Liver Disease Prediction – A Comparative Study

Using Machine Learning Algorithms

Soumyadip Chanda1 and Supriya Sarkar2 and Saurabh Banerjee3

1,2
Undergraduate Researcher, NSHM Institute of Engineering and Technology, Durgapur, India
3
Assistant Professor, NSHM Institute of Engineering and Technology, Durgapur, India
[Link]@[Link]

Abstract. Liver diseases are the deadliest orders of several countries. These liver
disorders are increasing among the population due to drinking alcohol, inhaling
poisonous fumes and taking contaminated food and drugs, causing a huge loss of
our population. There is ongoing research on datasets of patients suffering from liver
diseases to develop predictive models that can help determine liver disorders. This
dataset was used for prediction by using classification algorithms, thus reducing the
burden on medical practitioners. This disease is spreading like an epidemic in India
where out of every five adults, one is getting affected. There are a couple of
expensive tests that are done to diagnose liver diseases. To make it affordable and
reliable for common people and also to reduce the work load in medical field, we
have studied on liver disease dataset. Here, several machine learning algorithms are
implemented to evaluate the best performance out of it.
Our purpose is to do a comparative study on different classifications techniques and
judge them under certain measurement techniques like - confusion matrix, accuracy,
precision, recall, f1 score, etc. Thus, we will pick up the best algorithm that is giving
the maximum accuracy and do analysis on it. In this paper, some machine learning
algorithms were implemented to evaluate their performance. The idea is to compare
different classification methods with aforementioned evaluation metrics. Thus, we
choose the best-performing algorithm in terms of having the highest accuracy and
do the analysis based on it.

Keywords: Liver Disease, Machine Learning, Logistic Regression, Predictive

modeling, Decision Tree, Healthcare analysis.

1. Objective of our Research

To identify the most effective machine learning algorithm for liver disease prediction
by comparing multiple models based on accuracy, precision, recall, specificity, F1-
Score, confusion matrix and execution time, ensuring optimal performance for early
diagnosis in comparison with existing traditional methods.

2. Introduction
Chronic liver disease has increased over the years in many countries, with India being
a major center. Recent lifestyle behaviors such as consumption of junk food and
alcohol now affect one in five adults. A human body contains the most vital
component in the organ referred to as the liver. The liver decomposes insulin. The
2 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

liver in glucuronidation metabolizes bilirubin, assisting its excretion into bile. It is

also responsible for the decomposition and removal of numerous unwanted
substances. It plays a significant part in modifying harmful stuffs. Also, it plays a
significant role in the destruction of drugs. This process is called ‘Drug Metabolism’.
It has a weight of 1.3 kilograms. Human liver is comprised of two large sections
known as the right lobe and the left lobe. It lies above the stomach, close to the
pancreas, with other organs aiding in digestion and nutrients absorption. The role of
the Liver is to eliminate dangerous elements from stomach and other body parts
which flow via blood. Human liver diseases occur when the functioning of the liver is
compromised. These diseases have complicated and diverse development due to
several factors that affect susceptibility. Some factors include sex, race, genetic
makeup, diet, virus, toxins and alcohol, BMI and other related illnesses (diabetes,
etc.). Liver disease remains one of the most serious health issues affecting a huge
portion of the entire population around the globe. Early diagnosis and accurate
analysis of liver diseases can prevent the patients from having some lethal damage so,
accurate outcomes can reduce the pressure on the healthcare system. In developed
countries, a lot of young people and one-third of adults have to suffer from non-
alcoholic fatty liver disease – a rapidly increasing health issue. However, the machine
learning methods have been demonstrated with great results in the area of liver
disease forecasting and classification using patient data. These methods can identify
patterns and predict outcomes by using complex algorithms to examine and learn
from several large datasets. Prediction and classification of liver diseases are rapidly
progressing fields of study using ML methods, where each day sees efforts to
improve precision and reduce medical expenses.

