Supervised Machine Learning Models for Liver Disease Risk Prediction

Abstract

The liver constitutes the largest gland in the human body and performs many different functions. It processes what a person eats and drinks and converts food into nutrients that need to be absorbed by the body. In addition, it filters out harmful substances from the blood and helps tackle infections. Exposure to viruses or dangerous chemicals can damage the liver. When this organ is damaged, liver disease can develop. Liver disease refers to any condition that causes damage to the liver and may affect its function. It is a serious condition that threatens human life and requires urgent medical attention. Early prediction of the disease using machine learning (ML) techniques will be the point of interest in this study. Specifically, in the content of this research work, various ML models and Ensemble methods were evaluated and compared in terms of Accuracy, Precision, Recall, F-measure and area under the curve (AUC) in order to predict liver disease occurrence. The experimental results showed that the Voting classifier outperforms the other models with an accuracy, recall, and F-measure of 80.1%, a precision of 80.4%, and an AUC equal to 88.4% after SMOTE with 10-fold cross-validation.

1. Introduction

The liver is located in the upper part of the gastrointestinal tract of the human body, and its weight in men ranges between 1400–1800 g and in women between 1200–1400 g. It performs important functions related to digestion, metabolism, releasing toxins, immunization and nutrient storage. That is why some liver diseases can even lead to death [1,2].

Liver diseases are categorized based on their aetiology and effect on the liver. The aetiology may include infection, injury, exposure to drugs or toxic substances, a process, or a genetic abnormality (such as hemochromatosis). The above causes can lead to hepatitis, cirrhosis, and stones that can increase in size and cause blockages, fatty infiltration and, in rare cases, liver cancer. Genetic abnormalities can also interfere with vital functions of the liver and lead to the deposition and concentration of harmful components, such as iron or copper [3,4,5].

Non-alcoholic fatty liver disease (NAFLD) is one of the main liver diseases and is characterized by an accumulation of lipids in the liver. If there is inflammation and injury to the liver cells, it is called “non-alcoholic steatohepatitis” [6]. Cirrhosis is also one of the most serious liver diseases. This disease causes healthy tissue to be replaced by scar tissue. Thus, the liver is permanently injured and cannot function properly. The main causes of cirrhosis of the liver include alcoholism, non-alcoholic fatty liver disease, chronic hepatitis C, and chronic hepatitis B [7].

There are two main forms of hepatitis, acute hepatitis [8], where the liver becomes inflamed at a rapid rate, and chronic hepatitis [9], where the liver becomes inflamed and destroyed slowly over a long period of time. Although hepatitis can be caused by any of the above, it is most often caused by infection with a virus in a group called hepatitis viruses. These viruses have been named, in the order they were discovered, as hepatitis A, B, C, D, and E viruses [10].

Hepatitis A is caused by the HAV virus. It is usually spread by contaminated food or water. It can also be transmitted through sexual contact and more rarely through blood. [11]. Hepatitis B has similar characteristics. However, in this case, it is transmitted by contact with any bodily fluid of an infected person. There are two forms, acute and chronic. Left untreated, the latter can turn into liver cancer or cause liver failure [12]. Hepatitis C has a similar pattern to the previous two. However, it is transmitted through contact with the blood of an infected person. Early symptoms can appear even ten years after infection. As with hepatitis B, there is an acute and chronic form [13]. The hepatitis D virus is very rare and only affects people with hepatitis B [14], and hepatitis E is spread by drinking water contaminated by the faeces of an infected person [15].

Liver disease is usually associated with alcoholism or hepatitis, but obesity and diabetes are becoming a growing threat of potentially fatal liver damage. Advanced fatty liver disease increases a person’s risk of death by nearly sevenfold. It is a silent “killer” and if the symptoms associated with fatty liver damage appear, the situation is already difficult [16,17,18].

In order to prevent liver disease, patients should not drink large amounts of alcohol. However, in the case where a patient is diagnosed with hepatitis B or C, alcoholic hepatitis, etc., the simple recommendation for these patients is to not consume any alcohol at all. Additional precautions include the use of a condom during sexual intercourse, the avoidance of sharing syringes or needles, vaccination for hepatitis A and B, and the protection of the skin from toxic chemicals. Finally, exercise, a healthy diet and maintaining a normal body weight contribute to the proper functioning of the liver [19,20].

