Radiomics‐based classification models for HBV‐related cirrhotic patients with covert hepatic encephalopathy

Abstract Introduction The significant abnormalities of precuneus (PC), which are associated with brain dysfunction, have been identified in cirrhotic patients with covert hepatic encephalopathy (CHE). The present study aimed to apply radiomics analysis to identify the significant radiomic features in PC and their subregions, combine with clinical risk factors, then build and evaluate the classification models for CHE diagnosis. Methods 106 HBV‐related cirrhotic patients (54 had current CHE and 52 had non‐CHE) underwent the three‐dimensional T1‐weighted imaging. For each participant, PC and their subregions were segmented and extracted a large number of radiomic features and then identified the features with significant discriminative power as the radiomics signature. The logistic regression analysis was employed to develop and evaluate the classification models, which are constructed using the radiomics signature and clinical risk factors. Results The classification model (R‐C model) achieved best diagnostic performance, which incorporated radiomics signature (4 radiomic features from right PC), venous blood ammonia, and the Child‐Pugh stage. And the area under the receiver operating characteristic curve values (AUC), sensitivity, specificity, and accuracy values were 0.926, 1.000, 0.765, and 0.848, in the testing set. Application of the radiomics nomogram in the testing set still showed a good predictive accuracy. Conclusions This study presented the radiomic features of the right PC, as a potential image marker of CHE. The radiomics nomogram that incorporates the radiomics signature and clinical risk factors may facilitate the individualized prediction of CHE.


| INTRODUC TI ON
Covert hepatic encephalopathy (CHE), which consists of minimal hepatic encephalopathy (MHE) and grade 1 West Haven Criteria hepatic encephalopathy (HE), is characterized by the presence of mild cognitive impairments particularly in attention, visuospatial coordination, executive functions, and psychomotor speed. It is estimated that the prevalence of CHE in patients with cirrhosis is 30%-85% (Ampuero et al., 2018;Ortiz et al., 2005;Vilstrup et al., 2014). In addition, CHE has significant impact on cirrhotic patients, including declined the work performance and/or increased the traffic violations and accidents (Labenz et al., 2018;Ortiz et al., 2005).
Furthermore, CHE is associated with increased progression to overt HE (OHE), which have a negative impact on patient's quality of life (Ampuero et al., 2018;Labenz et al., 2018;Ortiz et al., 2005) and cause a high mortality (Cui et al., 2018). Therefore, testing for CHE in patients with HBV-related cirrhosis is beneficial to the patients and it is recommended (Vilstrup et al., 2014). Due to its subtle clinical symptoms, diagnosis of CHE mainly depends on specialized neurophysiologic, computerized and paper-pencil tests such as psychometric hepatic encephalopathy score (PHES). These tests seem simple, but they all should be performed by experienced examiners, and the test results are usually affected by the patient's age and literacy (Ortiz et al., 2005;Vilstrup et al., 2014). Owing to the complexity of diagnostic strategies, the subjectivity of evaluations and the lack of sufficient attentions (Patidar & Bajaj, 2015), most CHE patients do not receive timely diagnosis and appropriate treatment and are faced with risk of accidents. Thus, validating a noninvasive and objective method of diagnosis would be beneficial for both selecting therapeutic strategies and prognosis in clinical practice. Now, a variety of magnetic resonance imaging (MRI) technology has been widely used in HE researches. MRI can identify abnormalities in brain structures and functions in patients with HE.
Many researches showed that patients with CHE or OHE had significant alterations in PC, including gray and white matter volume Iwasa et al., 2012;Montoliu et al., 2012;Wu et al., 2015), functional connectivity (Chen et al., 2013Yang et al., 2018), and diffusion properties Lin et al., 2012;Qi et al., 2012). In addition, some researchers found that the alterations of the volume of PC were correlated with the ammonia levels and the extent of cognitive impairment Montoliu et al., 2012;Wu et al., 2015). Therefore, the alterations of PC might be one of important neuropathological mechanisms of cognitive dysfunction and be relevant to the early diagnosis of CHE.
Radiomics is an auxiliary detection and diagnostic technique that converts medical images into high-dimensional mineable data. It is intended to develop decision support tools. Radiomic data and available clinical factors can increase the power of the decision support models, which may potentially improve diagnostic and predictive accuracy and evaluation of prognosis for disorders (Gillies et al., 2016).
Nowadays, radiomics analysis was not only applied in the tumor field such as classifying tumors (Aerts et al., 2014) and predicting their outcomes (Huynh et al., 2016), but also was used in the nontumor field of Alzheimer's disease (AD) Li et al., 2019), Parkinson's disease (Wu et al., 2019), attention-deficit hyperactivity disorder , and Autism Spectrum Disorder (Chaddad et al., 2017;Heinsfeld et al., 2018). Those studies had shown that objective and quantitative features could potentially provide a new approach to develop classifiers, which may facilitate the individualized diagnostic biomarkers. But currently, CHE has no research use radiomics analysis.
The aim of this study is to extract quantitative features from PC and their subregions and combine with clinical risk factors to develop and evaluate classification models for HBV-related cirrhotic CHE in a framework of radiomics analysis (Figure 1).

