A diagnostic model for autoimmune hepatitis based on noninvasive clinical data

Background ： All of these diagnostic criteria of AIH (autoimmune hepatitis) include histopathology which is significant for the final diagnosis of AIH patients. Nevertheless, some patients failed to perform the examination in time because of fear of the trauma risk of liver biopsy. We aim to develop an AIH diagnostic model without liver biopsy in adult and validate its performance. Methods ： We collected demographics, blood and liver histological data of ULI (unknown liver injury)patients in two independent adult cohorts. In the training cohort, we used logistic regression to develop a model and validated internally its performance by bootstraps methods. In the validation cohort, We operated the receiver operating characteristic (ROC) curves, decision curve analysis and calibration plots to evaluate model externally . Results ： We developed a nomogram to predict the risk of AIH using three risk factors, GGP (percentage of gamma globulin), FBG (fibrinogen) and AA (AIH-related autoantibodies). Area under curve (AUC) for the training and validation cohorts were 0.862 and 0.870 respectively. The calibration plots and decision curve analysis plot suggested that the model had an acceptable accuracy and a great clinical utility. Conclusions ： The diagnostic model we operated can be well used to predict of AIH for ULI patients without liver biopsy especially for patients who are afraid of liver puncture.


Introduction
Autoimmune hepatitis (AIH) is an inflammatory injury of the liver caused by abnormal autoimmunity 1 . It is generally believed that the incidence of AIH may be related to multiple genes, environmental factors and other factors, but it is still difficult to accurately quantify the risk factors 2, 3 .
In the early stage of AIH onset, the disease symptoms are often mild and may include only mild fatigue, poor stomach appetite, jaundice, nausea, skin itching, joint pain, etc. These symptoms sometimes relieve themselves and then relapse [4][5][6] . The condition can last for years without the patient knowing. As a result, some AIH patients with early manifestations of so-called unknown liver injury (ULI) may not receive a timely, accurate diagnosis and treatment and eventually develop severe diseases, such as cirrhosis or liver failure, when they first visit the doctor [7][8][9] .
AIH patients had a higher mortality risk than normal people, and the mortality hazard ratio was 5.76 (95% CI 4.66 to 7.11) in the first year after diagnosis. After this initial phase, the mortality hazard stabilized at approximately two 10 . Timely diagnosis and immunosuppressive therapy can bring great clinical benefits to patients with AIH, while untreated AIH has a 5-year mortality above 5% 11 . Early diagnosis may be different because the clinical symptoms are heterogeneous, and a single diagnosis test is not applicable for all patients. The international autoimmune hepatitis group (IAIHG) proposed and revised the AIH diagnostic criteria in 1993 and 1999, respectively 12,13 . Due to its complexity, the IAIHG simplified the scoring system in 2008 14,15 .
However, the median sensitivity of the simplified scoring system in the diagnosis of "clear" AIH (≥ 7 points) was approximately 75.5% (range 15% to 87%). To a large extent, it may omit some atypical patients, such as patients with low or negative autoantibody titers or low or even normal serum IgG levels 11 . All diagnostic criteria include histopathology, which is of great significance for the final diagnosis in AIH patients [16][17][24][25] . Nevertheless, it is unfortunate that some ULI patients always refuse to perform the examination because of fear of the invasiveness and unsafety of liver biopsy in clinical practice, which will undoubtedly further reduce the practical application value of the 2008 scoring criteria. Based on the above two deficiencies, we aimed to develop an AIH prediction nomogram model without liver biopsy in adults based on a small amount of noninvasive clinical data of patents with wellestablished diagnoses and validate its accuracy to provide a valuable supplement to the IAIHG simplified scoring system in 2008.

