0.40) | 0.20 (0.00, 0.40) | 0.050 |
| EO (10ˆ9/L) | 0.02 (0.00, 0.11) | 0.03 (0.00, 0.11) | 0.02 (0.00, 0.11) | 0.326 |
| EO (%) | 0.21 (0.00, 1.20) | 0.30 (0.00, 1.25) | 0.20 (0.00, 1.20) | 0.568 |
| RBC (10ˆ12/L) | 2.81 (2.43, 3.31) | 2.79 (2.52, 3.20) | 2.81 (2.42, 3.31) | 0.683 |
| HCT (%) | 26.30 (22.70, 30.40) | 26.55 (23.48, 29.30) | 26.20 (22.60, 30.50) | 0.607 |
| RDW-CV (%) | 14.50 (13.40, 16.20) | 14.45 (13.57, 16.15) | 14.50 (13.40, 16.20) | 0.841 |
| HGB (g/L) | 85.00 (74.00, 98.00) | 85.00 (76.00, 95.00) | 85.00 (74.00, 99.00) | 0.832 |
| MCV (fL) | 93.20 (89.40, 96.70) | 93.35 (90.45, 96.90) | 93.20 (89.30, 96.70) | 0.554 |
| MCH (pg) | 30.50 (29.30, 31.70) | 30.20 (29.20, 31.13) | 30.50 (29.30, 31.70) | 0.088 |
| MCHC (g/L) | 327.00 (316.00, 338.00) | 322.00 (314.75, 335.00) | 328.00 (317.00, 339.00) | 0.005 |
| MPV (fL) | 11.00 (10.09, 12.10) | 11.30 (10.30, 12.30) | 11.00 (10.00, 12.00) | 0.036 |
| PLT (10ˆ9/L) | 124.00 (37.00, 225.00) | 140.00 (53.00, 218.00) | 121.00 (36.00, 225.00) | 0.145 |
| Plateletcrit (%) | 0.14 (0.04, 0.24) | 0.18 (0.06, 0.24) | 0.14 (0.04, 0.24) | 0.085 |
| PDW (%) | 15.00 (12.00, 16.80) | 14.50 (12.40, 16.50) | 15.00 (12.00, 16.80) | 0.689 |
| CRP (mg/L) | 54.62 (16.83, 91.00) | 62.52 (26.53, 91.00) | 52.98 (16.18, 91.00) | 0.113 |
| PCT (ng/mL) | 0.28 (0.12, 0.95) | 0.47 (0.17, 1.47) | 0.27 (0.11, 0.89) | <0.001 |
| ALT (U/L) | 27.70 (16.60, 50.42) | 30.70 (18.60, 49.80) | 27.30 (16.40, 50.45) | 0.223 |
| AST (U/L) | 29.25 (19.50, 47.12) | 33.90 (22.00, 52.00) | 28.80 (19.30, 46.90) | 0.044 |
| ALP (U/L) | 91.00 (68.00, 133.75) | 103.00 (79.00, 131.00) | 90.00 (67.00, 134.00) | 0.073 |
| GGT (U/L) | 54.35 (28.70, 108.08) | 58.60 (36.00, 92.70) | 53.10 (27.80, 109.25) | 0.247 |
| TBIL (μmol/L) | 12.70 (8.70, 19.30) | 14.30 (10.20, 20.80) | 12.50 (8.60, 19.17) | 0.009 |
| DBIL (μmol/L) | 5.53 (3.70, 9.20) | 7.00 (4.50, 10.70) | 5.50 (3.70, 9.00) | <0.001 |
| IBIL (μmol/L) | 6.80 (4.60, 9.90) | 7.30 (5.00, 10.25) | 6.70 (4.50, 9.90) | 0.109 |
| TP (g/L) | 59.30 (53.90, 64.30) | 59.45 (53.88, 62.92) | 59.30 (53.90, 64.40) | 0.722 |
| ALB (g/L) | 32.70 (29.30, 36.05) | 32.20 (29.20, 35.55) | 32.80 (29.35, 36.10) | 0.349 |
| GLB (g/L) | 25.85 (22.00, 29.80) | 26.05 (22.93, 30.20) | 25.80 (22.00, 29.72) | 0.443 |
| A/G ratio | 1.30 (1.10, 1.50) | 1.25 (1.00, 1.50) | 1.30 (1.10, 1.50) | 0.268 |
| CK (U/L) | 37.00 (19.00, 113.50) | 70.00 (19.90, 163.50) | 36.00 (19.00, 110.50) | 0.006 |
| Urea (mmol/L) | 8.70 (6.08, 14.12) | 10.78 (7.18, 16.78) | 8.54 (5.97, 13.63) | <0.001 |
| Cr (μmol/L) | 61.80 (46.90, 91.10) | 71.10 (52.20, 119.80) | 61.30 (46.50, 88.60) | <0.001 |
| FIB (g/L) | 3.89 (2.90, 5.06) | 4.04 (2.69, 5.31) | 3.88 (2.92, 5.06) | 0.750 |
| D-dimer (mg/L) | 2.46 (0.87, 6.59) | 3.82 (1.71, 7.04) | 2.31 (0.82, 6.37) | <0.001 |
| APTT (s) | 31.10 (28.20, 35.10) | 32.60 (28.10, 36.50) | 31.05 (28.20, 34.90) | 0.253 |
| PT (s) | 13.60 (12.60, 15.10) | 13.90 (12.90, 15.80) | 13.60 (12.55, 15.00) | 0.001 |
| TT (s) | 16.00 (15.20, 16.90) | 16.00 (15.17, 16.70) | 16.00 (15.20, 17.00) | 0.737 |
| INR | 1.19 (1.10, 1.32) | 1.23 (1.13, 1.40) | 1.19 (1.10, 1.32) | 0.001 |
Categorical variables are presented as n (%) and continuous variables as median (Q1, Q3). The DILI and non-DILI groups were compared using the chi-square test or Fisher’s exact test, as appropriate, for categorical variables and the Mann-Whitney U test for continuous variables. DILI, drug-induced liver injury; BMI, body mass index; CCI, Charlson comorbidity index; ICU, intensive care unit; WBC, white blood cell; NE, neutrophil; LY, lymphocyte; MO, monocyte; BA, basophil; EO, eosinophil; RBC, red blood cell; HCT, hematocrit; RDW-CV, red cell distribution width-coefficient of variation; HGB, hemoglobin; MCV, mean corpuscular volume; MCH, mean corpuscular hemoglobin; MCHC, mean corpuscular hemoglobin concentration; MPV, mean platelet volume; PLT, platelet; PCT, procalcitonin; CRP, C-reactive protein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltransferase; TBIL, total bilirubin; DBIL, direct bilirubin; IBIL, indirect bilirubin; TP, total protein; ALB, albumin; GLB, globulin; A/G ratio, albumin-to-globulin ratio; CK, creatine kinase; Cr, serum creatinine; FIB, fibrinogen; APTT, activated partial thromboplastin time; PT, prothrombin time; TT, thrombin time; INR, international normalized ratio.
