Abstract
Objectives
All patients with cirrhosis should be periodically examined for esophageal varices (EV), however, a large percentage of patients undergoing screening, do not have EV or have only mild EV and do not have high-risk characteristics. Therefore, developing a non-invasive method to predict the occurrence of EV in patients with liver cirrhosis as a non-invasive method with high accuracy seems useful. In the present research, we compared the performance of several machine learning (ML) methods to predict EV on laboratory and clinical data to choose the best model.
Methods
Four-hundred-and-ninety data from the Liver and Gastroenterology Research Center of Shahid Beheshti University of Medical Sciences in the period 2014–2021, were analyzed applying models including random forest (RF), artificial neural network (ANN), support vector machine (SVM), and logistic regression.
Results
RF and SVM had the best results in general for all grades of EV. RF showed remarkably better results and the highest area under the curve (AUC). After that, SVM and ANN had the AUC of 98%, for grade 3, the SVM algorithm had the highest AUC after RF (89%).
Conclusions
The findings may help to better predict EV with high precision and accuracy and also can help reduce the burden of frequent visits to endoscopic centers. It can also help practitioners to manage cirrhosis by predicting EV with lower costs.
Introduction
Chronic liver disease (CLD) is one of the important causes of deaths around the world [1], which may lead to excessive use of health care services [2]. CLD may cause to cirrhosis of the liver [3]. Esophageal varices (EV) is one of the current consequences of ascites, hepatic encephalopathy and renal disorders [4] and especially a serious consequence of cirrhosis [5, 6]. The lifetime prevalence of EV in people with cirrhosis is between 60 and 80% and the mortality due to the acute bleeding of EV ranges between 15 and 55% [7, 8]. In patients with cirrhosis who survive the initial bleeding of EV, it is 33–76% more likely to have a recurrence in the following years [8] which may increase the probability of mortality. According to the guidelines frequent endoscopy is recommended for patients with cirrhosis, as it is the only approach that can directly observe varicose veins and measure their size [9]. Therefore, all patients with cirrhosis should be periodically examined for EV by esophago gastro duodenoscopy (EGD) [7, 10].
EV are classified as having: no (Grade 0), small (Grade 1), medium (Grade 2), and large (Grade 3) degrees of EV, according to the criteria of the Japanese Research Society for Portal Hypertension [11]. High-risk EV are defined as the grades 2 and 3 varicose veins [12]. However, a large percentage of patients suffering from cirrhosis undergoing screening, do not have EV or have only mild EV and do not report high-risk characteristics [13]. Hence, a large number of negligible, inconvenient, and costly endoscopies are performed which may cause severe consequences such as pain, and bleeding of the esophageal and gastric varices [14]. Therefore, these risks may be prevented by developing accurate non-invasive tools to predict EV. Recently, some studies tried to develop models that can predict EV [15], [16], [17]. To the best of our knowledge, most studies have used statistical methods to predict the complications of liver cirrhosis and are not accurate and none of which has been applied in clinical practice today [18]. Therefore, there is a significant requirement to develop a non-invasive tool to better predict the EV and to reduce the burden of unnecessarily high EGD [19].
Machine learning (ML) is recognized as a major part of health care research in general and for clinical medicine in particular [20], [21], [22]. Therefore, developing an ML-based approach to predict the grades of EV in patients with liver cirrhosis as a non-invasive method with high accuracy seems useful. The main purpose of this study was to compare the prediction performances of the degrees of EV in patients with cirrhosis using ML techniques. In the present research, we employed ML methods to develop a prediction based solely on laboratory and clinical data to help physicians accurately predict EV grades that need treatment. The results of this study will assist in preventing unnecessary risks and costs for patients who are less likely to need endoscopy.
Materials and methods
Study population
In the present study, the data was gathered from 490 distinct patients from the Liver and Gastroenterology Research Center of Shahid Beheshti of Medical Sciences in the period of 2014–2021, the dataset consisted of 26 variables of patients with cirrhosis, that 254 patients had no EV, and 236 patients had a history of EV, among whom 72 patients had grade 1, 55 patients had grade 2, and 109 patients had grade 3 which are represented in Figure 1. The complete list of features and their definitions are stated in [22] and also Table 1. Two independent medical experts labeled the data and any disagreements were resolved by asking the help of a third practitioner.
Number of patients in each class of grades for esophageal varices (EV).