Study design and data source
This research used a cross-sectional study design. We collected the data from the electronic case system, and data were primarily collected retrospectively from December 27, 2005 to January 8, 2020 in three hospitals that belong to territory medical institutions. The study was conducted at the Shanghai Sixth People's Hospital in Xuhui district, Shuguang Hospital in Pudong district and Changzheng Hospital in Huangpu district. Data from the Shanghai Sixth People's Hospital were used to develop a nomogram model, and data from the other two hospitals were used for model validation.

Case selection and classification
Case inclusion criteria: All patients were diagnosed with ULI who were intended to be diagnosed with AIH and had liver biopsy for the first time.
Exclusion criteria: Age less than 18 years. The condition had been clinically diagnosed as viral hepatitis, alcoholic liver disease, drug-induced liver damage, biliary liver damage, or another clear basis for liver damage caused by tumors, cardiac insufficiency, genetic metabolism and other factors.
All cases were classified as AIH and nonautoimmune hepatitis (NAIH), according to the diagnostic criteria for autoimmune hepatitis developed by the international autoimmune hepatitis group in 1999.
With these criteria, post-treatment scores were taken into account, if possible. Therefore, we established a threshold of >15 in the pre-treatment group and >17 in the post-treatment group for a positive case.
Two conditions were considered necessary for those whose scores were likely to be AIH (10 to 15 score in the pre-treatment group and 12 to 17 in the post-treatment group): (1) treatment response was defined as symptom relief, transaminases and IgG normalization and inflammation of liver biopsy disappeared; and (2) liver histology describing features compatible with AIH (chronic hepatitis with lymphocytic infiltration), and if they satisfied one of two necessary conditions, they were also classified as AIH.

Data collection
For the collection of noninvasive clinical data and liver biopsy data, we mainly refer to the diagnostic criteria for autoimmune hepatitis developed by the international autoimmune hepatitis group in 1999 According to Peduzzi et al., the number of events per predictor should be 10 or more. As a result, we aimed to obtain at least 30 AIH patients in the training cohort by a rough estimate.

Study design and statistical analysis
All participants stratified by the training and validation cohort were presented as the means (standard deviations) or medians (interquartile ranges) for continuous variables and as frequencies (percentage) for categorical variables. First, we used univariate and multivariate logistic regression analysis to acquire the risk factors and their determination coefficients for AIH. In the model-development phase, we performed a best selection process and formulated a nomogram in the training cohort, according to the Akaike information criterion (AIC). Second, the nomogram was internally validated using the bootstrap method, which assesses how accurately the model will predict AIH in a new sample of participants. We operated the receiver operating characteristic (ROC) curves and calibration plots by training and validation cohorts to indirectly evaluate the discrimination and accuracy of the predictive model. Third, we calculated the optimal cutoff values of diagnosis using Youden's index (sensitivity+specificity-1) and directly presented the sensitivity and specificity to evaluate the model by validation cohort, compared with the 2008 scoring criteria. IBM© SPSS 20, Stata 15 (64-bit) and R x64 3.6.2 were used for all analyses; p<0.05 was considered significant; and all statistical tests were twotailed.

Study population and baseline characteristics
A total of 76 individuals were excluded because they fulfilled our exclusion criteria or had incomplete and missing data. Overall, 131 and 111 patients were included in the final analysis as the training and validation cohorts, respectively. All data came from the complete data. Table 1 shows the baseline characteristics of AIH and non-AIH patients in the training cohorts. Approximately 52 (40%) participants showed significant differences: patients with AIH showed a high-baseline age, GGP, FBG and AA. In the univariate logistic regression, on the basis of the p-value results, age (P=0.005), GGP (P=0.001), AA (P=0.001) and FBG (p=0.034) were significantly associated with a high risk of AIH.