F I G U R E 1
The workflow of data processing 2 | ME THOD

| Participants
A total of 106 patients with HBV-related cirrhosis diagnosed by liver biopsy or clinical criteria were consecutively recruited from February 2018 to June 2019 and written informed consents were obtained from all participants, and this study was approved by the local ethics committee (ethics reference number: 2018-043).
Exclusion criteria was followings: (a) current overt HE or other neuropsychiatric disorders; (b) liver malignancy; (c) history of brain surgery; (d) alcohol abuse within 6 months prior to the study; (e) active infection; (f) recent (<4 weeks) gastrointestinal bleeding; (g) metabolic diseases or endocrine diseases (e.g. diabetes mellitus or and thyroid dysfunction); (h) history of taking psychotropic medications; (i) age ≤ 18 years or ≥75; and (j) MRI contraindications.

| Paper-pencil testing and diagnosis of CHE
The diagnosis of CHE was made according to the practice guideline of the 14th International Society for Hepatic Encephalopathy (Vilstrup et al., 2014). Patient who showed abnormal scores in the number connection test A (NCT-A) and digit symbol test (DST) was defined as CHE. The abnormal scores were defined as exceeding the reference value by two standard deviations, referred to the normal value of a domestic expert consensus of China (Xing, 2009

| Clinical staging and laboratory examinations
Data including the Child-Pugh stage (based on albumin, total serum bilirubin, prothrombin time, and ascites) and venous blood ammonia were obtained within one week prior to MRI scan to assess the severity of liver disease for each subject.

| MR imaging
MRI examinations were performed using a 3.0 T MR scanner Age (

| Segmentation of PC and their subregions
10 regions of interest (ROI) segmentation of the PC was carried out using the statistical parametric mapping (SPM8, www.fil.ion.

| Quantitative radiomic features calculate
The calculation of radiomic features were carried out by using inhouse MATLAB scripts (http://atlas.brain netome.org/) . 423 radiomic features were calculated from each ROI.
There are 10 ROIs, a total of 4,230 (423 × 10) radiomic features for the further analysis. The features included follows (1) 14 intensity features, which calculated from the histogram of voxel intensity, LASSO is a shrinkage and selection method for linear regression.
It minimizes the usual sum of squared errors, with a bound on the sum of the absolute values of the coefficients (Meng et al., 2020).

The optimization objective for LASSO is:
where x i is the i-th patient's feature vector, y i is the classification variable, is the weight vector of the linear model, and > 0 is a penalty term, which controls the value of shrinkage.
The selection step was embedded in a 10-fold cross-validation framework to obtain unbiased estimates of classification error and then chose the to get the minimum criteria (Varma & Simon, 2006).

| Classification models construction
Two classification models were developed to diagnose CHE. First, the logistic regression analysis was employed to construct the radiomics signature model (R model), which only use the radiomic features of the radiomics signature. Then, multivariable logistic regression analysis was used to develop another model (R-C model), which incorporated the radiomics score (Radscore) and clinical risk factors. A linear combination was applied on radiomics signature to get the Radscore of each subject. The formula is as follows: 0 represents a constant, and there 0 = −0.553, i is the coefficient of radiomic feature, x i is the value of the feature.
The diagnostic performance of the two models were assessed in the training set by using the area under the curve (AUC) of the receiver operating characteristic curve (ROC), and then, they were evaluated in the testing set. At the same time, Delong test was used to observe whether the models are under-fitting or over-fitting.

| Development of radiomics nomogram
To provide the clinician with a quantitative tool to predict individual probability of CHE, the radiomics nomogram was built on the basis of the model with the highest predictive efficiency. And the calibration curves of testing set and training set were plotted to assess the calibration of the radiomics nomogram. And the Hosmer-Lemeshow test was employed to assess the goodness of fit of calibration curve for radiomics nomogram (Huang et al., 2016).

| Correlation analysis
The correlations among cognitive test scores (NCT-A and DST), venous blood ammonia, the Child-Pugh stage, demographic characteristics (age, gender, and education level), and radiomic features were studied via Spearman correlation analysis.

| Demographic characteristics and paper-pencil testing
The demographics, neuropsychological tests, and biochemical parameters of training set and testing set are summarized in Table 1.
Compared with the nCHE, CHE spent more time to complete the NCT-A and had less correct number of DST (p < .001). And the venous blood ammonia of CHE was significantly higher than that of nCHE (p < .001). More CHE patients had high-level Child-Pugh stage both in training and testing set (p < .001). There were no significant differences in age, gender and education level between the CHE and nCHE (p > .05) in the training set and testing set.

| Classification models
The results showed that the diagnostic performance of R-C model was superior than the R model, as shown in Table 2. The R-C model had higher AUC (0.926 95% CI, 0.780-0.988) and specificity (0.765) and with the same accuracy (0.848) and sensitivity (1.000) than the R model, in the testing set. The AUC, accuracy, sensitivity, and specificity in testing set of the R model were 0.846 (95% CI, 0.678-s0.947), 0.848, 1.000, and 0.706, respectively. The ROC curve of models is shown in Figure 3a,b, and the coefficients value of Radscore coefficient for each radiomic feature is shown in Figure 4.