Input variables of the study.
| S. No. | Variable | Mean (SD) | Description |
|---|---|---|---|
| 1 | Age | 63.5 (15.76) | – |
| 2 | Sex | Categorical | |
| 1—Male | |||
| 0—Female | |||
| 3 | Etiology | Categorical | |
| 1—HBV | 1—Hepatitis B virus | ||
| 2—HCV | 2—Hepatitis C virus | ||
| 3—Alcohol | 3—Alcohol consumption | ||
| 4—NASH | 4—A form of fatty liver | ||
| 5—Autoimmune | 5—Same as the medical term | ||
| 6—PBC | 6—Primary biliary cholangitis | ||
| 7—PSC | 7—Primary sclerosing cholangitis | ||
| 8—Metastasis | 8—Same as the medical term | ||
| 9—Cryptogenic | 9—Cryptogenic liver disease | ||
| 10—Wilson | 10—Wilson’s disease | ||
| 11—Autoimmune+PBC | |||
| 12—HBV+NASH | |||
| 13—Autoimmune+PSC | |||
| 14—Myelofibrosis | |||
| 15—Unknown | |||
| 4 | Ascites | Categorical | The degrees of ascites among patients |
| 0—No | |||
| 1—Mild-slight | |||
| 2—Moderate | |||
| 3—Severe | |||
| 5 | SBP | Categorical | Spontaneous bacterial peritonitis or ascites fluid infection |
| 1—YES | |||
| 0—NO | |||
| 6 | Encephalopathy | Categorical | Same as the medical term |
| 0—NO | |||
| 1—GRADE 1–2 | |||
| 2—GRADE 3–4 | |||
| 7 | Bleeding | Categorical | Varices bleeding |
| 1—YES | |||
| 0—NO | |||
| 8 | Band.ligation | Categorical | Same as the medical term |
| 1—YES | |||
| 0—NO | |||
| 9 | Child_Pugh score | Categorical | ates which are used to indicate the severity of long-term liver diseases |
| 5–14 | |||
| 10 | Child.class | Categorical | The classes for the previous item |
| A-1 | |||
| B-2 | |||
| C-3 | |||
| 11 | MELD.score | Categorical | A score used as the model for end-stage liver disease |
| 6–40 | |||
| 12 | WBC | 435.94 (16.54) | White blood cell |
| 13 | Hb | 2.57 (0.27) | Hemoglobin |
| 14 | MCV | 7.89 (22.7) | Mean corpuscular volume |
| 15 | MCH | 3.36 (6.37) | Sign of macrocytic anemia |
| 16 | PLT | 11.7 (42.8) | Platelets in liver disease |
| 17 | Bili.T | 1.96 (1.23) | Conjugated bilirubin |
| 18 | Bili.D | 1.86 (1.23) | Direct bilirubin |
| 19 | AST | 7.53 (32.9) | Aspartate aminotransferase |
| 20 | ALP | 23.95 (101.82) | Alkaline phosphatase |
| 21 | PT | 2.84 (7) | Prothrombin time |
| 22 | INR | 1.84 (0.67) | International normalized ratio |
| 23 | ALB | 2.17 (6.45) | Albumin; protein made by the liver |
| 24 | K | 2.05 (0.97) | Vitamin K |
| 25 | Cr | 1.08 (0.72) | Serum creatinine |
| 26 | Na | 136.57 (30.32) | Serum sodium level |
-
HBV, hepatitis B virus; HCV, hepatitis C virus; NASH, non-alcoholic steato-hepatitis; PBC, primary biliary cholangitis; PSC, primary sclerosing cholangitis; SBP, spontaneous bacterial peritonitis; MELD, model for end-stage liver disease; WBC, white blood cell; Hb, hemoglobin; MCV, mean corpuscular volume; MCH, mean corpuscular hemoglobin; PLT, platelets score; Bili.T, conjugated bilirubin; Bili.D, direct bilirubin; AST, aspartate amino-transferase; ALP, alkaline phosphatase; PT, prothrombin time; INR, international normalized ratio; ALB, albumin; K, vitamin K; Cr, creatinine; Na, sodium.
Dataset preprocessing
This part included data cleaning, and resolving missing data. First, all the missing values that contained less than 50% of missing values were replaced with the average of each column and the scale of all the features were standardized using “StandardScaler” function from preprocessing library of python since there were big differences in the ranges of different laboratory features. The analyses of this study were performed using Python programming language in Anaconda open-source environment. Also, TRIPOD checklist for the prediction model development was used (Transparent Reporting of Multivariate Prediction Model for Individual Prognosis or Diagnosis Statement) (Supplementary Material).