Development of an AIH-predicting nomogram
According to the AIC, we performed a nomogram of the stepwise model in predicting the risk of AIH

Validation of the nomogram model
In the internal validation, we used a bootstrap method based on 500 replications to internally verify the model. The C-statistic is 0.862 (＞0.8), which means that the model is excellent. The calibration plots (Figure 2a) revealed that the estimated regression line fit well with the ideal line.
In the external validation, we calculated the predicted p-value, according to the model formula, and used the YOUDEN index (sensitivity+specificity-1) to determine the cutoff value in the training cohort, and the calculated cutoff value was 0.404. According to the cutoff value, the calibration plots ( Figure   2b) revealed that the predicted risks were slightly lower than the observed risks, with a brier score of 0.13, calibration intercept of -0.50, and calibration slope of 1.04. The AUCs of the ROC curve ( Figure   3a) were 0.870 (95% confidence interval 0.808 to 0.942), and the decision curve analysis (Figure 3b) showed good clinical usefulness.

Figure 2
The calibration plot curves of the model in the training cohort (a) and validation cohort (b). The main concern is the pvalue of the unreliability index, which was p>0.05 (p-a=0.1310. p-b=0.6463) for the two groups. It is shown that the null hypothesis is not rejected (H0: intercept=0, slope=1); that is, the fitted line coincides with the 45 degree line, which means the prediction is relatively accurate.

Figure3
The ROC curves (a) and decision curve analysis (b) of the model for AIH risk in the validation cohort. a: The AUC of each cohort is more than 0.8, which means that the model is excellent.b: The blue line represents the net benefit when all participants were considered as AIH, while the red line represents the net benefit when no participants were considered as AIH. The decision curves of the training cohort were mostly higher than those of the "treat none line" (red line) and the "treat all line" (blue line), which shows good clinical usefulness. ROC: receiver operating characteristic curves; AUC: area under the curve.

