Drug-induced liver injury (DILI) is a potential and usually unpredictable response that may occur due to the use of many prescribed medicines, herbal products or diet supplements. Previous research on DILI has not been clear enough, and new drugs are constantly being used in clinical practice. Although DILI is generally rare (around 0.3–0.33 ‰ in the healthcare system) (1–3), it is a potential life-threatening adverse drug reaction that accounts for over 50% of acute liver failure cases in Western countries (4). Therefore, DILI has become a global new issue. During long-term follow-up, it was found that a small number of cases may progress to chronic liver injury, which was manifested as asymptomatic elevated liver biochemistry, aggravated chronic hepatitis, and even cryptogenic cirrhosis (5–7). In the advanced stage of DILI, patients have a higher risk of extrahepatic metabolic disorders and complications, so early diagnosis of DILI is of great significance.

The diagnosis of DILI in clinical practice is mainly based on comprehensive analysis of detailed medical history, clinical symptoms, clinical serum biochemical indicators and so on. When patients diagnosed with liver injury suspect a history of drug exposure, non-drug related liver disease should be excluded first. Considering these factors, it is recommended to use the Roussel Uclaf Causal Relationship Assessment Method (RUCAM) scale for scoring and evaluating the possibility of a diagnosis of DILI, but it is still unable to make a definite diagnosis. Liver biopsy (LB) is considered as the “gold standard” for diagnosis, differential diagnosis, and evaluation of liver inflammation and fibrosis. However, this method is an invasive procedure with potential complications and sampling errors (8). More importantly, as an invasive method, it cannot guarantee the dynamic observation of liver fibrosis and inflammation. So far, there is a lack of specific biomarkers for evaluating the severity of inflammation and fibrosis in DILI. While in recent years, researches on non-invasive diagnostic indicators of liver inflammation and fibrosis have attracted attentions. Non-invasive liver fibrosis models composed of simple parameters have been developed to predict liver fibrosis in patients with viral hepatitis, such as fibrosis index based on fibrosis-4 index (FIB-4), aspartate transaminase (AST)-to-platelet ratio index (APRI), gamma-glutamyl transpeptidase-to-platelet ratio (GPR) (9–14). APRI and FIB-4 are recommended methods to determine liver fibrosis and cirrhosis in the World Health Organization (WHO) chronic hepatitis B (CHB) guidelines (15). GPR has been found to be more accurate than APRI and Fib-4 in evaluating liver fibrosis in patients with CHB virus infection (14). Recently, a growing number of studies indicate that a new model, fibrosis-5 index (FIB-5), which combines Alb, platelet count (PLT), alkaline phosphatase (ALP), AST and alanine aminotransferase (ALT) values, is superior to FIB-4 for diagnosing significant hepatic fibrosis in patients with chronic HBV and HCV infections (16, 17). Moreover, previous studies have suggested that GPR and APRI can also be used to predict the severity of liver inflammation in CHB patients (18–20). In most liver diseases caused by various etiologies, liver fibrosis will develop after sustained chronic inflammatory damage to the liver parenchyma. Given that acute DILI lasting for more than 6 months may progress to chronic, which is similar to the disease progression of hepatitis B or C (7), it can be reasonably inferred that these serological markers that can predict liver fibrosis and inflammation in chronic hepatitis B or C infection may also be used for DILI. However, research in this area is very limited up to now. Therefore, in this study, we elucidated the high-risk factors for DILI and developed non-invasive column chart models to predict the histological stages of inflammation and fibrosis. In addition, we also compared the diagnostic accuracy of our new models with GPR, APRI, FIB-4, and FIB-5 to determine the optimal predictive model.

DILI is the most common cause of acute liver failure in Western countries (4), so accurately assessing the severity of liver disease in DILI patients is crucial. LB is the gold standard for evaluating liver diseases. However, LB has some drawbacks, such as invasiveness, high cost, potential life-threatening complications, possible sampling errors, and inconsistent abilities and experience among pathologists. Therefore, several models were proposed for predicting liver inflammation and fibrosis. At present, some prediction models constructed using conventional biochemical indicators are mainly used to predict liver fibrosis, and few models are used to predict liver inflammation. Most models seem to perform well in chronic viral hepatitis, but there are rarely applied to DILI patients.

APRI and FIB-4 have been recommended in assessing hepatic fibrosis by CHB consensus guidelines (15, 25). Several studies revealed that the AUCs of APRI for predicting severe fibrosis and cirrhosis in patients with chronic hepatitis C were 0.77 to 0.88 and 0.83 to 0.94, respectively (16, 26). However, in our study on DILI, the AUC of APRI in predicting moderate–severe fibrosis was 0.667, which was lower than previous research. This may be due to the fact that APRI tends to show more accurate results in cirrhosis, but the patients included in our study did not show DILI induced cirrhosis in pathological results. A study of Shoaei et al. (18) on patients with CHB demonstrated that the AUC of APRI for predicting liver inflammation was 0.66, which was lower than ours (0.778). The possible reasons are as follows: First, the criteria of liver pathology score between these two studies was different. In previous studies, Shoaei SD et al. used the Knodell histological activity index to assess the level of necrotic inflammation. While in our study, the Scheuer scoring criteria was applied. Second, mean AST of moderate–severe inflammation group in our study (623 U/L) was higher than that in the study above (97.82 U/L). Due to the limited number of studies reporting the application of APRI in predicting liver inflammation, these results need to be confirmed through large-scale researches.