Predictive model
Classification models including random forest (RF), artificial neural network (ANN), support vector machine (SVM) and logistic regression (LR) were applied to develop predictive models. We considered these four models for the following reasons: RF is a prediction algorithm that integrates large sets of decision trees and has a very accurate performance in various fields including medical diagnosis [23].
ANNs are computational models that emulate biological neural networks, which are very powerful for nonlinear problems and have accurate predictions in many clinical applications [24]. This model consists of a number of artificial nerve units called perceptron. It automatically learns from the training set with a number of samples until each input matches the output to get the best prediction [25].
SVM is a supervised learning approach which admirably works well for some linear and nonlinear problems. SVM classifier is an optimization-based strategy to find a discriminative hyperplane with maximum margin in a linearly separable feature space [26]. And LR is a discrete model that belongs to multivariate analysis methods and is widely used for empirical analysis in biostatistics and clinical medicine, and in many applications show competitive performance in comparison with other ML models [27]. The hyperparameter selection of the models was performed manually by the researchers by objectively searching different values until the best prediction performance was achieved. The list of hyperparameters for each algorithm is mentioned in Table 2.
Hyperparameters of the models.
| Algorithm | Hyperparameters |
|---|---|
| SVM | SVC (C=1, kernel=‘linear’) |
| RF | n_estimators=100 |
| LR | penalty=‘l2’, dual=False, tol=0.0001, C=1.0, fit_intercept=True, intercept_scaling=1, class_weight=None, random_state=None, solver=‘lbfgs’, max_iter=200, multi_class=‘auto’, verbose=0, warm_start=False, n_jobs=None, l1_ratio=None |
| ANN | activation=‘relu’, solver=‘adam’, alpha=1e-5, hidden_layer_sizes= (3, 1), random_state=1 |
-
SVM, support vector machine; RF, random forest; LR, logistic regression; ANN, artificial neural network.
Model evaluation
We evaluated the performance of the prediction models by using a stratified 10-fold cross-validation. This is a preferred validation technique in ML in using small datasets, without overfitting or overlapping between the test and validation dataset [28]. In this approach the dataset is randomly divided into 10 equal groups (or folds) of which one-fold is used as the validation set, and the remaining folds as the training set. Therefore, each fold is employed once for testing and training without being overlapped. The validation results after 10 times of running the model is then combined to provide the overall performance.
The receiver operating characteristics (ROC) is a well-known performance measure calculated based on sensitivity and specificity to evaluate and compare different classification models. Besides, to assess the performance of each prediction model, important performance indicators including accuracy, recall, precision, and f1-score were calculated.
Results
In the present study, we examined the learning performance of four algorithms in predicting EV in patients with cirrhosis. Table 3 shows the performance of the classification models. All continuous variables were normalized using scaling method in python. According to Table 3 for the prediction of the grade 0 of the EV, ANN and SVM had 100% precision, and RF had the highest F1-score equal to 0.92. For grade 1, RF, ANN, and SVM had the highest F-scores of 0.97, regarding precision RF had the precision of 1, and after that ANN and SVM had the precisions of 0.96 and 0.95 respectively. For the grade 2, all performance measures of RF were equal to 100%, which was the best model in predicting this grade, SVM also had the precision of 0.97 and F-score of 0.99 as the second-best model for this grade. Finally, for the grade 3, RF had all the measures equal to 1, after that ANN had F1-score of 0.97 and precision equal to 1. For the statistical method of LR, it had the precision and F-score of 0.94 and 0.79 respectively for grade 0. For grade 1, the precision and F-score were found 0.89 and 0.815 respectively, for grade 2 it had the precision and F-score of 0.92 and 0.87 respectively, and for the grade 3, the precision and F-score were 0.75 and 0.89.