2.2 Overview of Liver Disease

Liver disease is a condition whereby the liver does not perform its tasks normally
and, as a consequence, leads to illness. When the patient is suffering from liver
disease, the liver does not perform its essential functions which can significantly
impact health because the liver performs many necessary functions in the body. This
all-encompassing term encompasses several complications that impair the
functioning of the liver. If only one-fourth of the liver is active while the remaining
part is damaged, then it will quite obviously lose some of its potency. The largest
organ in the human body is the liver, categorized as a gland, as it produces and
secretes bile, among other responsibilities. It resides in the upper abdominal cavity
and remains covered with the rib cage. Two distinct lobes of the human liver are
joined by small lobules. There are two separate sources of blood supply in the liver
cells. A blood vessel (commonly known as the hepatic artery) transports oxygenated
blood from the heart, whereas another blood vessel (commonly known as the portal
vein) brings nutrients-rich blood from the intestines of our body. In general, veins are
responsible for transporting blood from various parts of the body, back to the heart.

2.3 Sources of Liver Disease

There are several factors that can cause human liver diseases and can be listed as
follows –
 Liver Disease Prediction 3

2.1.1 Infection - Infections caused by parasites and virus in the liver contribute to
inflammation or swelling and, subsequently impede the working condition of the
liver. The virus that can cause damage to the liver is mainly exposed through either
semen or blood and is therefore mostly from contaminated food, polluted water, or
touching a person, who is already infected. Types of infections in liver that may
affect anyone are – Hepatitis A, B and C.

2.1.2 Immunity-related disorder - In some diseases, the body's immune system gets
anomalous and begins to attack other parts of the body. In such conditions, the liver
gets affected too. These health issues might possibly arise because of Auto-immune
Hepatitis. This can also be Primary Sclerosing Cholangitis or Primary Biliary
Cholangitis.

2.1.3 Genetic inheritance - An abnormally inherited gene can lead to the buildup of
various substances in our liver that might result in liver damage. Some examples of
genetically inherited liver disorders are: Hemochromatosis, Alpha-1 Antitrypsin
Deficiency and Wilson's Disease.

2.4 Chemicals Compounds in Liver

Human liver contains several chemicals such as Albumin, Bilirubin, Globulin,
Aspartate Aminotransferase, Alkaline Phosphatase, etc., necessary for proper
functioning of the liver.
2.4.1 Bilirubin - It is yellowish a compound produced during the normal catabolism
of heme in vertebrates. It is excreted through bile and urine. When its levels are high
in the body, one develops illnesses. The yellowish appearance of bruises and
yellowish discoloration while suffering from jaundice is because of bilirubin.
Derivatives such as stercobilin give the brown color to stool. Urobilin, another
derivative, is responsible for giving urine its straw-yellow color.

2.4.2 Alkaline Phosphatase - It is present in every tissue of the human bodies, with
its highest concentrations found in human liver, bile duct, intestinal mucosa, bones,
placenta and kidney. There are two types of alkaline phosphatase isozymes that are
present in serum - hepatic (in liver) and skeletal. At early age, the predominant
source of alkaline phosphatase in human body is skeletal. These alkaline phosphates
can be found in most mammals including humans.
 ALPI: A protein in the intestine with a weight of 150 kDa.
 ALPL: This enzyme is traced in bones, liver and kidney, but not specifically to
any tissue.
 ALPP: Known as the Regan isozyme and is derived from the placenta.
 GCAP : This is a type of germ cell in our body.

2.4.3 Aspartate Aminotransferase - It is an enzyme whose levels are lifted up in

human liver as well as in the heart. The muscles and kidneys contain this enzyme,
although in very minimal quantites. It occurs in trace amounts in human blood.
Whenever cells from the muscles or liver are broken, this enzyme is secreted into the
4 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

main bloodstream. Thus, conducting an AST test can help us monitor or detect any
damage or dysfunction in the liver.

2.4.4 Albumin - These are proteins that are globular in nature. Albumin is abundant
in serum and constitutes the major portion of protein in our blood. Its primary role is
to regulate as well as sustain oncotic pressure in blood. Additionally, it binds with
cations, bilirubin and fatty acids.

