Machine learning-based system for prediction of ascites grades in patients with liver cirrhosis using laboratory and clinical data: design and implementation study

Introduction

Among all the leading causes of mortality, chronic liver diseases have increased in the past two decades worldwide [1]. Chronic liver diseases may lead to liver cirrhosis, in which healthy liver cells are transformed into fibrous scar tissue [2, 3]. It represents an irreversible histological change for a variety of chronic liver diseases [4, 5]. The major consequences of cirrhosis include gastrointestinal varices, ascites, variceal hemorrhage, and hepatic encephalopathy [6, 7]. Ascites are one of the most important complications of cirrhosis [8], which may occur in 50% of patients with cirrhosis after 10 years of follow-up [9]. Ascites is a serious turning point in the natural history of cirrhosis because it is associated with two years mortality of 50% [5, 8]. that may need to consider liver transplantation as a treatment option [8]. There have been several changes in the clinical management of ascites cirrhosis in recent years. Although the development of ascites usually indicates advanced liver disease, the clinical history of patients with cirrhosis and ascites are widely different [10]. Ascites often have been classified into three grades, consisting Grade 1 which shows mild ascites only detectable by ultrasound examination, Grade 2 which is moderate ascites and Grade 3 which shows the severe ascites with marked abdominal distension [11, 12]. All cirrhotic patients are recommended to be examined for the incidence of ascites [11]. Patients with cirrhosis who show ascites often are more likely to be considered for liver transplantation [11]; therefore, the precise prognostic models for the ascites prediction are necessary [5].

Machine learning is recognized as a major part of health care research and has played a key role in the treatment and management of diseases [13]. Machine learning can significantly contribute to the disease’s diagnosis and predictions [14]. By learning the procedure, the internal parameters can be tuned up to achieve the classification model with high performance [15]. Therefore, machine learning as an accurate and non-invasive approach with high accuracy to identify the grades of ascites in patients with liver cirrhosis seems useful. Several studies have been conducted to predict ascites in patients with cirrhosis some of them used statistical methods [16] which have less power to deal with large and complex nonlinear data [17], or used the presence of ascites as a key prognostic factor to predict other consequences of cirrhosis [18, 19]. Several studies used machine learning as an approach to predict the survival of cirrhosis [20] or to predict other consequences of cirrhosis [21], [22], [23]. As a result, effective and high-performance approaches are required to predict the grades of ascites in patients with cirrhosis. These approaches include the use of data mining and machine learning techniques.

Therefore, in the present study, we used modern machine learning methods to develop a prediction model solely based on routine laboratory and clinical data to help physicians accurately diagnose and predict different degrees of ascites. This machine learning-based approach can help to avoid unnecessary risks and costs for patients with acute stages of the disease.

Materials and methods

The primary aim of the present study was to implement four machine learning models to predict the grades of ascites in patients with cirrhosis. For developing the present study, we used ANACONDA3–5.2.0 64 bit, free and open-source platform distribution of Python programming language with numerous modules, packages, and rich libraries that provide various methods for classification problems. For validating the best performance measures of the selected models, ROC curve, precision, recall and f1-score were compared for the methods. We employed four popular learning models on our dataset, including k-nearest neighbors (KNN), support vector machine (SVM), neural network and random forest (RF) classification algorithms.

Dataset

This study was based on the data obtained from the Research Institute for Gastroenterology and Liver Diseases, Shahid Beheshti University of Medical Sciences from 2016 to 2021. Table 1 represents the available variables in the study. The data were collected from 492 unrelated Iranian subjects who had been diagnosed with cirrhosis. Twenty variables for predicting ascites and comparing the performance of the applied techniques were considered. For the preprocessing phase, we normalized the data inputs for KNN, SVM, ANN and RF models. We also had low missing variables, and we replaced them with the average of each column. Also, the scales of the data were normalized to avoid the bias in the results. The target data comprised of non-ascites patients, patients with the grades that have been represented in Table 2 in each class of ascites.