Selection of predictor variables
Variable selection was performed only within the imputed training datasets. Grouped least absolute shrinkage and selection operator (LASSO) logistic regression initially identified 13 candidate predictors with nonzero grouped coefficients: trauma as the principal admission diagnosis (trauma), duration of TGC therapy (duration), baseline hepatic impairment (hepatic_impairment), number of concomitant hepatotoxic medications (drugs), neurological disease as the principal admission diagnosis (nervous), receipt of hemodialysis (hemodialysis), alcohol drinking history (drink), tracheal intubation (trachealtube), TGC maintenance dose (maintenancedose), ICU admission, blood urea (urea), mean corpuscular hemoglobin concentration (MCHC), and prothrombin time (PT). Grouped coefficient summaries are shown in Table S2.
Spearman’s rank correlation analysis was then used to evaluate multicollinearity and redundancy among these candidate predictors (Figure 2A). ICU admission was highly correlated with tracheal intubation. Because ICU admission can be influenced by bed availability, local admission thresholds, and regional differences in critical care practice, whereas tracheal intubation represents a more consistently defined clinical intervention, ICU admission was removed from further consideration. The remaining 12 predictors were entered into bootstrap stability selection.
Figure 2.
Variable selection procedures
(A) Spearman’s correlation heatmap. Based on the 20 imputed training datasets, the heatmap shows the Spearman’s correlations among variables initially selected by grouped LASSO. Circle size represents the absolute value of the Spearman’s correlation coefficient, and color indicates the direction of correlation.
(B) Bootstrap stability selection frequency. After 200 bootstrap resampling iterations, the selection frequency of each candidate predictor was calculated. The bar length represents the proportion of bootstrap samples in which each variable was retained. The red vertical dashed line indicates the prespecified stability threshold of 0.90.
After 200 bootstrap resampling iterations, seven predictors had a selection frequency greater than 0.90 and were retained as stable predictors for subsequent model development and validation: number of concomitant hepatotoxic medications, neurological disease as the principal admission diagnosis, trauma as the principal admission diagnosis, baseline hepatic impairment, duration of TGC therapy, alcohol drinking history, and receipt of hemodialysis (Figure 2B).
Model performance and final model selection
Using the seven selected predictors, we trained eight supervised ML models across the multiply imputed training datasets. Overall model performance in the training datasets is summarized in Table S3, and the selected hyperparameter configurations for each imputed dataset are reported in Table S4. Based on pooled area under the receiver operating characteristic curve (AUC) values greater than 0.75 from training-set out-of-fold cross-validated predictions across imputations, Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Gradient Boosting (GB), decision tree (DT), and random forest (RF) were carried forward for ensemble modeling.
The receiver operating characteristic (ROC) curves of all single models in the internal validation cohort are shown in Figure 3. Among single models, XGBoost achieved the highest AUC (0.935; 95% confidence interval [CI], 0.902–0.960). In addition, XGBoost had the lowest Brier score among all candidate and ensemble models (0.057; 95% CI, 0.042–0.073) and the highest average precision in precision-recall analysis (0.491). At the locked threshold of 0.153, XGBoost yielded an accuracy of 0.860, sensitivity of 0.966, specificity of 0.849, positive predictive value (PPV) of 0.384, and negative predictive value (NPV) of 0.996 (Figure 4 and Table 2). Pairwise DeLong testing showed no significant difference between XGBoost and the stacking ensemble (Δ AUC, 0.004; p = 0.308), between XGBoost and the soft-voting ensemble (Δ AUC, 0.011; p = 0.081), or between the stacking and soft-voting ensembles (Δ AUC, 0.006; p = 0.140; Table S5). Considering predictive performance, calibration, parsimony, and interpretability, XGBoost was selected as the final model for predicting TGC-associated DILI.
Figure 3.
Receiver operating characteristic curves of eight single ML models in the internal validation cohort
Figure 4.
Performance comparison of the machine learning models in the internal validation set
(A) Receiver operating characteristic curves.
(B) Precision-recall curves.
(C) Calibration plots.
(D) Decision curve analysis.
Table 2.
Performance of machine learning models in the internal validation cohort
| Model | AUC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | Brier (95% CI) | Calibration Intercept | Calibration Slope | Best Threshold |
|---|---|---|---|---|---|---|---|---|---|---|
| Logistic Regression | 0.771 (0.689–0.848) | 0.701 | 0.793 | 0.692 | 0.200 | 0.972 | 0.090 (0.075–0.107) | −0.593 | 1.471 | 0.232 |
| Decision Tree | 0.869 (0.818–0.916) | 0.646 | 1.000 | 0.612 | 0.200 | 1.000 | 0.112 (0.096–0.127) | −1.586 | 1.431 | 0.258 |
| Random Forest | 0.819 (0.753–0.877) | 0.674 | 0.931 | 0.649 | 0.205 | 0.990 | 0.073 (0.051–0.096) | −0.283 | 0.792 | 0.042 |
| Gradient Boosting | 0.923 (0.885–0.954) | 0.851 | 0.931 | 0.843 | 0.365 | 0.992 | 0.063 (0.045–0.084) | 0.123 | 0.986 | 0.073 |
| Support Vector Machine | 0.777 (0.703–0.850) | 0.436 | 1.000 | 0.381 | 0.136 | 1.000 | 0.075 (0.053–0.099) | 0.605 | 1.275 | 0.055 |
| K-Nearest Neighbors | 0.782 (0.686–0.868) | 0.680 | 0.828 | 0.666 | 0.194 | 0.975 | 0.073 (0.052–0.096) | −1.376 | 0.258 | 0.064 |
| LightGBM | 0.922 (0.887–0.950) | 0.811 | 1.000 | 0.793 | 0.319 | 1.000 | 0.066 (0.048–0.087) | −0.306 | 0.761 | 0.039 |
| XGBoost | 0.935 (0.902–0.960) | 0.860 | 0.966 | 0.849 | 0.384 | 0.996 | 0.057 (0.042–0.073) | 0.063 | 1.253 | 0.153 |
| Stacking Ensemble | 0.930 (0.901–0.958) | 0.872 | 0.931 | 0.866 | 0.403 | 0.992 | 0.064 (0.045–0.085) | 0.512 | 1.199 | 0.071 |
| Soft Voting Ensemble | 0.924 (0.890–0.952) | 0.845 | 0.966 | 0.833 | 0.359 | 0.996 | 0.061 (0.045–0.076) | 0.060 | 1.408 | 0.155 |
AUC, the area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value.