2.4.5 Globulin - These are also protein globules, It is insoluble in pure water but gets
dissolved in the solutions of weak salts. The liver produces certain globulins. In
normal human blood, the rate of globulin absorption generally lies between 2.6 and
3.5 grams per deciliter. Several kinds of globulins exist in human body - alpha 1,
alpha 2, beta and gamma. Excessive amount of chemical production in the kidneys
can often lead to imbalances in our bodies, which in turn may result in liver diseases
also.

3. Literature Review

In the thesis by M.B. Priya (2018)[1], liver patient datasets are investigate for
building classification models in order to predict liver disease. This thesis
implemented a feature model construction and comparative analysis for improving
prediction accuracy of Indian liver patients in three phases. In first phase, min max
normalization algorithm is applied on the original liver patient datasets collected
from UCI repository. In liver dataset prediction second phase, by the use of PSO
feature selection, subset (data) of liver patient dataset from whole normalized liver
patient datasets is obtained which comprises only significant attributes. Third phase,
classification algorithms are applied on the data set. In the fourth phase, the accuracy
will be calculated using root mean Square value, root mean error value. J48
algorithm is considered as the better performance algorithm after applying PSO
feature selection. Finally, the evaluation is done based on accuracy values. Thus
outputs shows from proposed classification implementations indicate that J48
algorithm performances all other classification algorithm with the help of feature
selection with an accuracy of 95.04%.

A.K.M Sazzadur Rahman (2019)[2], in his study, used six ML algorithms - Logistic
Regression, K Nearest Neighbors, Decision Tree, Support Vector Machine, Naïve
Bayes, and Random Forest. The performances of different classification techniques
were evaluated on different measurement techniques such as accuracy, precision,
recall, f-1 score, and specificity. The analysis result shows that the Linear Regression
has achieved the highest accuracy. Moreover, their study mainly focused on the use
of clinical data for liver disease prediction and explores different ways of
representing such data through analysis.

P.C. Sen (2020)[3] in his research, has shown the implementation of machine
 Liver Disease Prediction 5

learning for prediction of liver diseases in human bodies. Supervised learning is one
of two broad branches of machine learning that makes the model enable to predict
future outcomes after they are trained based on past data where we use input/output
pairs or the labeled data to train the model with the goal to produce a function that is
approximated enough to be able to predict outputs for new inputs when introduced to
them. Supervised learning problems can be grouped into regression problems and
classification problems. A regression problem is when outputs are continuous
whereas a classification problem is when outputs are categorical. This paper tries to
compare different types of classification algorithms precisely widely used ones on
the basis of some basic conceptions

Chronic liver disease is one of the principal causes of death affecting large portions
of the global population. An accumulation of liver-damaging factors deteriorates this
condition. Obesity, an undiagnosed hepatitis infection, alcohol abuse, coughing or
vomiting blood, kidney or hepatic failure, jaundice, liver encephalopathy, and many
more disorders are responsible for it. Thus, immediate intervention is needed to
diagnose the ailment before it is too late. The work by M. Sarker (2021)[4] aims to
evaluate several machine learning algorithm outputs, namely logistic regression,
random forest, XGBoost, support vector machine (SVM), AdaBoost, K-NN, and
decision tree for predicting and diagnosing chronic liver disease. The classification
algorithms are evaluated based on various measurement criteria, such as accuracy,
precision, recall, F1 score, an area under the curve (AUC), and specificity. Among
the algorithms, the random forest algorithm showed better performance in liver
disease prediction with an accuracy of 83.70%. Furthermore, the random forest
algorithm also showed better precision, F1, recall, and AUC metrics. Hence, random
forest is considered the best algorithm for early liver disease prediction.

There has been a rapid growth in the use of automatic decision-making systems and
tools in the medical domain. By using the concepts of big data, deep learning, and
machine learning, these systems extract useful information from large medical
datasets and help physicians in making accurate and timely decisions regarding
predictions and diagnosis of diseases. In this regard, the study by Neha Tanwar
(2021)[5] provides an extensive review of the progress of applying Artificial
Intelligence in forecasting and detecting liver diseases and then summarizes related
limitations of the studies followed by future research.