Traditionally, health professionals make a medical report concerning a patient’s condition based on histopathological exams. With the advances in information and communication technologies, especially in artificial intelligence (AI) and machine learning (ML), efficient data collection, processing, and visualization methods have arisen. Clinicians combining the outcomes of AI and ML models with the findings of clinical methods can further improve their decisions on disease detection. Undoubtedly, ML techniques have significantly contributed to the early prediction of disease complications in diabetes (handling it as a classification problem [21,22] or regression task for the short-term glucose prediction [23,24]), cholesterol [25], hypertension [26,27], hypercholesterolemia [28], chronic obstructive pulmonary disease (COPD) [29], covid-19 [30], stroke [31], chronic kidney disease (CKD) [32], lung cancer [33], sleep disorders [34,35], cardiovascular diseases (CVDs) [36], etc.

In particular, liver disease occurrence will concern us in the context of this research work. The main contributions of the methodology adopted are the following:

• Data preprocessing is performed with the synthetic minority oversampling technique (SMOTE). In this way, the dataset’s instances are distributed in a balanced way allowing us to design efficient classification models and predict the occurrence of liver disease.
• Features’ importance evaluation based on the Pearson Correlation, Gain Ratio and Random Forest is carried out.
• A comparative evaluation of many ML models’ performance is illustrated considering well-known metrics, such as Precision, Recall, F-Measure, Accuracy and AUC. The experimental results indicated that the Voting classifier prevailed over the other models constituting the proposition of this research work.
• Considering published papers that were based on the same dataset with the same features we relied on, our main proposal (i.e., Voting classifier) outperformed in terms of accuracy.

The rest of the paper is organized as follows. Section 2 provides a dataset description and an analysis of the methodology followed. Furthermore, in Section 3, we evaluate the experimental research results. In addition, Section 4 discusses related works on the topic under examination. Finally, conclusions and future directions are noted in Section 5.

2. Materials and Methods

Here, we will provide the dataset we relied on and the main steps of the adopted methodology for liver disease risk prediction, namely, class balancing and features’ ranking in the balanced data. Finally, we note the ML models we based on for the experimental results.

2.1. Dataset Description

Our research was based on the Indian Liver Patients’ Records dataset [37]. The specific dataset includes 579 participants, of which the number of men (male) is 439 (75.8%) and women (female) 140 (24.2%). The target class indicates if the participant has been diagnosed with liver disease or not. The number of participants diagnosed with liver disease is 414 (71.5%). A brief overview of the dataset’s characteristics is shown in

Table 1. Dataset Description.

Feature	Type	Description
Gender [38]	nominal	This feature illustrates the participant’s gender.
Age (years) [39]	numeric	The age range of the participants is 4–90 years.
Total Bilirubin—TB (mg/dL) [40]	numeric	This feature captures the participant’s total bilirubin.
Direct Bilirubin—DB (mg/dL) [40]	numeric	This feature captures the participant’s direct bilirubin.
Alkaline Phosphatase—ALP (IU/L) [41]	numeric	This feature captures the participant’s alkaline phosphatase.
Alanine Aminotransferase—SGPT (U/L) [42]	numeric	This feature captures the participant’s alanine aminotransferase.
Aspartate Aminotransferase—SGOT (U/L) [42]	numeric	This feature captures the participant’s aspartate aminotransferase.
Total Protein—TP (g/L) [43]	numeric	This feature captures the participant’s total protein.
Albumin—ALB (g/dL) [44]	numeric	This feature captures the participant’s albumin.
Albumin and Globulin Ratio—AGR [45]	numeric	This feature captures the participant’s albumin and globulin Ratio.
Liver Disease	nominal	This feature stands for whether the participant has been diagnosed with liver disease or not.

.

2.2. Liver Disease Risk Prediction

Nowadays, clinicians and health carers exploit machine-learning models to develop efficient tools for the risk assessment of a disease occurrence based on several risk factors. Here, the long-term risk prediction of liver disease is formulated as a classification problem with two possible classes c = “Liver-Disease” (LD) or c = “Non-Liver-Disease” (Non-LD). The trained ML models will be able to predict the class of a new unclassified instance, either as LD or Non-LD based on the input features’ values and thus the risk of occurring liver disease.