Subgroup threshold analysis
In prespecified subgroup analyses stratified by ICU admission, the ICU-specific threshold was lower (0.083), whereas the non-ICU-specific threshold was higher (0.171). The model showed higher discrimination in the non-ICU subgroup than in the ICU subgroup (AUC, 0.956 vs. 0.910). When ICU-specific and non-ICU-specific thresholds were combined, sensitivity was 0.931, specificity was 0.826, and NPV was 0.992 (Table 3). Compared with the overall thresholds, subgroup-specific thresholds shifted the balance between sensitivity and specificity but did not provide a meaningful overall improvement. Therefore, the overall threshold was retained for subsequent analyses.
Table 3.
Subgroup performance of the XGBoost model in the internal validation cohort
| Analysis group | Thresholding strategy | Decision threshold | AUC | Accuracy | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|---|---|---|
| Overall cohort | Overall threshold | 0.153 | 0.935 | 0.860 | 0.966 | 0.849 | 0.384 | 0.996 |
| ICU subgroup | ICU-specific threshold | 0.083 | 0.910 | 0.739 | 0.950 | 0.707 | 0.328 | 0.989 |
| Non-ICU subgroup | Non-ICU-specific threshold | 0.171 | 0.956 | 0.920 | 0.889 | 0.922 | 0.381 | 0.994 |
| Subgroup-specific combined strategy | ICU/non-ICU-specific thresholds | 0.083/0.171 | 0.935 | 0.835 | 0.931 | 0.826 | 0.342 | 0.992 |
All thresholds were selected using the maximum F-beta score (beta = 3) criterion. AUC, the area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value.
External validation of the final model
The final XGBoost model was externally validated in the independent cohort of 306 patients. The model maintained good discrimination, with an AUC of 0.850 (95% CI, 0.778–0.914) on the ROC curve (Figure S2A). The precision-recall curve showed an average precision of 0.490 (Figure S2B). Overall calibration was acceptable, with a Brier score of 0.094 (Figure S2C). Decision curve analysis showed a positive net benefit for XGBoost compared with treat-all and treat-none strategies across low to moderate threshold probabilities (Figure S2D). At the locked threshold of 0.153, external validation performance was characterized by high sensitivity and lower specificity, with an accuracy of 0.516, sensitivity of 0.921, specificity of 0.459, PPV of 0.194, and NPV of 0.976. This performance pattern was consistent with the sensitivity-oriented screening purpose of the model.
False-positive analysis in the external validation cohort
To better characterize false-positive predictions in the external validation cohort, we compared false-positive and true-negative cases among patients without DILI (Table S7). The analysis included the final model predictors and additional variables with a preliminary univariable p < 0.20. False-positive cases had substantially longer TGC exposure than true-negative cases (median, 168.0 h [interquartile range (IQR), 144.0–216.0] vs. 96.0 h [IQR, 72.0–216.0]; false discovery rate (FDR)-adjusted q < 0.001). Neurological disease as the principal admission diagnosis, gastrointestinal disease as the principal admission diagnosis, and skin infection showed nominal univariable differences before FDR correction, whereas alanine aminotransferase (ALT) and gamma-glutamyltransferase (GGT) showed borderline univariable imbalances. None of these variables remained statistically significant after FDR correction. These findings indicated that longer TGC exposure was the only measured factor that remained significantly different between false-positive and true-negative cases in the external validation cohort.
Model interpretation and visualization
SHAP were used to interpret the final XGBoost model. The global SHAP bar plot showed that duration of TGC therapy was the leading predictor, followed by the number of concomitant hepatotoxic medications, neurological disease as the principal admission diagnosis, alcohol drinking history, baseline hepatic impairment, receipt of hemodialysis, and trauma as the principal admission diagnosis (Figure 5A). The SHAP beeswarm and dependence plots showed that longer TGC exposure was consistently associated with higher predicted risk, with a nonlinear pattern across the observed range (Figures 5B and 5C). A greater number of concomitant hepatotoxic medications also tended to increase predicted risk. Among binary predictors, trauma as the principal admission diagnosis, alcohol drinking history, baseline hepatic impairment, and receipt of hemodialysis were generally associated with higher predicted risk, whereas neurological disease as the principal admission diagnosis was associated with lower predicted risk in the SHAP dependence plot. These model explanations should be interpreted as model-based associations rather than causal effects.
Figure 5.
Global model explanation by the SHAP method
(A) SHAP summary bar plot. This plot ranks the seven predictors by mean absolute SHAP values, showing their overall contributions to the model.
(B) SHAP summary dot plot. Higher SHAP values indicate a higher predicted probability of TGC-associated DILI. Each dot represents the SHAP value for a single patient for the corresponding feature. The color indicates the feature value (higher vs. lower), and vertical stacking reflects the density of points.
(C) SHAP dependence plots. Each plot illustrates how an individual feature influences the model output, with each dot representing one patient. SHAP values are shown on the y axis and actual feature values on the x axis. SHAP values above zero push the prediction toward the “DILI” class. TGC, tigecycline; DILI, drug-induced liver injury; SHAP, Shapley additive explanations.
At the individual level, we generated SHAP waterfall plots for representative patients (Figure 6). Figure 6A shows a patient from the TGC-associated DILI group. The waterfall plot decomposes the model output into contributions from each predictor; features with positive SHAP values move the prediction toward the DILI class on the model-output scale. Figure 6B shows a patient from the non-DILI control group. In this case, the large negative contribution from shorter TGC duration moved the model output away from the DILI class despite positive contributions from alcohol drinking history and baseline hepatic impairment.
Figure 6.
Local model explanation using the SHAP method
(A) Waterfall plot for a representative patient classified as high risk for TGC-associated DILI.
(B) Waterfall plot for a representative patient classified as low risk. SHAP values show how each feature moves the model output from the baseline value E[f(X)] to the individual value f(x) on the model-output scale. Positive SHAP values move the prediction toward the DILI class, whereas negative values move it away from the DILI class. TGC, tigecycline; DILI, drug-induced liver injury; SHAP, Shapley additive explanations.
Implementation of the web calculator
Figure 7 shows the deployed XGBoost prediction model implemented as an interactive web-based calculator for preliminary clinical risk estimation. Users enter the values of the seven predictor variables, and the application computes the individualized probability of TGC-associated DILI and reports the corresponding risk category. The web application is available at https://predictionmodeltgcdili.streamlit.app/.
Figure 7.
Web-based calculator for predicting TGC-associated DILI using the final XGBoost model
After the patient’s values for the seven predictors are entered, the application automatically outputs the individualized probability and corresponding risk classification. TGC, tigecycline; DILI, drug-induced liver injury.
Discussion
This multicenter study developed and externally validated an explainable ML model for early screening of TGC-associated DILI using seven routinely available clinical predictors. We selected XGBoost as the optimal algorithm because it demonstrated superior discrimination, calibration, and clinical utility across the evaluation metrics. Because the model was designed for screening rather than definitive diagnosis, threshold selection prioritized sensitivity using an F-beta score (beta = 3). We then applied SHAP to generate explanations at both the population and individual patient levels. Finally, we implemented the model as a web-based calculator to facilitate point-of-care use.