Discussion
In summary, we developed a nomogram to predict the risk of AIH using three risk factors (GGP, FBG and AA) through an analysis of the training cohort (AUC=0.862). The internal and external validation results showed predictive value as a remarkable tool. To the best of our knowledge, this is the first study to design a nomogram that can accurately predict the risk of AIH with noninvasive data. The intuitiveness of visualization, which is the merit of nomograms, helps to understand the risk factors for AIH in ULI patients. It may be useful for clinical reference.
Liver biopsy is of great significance for the final diagnosis in ULI patients. In fact, liver biopsy data are one of the key parameters in all three versions of the diagnostic criteria for AIH developed by the international autoimmune hepatitis group [12][13][14][15] . In particular, the simplified diagnostic criteria in 2008 have been widely used because of repeated and effective verification 1, 3-7 . Nevertheless, AIH can have an acute onset, and it is usually difficult to distinguish it from other acute liver injuries caused by viral hepatitis or drug-induced hepatitis 26 . The 2008 simplified scoring system of AIH is not suitable for immediate determination because it is based on data from patients with chronic hepatitis and does not take into account acute manifestations. Nevertheless, some ULI patients always refuse to perform the examination because of fear of invasiveness, which will undoubtedly further reduce the practical application value of the 2008 scoring criteria. In the validated population, we used our model and the 2008 simple diagnostic standard to predict AIH and calculate sensitivity and specificity. The results show that our model has little difference in sensitivity (85.7% vs. 89.3%). Moreover, our model was based on the concept of individualization, in which the continuous nature of risk factors was preserved to increase the degree of uniqueness of an individual. In future research, we aim to test our model in other population samples and design a random clinical trial to validate our model. If they could predict AIH with high accuracy so that liver biopsy, which is associated with invasive risk, can be avoided, it could be much more valuable clinically.
In our study, we found that a new indicator, fibrinogen (FBG), is associated with AIH. Clinically, we also observed that the FBG of most AIH patients is lower than that of normal people. A study showed that the process of fibrinogen-degradation products (FDP) had a negative effect on AIH diagnosis, which is roughly consistent with our paper's conclusion 18 . It is easy to understand that FBG represents the synthetic function of the liver. When patients with AIH have impaired synthetic function of the liver, FBG will decrease. Of course, more research is needed to verify this hypothesis. AIH can occur at any age according to its mechanism, but we found that age is a risk factor for AIH (p=0.005). According to some studies 11,[19][20] , the distribution of age at onset was thought to be bimodal, with peaks around puberty and between the fourth and sixth decades of life. These studies are consistent with our research results. Although age was not included in our model in the end, we suggest that patients of any age who may have autoimmune hepatitis need to be considered, but we are particularly concerned about patients between 40 and 60. Moreover, in the study, it was found that more female ULI patients were diagnosed with autoimmune hepatitis than men. However, there was no statistically significant difference between the two genders (p=0.071). Of course, our study's sample size is relatively small, and the conclusion should be further validated.
For autoantibodies, we chose ANA, SMA, LKM and SLA/LP as our analysis, according to the 1999 criteria of AIH. However, we did not stratify the antibody results into different levels because we considered that the current detection methods and criteria in many hospitals are not completely unified.
As noted in the "Brighton report", several other autoantibodies are of relevance to AIH, namely, those reacting with ASGP-R (the hepatic asialoglycoprotein receptor) 21,22 and LCI (a soluble liver antigen) 23 in hepatocyte plasma membranes, but those are still only available in a few special laboratories. We believe that significant differences may occur when an additional study sample size is accumulated.
In the 1999 criteria, immunoglobulin (protein electrophoresis was selected to detect 100 percent of globulin in this study) or IgG was an indicator of chronic immune response. Compared with IgG, in the single-factor analysis, GGP and IgG were related to AIH. We also found that there was a linear correlation between GGP and IgG, GGP = IgG * 0.752 + 9.646 (r = 0.674). Therefore, we converted the IgG values of 47 patients in Shuguang Hospital into GGP values for further analysis because the hospital's GGP data were not tested. In fact, many studies have confirmed the very close relationship between GGP and IgG, and one of them is often used to replace the meaning of the other 14 .
There are some problems in the research that we must mention. In the training cohort, for those whose scores are likely to be AIH (10 to 15 score in the pre-treatment group and 12 to 17 in the post-treatment group), we defined them as AIH if they satisfied one of two necessary conditions. As an additional confirmation of the diagnosis, not all patients' subsequent responses to immunosuppressive therapy were observed in this study. This design makes our statistical analysis easier; on the other hand, it makes the sensitivity and specificity data shown by our model somewhat conservative. In this study, although we showed some patients with overlap syndrome of AIH and PBC, we did not have a good method to distinguish AIH and PBC. In the training cohort, we found that 5 patients were AMA positive in 12 AIH and PBC overlap syndrome patients, 50% of patients were AMA positive in PBC patients, and pANCA appeared to be positive in AIH but negative in PBC. It is worth noting that some research shows that pANCA may be related to the identification of AIH and PBC [27][28] . Of course, how to distinguish them still needs further research. In addition, although the case data we collected covered a span of 15 years and involved three different hospitals, the low incidence of AIH and the very small proportion of AIH patients willing to receive liver biopsy resulted in only 242 cases enrolled. In fact, that is why we aim to operate another supplementary diagnostic method of AIH without liver biopsy. However, data heterogeneity needs to be further improved.
In conclusion, the new model we developed shows great risk predictive value for ULI patients who are likely to be diagnosed with AIH. Furthermore, our model can be used to perform more sophisticated data analysis on the general screening population to detect hidden autoimmune hepatitis patients earlier.
In any case, it is important to continue to train and validate our nomogram model for further optimization by conducting subsequent larger multicenter studies and external validation studies.

Ethics approval and consent to participate
All patients undergoing liver biopsy were informed in advance of the relevant risks and purposes by the attending physician and signed the informed consent of this study. This study was approved by the Ethics Committees of Shanghai Sixth People's Hospital with approval number 2020-KY-032(K). And we can confirm that all methods were performed in accordance with relevant guideline and regulations.

Consent for publication
Informed consent for publication was obtained from all participants.