2.2.1. Data Preprocessing

The accurate identification of LD and Non-LD instances may be impacted by their unbalanced distribution in the dataset. Here, an oversampling method is applied, namely SMOTE [46], which, based on a 5-NN classifier, creates synthetic data [47] on the minority class. The instances in the Non-LD class are oversampled such that the subjects in the two classes are uniformly distributed. After the implementation of SMOTE, the number of participants is 828, of which the number of men (male) is 615 (74.3%) and women (female) 213 (25.7%). Now the dataset is balanced, and the target class includes 414 LD and 414 Non-LD instances. Finally, in

Table 2. Statistical description of the numerical features in the balanced dataset after SMOTE.

	Min	Max	Mean ± Stdv
Age	4	90	43.55 ± 16.28
TB	0.4	75	2.65 ± 5.32
DB	0.1	19.7	1.16 ± 2.42
ALP	63	2110	267.26 ± 212.62
SGPT	10	2000	66.78 ± 155.16
SGOT	10	4929	88.78 ± 245.07
TP	2.7	9.6	6.50 ± 1.02
ALB	0.9	5.5	3.19 ± 0.76
GR	0.3	2.8	0.98 ± 0.30

, we present the statistical characteristics of the features in the balanced dataset, focusing on the minimum, maximum, mean and standard deviation.

2.2.2. Features Analysis

Three ranking methods [48] have been selected to evaluate the contribution of a feature in the target class (LD). Their results are illustrated in

Table 3. Features’ importance evaluation based on the Pearson Correlation, Gain Ratio and Random Forest.

Feature	Pearson Rank	Feature	Gain Ratio	Feature	Random Forest
DB	0.3205	DB	0.1421	DB	0.2895
TB	0.2874	TB	0.1373	ALB	0.2883
ALP	0.246	SGOT	0.1005	TB	0.2848
SGPT	0.2141	ALP	0.0867	Age	0.2625
AGR	0.2046	AGR	0.0822	SGPT	0.2613
SGOT	0.2017	SGPT	0.0701	AGR	0.2599
ALB	0.1836	ALB	0.0408	SGOT	0.2393
Age	0.1596	Age	0.0372	TP	0.2161
Gender	0.0857	Gender	0.0065	ALP	0.1936
TP	0.0443	TP	0	Gender	0.0255

.

Initially, Pearson correlation analysis [49] is executed and its outcomes are first illustrated in Figure 1. This coefficient is utilized to capture the strength and direction of the association between two features and/or a feature and the interested class. Its values vary between −1 and 1. Focusing on this coefficient, we observe a strong correlation of 0.88 between TB and DB. Moreover, a strong association is noted between SGOT and SGPT, and ALB and TP of rank 0.79 and 0.76, respectively. Furthermore, a moderate relationship of rank 0.65 is shown to have AGR with ALB. Evaluating the features’ contribution in the target class, DB is first in order of importance; however, a low association of 0.32 is recorded. The same holds for the TB, ALP, SGPT, AGR and SGOT. A weaker association of 0.1836 and 0.1596 is shown to have liver disease variable with the features of age and ALB, correspondingly. An absence of correlation seems to occur with the gender and TP features where the rank is lower than 0.1.

Next, we employed Gain Ratio (GR) method [50], which is calculated as $GR\left(X\right)=\frac{H\left(C\right)-H\left(C\right|X)}{H\left(X\right)}$, where $H\left(X\right)=-{\sum }_{x\in V_X}{p}_{x}lo{g}_2\left(p_x\right)$ (with $p_x$ denoting the probability of feature X takes value equal to $x\in V_X$), $H\left(C\right)=-{\sum }_{c\in C}{p}_{c}lo{g}_2\left(p_c\right)$ (with $p_c$ be the probability of selecting an instance in class c with two possible states Non-LD, LD) and $H\left(C\right|X)$ are the entropy of feature X, the entropy of class variable C and the conditional entropy of feature X given class C, respectively. This method assigned the highest scores to DB, TB and SGOT.

Thirdly, the Random Forest classifier was considered to measure the features’ importance based on Gini impurity [51], which measures a candidate feature’s ability to optimally split the instances into two classes. In this method, gender was assigned the lowest ranking, close to zero. As we see, the importance of the rest features is very close to each other (such as in the case of DB, ALP, and TB).

In conclusion, the models’ training and validation will be based on all of these features, as they are the most important for liver disease screening by physicians.

2.3. Machine Learning Models

In this research article, we experimented with various ML models to uncover which one outperforms the others by evaluating their prediction performance. Specifically, we focused on naive Bayes (NB) [52] and Logistic Regression (LR) [53] models, which are probabilistic classifiers. In addition, we used the well-known kernel-based (linear, non-linear) classifier Support Vector Machine (SVM) [54].