Summary results of the four classification models.
| RF | ANN | SVM | LR | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Varices grades | 0 | 1 | 2 | 3 | 0 | 1 | 2 | 3 | 0 | 1 | 2 | 3 | 0 | 1 | 2 | 3 |
| Recall | 1 | 0.95 | 1 | 1 | 0.67 | 0.98 | 0.4 | 0.94 | 0.8 | 1 | 0.98 | 1 | 0.86 | 0.83 | 0.81 | 0.84 |
| Precision | 0.86 | 1 | 1 | 1 | 1 | 0.96 | 0.21 | 1 | 1 | 0.95 | 1 | 0.92 | 0.94 | 0.89 | 0.92 | 0.75 |
| F-1 | 0.92 | 0.97 | 1 | 1 | 0.8 | 0.97 | 0.7 | 0.97 | 0.89 | 0.97 | 0.99 | 0.96 | 0.79 | 0.815 | 0.87 | 0.89 |
-
SVM, support vector machine; RF, random forest; LR, logistic regression; ANN, artificial neural network.
In terms of recall, the four algorithms RF, ANN, SVM, and LR in grade 1 were 1, 0.98, 1, and 0.86, respectively. Also, for grade 1 of EV the f1-score for the RF algorithm had the highest rate, for grade 2 of EV the RF, SVM, and neural networks algorithms had the highest rate f1-scores respectively and for grade 3 of EV, RF, and ANN algorithms had the most f1-scores respectively (1 and 0.97). Meanwhile, RF and SVM had the best results in general for all grades of EV.
The ROC curves for the four algorithms have been represented in Figures 2–5. Among the four approaches, RF shows remarkably better results and the highest area under the curve (AUC). After that, SVM and ANN had the AUC of 98%, for grade 3, the SVM algorithm had the highest AUC after RF which was 89%.
Multiclass receiver operating characteristics (ROC) for support vector machine (SVM).
Multiclass receiver operating characteristics (ROC) for random forest.
Multiclass receiver operating characteristics (ROC) for artificial neural network.
Multiclass receiver operating characteristics (ROC) for logistic regression.
Discussion
EV is a common and serious complication of cirrhosis that is associated with increased mortality and morbidity. In recent years, several attempts [29], [30], [31] have been made to predict this complication mostly on EV bleeding or re-bleeding prediction with non-invasive models, but they have not been sufficiently accurate or considered the prediction of EV bleeding among cirrhotic patients. Although the implementation and evaluation of ML models has increased rapidly in recent years, a promising model for predicting EV grades in routine clinical care has not yet been used. ML models in predictive problems mostly are compared to traditional statistical models which are regression models specially LR model [32]. Therefore, the performance of different models along with the ease of use and interpretation of models should be considered. Also, the application of ML models in the analysis of clinical variables from medical records is an efficient approach to discovering the associations between variables that are often difficult to identify [28]. In the present study, we developed and evaluated classification models for EV grading prediction. To the best of our knowledge, this is the first study that attempts to predict EV grading using different models of ML relying on clinical and laboratory data. The performance of the classification methods was evaluated by accuracy, recall, precision, f1-score, and ROC. The RF model showed better performance among other classification models. RF model, showed a higher performance with 99% average accuracy. And also, RF was the most powerful model to predict grades 2 and 3, with the AUC of 0.99 and 1.00 respectively. For the grade 0, we had the lowest AUC of 45% for the SVM algorithm. The highest AUCs for ANN and SVM were related to the prediction of the grade 1, with the 0.88 and 0.98 respectively, however the lowest AUC for ANN model was related to the grade 2, which may be due to the lower number of patients in this class. SVM had the best AUC for grades 1 and 3.
In a study conducted by Sharma and Aggarwal [15], patients underwent rigorous clinical examination, blood tests (hematology, liver function tests), and ultrasounds, and were tested using multivariate LR analysis to predict variables. The performance of the study was evaluated with ROC applying the prediction function. The obtained result from this analysis had a curve area below 0.760. In the present study, the average AUC of the LR method was equal to 87%. The difference between these two studies may be due to the use of 10-fold cross validation approach in the present study. In the study of Karaja et al. [16] with the aim of evaluating the predictors of EV and varicose bleeding using non-invasive markers in Albanian patients diagnosed with liver cirrhosis, the sensitivity of 72%, specificity of 58% and the AUC = 66% was reported, which does not seem to be significant. Finally, according to another study [17], the validation of non-invasive tests to predict EV in children was indicated as a priority because repeated endoscopic evaluations are very invasive and costly. According to their report, the positive predictive values for clinical prediction rule and platelet count were 0.87 and 0.86, the negative predictive values were 0.64 and 0.63, the positive probability ratios were 3.06 and 2.76, and the negative probability ratios were 0.64 and 0.63, respectively. Based on positive and negative predictive values, the most accurate non-invasive tests were clinical prediction rule and platelet count. However, it was not very accurate and the amount of data used in the evaluation was not very representative, which could affect the accuracy. One of the strengths of the present study is the use of local patient data, as well as applying powerful ML methods to increase the accuracy of predictions and selecting the best predictive algorithm. A study [15] conducted artificial intelligence methods to predict EV in patients with cirrhosis using endoscopy reports, physician notes during inpatient and outpatient encounters, laboratory results, and abdominal radiology reports. Their method was gradient boosting ML approach and they reported the f1-score of 0.7 and accuracy of 63% which our finding showed significantly better results. Two other studies [1, 13] attempted to predict EV among patients with cirrhosis and hepatitis C, but the maximum AUC for them reported was 84%. Our results also proved that in general ML methods obtain slightly better results than the regression method that is consistent with the two above mentioned studies. RF algorithm had the best AUC among the four classes of EV that can be considered as a strong method to predict EV among patients with cirrhosis.