Table 1:

Description of database characteristics.

No.	Variable	Mean, SD	Annotations	Description
1	Age	63.5 (15.76)		Demographic variable
2	Sex	Categorical	1 – Male 0 – Female	Demographic variable
3	Etiology	Categorical	1 – HBV (hepatitis B virus) 2 – HCV (hepatitis C virus) 3 – Alcohol 4 – NASH 5 – Autoimmune 6 – PBC (primary biliary cholangitis) 7 – PSC (primary sclerosing cholangitis) 8 – Metastasis 09 – Cryptogenic 10 – Wilson 11 – Autoimmune+PBC 12 – HBV+NASH 13 – Autoimmune+PSC 14 – Mylofibrosis 15 – Unknown	Etiology identifies the first cause of cirrhosis in patients
4	Ascites	Categorical	0 – No 1 – Mild-slight 2 – Moderate 3 – Severe	The degrees of ascites among patients, the target variable
5	SBP	Categorical	1 – Yes 0 – No	Spontaneous bacterial peritonitis or ascites fluid infection
6	Encephalopathy	Categorical	0 – No 1 – Grade 1–2 2 – Grade 3–4	Same as the medical term
7	Bleeding varice	Categorical	1 – Yes 0 – No	Same as the medical term
8	Band. Ligation	Categorical	1 – Yes 0 – No	Same as the medical term
9	Child_Pugh score	Categorical	5–14	Rates which are used to the severity of long-term liver diseases
10	Liver transplantation	Categorical	A B C	Same as the medical term
11	MELD.score	Categorical	6–40	A score used as the model for end-stage liver disease
12	WBC	435.94 (16.54)		White blood cell
13	Hb	2.57 (0.27)		Hemoglobin
14	MCV	3.36 (6.37)		Mean corpuscular volume
15	PLT	11.7 (42.8)		Platelets in liver disease
16	Bili.D	1.86 (1.23)		Conjugated bilirubin or direct bilirubin
17	AST	7.53 (32.9)		Aspartate aminotransferase
18	ALP	23.95 (101.82)		Alkaline phosphatase
19	PT	2.84 (7)		Prothrombin time
20	K	2.05 (0.97)		Vitamin K

Table 2:

Number of people with different degrees of ascites in the study.

Classes	Number
None	39
Ascites class A	160
Ascites class B	189
Ascites class C	97

Figure 1 demonstrates the correlation heatmap between the variables and the target variable. The highly correlated variables should not be included in the models because they may affect the model prediction, however, as the heatmap demonstrates there are no highly correlated features in the study, so all the variables are included in the study.

Figure 1:

Correlation heatmap between variables of the study.

Results

According to the data received from the research institute, four types of data analysis have been performed. The algorithms used to predict ascites were KNN, SVM, RF, and MLPNN, which achieved the comparison of their result in terms of precision, recall and f1-score are shown in Table 3.

According to Table 3, SVM had the best performance for the grade 2, of which the precision, recall and f1-score were 0.99, 0.98 and 0.99 respectively. Also, SVM had acceptable performance for predicting the grades 2 and 3 of ascites with the f1-score of 0.90 and 0.96 respectively. The grade 0 did not have suitable prediction performances. ANN had a good performance for all of the measures, the f1-score for grades 0 to 3 were: 0.86, 0.97, 1.00 and 0.97 respectively. ANN could not predict grades 0 and 3 of ascites and all the measures were equal to zero for these grades, the best prediction performance for this algorithm was for grade 1 with the recall equal to 1. RF algorithm, had the best prediction performance for grades 2 and 3 with the f1-score of 0.99 and 0.98, and for grade 3 the f1-score was 0.93.

To compare the predictive power of each algorithm in predicting the defined grades of ascites, the ROC curves were presented.

Figure 2 shows that SVM had the highest area under the curve for grades 3 (0.95).