Recently liver diseases are becoming most lethal disorder in a number of countries.
The count of patients with liver disorder has been going up because of alcohol intake,
breathing of harmful gases, and consumption of food which is spoiled and drugs.
Liver patient data sets are being studied for the purpose of developing classification
models to predict liver disorder. This data set was used to implement prediction and
classification algorithms which in turn reduce the workload on doctors. In the work
by Srilatha Tokala (2023)[6], the author proposed apply machine learning algorithms
to check the entire patient’s liver disorder. Chronic liver disorder is defined as a liver
6 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

disorder that lasts for at least six months. As a result, they have used the percentage
of patients who contract the disease as both positive and negative information. They
have processed liver disease percentages with classifiers, and the results are
displayed as a confusion matrix. They proposed several classification schemes that
can effectively improve classification performance when a training data set is
available. Then, using a machine learning classifier, good and bad values are
classified. Thus, the outputs of the proposed classification model show accuracy in
predicting the result.

4. Dataset Description

The publicly available medical records of liver patients provide the data for this paper
and it comprises data of 30,691 individuals (22,888 males and 7,803 females) and 11
columns (attributes – Patient’s age, Gender of the patient, Total Bilirubin count in
body, Direct Bilirubin count in body, Alkphos Alkaline Phosphatase, Sgpt Alamine
Aminotransferase, Sgot Aspartate Aminotransferase, Total Proteins count, Albumin,
Albumin and Globulin Ratio and the patient status i.e., the Result/Target Variable).

Fig. 1. Frequency Distributions of the attributes present

4.1 Features Importance Analysis

As a feature engineering step is highly needed to understand, combine or transform

certain features and the relationships between them, we have used a correlation
matrix. This helps us identify the underlying dependencies among features and
determine their relevance to the target variable. Here, at first, we have imported the
entire dataset and examined the feature distributions. After that, we computed the
 Liver Disease Prediction 7

necessary correlation matrix using [Link]() [here, ‘df’ refers to the dataset used],
which is used to interpret the feature relationships (highly correlated, weakly
correlated or negatively correlated). This is very useful in order to analyze
correlations between the individual features and the target variable and identify the
redundant features.
Correlation values range from -1 to 1, indicating the strength and direction of the
relationship between the features as follows –
4.1.1. 1.0 : Perfect positive correlation (features increase together)
4.1.2. -1.0 : Perfect negative correlation (one feature increases while the other
decreases)
4.2.3. 0 : No Correlation (features are independent)

Fig. 2. Correlation Matrix

4.2 Key Insights

We can infer the following observations from the above diagram (Fig.2.) –
4.2.1 Highly Correlated Features – Features like Total Bilirubin (TB) and Direct
Bilirubin (DB) typically show very high correlation [0.876] since DB is a part of TB.
Also, Albumin (ALBA) is highly correlated with features like Total Proteins (TP)
[0.764] and A/G Ratio [0.674]. These features might be redundant and could lead to
multicollinearity issues in models like Logistic Regression.
4.2.2 Negatively Correlated Features – Features like TB [-0.223], DB [-0.247],
AAP [-0.18], SgptAA [- 0.164], SgotAA [-0.157] show negatively weak correlation
with the target variable. This suggests that higher values of these features might
reduce the likelihood of liver disease and also other factors might be influencing the
disease outcome. Fig. 2.
4.2.3 Weak Correlated Features Correlation
– Age [-0.0047] and Gender [-0.0032] have very
weak correlation, indicating age andMatrix
gender might not be significant predictors.

4.3 Inferences

4.3.1 Feature Selection – TB and DB are highly correlated, so one might be dropped
to avoid redundancy.
4.3.2 Weak Correlations with Target – Advanced feature importance methods (e.g.,
SHAP, etc.) can be used to confirm the most significant predictors.
8 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

4.3.3 Potential Feature Engineering – Combining correlated features like TP, ALBA
or using feature transformation techniques might improve model performance.