Moreover, we used Decision-Tree-based models such as, J48 [55], Random Tree (RT) [56] and Reduced Error Pruning Tree (RepTree) [57]. From Ensemble ML algorithms [58], Bagging [59], Random Forest (RF) [60], Rotation Forest (RotF) [61], AdaBoostM1 [62], Voting [63] and Stacking [64] were exploited. Finally, a simple Artificial Neural Network (ANN) called multilayer perceptron (MLP) [65], and k-nearest neighbors (kNN) [66], a distance-based classifier, were evaluated.

2.4. Evaluation Metrics

To evaluate the ML models’ performance we utilised the most commonly used metrics in the relevant literature, such as Accuracy, Precision, Recall, F-Measure, and AUC [67,68]. The confusion matrix consists of the elements true positive (TP), true negative (TN), false positive (FP) and false negative (FN). The aforementioned metrics are defined as follows.

• Accuracy: summarizes the performance of the classification task and measures the number of correctly predicted instances out of all the data instances.

  $$\begin{array}{c}\hfill \mathrm{Accuracy}=\frac{\mathrm{TN}+\mathrm{TP}}{\mathrm{TN}+\mathrm{TP}+\mathrm{FN}+\mathrm{FP}}\end{array}$$

  (1)
• Recall: corresponds to the proportion of participants who were diagnosed with LD and were correctly classified as positive, concerning all positive participants.

  $$\begin{array}{c}\hfill \mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\end{array}$$

  (2)
• Precision: indicates how many of those who were identified as LD belong to this class.

  $$\begin{array}{c}\hfill \mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}\end{array}$$

  (3)
• F-Measure: is the harmonic mean of the Precision and Recall and sums up the predictive performance of a model. The desired metrics will be calculated with the help of the Confusion matrix.

  $$\begin{array}{c}\hfill \mathrm{F}\text{-}\mathrm{Measure}=2\frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}\end{array}$$

  (4)
  In order to evaluate the distinguishability of a model, the AUC is exploited. It is a metric that varies in [0, 1]. The closer to one, the better the ML model performance is in distinguishing LD from Non-LD instances.

3. Results

In this section, the experiment settings and the acquired outcomes will be described.

3.1. Experimental Setup

For the evaluation of our proposed ML models, we relied on the Waikato Environment for Knowledge Analysis (Weka) [69], which is an open-access software developed at the University of Waikato, New Zealand. In addition, the experiments were performed on a computer system with the following specifications: 11th generation Intel(R) Core(TM) i7-1165G7 @ 2.80 GHz, RAM 16 GB, Windows 11 Home, 64-bit OS and x64 processor.

Various ML models, such as NB, SVM, LR, ANN, kNN, J48, RF, RT, RepTree, RotF, AdaBoostM1, Stacking, Bagging, and Voting were assessed in terms of Accuracy, Precision, Recall, F-Measure and AUC. In Stacking, we consider as base classifiers the RF and AdaBoostM1 models, and as meta classifier the LR model. Concerning Voting, the same base classifiers as Stacking were assumed, and the final prediction was derived from the average probabilities (soft Voting). The Bagging had as a base classifier the RF. In

Table 4. Machine learning models’ settings.

Model	Parameters
NB	useKernelEstimator: False useSupervisedDiscretization: True
SVM	eps = 0.001 gamma = 0.0 kernel type: linear loss = 0.1
LR	ridge = ${10}^{-8}$ useConjugateGradientDescent: True
ANN	hidden layers: ‘a’ learning rate = 0.1 momentum = 0.2 training time = 200
J48	reducedErrorPruning: True savelnstanceData: True useMDLCorrection: True subtreeRaising: True binarysplits = True collapseTree = True
RT	maxDepth = 0 minNum = 1.0 minVarianceProp = 0.001
RepTree	maxDepth = 1 minNum = 2.0 minVarianceProp = 0.001
RF	breakTiesRandomly:True numIterations = 100 numFeatures = 0
RotF	classifier: Random Forest numberOfGroups: False projectionFilter: PrincipalComponents
kNN	k = 1 Search Algorithm: LinearNNSearch with Euclidean cross-validate = True
AdaBoostM1	classifier: Random Forest resume: True useResampling: True
Stacking	classifiers: Random Forest and AdaBoostM1 metaClassifier: Logistic Regression numFolds = 10
Voting	classifiers: Random Forest and AdaBoostM1 combinationRule: average of probabilities
Bagging	classifiers: Random Forest printClassifiers: True storeOutOfBagPredictions: True

, we illustrate the optimal parameters’ settings of the ML models that we experimented with. Finally, we applied 10-fold cross-validation in order to measure the models’ efficiency in the balanced dataset of 828 instances after SMOTE.