Figure 2:

The ROC curve for support vector machine (SVM).

According to Figure 3, the ROCs for grades 2 and 3 are the highest with 0.99 and 0.98 respectively and for grades 0 and one are 0.83 and 0.92.

Figure 3:

The ROC curve for k-nearest neighbors (KNN).

Figure 4 represents that the highest ROCs for ANN method are for grades 1 and 3, that are 0.95 and 0.97.

Figure 4:

The ROC curve for artificial neural network (ANN).

According to Figure 5, the ROC for all of the grades is more than 0.95 and the best ROC belongs to predicting grade 3 of ascites.

Figure 5:

The ROC curve for random forest (RF).

Discussion

The purpose of the present study was to develop and compare predictive models for grades of ascites among cirrhotic patients. Most of the time, the cause of ascites is not difficult to determine. However, “on occasion, ascites may develop as a seemingly isolated finding in the absence of a clinically evident underlying disease” [28]. A better understanding of prognosis may cause to improve patients’ preferences, and inform decision-making across many medical conditions. Singal et al. [19] conducted predictive models for hepatocellular carcinoma (HCC) using machine learning algorithms using a novel methodology to improve HCC risk prognostication among patients with cirrhosis. They obtained a sensitivity of 70.7% and a specificity of 41.6% to identify HCC patients. Lee et al. [29] compared the performance of machine learning approaches with that of logistic regression analysis to Predict Acute Kidney Injury (AKI) after liver transplantation. Several studies conducted machine learning methods to predict other consequences of cirrhosis such as esophageal varices bleeding [3, 21], or considered ascites as a risk factor to predict survival of cirrhosis [16, 20]. However, to the best of our knowledge, no study conducted to predict the grades of ascites in patients with cirrhosis which has a great effect in the management of cirrhotic patients. In the present study the machine learning techniques were compared regarding ROC and precision, recall and f1-score measures. Among the applied algorithms KNN had the best performance for predicting the grade 0 with the average accuracy of 0.98, and ANN had the worst prediction performance with the precision of 0. The reason may be due to the fact that this class had the lowest number of patients compared to other classes in the test set which was seven patients. With more data in this class we can improve the performance of all algorithms. Regarding f1-score, KNN had the highest performance with 97%, for grade 2 also KNN had all the measures equal to 100%, also the next best algorithms were RF and SVM with the f1-score of 99%. For grade 3, the best algorithm was RF with f1-score of 98% and after that KNN had the f1-score of 97%. Therefore, KNN with an average accuracy of 98% was selected as the best performing approach and after that RF had the best performance.

One of the strengths of the present study was the use of several prominent ML methods instead of classical statistical analysis methods to predict the grades of ascites and achieving an accuracy of more than 90%. The algorithms used in this field were KNN, SVM, RF and ANN, with the average accuracies 98, 93, 93% and 0.83 respectively. Therefore, the results of this study may be used to improve patient management programs for patients who suffer from cirrhosis. One of the limitations of this study was the insufficient number of data samples especially in grade 0 which was the persons without ascites, to accurately predict different degrees of ascites, which in KNN approach, this degree of ascites was predicted with the highest accuracy and in ANN method with the worst accuracy. Due to the low number of samples in this class, these approaches may require more validation than other classes. Among the data that is routinely used for the detection and management of ascites, the laboratory parameters include [16, 30] serum bilirubin and creatinine levels, INR and MELD score, and also variables such as CHILD-Pugh_score, however in different contexts gastroenterologists may use different sets of data. Therefore, another limitation could be that not all variables used in this paper may be considered in the evaluation of patients with cirrhosis, so it may affect the generalizability of this work.

In conclusion, machine learning may result considerable information and insights. Well-known machine learning methods, for example, KNN, SVM, RF and ANN, were considered. In the present study, KNN and RF had the better performances and could be considered for future research in this field to improve the accuracy of predictions.