5. Methodology
We used various machine learning algorithms in our study for predicting liver disease.
This allowed us to detect the possibility of having liver disease based on the
individual characteristics of the patients, gaining a better understanding and the
effectiveness of each of these algorithms. We used measurements such as confusion
matrix, precision, accuracy and recall to evaluate performance. These tell us about the
overall success rate of the models while providing information on how good the
models are at correctly classifying positive cases of liver disease.
Experimental study demonstrated that some of the algorithms performed out there in
particular aspects. For instance, the accuracy of Decision Tree as well as Random
Forest models was remarkably high. The Gradient Boosting Classifier was remarkable
for its effective recall performance. Logistic Regression as well as KNN showed
strong performance, suggesting that these are both really applicable in different
measurements. This analysis highlights the need to evaluate different algorithms to
get an optimal model of predicting liver disease based on the dataset and goals of
analysis.

5.1 Data Preprocessing

5.1.1 Removing Missing Values – Missing values can arise because of incorrect data
entry, sensor failures or incomplete records. Handling them is essential to prevent
biases in model training.
Steps :
 Checked for Missing Values : Used ‘[Link]()’ [‘df’: dataset] to
identify missing values in each column.

Strategies Used :
 If missing values were few (<5%), we used mean/median imputation for
numerical features (e.g., replacing missing albumin levels with the median).
 If missing values were many (>30%), we considered dropping the column if
it was not that crucial.
 For categorical features like Gender, we used mode imputation

5.1.2 Handling Duplicate Data - Duplicate records can distort model predictions and
inflate accuracy falsely.
Steps :
 Used ‘[Link]().sum()’ to check for duplicate rows.
 Removed exact duplicates using ‘df.drop_duplicates(inplace=True)’.
 Verified that removing duplicates did not lead to data loss, ensuring class
balance was maintained.
 Normalization of numerical features standardized the input data.
 Liver Disease Prediction 9

 The entire dataset is divided into two portions - training set contains 80
percent of the data and testing set contains 20 percent of the data.

5.1.3 Handling Outliers – Outliers are the extreme values that can skew model
performance, especially in medical datasets where biochemical indicators vary within
a range.
 Visualization : Used box plots and histograms to detect outliers in several
features
 Identification : Calculated the IQR and removed the extreme values
5.1.4 Handling Class Imbalances – We had found out that the dataset we were
working with, was imbalanced, where, the training class of X features had
significantly more training samples than the training class of Y labels. Such
imbalances could lead to biased model predictions, where the classifier favours the
majority class more. In order to handle such imbalances, we have used ADASYN
(Adaptive Synthetic Sampling for oversampling because of the following reasons –
 It generates synthetic samples for the minority class (Y) instead of just
duplicating existing ones.
 It focuses more on the difficult-to-learn samples, improving model
generalization.
 It helps in reducing bias towards the majority class.
Steps :
 Checked the overall class distribution primarily.
 Imported ADASYN using imblearn.over_Sampling.ADASYN and applied to
the training data
 Verified the new class distribution to ensure balance.
Outcomes :
 The dataset became balanced, reducing the overall bias in model
predictions.
 The model learnt more effectively, improving classification performance for
both the classes.

5.2 Machine Learning Models Applied

5.2.1 Logistic Regression (LR)

It is a mathematical model (Fig.3.) used to describe the possibility of a specific
outcome or an event happening: pass or fail. In the most basic form, it is a logistic
approach to predicting a binary outcome, though with many complex extensions to
choose from. A binary logistic regression model will consist of both dependent and
independent variables that encompass two possible outcomes, exemplified by
pass/fail status, functioning as an indicator variable with designated values of 0 and 1.

Fig. 3. Linear Regression Residual Plot

10 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

5.2.2 K-Nearest Neighbors (KNN)

It is a classification algorithm based on distance. The K-Nearest Neighbors algorithm
is an easy, non-parametric and flexible approach used in classification and regression.
The 'k' closest data points to a given input point are determined based on the chosen
distance metric: such as Euclidean distance, Manhattan distance, etc. The algorithm
labels the input point based on the majority label among its neighbors in classification
or averages their values in regression. KNN needs little training, making it ideal for
small to medium datasets. Although it can be very computationally intensive method
with reduced efficiency with large datasets because all the training information is
reserved and distance is computed at every prediction, the choices of 'k' and distance
metric may affect the accuracy and performance of KNN.