3.2. Performance Evaluation

In the context of performance evaluation, we will first present how SMOTE method contributed to the design of efficient models. The impact of the specific class balancing method is clearly reflected in the Recall and Precision metrics which are captured in

Table 5. Performance of ML models before and after SMOTE into class “No” with 10-fold cross-validation.

Class “No”	Precision		Recall
	No SMOTE	SMOTE	No SMOTE	SMOTE
NB	0.429	0.671	0.679	0.831
SVM	0.000	0.648	0.000	0.908
LR	0.544	0.659	0.261	0.826
MLP	0.415	0.653	0.297	0.787
1-NN	0.405	0.670	0.467	0.744
J48	0.391	0.647	0.055	0.792
RF	0.495	0.763	0.291	0.848
RT	0.475	0.717	0.509	0.717
RepTree	0.358	0.687	0.206	0.763
RotF	0.564	0.726	0.321	0.884
AdaBoostM1	0.529	0.756	0.442	0.853
Stacking	0.494	0.770	0.255	0.831
Bagging	0.494	0.749	0.267	0.853
Voting	0.515	0.791	0.309	0.875

. These outcomes are related to the minority class (namely healthy participants) and show that ML models highly benefited from the application of SMOTE technique. Then, in

Table 6. Performance evaluation of ML models after SMOTE with 10-fold cross-validation.

	Accuracy	Precision	Recall	F-Measure	AUC
NB	0.711	0.724	0.711	0.707	0.771
SVM	0.708	0.748	0.708	0.708	0.708
LR	0.700	0.713	0.700	0.700	0.754
MLP	0.690	0.701	0.690	0.690	0.742
1-NN	0.688	0.691	0.688	0.687	0.693
J48	0.732	0.733	0.732	0.732	0.735
RF	0.794	0.798	0.793	0.793	0.877
RT	0.718	0.717	0.717	0.717	0.717
RepTree	0.708	0.710	0.708	0.707	0.761
RotF	0.775	0.789	0.775	0.773	0.869
AdaBoostM1	0.795	0.797	0.795	0.795	0.879
Stacking	0.794	0.795	0.793	0.793	0.881
Bagging	0.785	0.791	0.785	0.784	0.872
Voting	0.801	0.804	0.801	0.801	0.884

, we illustrate the average performance of the models under consideration after applying SMOTE with 10-fold cross-validation. At this point, it should be noted that the uniform class distribution helped to improve the correct identification of healthy instances while keeping the average performance at higher levels. That means that the investigated models maintained their efficiency for the patients’ correct and accurate identification as well. The Voting Ensemble method outperforms in comparison to the other models with an Accuracy, Recall, and F-measure of 80.1%, a Precision of 80.4%, and an AUC equal to 88.4%. Very good performance is also presented by the AdaboostM1 model, which has as a base classifier the RF, with an Accuracy, F-Measure and Recall of 79.5%, a Precision of 79.7%, and an AUC equal to 87.9%. Finally, the RF and the Stacking Ensemble model operate at similar levels of accuracy (79.4%).

In addition to the above metrics, in Figure 2, we provide models’ evaluation based on AUC ROC curves after SMOTE with 10-fold cross-validation. From the illustration, we observe that the RF, RotF, AdaboostM1, Stacking, Bagging and Voting classifiers present better performance than the other models, which is also confirmed and numerically.

Moreover, in

Table 7. Illustration of proposed models from works based on the same dataset [37].

Research Work	Proposed Model	Accuracy
Present Work	Voting	80.10%
[70]	NB	75.54%
[71]	MAMFFN	75.30%
[72]	SVM	75.04%
[73,74,75,76,77]	LR	75%
[78]	kNN	74.67%
[79]	AdaBoostM1	74.36%
[80]	RT	74.20%
[81]	KNNWFST	74%
[82]	Back Propagation	73.20%
[83]	Gradient Tree Boosting	72%
[84,85]	RF	71.87%
[86]	CHIRP	71.30%
[87]	DT	69.40%
[88]	Bagging	69.30%

, we provide the proposed models from published research works based on the same dataset [37] with the same features that we relied on. Our proposed model (i.e., Voting) performs better compared to the other works achieving an accuracy of 80.10%.