Limitation of the study
Among the limitations of the present study, we should mention the lack of enough available data which may cause overfitting and affects the representativeness of the study. By validating the models by 10-fold cross validation we attempted to avoid overfitting of the model on training set, this helps to improve predictions over unseen data. Also, using multi-center data gathering and also considering a greater number of data may improve further generalizations which should be considered for future research. Finally, the strengths of this study include the prediction of grading of EV using clinical and laboratory datasets which are noninvasive and may reduce unnecessary costs.
Conclusions
As a result, according to several measures considered for the performance evaluation of the models, the RF algorithm can be introduced as the best algorithm in this problem. The results of the present study may help to better predict EV in patients with cirrhosis relying on clinical and laboratory data with a high precision and accuracy and also this prediction can help reduce the burden of frequent visits to endoscopic centers and reduce the possible complications and risks for patients with cirrhosis that could be bleeding of varices. It can also help practitioners to better manage cirrhosis by predicting EV in patients with lower costs.–
-
Research funding: None declared.
-
Author contributions: Study design: Farkhondeh Asadi and Azadeh Bayani. Database synthesis: Azadeh Bayani, Behzad Hatami, and Azamossadat Hosseini. Initial manuscript draft: Azadeh Bayani. Critical revision of the manuscript: Azadeh Bayani, Mehrad Aria, Mohammad Reza Zali, and Kaveh Kavousi. All authors have accepted responsibility for the entire content of this manuscript and approved its submission.
-
Competing interests: Authors state no conflict of interest.
-
Informed consent: An informed consent was obtained from all participants. The raw data did not contain any personal identifying information that can be linked to particular individuals, and was anonymized before its use.
-
Ethical approval: The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Medical Ethics Committee of Shahid Beheshti University of Medical Sciences due to the retrospective nature of the study and that only anonymous data were used (ethics code: IR.SBMU.RETECH.REC.1400.588).
References
1. Abd El-Salam, SM, Ezz, MM, Hashem, S, Elakel, W, Salama, R, ElMakhzangy, H, et al.. Performance of machine learning approaches on prediction of esophageal varices for Egyptian chronic hepatitis C patients. Inform Med Unlocked 2019;17:100267. https://doi.org/10.1016/j.imu.2019.100267.Search in Google Scholar
2. Asrani, SK, Devarbhavi, H, Eaton, J, Kamath, PS. Burden of liver diseases in the world. J Hepatol 2019;70:151–71. https://doi.org/10.1016/j.jhep.2018.09.014.Search in Google Scholar PubMed
3. Nayak, A, Kayal, EB, Arya, M, Culli, J, Krishan, S, Agarwal, S, et al.. Computer-aided diagnosis of cirrhosis and hepatocellular carcinoma using multi-phase abdomen CT. Int J Comput Assist Radiol Surg 2019;14:1341–52. https://doi.org/10.1007/s11548-019-01991-5.Search in Google Scholar PubMed
4. Fukui, H, Saito, H, Ueno, Y, Uto, H, Obara, K, Sakaida, I, et al.. Evidence-based clinical practice guidelines for liver cirrhosis 2015. J Gastroenterol 2016;51:629–50. https://doi.org/10.1007/s00535-016-1216-y.Search in Google Scholar PubMed
5. Yang, J, Zeng, R, Cao, J, Wu, C, Chen, T, Li, R, et al.. Predicting gastro-oesophageal variceal bleeding in hepatitis B-related cirrhosis by CT radiomics signature. Clin Radiol 2019;74:976.e1–e9. e979. https://doi.org/10.1016/j.crad.2019.08.028.Search in Google Scholar PubMed