5.2.3 Decision Tree (DT)

This is a decision rule-based model that does not rely on any parameters. In machine
learning, Decision Tree is an algorithm based on supervised learning to classify and
perform regression. It operates by dividing the data into branches based on the values
of characteristics, creating a decision tree model (Fig.4). Each internal node is an
attribute, each branch is a decision rule and each leaf node is a class or outcome.
Building the tree is making divisions that maximize a criterion like Gini impurity or
entropy, which enhances the uniformity of the data along the branches related to the
target variable. Decision Trees are very interpretable and can also handle numeric as
well as categorical data, but are generally prone to overfitting. Other frequently used
strategies include the use of trimming, limiting tree depth and use of ensemble
techniques, such as Random Forest.

'

Fig. 4. Predictive Modeling with Decision Trees

5.2.4 Random Forest (RF)

The random forest algorithm constructs different subsets of data and combines the
predictions for each of the decision trees that have been constructed. At the final
stage, it makes an election of the best solution through a voting process. The
information has been organized and set apart into testing data and training datasets.
Twenty percent of the data has been allocated to testing. A person draws lines
connecting data points within a subgroup, then subgroups to form a group and
continues this pattern. The structure is similar in shape to a tree. The hyperplane finds
the maximum distance to the closest training data point which provides a reliable
mean to discriminate classes.
 Liver Disease Prediction 11

5.2.5 Gradient Boosting Classifier (GBC)

A boosting method which seeks to decrease error, enhancing precision. The GBC is
one of the robust machine learning algorithms used for classification purposes. It
builds a series of ensemble of weak learners (Fig.4). Each succeeding tree is a
decision tree that aims at correcting the mistakes of earlier trees by minimizing the
loss function, such as log-loss in classification. The model improves incrementally by
giving more emphasis to the samples that were misplaced by previous models. This
therefore gives a strong model that often reaches high accuracy and manages complex
patterns in the data appropriately. For its steep computational cost, Gradient Boosting
produces quite solid predictive performance, which makes it widely used in
healthcare, finance and risk analysis circles.

Fig. 5. Gradient boosting – An ensemble of weak learners

5.3 Confusion Matrix

In machine learning, a “confusion matrix” is another name for a statistical tool that is
employed when measuring the efficiency of the classification model. It helps in
predicting and comparing how classifications were made and how they were achieved
during the evaluation of different classes. The matrix for the cross tab includes four
key components (Fig. 5) as follows :
5.3.1 True Positives (TP) : The actual results predicted for positive class that were
predicted as true positive class.
5.3.2 True Negatives (TN) : The true outcomes predicted as negative who were
classified to be true negative class.
5.3.3 False Positives (FP) : Missed positives where the actual prediction should be
false but it predicts as true.
5.3.4 False Negatives (FN) : Missed negatives predicted (Type II Error where
negatives are wrong emphasized).

This structure helps to obtain the most important quantitative indicators, including
accuracy, precision, sensitivity, or recall and the F1 score. This is particularly
pertinent with respect to such datasets that contain imbalanced cases, where reliability
takes this metric alone.

Fig. 6. Confusion Matrix- Diagnostic Accuracy

12 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

5.4 Accuracy
The accuracy metric in machine learning measures the correctness with which a
model predicts results. Often, the accuracy is measured in terms of a ratio of the
number of instances correctly predicted to the number of observations in general -
hence often reported in percentage form. It is quite popular as it is easy to interpret
and gives a quick overview of how well a model performs. Accuracy should be
supplemented with more metrics, such as support, precision, recall and F1-score in the
case of a class-imbalanced distribution while attempting to carry out a comprehensive
evaluation.

Acurracy=
5.5 Precision
It is a measure of the accuracy that how frequently does the model gives positive
predictions. This is more appropriately defined as the amount of correct positive
identifications relative to all positive identifications, including both correct and
incorrect ones. Precision is crucial when the economic term of ‘false positives’ is
considerably high. Nevertheless, to have any meaningful appreciation of what a
model really does, other measures such as recall, etc. have to be used together with
the precision measure.
Precision=
5.6 Recall
Recall in machine learning evaluates the ability of a model to achieve the good
recognition of all relevant instances of a given class, especially in classification tasks.
The following formula is obtained based on the number of true positive divided by the
sum of the true positive predictions and false negative ones. High recall refers to how
well the model picks out most of the true positive cases and it is valuable in situations
where one wants to reduce false negatives. However, recall cannot be allowed to
comprise accuracy to maintain relevance in the predictions.