DILI is a potentially fatal adverse drug reaction. Prospective cohorts have reported substantial morbidity, and a clinically meaningful proportion of affected patients have died or required liver transplantation after DILI onset.15 Nevertheless, early risk assessment in routine care remains limited and relies largely on serial liver tests and clinical judgment. Interpreting abnormal liver biochemistry during TGC therapy is challenging because these abnormalities often reflect sepsis-associated cholestasis, hypoxic hepatitis, or multi-organ dysfunction rather than intrinsic TGC hepatotoxicity. To reduce misclassification, we excluded patients with Child-Pugh class C hepatic dysfunction, septic shock, multiple organ failure, autoimmune hepatitis, or acute viral hepatitis, and adjudicated causality in the remaining patients using the updated RUCAM.
Although several studies evaluated the hepatic safety of TGC, they have not produced a robust screening tool for TGC-associated DILI because of limited cohort sizes, lack of external validation, and incomplete outcome adjudication. Shi et al.7 and Yu et al.8 used the updated RUCAM and multivariable logistic regression to examine the association between TGC exposure and liver injury and to identify independent risk factors. These studies, however, concentrated on identifying risk factors and did not produce individualized risk estimates or systematically compare alternative modeling approaches. Nomogram-based scores16,17 and a single ML model10 for TGC-associated DILI have been reported, but they were derived from non-multicenter cohorts with modest sample sizes and lacked structured causality assessments, which can increase bias and limit generalizability.
XGBoost is an optimized gradient-boosting ensemble algorithm that combines multiple weak learners to improve predictive performance and generalization.18,19 After comparing model discrimination, calibration, and screening-oriented performance, we selected XGBoost as the final model and used SHAP to quantify predictor contributions at both population and individual-patient levels. Because the model was designed for screening rather than definitive diagnosis, threshold selection prioritized sensitivity. The final operating threshold was determined using the maximum F3 score based on training-set out-of-fold predictions. The F3 rule was used as a transparent sensitivity-weighted operating rule, and clinical utility was evaluated separately using decision curve analysis.20 In the internal validation cohort, the final XGBoost model achieved an AUC of 0.935 and a sensitivity of 0.966 at the locked threshold of 0.153. In the external validation cohort, discrimination remained acceptable, with an AUC of 0.850, and sensitivity remained high at 0.921, although specificity decreased to 0.459. In the external validation cohort, false-positive cases had longer TGC exposure than true-negative cases, suggesting that prolonged exposure may have contributed to false-positive classification.
The decline in external specificity should be interpreted in the context of between-cohort differences. Compared with the development cohort, the external validation cohort included older patients, a higher proportion of men, more ICU admissions, a much higher proportion of respiratory disease as the principal admission diagnosis, and differences in TGC treatment duration and dosing patterns. These differences suggest possible dataset shift related to time, geography, hospital practice, and case mix. Therefore, the model should be regarded as an early warning tool that may support monitoring decisions rather than a fixed diagnostic rule. Local validation and, if necessary, recalibration, should be performed before broad implementation in new regions, hospital systems, or patient populations.
SHAP values for TGC treatment duration rose in an approximately monotonic pattern, indicating steadily increasing predicted risk with longer courses. This finding is consistent with prior TGC cohorts, which repeatedly identify prolonged therapy as an independent risk factor and show that risk becomes more apparent after about 1–2 weeks of exposure.7,17 Experimental data for the tetracycline class support plausible mechanisms such as mitochondrial dysfunction, oxidative stress, and pro-inflammatory signaling.21 Clinically, these patterns support intensified liver test monitoring and regular benefit-risk reassessment once TGC is continued beyond the first week.
SHAP values also rose stepwise with the number of concomitant hepatotoxic agents, particularly when two or more were co-prescribed, consistent with prior reports.7 In practice, nonessential hepatotoxic drugs should be de-prescribed and lower-risk alternatives chosen whenever feasible.
Baseline hepatic impairment showed a strong positive association, consistent with prior studies indicating that chronic liver disease or baseline liver abnormalities increase susceptibility to TGC-associated DILI.7,15 Because TGC is primarily processed hepatically with predominant biliary clearance, impaired liver function likely reduces clearance and raises drug exposure, which plausibly increases the risk of toxicity.22
Neurological disease as the principal admission diagnosis showed a negative SHAP contribution. This finding should be considered hypothesis-generating rather than evidence of a causal protective effect. Only four patients in the TGC-associated DILI group had neurological disease as the principal admission diagnosis, so the signal may reflect the case mix and treatment pathways in this cohort rather than a biologic protective mechanism. It should not be used to reduce liver monitoring when other high-risk features are present.
Trauma as the principal admission diagnosis and receipt of hemodialysis both contributed positively to the SHAP score. Trauma-related admission may increase susceptibility via hemorrhagic shock, ischemia-reperfusion injury, immune activation, and polypharmacy.23 Although pharmacokinetic studies suggest conventional hemodialysis does not substantially alter TGC exposure,24 dialysis often indicates chronic inflammation and complex co-medication, which may lower the threshold for hepatic injury. Given the limited direct evidence, these findings should be considered hypothesis-generating. Nevertheless, closer monitoring remains prudent when TGC is used in patients admitted for trauma or receiving hemodialysis.
Alcohol drinking history was likewise associated with increased risk, consistent with updated RUCAM and epidemiological studies that identify alcohol as a host risk factor for DILI.25
Although prior studies have identified TGC maintenance dose as an independent risk factor and predictor of DILI,16,17 it did not persist in our final model. High-dose therapy (100 mg q12h) was associated with DILI in univariable analyses (p < 0.01). However, forcing dose category into the multivariable ML models slightly worsened discrimination and calibration, and its SHAP contribution remained small, suggesting limited incremental value beyond other covariates. We therefore excluded maintenance dose to preserve parsimony and minimize over-fitting. In the exploratory dose-stratified interaction analysis, high-dose TGC did not show statistically significant interactions with any of the seven final predictors after FDR correction (Table S8). The safety profile of high-dose TGC remains heterogeneous26; specific toxicities, such as coagulation abnormalities, have been reported in critically ill patients receiving high-dose regimens.27 In our cohort, high-dose prescribing was uncommon, and the validation sample was modest, which may have reduced power to detect a stable dose-hepatotoxicity relationship. However, for patients receiving high-dose TGC, especially those with baseline hepatic impairment or multiple concomitant hepatotoxic medications, liver biochemistry, and coagulation parameters should be assessed at baseline and monitored repeatedly during treatment, with regular reassessment of the benefit-risk balance.