4. Discussion

In this section, works based on the Indian Liver Patients’ Records dataset [37] are presented in order to predict liver disease occurrence by applying various ML models. Specifically, [84,85] proposed the Random Forest model reaching an accuracy of 71.87%, while [87] suggested the Decision Tree model achieving an accuracy of 69.40%. In addition, [80] recommended the Random Tree model with an accuracy of 74.20% and [79] showed that AdaBoostM1 is the best model achieving an accuracy of 74.36%, [88] proposed the ensemble Bagging method achieving an accuracy of 69.30%, [78] proposed the k-nearest neighbour model with an accuracy of 74.67%, and in [71] the authors proposed a hybrid ML model called Mathematical Approach on Multilayer Feedforward Neural Network with Backpropagation (MAMFFN) achieving an accuracy of 75.30%.

Furthermore, the authors in [73,74,75,76,77] proposed the Logistic Regression model achieving an accuracy of 75%. The Composite Hypercube on Iterated Random Projection (CHIRP) model proposed by the authors in [86] with an accuracy equal to 71.30%. A classic multi-layered neural network based on a backpropagation algorithm is the authors’ proposal in [82] with 73.2% accuracy. Additionally, the authors at [81] tested that the k-nearest neighbour with feature selection techniques (KNNWFST) outperforms all other comparables. The Gradient Tree Boosting classifier for a balanced dataset showed the best accuracy with 72% in [83]. Finally, the naïve Bayes in [70] performs better with 75.54% accuracy, and in [72], the Support Vector Machine achieves 75.04% accuracy.

In this work, exploiting the same dataset with [70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88], we decided to study more efficient models resorting to ensemble learning. The models in this category were trained and tested on the same biochemical features before and after class balancing assuming 10-fold cross-validation. We focused on the SMOTE-based balanced dataset and the ensemble models which the previous works did not consider, along with a graphical capture of AUC ROC curves. Comparing the performance of the single models in the studies [70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88] with the ones examined here, our trained and tested classifiers prevailed in terms of accuracy. The Voting method, which is the main suggestion of this study, prevailed both against the models in the same family and the single ones as shown in Table 7. Quality of life is an important variable. Hence, from a clinical and technological research perspective, designing a machine learning model based on biochemical data that capture human health status has become an imperative need.

Concluding this section, the current study will help clinicians and researchers monitor liver disease, design high-performance personalized models, providing the flexibility to incorporate both quality-of-life features that show the well-being of patients, and indicate the difficulties associated with this condition. Summarizing the evaluation of the proposed models, we have to point out some limitations of our article. The present research work was based on the Indian Liver Patients’ Records dataset, which, although a well-known publicly available dataset [37], has no indications on what were the exact definitions and criteria used for diagnosing patients with liver disease. Finally, exploiting data coming from a hospital unit or institute could give us a greater variety of features to better evaluate the ML models. However, it should be emphasized that gaining access to sensitive medical data is difficult due to privacy reasons.

5. Conclusions

Liver disease is a serious condition that threatens human life and requires urgent medical attention. Health professionals are based on pathological methods to make a medical report concerning a patient’s condition. Early prediction of liver disease using machine learning techniques was the point of interest in this study.

Specifically, plenty of ML models, such as NB, SVM, LR, ANN, kNN, J48, RF, RT, RepTree, RotF, AdaBoostM1, Stacking, Bagging, and Voting, were evaluated in terms of Accuracy, Precision, Recall, F-Measure and AUC, in order to predict liver disease occurrence. From the experimental results, the Voting classification method outperforms the other ones with an Accuracy, Recall, and F-measure of 80.1%, a Precision of 80.4%, and an AUC equal to 88.4% after SMOTE with 10-fold cross-validation and, thus, it constitutes the main proposition of this study. Finally, our proposed model (i.e., Voting) shows better accuracy compared to matched published research works based on the dataset [37] with the same features we relied on.

In future work, we aim to re-consider the liver disease manifestation methodology following, first, the ROPE (region of practical equivalence) analysis to test whether a feature is significant (in the sense of important enough for the liver disease risk prediction) and, secondly, extend the machine learning framework by using deep learning methods and comparing the results on the aforementioned metrics.