There have been various conclusions in previous studies regarding the predictive efficacy of different models in diagnosing liver cirrhosis. The AUC of APRI was slightly higher than that of FIB-4 for evaluating cirrhosis, while there was no difference for fibrosis (27). European Association for the Study of the Liver (EASL)-Asociación latin oamericana para el estudio del hígado (ALEH) clinical practice guidelines showed that FIB-4 owned a good accuracy for diagnosing advanced fibrosis. Studies showed that GPR was a more accurate model than APRI and FIB-4 in predicting different fibrosis levels of CHB patients (14, 28). Shiha et al. (17) studied 604 patients with CHB and found that FIB-5 was superior to FIB-4 in distinguishing between significant and non-significant fibrosis. Similarity, Metwally et al. (16) concluded that FIB-5 was more specific than FIB-4 in evaluating hepatic fibrosis in CHB patients. In our study, we reached similar conclusions. We found that the AUC value of FIB-5 was higher than that of GPR, APRI and FIB-4 in the diagnosing moderate–severe inflammation (G2–G4) and fibrosis (S2–S4) for DILI.

In this study, backward stepwise regression model, best subset model and logistic regression model for predicting significant liver inflammation and fibrosis were created with the available laboratory parameters from DILI. Nomograms were also built by the selected independent variables to predict the risk of moderate–severe inflammation and fibrosis in DILI. The four factors including RDW, PLT, Hb and age built the backward stepwise regression model to predict severe fibrosis. Previous studies indicated that RDW levels were positively relevant to the liver disease severity with various etiologies (29–31), and lower Hb, higher bilirubin and thrombocytopenia are independent predictors of poor outcomes (32, 33), which is consistent with our findings. To compare the clinical usability of these models, Brier score, calibration curve and DCA were performed. By testing the above methods, we found that the model constructed by backward stepwise regression is very effective in predicting moderate–severe inflammation and fibrosis in DILI. Compared with other traditional criteria, such as GPR, APRI, FIB-4 and FIB-5, backward stepwise regression model owns extraordinary predictive power with the AUC more than 0.85. To our knowledge, it is the first time to perform models constructed by various methods and compared them with the existing models to predict significant liver inflammation and fibrosis for DILI. Considering the accuracy range of LB, an AUC > 0.90 cannot be achieved even for a perfect biomarker (34). EASL-ALEH clinical practice guidelines has reported that the AUC below 0.80 is generally considered too imprecise and of no value in clinical practice (35). Our nomogram models are performed by combining several clinical characteristics and have been proved to be highly effective in predicting severe liver inflammation and fibrosis with AUC values of 0.856 and 0.889, respectively. In addition, our nomogram models also have some advantages when compared with other traditional criteria. First of all, nomogram is a user-friendly method that guides clinical decision-making for doctors, as they can transform statistical prediction models into single numerical estimates of event probabilities. By summing the scores of different predictive variables, clinicians can quickly estimate an individual risk of patients without complex calculations. Nomogram integrates multiple clinical variables (e.g., general information, laboratory results), providing a more precise prediction compared to single-factor assessments. Secondly, our models are more accurate to predict severity of liver pathology of patients compared to traditional models.

There are also several limitations in our study. Firstly, potential selection bias is inevitable as our analysis is a retrospective analysis with a single center study. Therefore, the conclusions drawn from this study need to be validated through large-scale prospective studies. Secondly, we did not investigate the previous medication use, which may affect the AST values included in this study. Therefore, similar potential confounding factors should be minimized as much as possible in future large-scale analysis. Thirdly, due to the relatively small sample size, the diagnostic efficiency of the model cannot be verified through the test set, so the predicting value of models need to be validated through large-scale research in the future.

However, our analysis has several highlights. Most studies on non-invasive liver models have been mainly focused on patients with chronic HBV and HCV infection, while little attention was focused on exploring predictive factors for inflammation and fibrosis grades in DILI patients. To our knowledge, this is the first study conducted in DILI patients aimed at determining the clinical usefulness of non-invasive serum markers based models of GPR, APRI, Fib-4, and Fib-5 in inflammation and fibrosis severity, and comparing them with liver biopsy results. Furthermore, the models constructed in our study were compared with GPR, APRI, FIB-4, and FIB-5 models to analyze the diagnostic effectiveness of predicting the severity of inflammation and fibrosis in DILI patients. In conclusion, our models will improve diagnostic accuracy and clinical management, and our findings will help clinicians identify DILI patients with potential risk of severe liver inflammation and fibrosis.

In summary, our study proposed clinical predicting models for evaluating liver inflammation and fibrosis in DILI. The results indicate that backward stepwise regression model possesses a better clinical application value. Our research can help assess the severity of liver inflammation and fibrosis in clinical practice, while more cohort studies are still needed to validate the accuracy of our predictive model.