Future research should build on these findings in several ways. First, larger multicenter cohorts, including hospitals from different regions and care settings, are needed to further externally validate the model and evaluate its performance in broader clinical populations. Second, dedicated analyses of predefined subgroups, including strata defined by age, sex, comorbidity burden, underlying liver disease and infection severity, are important for characterizing heterogeneity in model performance and addressing fairness and equity concerns. Third, future studies should develop dynamic prediction models that incorporate serial liver tests, coagulation parameters, changes in TGC exposure, and adjustments in concomitant hepatotoxic medications. Such time-updated models could provide sequential risk estimates at key decision points during therapy and may better support ongoing benefit-risk assessment and adaptive treatment strategies. From an implementation perspective, further work is required to embed the model into routine workflows. The current web-based calculator could be refined and linked to hospital information systems so that key predictors are automatically retrieved from electronic records and individual risk estimates are generated upon prescription and during treatment.
In conclusion, we developed and externally validated an explainable ML model to predict the risk of TGC-associated DILI in adult inpatients. The final XGBoost model was based on seven routinely available clinical predictors and maintained high sensitivity when used as a screening tool. To facilitate bedside use, we implemented the model as an interactive web-based calculator.
The intended population is adult inpatients receiving i.v. TGC in tertiary-care hospitals. The model is not intended for community or outpatient settings, nor for patients with baseline Child-Pugh class C hepatic dysfunction, septic shock, multiple organ failure, autoimmune hepatitis, or acute viral hepatitis. Without additional external validation, it should not be directly extrapolated to other regions or countries. A positive prediction should not be interpreted as confirmed TGC-associated DILI or used as a stand-alone basis for changing TGC therapy. Instead, it indicates a higher predicted risk and should prompt closer monitoring of liver biochemistry and coagulation function, reassessment of concomitant hepatotoxic medications, and careful evaluation of alternative causes of liver injury.
Limitations of the study
This study has several limitations. First, the retrospective design increases the risk of residual confounding, incomplete capture of clinically relevant risk factors, and information bias. Although we applied the updated RUCAM and, whenever possible, excluded patients with clear alternative causes of liver injury, misclassification of liver injury etiology cannot be entirely excluded. Second, the study population comprised adult inpatients from Chinese tertiary hospitals, who carry a high burden of severe infection and multidrug-resistant organisms. The model’s performance in other healthcare systems, in community or outpatient settings, and inpatients with less severe disease therefore remains uncertain. Third, some potentially important predictors and patient-level characteristics, such as detailed pharmacokinetic measurements, gender identity information beyond recorded biological sex, and comprehensive evaluations for viral hepatitis and autoimmune liver disease, were not systematically available in routine electronic records and could not be incorporated. This limitation may have introduced unmeasured residual confounding and reduced the model’s ability to capture mechanistic determinants of risk. Fourth, although we performed both internal and external validations, overfitting remains possible, particularly for rare combinations of risk factors. Consequently, larger prospective multicenter cohorts are required to evaluate the model’s robustness and transportability. Fifth, although we explored ICU-specific and non-ICU-specific thresholds, these subgroup-specific thresholds changed the balance between sensitivity and specificity but did not provide a meaningful overall improvement over the global threshold. Therefore, the overall threshold was retained for the main analysis. We did not comprehensively evaluate performance across other clinically relevant strata. In populations with substantially different case mix, local validation, recalibration, or threshold adjustment may be necessary before clinical implementation. Future work should address these issues, including prospective implementation of the web-based calculator and integration of the model into hospital information systems.
Resource availability
Lead contact
Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Dr. Wenkui Sun (sunwenkui@njmu.edu.cn).
Materials availability
This study did not generate new unique reagents.
Data and code availability
-
•
The de-identified clinical data reported in this study are publicly available at Mendeley Data: https://doi.org/10.17632/8fmyrxmnkd.1.
-
•
This paper does not report original code.
-
•
Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.
Acknowledgments
The study was funded by the Qinglan Project of Jiangsu Province of China (2024-14). The funding sources had no influence on the study design, data collection, manuscript preparation or the decision to publish the results.
Author contributions
Y. Zhang performed the data analyses, developed and validated the ML models, and drafted the manuscript; P.J. contributed to data curation, provided data, and participated in manuscript writing and revision; P.J., S.M., C.F., and T.S. provided data and participated in data collection and verification; D.W., W.L., Y. Zheng, X.Z., and Y.D. participated in data collection and clinical data interpretation; Q.Q. coordinated the study and contributed to manuscript revision; W.S. conceived and designed the study, supervised the project, and contributed to data interpretation and critical revision of the manuscript; Y. Zhang and P.J. accessed and verified the underlying data. All authors read and approved the final manuscript. W.S. had final responsibility for the decision to submit for publication.
Declaration of interests
The authors declare no competing interests.
STAR★Methods
Key resources table
| REAGENT or RESOURCE | SOURCE | IDENTIFIER |
|---|---|---|
| Deposited data | ||
| De-identified clinical cohort dataset | The First Affiliated Hospital with Nanjing Medical University; Suqian First Hospital; Third Affiliated Hospital of Soochow University | Mendeley Data: https://doi.org/10.17632/8fmyrxmnkd.1 |
| Software and algorithms | ||
| IBM SPSS Statistics (v27.0) | IBM Corporation | https://www.ibm.com/products/spss-statistics |
| R software (v4.5.0) | R Core Team and R Foundation for Statistical Computing | https://www.r-project.org/ |
| Python (v3.10.4) | Python Software Foundation | https://www.python.org/ |
| NumPy (v1.26.4) | NumPy Developers | https://numpy.org/ |
| pandas (v2.3.2) | pandas development team | https://pandas.pydata.org/ |
| scikit-learn (v1.7.2) | scikit-learn developers | https://scikit-learn.org/stable/ |
| LightGBM (v4.6.0) | Microsoft; LightGBM developers | https://lightgbm.readthedocs.io/ |
| XGBoost (v3.1.2) | XGBoost developers | https://xgboost.readthedocs.io/ |
| statsmodels (v0.14.5) | statsmodels developers | https://www.statsmodels.org/ |
| SciPy (v1.15.3) | SciPy community | https://scipy.org/ |
| Matplotlib (v3.10.6) | Matplotlib development team | https://matplotlib.org/ |
| Grouped LASSO variable-selection procedure | Du et al., 202228 | https://doi.org/10.1080/10618600.2022.2035739 |
| Bootstrap stability selection procedure | Meinshausen et al., 201029 | https://doi.org/10.1111/j.1467-9868.2010.00740.x |
Experimental model and study participant details
We conducted a multicenter retrospective cohort study of adults who received TGC at the First Affiliated Hospital with Nanjing Medical University (January 1, 2018 to December 31, 2024), Suqian First Hospital (January 1, 2021 to August 31, 2025), and The Third Affiliated Hospital of Soochow University (January 1, 2021 to August 31, 2025). Eligible patients were identified through hospital information systems and medical records according to prespecified criteria. Patients were eligible for inclusion if they received TGC during hospitalization. Exclusion criteria were applied in a stepwise manner as follows: (1) TGC exposure for <72 h; (2) age <18 years; (3) missing key clinical or laboratory data required for the analysis; (4) baseline Child–Pugh class C hepatic dysfunction; and (5) septic shock, multiple organ failure, autoimmune hepatitis, or acute viral hepatitis. The final study cohort included all patients who met these eligibility criteria, and the patient selection process is illustrated in Figure 1.