Recall=

5.7 F1-Score
It is a crucial metric in machine learning, especially in classification, that helps to
maintain a fair balance between precision and recall. This is a measure that combines
precision and recall by the use of harmonic mean. This measure is particularly
valuable for description of imbalanced datasets where it punishes the model that
shines on one metric but fails on the other. An F1-score of 1 shows perfect precision
and recall while a score of 0 means the model lacks both.

F1-Score=
5.8 Support
In machine learning, support is a measure of how many times each class label has
occurred in the dataset. It shows how frequent each class is and is particularly useful
for imbalanced datasets with unequal sample sizes for different classes. Support is
often reported with other metrics, which include recall, precision and F1-score within
classification reports.
 Liver Disease Prediction 13

6. Results and Discussions

Fig.7. ROC curves of the models used in our project

In our project, we’ve used Logistic Regression, Decision Tree, Random Forest, K-
Nearest Neighbors and Gradient Boosting Classifier. A comparison of these models
has been done in terms of accuracy of both training and testing datasets using metrics
like precision, recall and confusion matrix. This will allow us to assess the strengths
and weaknesses of each model in predicting liver disease. As we can see in the above
figure (Fig. 6) that among the tested models, Random Forest and Gradient Boosting
Classifier have the best accuracies; therefore, they can be good selection for the
prediction of liver disease. Overall, Gradient Boosting Classifier has high
performance in accuracy and precision measures, pointing to its competence in both
detecting true positives and true negatives.

Both Decision Tree and Logistic Regression models showed noteworthy accuracy
despite some misclassifying cases. Logistic Regression demonstrated high sensitivity,
suggesting its efficiency in detecting individuals with liver disease. But this model
had a slightly lower precision, indicating potential inaccuracies in its predictions.
KNN performed decently but failed when compared to ensemble techniques such as
Random Forest and GBC, perhaps due to complex interrelation between features in
the dataset. The application of distance-based metrics in KNN might have led to a loss
of the ability to generalize in scenarios where feature values overlap.
Decision Tree and Logistic Regression models have certain observations with notable
accuracy, though a few were misclassified. In Logistic Regression model, sensitivity
results were high, hence suggesting the detection efficiency for liver disease patients
but the precision result was a bit lower, hence it may have errors in their prediction.
KNN performed satisfactorily but lagged behind in a comparison to ensemble
methods like RF and GBC, probably because the complex interrelation of features in
this data made it a complex problem. The distance-based measure on which KNN
relies on, might contribute to generality loss in the cases where values overlap.
14 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

Table 1. Comparison Chart of the models based on the metrics

Algorithm Accuracy Precision Recall F1-Score

LR 63.0% 83.0% 60.5% 70.0%
KNN 93.7% 97.6% 93.5% 95.5%
DT 98.4% 99.0% 98.8% 98.9%
RF 99.6% 99.8% 99.6% 99.7%
GBC 99.6% 99.7% 99.7% 99.7%

In summary, both Random Forest and Gradient Boosting Classifier have performed
quite efficiently and should be further improved to be applied in the prediction of
liver disease since they both exhibit balanced and good performance on accuracy,
precision, as well as recall. This potentiality of using machine learning to help early
diagnose and manage liver disease can be gained since these models work perfectly.
Gradient Boosting Classifier surpasses all other algorithms in terms of accuracy and
other performance metrics up to a 99.6% accuracy rate (Fig.7). Therefore, in this case,
GBC is the best algorithm that can be used for predicting liver disease.