6. Rigo, G, Merighi, A, Chahin, N, Mastronardi, M, Codeluppi, P, Ferrari, A, et al.. A prospective study of the ability of three endoscopic classifications to predict hemorrhage from esophageal varices. Gastrointest Endosc 1992;38:425–9. https://doi.org/10.1016/s0016-5107(92)70470-2.Search in Google Scholar PubMed
7. Garcia-Tsao, G, Abraldes, JG, Berzigotti, A, Bosch, J. Portal hypertensive bleeding in cirrhosis: risk stratification, diagnosis, and management: 2016 practice guidance by the American Association for the study of liver diseases. Hepatology 2017;65:310–35. https://doi.org/10.1002/hep.28906.Search in Google Scholar PubMed
8. Baiges, A, Hernández-Gea, V, Bosch, J. Pharmacologic prevention of variceal bleeding and rebleeding. Hepatol Int 2018;12:68–80. https://doi.org/10.1007/s12072-017-9833-y.Search in Google Scholar PubMed
9. Moctezuma-Velázquez, C, Abraldes, JG. Non-invasive diagnosis of esophageal varices after Baveno VI. Turk J Gastroenterol 2017;28:e65.10.5152/tjg.2017.16744Search in Google Scholar PubMed
10. Kim, TY, Kim, TY, Kim, Y, Lim, S, Jeong, WK, Sohn, JH. Diagnostic performance of shear wave elastography for predicting esophageal varices in patients with compensated liver cirrhosis. J Ultrasound Med 2016;35:1373–81. https://doi.org/10.7863/ultra.15.07024.Search in Google Scholar PubMed
11. Hino, S, Kakutani, H, Ikeda, K, Uchiyama, Y, Sumiyama, K, Kuramochi, A, et al.. Hemodynamic assessment of the left gastric vein in patients with esophageal varices with color Doppler EUS: factors affecting development of esophageal varices. Gastrointest Endosc 2002;55:512–7. https://doi.org/10.1067/mge.2002.122333.Search in Google Scholar PubMed
12. Tajiri, T, Yoshida, H, Obara, K, Onji, M, Kage, M, Kitano, S, et al.. General rules for recording endoscopic findings of esophagogastric varices. Dig Endosc 2010;22:1–9. https://doi.org/10.1111/j.1443-1661.2009.00929.x.Search in Google Scholar PubMed
13. Dong, TS, Kalani, A, Aby, ES, Le, L, Luu, K, Hauer, M, et al.. Machine learning-based development and validation of a scoring system for screening high-risk esophageal varices. Clin Gastroenterol Hepatol 2019;17:1894–901. e1891. https://doi.org/10.1016/j.cgh.2019.01.025.Search in Google Scholar PubMed
14. Mattos, ÂZ, Schacher, FC, John Neto, G, Mattos, AA. Screening for esophageal varices in cirrhotic patients – non-invasive methods. Ann Hepatol 2019;18:673–8. https://doi.org/10.1016/j.aohep.2019.06.003.Search in Google Scholar PubMed
15. Sharma, SK, Aggarwal, R. Prediction of large esophageal varices in patients with cirrhosis of the liver using clinical, laboratory and imaging parameters. J Gastroenterol Hepatol 2007;22:1909–15. https://doi.org/10.1111/j.1440-1746.2006.04501.x.Search in Google Scholar PubMed
16. Kraja, B, Mone, I, Akshija, I, Koçollari, A, Prifti, S, Burazeri, G. Predictors of esophageal varices and first variceal bleeding in liver cirrhosis patients. World J Gastroenterol 2017;23:4806. https://doi.org/10.3748/wjg.v23.i26.4806.Search in Google Scholar PubMed PubMed Central
17. Gana, JC, Turner, D, Mieli-Vergani, G, Davenport, M, Miloh, T, Avitzur, Y, et al.. A clinical prediction rule and platelet count predict esophageal varices in children. Gastroenterology 2011;141:2009–16. https://doi.org/10.1053/j.gastro.2011.08.049.Search in Google Scholar PubMed
18. de Franchis, R, Dell’Era, A. Variceal hemorrhage, 2014 edition. New York: Springer; 2014.10.1007/978-1-4939-0002-2Search in Google Scholar