The final study cohort included 1,946 adult patients, comprising 1,640 patients from The First Affiliated Hospital with Nanjing Medical University, 6 patients from Suqian First Hospital, and 300 patients from The Third Affiliated Hospital of Soochow University. The model development cohort included 1,640 patients, and the external validation cohort included 306 patients. The patient selection and grouping process is shown in Figure 1. Biological sex was extracted from the medical records and summarized in the baseline characteristics. Gender identity was not available in the deidentified medical records and therefore could not be analyzed.
The study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committees of the First Affiliated Hospital with Nanjing Medical University (2025-SR-268), Suqian First Hospital (2025-SR-0205), and Third Affiliated Hospital of Soochow University (F-IRB-SOP-00710). Owing to the retrospective design and use of deidentified data, the requirement for written informed consent was waived. The study is registered at ClinicalTrials.gov (NCT06990126).
Method details
Data collection
Data for all variables evaluated in this study were extracted from patients’ electronic medical records and hospital laboratory information systems. Demographic variables included age, sex, height, weight, body mass index (BMI), and history of alcohol use. Disease-related variables included the admitting department, comorbidities, site of infection, microbiological etiology, and invasive procedures performed during hospitalization. Treatment-related variables included TGC loading dose, maintenance dose, duration of therapy, and the number of concomitant hepatotoxic medications, defined as drugs with a reported incidence of liver injury greater than 1% in the prescribing information.7 Baseline hepatic impairment was defined as any baseline ALT, alkaline phosphatase (ALP), or total bilirubin (TBIL) value above the upper limit of normal. High-dose TGC was defined as a maintenance regimen of at least 100 mg every 12 hours. Laboratory variables included complete blood counts, serum liver and renal function tests, coagulation parameters, and inflammatory markers.
Definition of TGC-associated DILI
Suspected TGC-associated DILI was identified according to the EASL clinical practice guidelines.15 Patients were classified as having suspected TGC-associated DILI if, after initiation of TGC, any of the following liver biochemical thresholds were met: (1) alanine aminotransferase (ALT) ≥5 × the upper limit of normal (ULN); (2) alkaline phosphatase (ALP) ≥2 × ULN, particularly when accompanied by an increase in γ-glutamyl transpeptidase (GGT) in the absence of bone disease; or (3) ALT ≥3 × ULN together with total bilirubin (TBIL) ≥2 × ULN.
For all patients who met the biochemical criteria for suspected DILI, the R ratio was calculated on the day when these criteria were first fulfilled, using the following formula: R = [ALT/ALT ULN] / [ALP/ALP ULN]. In line with the Chinese DILI guideline, suspected cases were further classified into three patterns: hepatocellular injury (R ≥ 5), cholestatic injury (R ≤ 2), and mixed injury (2 < R < 5).
The causal relationship between TGC exposure and liver injury was evaluated using RUCAM. Final RUCAM scores were interpreted as follows: ≤0, excluded; 1–2, unlikely; 3–5, possible; 6–8, probable; and ≥9, highly probable. In this study, patients with a RUCAM score of at least 6 (probable or highly probable) were defined as having TGC-associated DILI. All other TGC-treated patients comprised the non-DILI control group.
Data processing
All records were cleaned and preprocessed before analysis. Free-text clinical entries were recoded into prespecified categorical or numeric variables to improve data quality and interpretability.
Within the development cohort, data were randomly split into a training set and an internal validation set in an 8:2 ratio. In addition, predictors with more than 20% missing values were excluded from further analyses to limit bias related to missingness. The proportion of missing data for each variable is summarized in Table S1. Missing values in the remaining variables were handled using multiple imputation, with 20 imputed datasets generated separately for the training set and the internal validation set.30 All variable selection procedures were performed only within the imputed training datasets.
Variable selection and sample size estimation
Based on the Riley sample size framework for binary prediction models and the common default assumption of an anticipated Nagelkerke R2 of 0.15 with shrinkage factor ≥0.90, the available number of events supported consideration of up to approximately 10 candidate predictor parameters.31 Grouped LASSO was used for initial predictor screening across the 20 imputed training datasets. The regularization penalty was determined by 5-fold cross-validation, with log loss used as the optimization metric. Following the one-standard-error rule, the largest penalty parameter with cross-validated log loss within one standard error of the minimum value was selected.28
To further reduce redundancy among variables, pairwise Spearman rank correlation analysis was performed for the variables selected by grouped LASSO. When highly correlated variables were identified, the predictor with stronger clinical interpretability was retained. Bootstrap stability selection29 was used to evaluate the reproducibility of variable selection. The selection procedure was repeated in 200 bootstrap resampling iterations, and the selection frequency of each candidate predictor was calculated as the proportion of iterations in which the predictor was selected. Predictors with a selection frequency >0.90 were defined as stable predictors and were included in the final model.
Model development, validation and performance comparison
All models were trained in the training set only. Because TGC-associated DILI was relatively rare in the training cohort, class imbalance was expected to reduce sensitivity for the minority class and distort model calibration. To mitigate this issue, we adopted a cost-sensitive learning strategy. Eight supervised ML algorithms were fitted to predict TGC-associated DILI: logistic regression (LR), DT, RF, k-nearest neighbors (KNN), GB, support vector machine (SVM), LightGBM, and XGBoost. For LR, DT, RF and SVM, alternative class-weight configurations were specified in hyperparameter grids (including ‘balanced’ schemes and manually up-weighted positive-class weights) so that misclassification of DILI cases incurred a higher penalty during model fitting. For LightGBM and XGBoost, the influence of the positive class was increased by tuning ‘scale_pos_weight’ to counteract the low event rate.
For each algorithm, hyperparameters were tuned separately in each imputed training dataset using stratified 5-fold cross-validation, with AUC as the primary optimization metric. The optimal hyperparameter setting identified in each imputed dataset was then used to refit the model on the full corresponding training dataset. The fitted models were applied to the matched imputed internal validation datasets, and patient-level predicted probabilities were averaged across imputations to obtain the final validation predictions. The locked modeling workflow was then applied to the external validation cohort to evaluate generalizability.