Fig.8. Performance of the algorithms used

6.1 Key Insights:

6.1.1 Tree-Based Models Dominate : Ensemble methods (RF, GBC) and Decision
Trees surpassed other algorithms, probably because of their capacity to manage
intricate, non-linear relationships within the data.
6.1.2 High Precision and Recall : RF and GBC reached nearly perfect precision
(>99%) and recall (>99%), reflecting a slight misclassification of both healthy and
diseased patients.
 Liver Disease Prediction 15

6.2.3 LR Limitations : Logistic Regression had difficulties with recall (60. 5%), not
recognizing numerous true positive cases, which is vital for medical diagnoses.
6.2.4 Sample Heading (Third Level). Only two levels of headings should be
numbered. Lower level headings remain unnumbered; they are formatted as run-in
headings.
The results of comparison among the models based on the evaluation metrics are -

Fig.9. Best Performing Algorithm- Gradient Boosting Classifier

7. Conclusion
This liver disease prediction paper incorporates several classification algorithms:
Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest and
Gradient Boosting Classifier for comparison of possible successfulness in liver
disease diagnosis at a higher accuracy rate. The performance assessment took place
using precision, accuracy, recall and confusion matrix that gives an overview of the
robustness of these models in general. Such models pointed out their specific benefits;
for example, Random Forest and Gradient Boosting had the highest recall values so
they can better identify positive instances while showing high accuracy for the
Decision Tree model. This comparison explains why certain algorithms should be
chosen over others which are specifically applied according to the distinct objectives
and characteristics of the provided data.
In general, this paper provides valuable knowledge about the prediction of liver
diseases in human bodies using certain machine learning models that seem to be
useful in improving medical diagnosis.
16 Soumyadip Chanda and Supriya Sarkar and Saurabh Banerjee

8. Future Work
For the future, there are multiple ways to improve and expand this model of
prediction of liver diseases. Machine learning methods that are more advanced than
those used in this paper could bring even greater accuracy in the predictions when the
datasets involved are larger in size. Ensemble learning with algorithms that combine
the strengths of different approaches may yield additional performance metric
improvements. Exploring techniques like recursive feature elimination for feature
engineering and selection could help point to the most important features, ultimately
streamlining models and enhancing performance. The model could also be adapted to
predict different kinds of liver diseases, which means the model has to use a
multiclass classification method.

In conclusion, patient records from a variety of demographics in the dataset could

improve the model's generalizability and relevance in a wider clinical setting.

References

1. M. B. Priya, P. L. Juliet, and P. Tamilselvi, “Performance analysis of liver disease

prediction using machine learning algorithms,” International Research Journal of
Engineering and Technology (IRJET), vol. 5, no. 1, pp. 206–211, 2018.
2. Rahman, A. Sazzadur, F. J. M. Shamrat, Zarrin Tasnim, Joy Roy, and Syed Akhter
Hossain. "A comparative study on liver disease prediction using supervised machine
learning algorithms." International Journal of Scientific & Technology Research 8, no.
11 (2019): 419-422.
3. P. C. Sen, M. Hajra, and M. Ghosh, “Supervised classification algorithms in machine
learning: A survey and review,” in Emerging technology in modelling and graphics.
Springer, 2020, pp. 99– 111.
4. M. Ghosh, M. Raihan, M. Sarker, M. Raihan, L. Akter, A. K. Bairagi, S. S. Alshamrani,
and M. Masud, “A comparative analysis of machine learning algorithms to predict liver
disease.” Intelligent Automation & Soft Computing, vol. 30, no. 3, 2021.
5. Neha Tanwar and Khandakar Faridar Rahman 2021 IOP Conf. Ser.:
Mater. Sci. Eng. 1022 012029.
6. Tokala, Srilatha; Koduru Hajarathaiah; Sai Ram Praneeth Gunda; Botla, Srinivasrao;
Nalluri, Lakshmikanth; Nagamanohar, Pathipati; Anamalamudi, Satish; Enduri, Murali
Krishna. International Journal of Advanced Computer Science and Applications; West
Yorkshire Vol. 14, Iss. 2, (2023). DOI:10.14569/IJACSA.2023.0140299.
7. "Liver Disease Patient Dataset 30K
data",[Link] disease-patient-dataset /data
(accessed Nov. 28, 2024).