| The radiomics nomogram and the calibration curve
The R-C model with the highest predictive efficiency was developed and presented as the radiomics nomogram ( Figure 5). The calibration curve for the radiomics nomogram was tested by the Hosmer-Lemeshow test, and the results showed no significant difference between the calibration curves and a perfect fit for predicting CHE, whether in the training set (p = .850) or the testing set (p = .475) ( Figure 6).

| Correlations analysis
Spearman correlation analysis suggested that all the 4 radiomic fea-

| D ISCUSS I ON
In our study, we identified significantly different radiomic features in PC between CHE and nCHE. After LASSO, we finally found 4 radiomic features including Variance, Median, GLN, and IMC1 which showed significant differences in PC of CHE when contrasted to nCHE. PC must have great changes in CHE. As a result of liver dysfunction, subsequently concentrations of the ammonia, reactive oxygen, and nitrogen, etc. rise in the blood. Those chemicals cross the blood-brain barrier, then effect on many signal transduction pathways (Wang et al., 2015) and trigger astrocyte swelling (Mínguez et al., 2006) and even cellular senescence (Görg et al., 2014). in cirrhotic patients without OHE (Chen, Lin, et al., 2017;Wu et al., 2015) as well. PC is a major association area, which has wide-spread connectivity with both cortical and subcortical structures. It was proven to be a critical area with multimodal and integrative functions including consciousness, visuospatial imagery, episodic memory retrieval, and self-processing operations (Cavanna & Trimble, 2006;Margulies et al., 2009). The underlying pathological changes of PC may cause impaired psychomotor speed, visual scanning efficiency, attention, and other functions, that is why patients with CHE spend more time to complete NCT-A. And the dysfunction of cognitive processing speed, visual perception, and working memory (Weissenborn, 2008) lead to the lower DST in CHE.
More interesting, most of our significant radiomic features of CHE were from the right PC, and this kept in line with previous studies revealing right PC was seem to have more obvious changes than the left side (Montoliu et al., 2012;Qi et al., 2013;Wu et al., 2015).
Right PC recall memories more (Freton et al., 2013) and have more prominent characteristics about people's social interactions (Petrini et al., 2014). This can explain CHE patients have declined work performance (Labenz et al., 2018;Ortiz et al., 2005) and life quality (Ampuero et al., 2018;Labenz et al., 2018;Ortiz et al., 2005). We strongly believe abnormalities of PC especially the right one can provide a new potential image marker for CHE.
For the radiomics model, previous CHE studies Wu et al., 2015) only revealed the abnormalities of PC but they did not construct a classifier to prove its diagnostic power. Radiomics classifier was proven to be a powerful diagnostic tool; it successfully verified the AD by radiomic features of hippocampus  and corpus callosum (Feng, Chen, et al., 2018). One recent radiomics study used hippocampus to recognize autism spectrum disorder (Chaddad et al., 2017)

| CON CLUS ION
In conclusion, our results highlight the importance of radiomic features of PC subregions, especial the right PC; this can be regarded as a potential image marker of CHE. The radiomics nomogram that incorporates the radiomics signature and clinical risk factors may facilitate the individualized diagnosis, which can be conveniently used to identify the cirrhotic patients with CHE.

This study was financially supported by the Medical Research Plan
Project of Chongqing Health Commission (No. 2017MSXM030). The authors would like to thank the GE Healthcare Life Sciences for supporting this study.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

AUTH O R CO NTR I B UTI O N S
Wei Zhang and Wei-jia Zhong contributed equally to study concept and design, revising manuscript, and study supervision. Sha Luo and Zhi-Ming Zhou were equally responsible for data collection, data F I G U R E 6 Calibration curves of radiomics nomogram-based prediction in the training set (a) and testing set (b). The calibration curves represent the calibration of the nomogram based on agreement between the predicted risk of CHE and actual CHE. A close fit between the dotted and solid lines indicates good predictive accuracy of the nomogram analysis, drafting the manuscript, and study design. Da-Jing Guo, Chuan-Ming Li, and Huan Liu contributed to the statistical analysis and revising manuscript. Wei Zhang obtained the funding to support this work. Xiao-Jia Wu, Shuang Liang, Xiao-yan Zhao, Ting Chen, Dong Sun, and Xin-Lin Shi involved in MRI and clinical data acquisition. All authors approved the final version of the manuscript.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/brb3.1970.

DATA AVA I L A B I L I T Y S TAT E M E N T
F I G U R E 7 Graph shows the correlations analysis. (a), the correlations analysis of the radiomic features, venous blood ammonia, Child-Pugh stage, and paper-pencil testing (NCT-A and DST) of all the cirrhotic patients; (b), positive correlations between the typical feature (DR241, right Lc1_Median HLL) and DST score; (c), negative correlations between the typical feature (DR241, right Lc1_Median HLL) and NCT-A score