Ensemble models were developed to assess whether combining classifiers improved predictive performance. Candidate base learners were selected according to the pooled AUC of their training-set out-of-fold predictions across imputations. Models with a pooled out-of-fold AUC greater than 0.75 were included in ensemble analysis. For the stacking ensemble, out-of-fold predictions from selected base learners were generated within each imputed training dataset using stratified cross-validation and were used as meta-features for a logistic-regression meta-learner. For the soft-voting ensemble, predicted probabilities from selected base learners were combined by weighted averaging, with weights proportional to each learner's pooled training out-of-fold AUC. Ensemble predictions were averaged across imputations for validation.
Because the model was designed for screening, threshold selection emphasized sensitivity. The final locked threshold was determined from the cross-imputation mean of the training-set out-of-fold predicted probabilities, using the maximum F3 score as the selection criterion. This threshold was fixed before validation and applied unchanged in both the internal and external validation cohorts. The overall, ICU-specific, and non-ICU-specific thresholds were derived and evaluated as supplementary thresholding strategies.
Model discrimination was assessed by AUC in the internal and external validation cohorts. Differences in AUC between the best-performing single model and the two ensemble models were compared using DeLong’s test. At the locked threshold, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were calculated. Calibration was evaluated using calibration plots, the Brier score, calibration intercept, and calibration slope. Clinical utility was assessed using decision curve analysis by estimating net benefit across a range of clinically relevant threshold probabilities.
False-positive prediction analysis
To characterize false-positive predictions, patients without TGC-associated DILI in the external validation cohort were classified as false-positive or true-negative cases according to the locked threshold of 0.153. The final model predictors were assessed a priori. To explore additional clinical features associated with false-positive classification, variables with an unadjusted p < 0.20 in preliminary comparisons were also examined. Continuous variables were compared using the Mann-Whitney U test, and categorical variables were compared using the chi-square test or Fisher's exact test, as appropriate. Multiplicity was controlled using the Benjamini-Hochberg FDR procedure.
Model interpretation and visualization
SHAP analysis was used to interpret the final model and improve model transparency. For each imputation-specific fitted final model, SHAP values were calculated in the internal validation cohort. To generate standard SHAP visualizations, we selected the imputed model whose internal-validation AUC was closest to the mean AUC across imputations as the representative model. SHAP bar plots, beeswarm plots, and dependence plots were then generated based on this representative model. Local SHAP waterfall plots were also created for representative high-risk and low-risk patients to illustrate individualized risk profiles and the contribution of each predictor to the model prediction.
Webpage deployment tool based on streamlit framework
To facilitate use of the model in clinical practice, the final prediction model was deployed as a web-based risk calculator using the streamlit framework. After entering the values of all required predictors, the application returns an individualized estimate of the probability of TGC-associated DILI.
Quantification and statistical analysis
Statistical analyses were performed using SPSS (version 27.0), R (version 4.5.0) and Python (version 3.10.4). Detailed software versions and random seeds for multiple imputation, cohort splitting, grouped LASSO, bootstrap stability selection, and model development are reported in Table S9. Categorical variables were summarized as counts and percentages [n (%)], and continuous variables were presented as mean ± standard deviation (SD) or median with interquartile range (IQR), depending on their distribution. Between group comparisons of continuous variables were conducted using the independent samples t-test for approximately normally distributed data and the Mann-Whitney U test for skewed data. Categorical variables were compared using the chi-squared test or Fisher’s exact test, as appropriate.
All statistical tests were two-sided, and p < 0.05 were considered statistically significant. For key performance measures, including AUC, 95% CIs were estimated using a nonparametric bootstrap with 1,000 resamples.
Footnotes
Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.116876.
Contributor Information
Qian Qian, Email: honeyhoney2007@163.com.
Wenkui Sun, Email: sunwenkui@njmu.edu.cn.
Supplemental information
References
- 1.Macesic N., Uhlemann A.-C., Peleg A.Y. Multidrug-resistant Gram-negative bacterial infections. Lancet. 2025;405:257–272. doi: 10.1016/S0140-6736(24)02081-6. [DOI] [PubMed] [Google Scholar]
- 2.Tasina E., Haidich A.-B., Kokkali S., Arvanitidou M. Efficacy and safety of tigecycline for the treatment of infectious diseases: a meta-analysis. Lancet Infect. Dis. 2011;11:834–844. doi: 10.1016/S1473-3099(11)70177-3. [DOI] [PubMed] [Google Scholar]
- 3.Moghnieh R.A., Abdallah D.I., Fawaz I.A., Hamandi T., Kassem M., El-Rajab N., Jisr T., Mugharbil A., Droubi N., Al Tabah S., et al. Prescription patterns for tigecycline in severely ill patients for non-FDA approved indications in a developing country: a compromised outcome. Front. Microbiol. 2017;8 doi: 10.3389/fmicb.2017.00497. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Giamarellou H., Poulakou G. Pharmacokinetic and pharmacodynamic evaluation of tigecycline. Expert Opin. Drug Metab. Toxicol. 2011;7:1459–1470. doi: 10.1517/17425255.2011.623126. [DOI] [PubMed] [Google Scholar]
- 5.Zuo C., Xu Q., Li X., Lyu Q., Shi X. Risk of liver function abnormalities with tigecycline: a retrospective observational study. Chin. J. Hosp. Pharm. 2024;44:2146–2153. doi: 10.13286/j.1001-5213.2024.18.12. [DOI] [Google Scholar]
- 6.Eguale T., Buckeridge D.L., Verma A., Winslade N.E., Benedetti A., Hanley J.A., Tamblyn R. Association of off-label drug use and adverse drug events in an adult population. JAMA Intern. Med. 2016;176:55–63. doi: 10.1001/jamainternmed.2015.6058. [DOI] [PubMed] [Google Scholar]