19. Patel, V, Armstrong, D, Ganguli, MP, Roopra, S, Kantipudi, N, Albashir, S, et al.. Deep learning in gastrointestinal endoscopy. Crit Rev Biomed Eng 2016;44:493–504. https://doi.org/10.1615/CritRevBiomedEng.2017025035.Search in Google Scholar PubMed
20. Robert, C. Machine learning, a probabilistic perspective. Chance; 2014;27:62–3.10.1080/09332480.2014.914768Search in Google Scholar
21. Hatami, B, Asadi, F, Bayani, A, Zali, MR, Kavousi, K. Machine learning-based system for prediction of ascites grades in patients with liver cirrhosis using laboratory and clinical data: design and implementation study. Clin Chem Lab Med 2022;60:1946–54. https://doi.org/10.1515/cclm-2022-0454.Search in Google Scholar PubMed
22. Bayani, A, Hosseini, A, Asadi, F, Hatami, B, Kavousi, K, Aria, M, et al.. Identifying predictors of varices grading in patients with cirrhosis using ensemble learning. Clin Chem Lab Med 2022;60:1938–45. https://doi.org/10.1515/cclm-2022-0508.Search in Google Scholar PubMed
23. Breiman, L. Random forests. Mach Learn 2001;45:5–32. https://doi.org/10.1023/a:1010933404324.10.1023/A:1010933404324Search in Google Scholar
24. Musleh, MM, Alajrami, E, Khalil, AJ, Abu-Nasser, BS, Barhoom, AM, Naser, SA. Predicting liver patients using artificial neural network. IJAISR 2019;3:1–11.Search in Google Scholar
25. Lin, S-W, Ying, K-C, Chen, S-C, Lee, Z-J. Particle swarm optimization for parameter determination and feature selection of support vector machines. Expert Syst Appl 2008;35:1817–24. https://doi.org/10.1016/j.eswa.2007.08.088.Search in Google Scholar
26. Yang, H, Du, Q, Chen, G. Particle swarm optimization-based hyperspectral dimensionality reduction for urban land cover classification. IEEE J Sel Top Appl Earth Obs Rem Sens 2012;5:544–54. https://doi.org/10.1109/jstars.2012.2185822.Search in Google Scholar
27. Lee, H-C, Yoon, SB, Yang, S-M, Kim, WH, Ryu, H-G, Jung, C-W, et al.. Prediction of acute kidney injury after liver transplantation: machine learning approaches vs. logistic regression model. J Clin Med 2018;7:428. https://doi.org/10.3390/jcm7110428.Search in Google Scholar PubMed PubMed Central
28. Wu, C-C, Yeh, W-C, Hsu, W-D, Islam, MM, Nguyen, PAA, Poly, TN, et al.. Prediction of fatty liver disease using machine learning algorithms. Comput Methods Progr Biomed 2019;170:23–9. https://doi.org/10.1016/j.cmpb.2018.12.032.Search in Google Scholar PubMed
29. Aleksić, A, Nedeljković, S, Jovanović, M, Ranđelović, M, Vuković, M, Stojanović, V, et al.. Prediction of important factors for bleeding in liver cirrhosis disease using ensemble data mining approach. Mathematics 2020;8:1887.10.3390/math8111887Search in Google Scholar
30. Yan, Y, Li, Y, Fan, C, Zhang, Y, Zhang, S, Wang, Z, et al.. A novel machine learning-based radiomic model for diagnosing high bleeding risk esophageal varices in cirrhotic patients. Hepatol Int 2021;16:423–32.10.1007/s12072-021-10292-6Search in Google Scholar PubMed
31. Trebicka, J, Gu, W, Ibáñez-Samaniego, L, Hernández-Gea, V, Pitarch, C, Garcia, E, et al.. Rebleeding and mortality risk are increased by ACLF but reduced by pre-emptive TIPS. J Hepatol 2020;73:1082–91. https://doi.org/10.1016/j.jhep.2020.04.024.Search in Google Scholar PubMed
32. Song, X, Liu, X, Liu, F, Wang, C. Comparison of machine learning and logistic regression models in predicting acute kidney injury: a systematic review and meta-analysis. Int J Med Inf 2021;151:104484. https://doi.org/10.1016/j.ijmedinf.2021.104484.Search in Google Scholar PubMed
Supplementary Material
The online version of this article offers supplementary material (https://doi.org/10.1515/cclm-2022-0623).
© 2022 Walter de Gruyter GmbH, Berlin/Boston