- 7.Shi X., Zuo C., Yu L., Lao D., Li X., Xu Q., Lv Q. Real-world data of tigecycline-associated drug-induced liver injury among patients in China: a 3-year retrospective study as assessed by the updated Roussel Uclaf Causality Assessment Method. Front. Pharmacol. 2021;12 doi: 10.3389/fphar.2021.761167. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Yu Z., Zhao Y., Jin J., Zhu J., Yu L., Han G. Prevalence and risk factors of tigecycline-induced liver injury: a multicenter retrospective study. Int. J. Infect. Dis. 2022;120:59–64. doi: 10.1016/j.ijid.2022.04.024. [DOI] [PubMed] [Google Scholar]
- 9.Jiang T., Huang X., Liu Q., Feng H., Huang Y., Lin J., Huang L., Chen S., Zhuang Y., Weng C. Risk factors for tigecycline-associated hepatotoxicity in patients in the intensive care units of 2 tertiary hospitals: a retrospective study. J. Clin. Pharmacol. 2022;62:1426–1434. doi: 10.1002/jcph.2099. [DOI] [PubMed] [Google Scholar]
- 10.Xiong Y., Liu G., Tang X., Xia B., Yu Y., Fan G. Prediction and analysis of toxic and side effects of tigecycline based on deep learning. Front. Microbiol. 2024;15 doi: 10.3389/fmicb.2024.1512091. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Torgersen J., Mezochow A.K., Newcomb C.W., Carbonari D.M., Hennessy S., Rentsch C.T., Park L.S., Tate J.P., Bräu N., Bhattacharya D., et al. Severe acute liver injury after hepatotoxic medication initiation in real-world data. JAMA Intern. Med. 2024;184:943–952. doi: 10.1001/jamainternmed.2024.1836. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Yang R., Jin Z., Jin X., Li Y., Wang D., Chao J. Advances in antituberculosis drug-induced liver injury: molecular mechanisms, genetic insights, and emerging animal-free prediction models. Toxicology. 2026;520 doi: 10.1016/j.tox.2025.154359. [DOI] [PubMed] [Google Scholar]
- 13.Azodi C.B., Tang J., Shiu S.-H. Opening the black box: interpretable machine learning for geneticists. Trends Genet. 2020;36:442–455. doi: 10.1016/j.tig.2020.03.005. [DOI] [PubMed] [Google Scholar]
- 14.Hoghooghi Esfahani H., Toyonaga S., Oyibo K. The application of explainable artificial intelligence in the prediction, diagnoses, treatment, and management of chronic diseases: a systematic review. Digit. Health. 2025;11 doi: 10.1177/20552076251355669. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Andrade R.J., Aithal G.P., Björnsson E.S., Kaplowitz N., Kullak-Ublick G.A., Larrey D., Karlsen T.H. EASL clinical practice guidelines: drug-induced liver injury. J. Hepatol. 2019;70:1222–1261. doi: 10.1016/j.jhep.2019.02.014. [DOI] [PubMed] [Google Scholar]
- 16.Zhang X., Li L., Zhang C., Zhang H., Huang H. A nomogram for predicting the risk of tigecycline-associated drug-induced liver injury in a Chinese population. Sci. Rep. 2025;15 doi: 10.1038/s41598-025-07116-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Wu S., Chen Y., Fan W., Wu X., Zhang C., Lin Y., Lin Q. Predicting tigecycline-related adverse events in infected patients: a machine learning approach with clinical interpretability. Front. Pharmacol. 2025;16 doi: 10.3389/fphar.2025.1697929. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Ali Z.A., Abduljabbar Z.H., Tahir H.A., Sallow A.B., Almufti S.M. Exploring the power of eXtreme Gradient Boosting algorithm in machine learning: a review. Acad. J. Nawroz Univ. 2023;12:320–334. doi: 10.25007/ajnu.v12n2a1612. [DOI] [Google Scholar]
- 19.Bentéjac C., Csörgő A., Martínez-Muñoz G. A comparative analysis of gradient boosting algorithms. Artif. Intell. Rev. 2021;54:1937–1967. doi: 10.1007/s10462-020-09896-5. [DOI] [Google Scholar]
- 20.Assel M., Vickers A. The F score ranks diagnostic tests and prediction models inconsistently with their clinical utility. Diagn. Progn. Res. 2025;9:30. doi: 10.1186/s41512-025-00214-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Koch D.T., Yu H., Beirith I., Schirren M., Drefs M., Liu Y., Knoblauch M., Koliogiannis D., Sheng W., De Toni E.N., et al. Tigecycline causes loss of cell viability mediated by mitochondrial OXPHOS and RAC1 in hepatocellular carcinoma cells. J. Transl. Med. 2023;21:876. doi: 10.1186/s12967-023-04615-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Yang X., Jin L., Luo X., An S., Wang M., Zhu H., Zhou Y., Liu H. Pharmacokinetic/pharmacodynamic target attainment of tigecycline in patients with hepatic impairment in a real-world setting. Ther. Drug Monit. 2023;45:786–791. doi: 10.1097/FTD.0000000000001115. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Lu W.-J., Lin K.-H., Tseng M.-F., Yuan K.C., Huang H.C., Sheu J.R., Chen R.J. New therapeutic strategy of hinokitiol in haemorrhagic shock-induced liver injury. J. Cell Mol. Med. 2019;23:1723–1734. doi: 10.1111/jcmm.14070. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Meagher A.K., Ambrose P.G., Grasela T.H., Ellis-Grosse E.J. The pharmacokinetic and pharmacodynamic profile of tigecycline. Clin. Infect. Dis. 2005;41:S333–S340. doi: 10.1086/431674. [DOI] [PubMed] [Google Scholar]
- 25.Danan G., Teschke R. RUCAM in drug and herb induced liver injury: the update. Int. J. Mol. Sci. 2015;17:14. doi: 10.3390/ijms17010014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Gong J., Su D., Shang J., Yu H., Du G., Lin Y., Sun Z., Liu G. Efficacy and safety of high-dose tigecycline for the treatment of infectious diseases. Medicine (Baltim.) 2019;98 doi: 10.1097/MD.0000000000017091. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Bai X.-R., Wang Z.-Z., Li W.-C., Wang Y.-G., Lou R., Qu X., Fan L., Zhang W., Wu Y.-C., Yan S.-Y., Zhang L. Clinical efficacy and safety of tigecycline based on therapeutic drug monitoring for carbapenem-resistant Gram-negative bacterium pneumonia in intensive care units. BMC Infect. Dis. 2023;23 doi: 10.1186/s12879-023-08815-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Du J., Boss J., Han P., Beesley L.J., Kleinsasser M., Goutman S.A., Batterman S., Feldman E.L., Mukherjee B. Variable selection with multiply-imputed datasets: choosing between stacked and grouped methods. J. Comput. Graph Stat. 2022;31:1063–1075. doi: 10.1080/10618600.2022.2035739. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Meinshausen N., Bühlmann P. Stability selection. J. R. Stat. Soc. Series B Stat. Methodol. 2010;72:417–473. doi: 10.1111/j.1467-9868.2010.00740.x. [DOI] [Google Scholar]
- 30.Pedersen A.B., Mikkelsen E.M., Cronin-Fenton D., Kristensen N.R., Pham T.M., Pedersen L., Petersen I. Missing data and multiple imputation in clinical epidemiological research. Clin. Epidemiol. 2017;9:157–166. doi: 10.2147/CLEP.S129785. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Riley R.D., Ensor J., Snell K.I.E., Harrell F.E., Martin G.P., Reitsma J.B., Moons K.G.M., Collins G., van Smeden M. Calculating the sample size required for developing a clinical prediction model. BMJ. 2020;368 doi: 10.1136/bmj.m441. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
-
•
The de-identified clinical data reported in this study are publicly available at Mendeley Data: https://doi.org/10.17632/8fmyrxmnkd.1.
-
•
This paper does not report original code.
-
•
Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.