IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 2, June 2024, pp. 1980~1991

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i2.pp1980-1991  1980

Comparative analysis of explainable artificial intelligence

models for predicting lung cancer using diverse datasets

Shahin Makubhai, Ganesh R. Pathak, Pankaj R. Chandre

Department of Computer Science and Engineering, MIT School of Engineering, MIT Art Design and Technology University,
Pune, India

Article Info ABSTRACT

Article history: Lung cancer prediction is crucial for early detection and treatment, and
explainable artificial intelligence (XAI) models have gained attention for their
Received Jun 27, 2023 interpretability. This study aims to compare various XAI models using diverse
Revised Oct 12, 2023 datasets for lung cancer prediction. Clinical, genomic, and imaging data from
Accepted Jan 1, 2024 multiple sources were collected, preprocessed, and used to train models such
as logistic regression (LR), support vector classifier (SVC)-linear, SVC-radial
basis function (RBF), decision tree (DT), random forest (RF), adaboost
Keywords: classifier, and XGBoost classifier. Preliminary results indicate that RF
achieved the highest accuracy of 98.9% across multiple datasets. Evaluation
Comparative analysis metrics such as accuracy, precision, recall, and F1 score were utilized, along
Diverse datasets with interpretability techniques like feature importance rankings and rule
Explainable artificial extraction methods. The study's findings will aid in identifying effective and
intelligence interpretable AI models, facilitating early detection and treatment decisions
Lung cancer prediction for lung cancer.
Support vector machines
This is an open access article under the CC BY-SA license.

Corresponding Author:
Shahin Makubhai
Department of Computer Science and Engineering, MIT School of Engineering
MIT Art Design and Technology University
Loni Kalbhor, Pune, India
Email: [email protected]

1. INTRODUCTION
Lung cancer is a significant cause of death worldwide, and early detection is crucial for successful
treatment. With the increasing availability of medical imaging data and the advances in machine learning
algorithms, there has been an increasing fascination with employing artificial intelligence (AI) models for the
diagnosis and prognosis of lung cancer [1]. However, the lack of interpretability and transparency in traditional
machine learning models can limit their applicability in medical decision-making, where the ability to explain
model predictions is essential. Explainable artificial intelligence (XAI) aims to cultivate machine learning models
to provide explanations for their predictions, enabling users to understand and trust the model's decision-making
process. In the context of lung cancer prediction, XAI models can help clinicians to interpret the results, identify
potential biases or errors, and make more informed decisions [2]. Lately, there has been an increasing amount of
research exploring XAI models for predicting lung cancer, utilizing various imaging methods such as
computerized tomography (CT) scans and magnetic resonance imaging (MRI) scans [3]. These models leverage
different approaches, such as decision trees (DT), neural networks, and deep learning, to develop accurate and
interpretable models. However, despite the promising results, the development of XAI models for lung cancer
prediction still faces several challenges [4], [5]. These encompass the necessity for extensive and varied datasets,
the challenge of maintaining a balance between accuracy and interpretability, and the intricate nature of the
fundamental biological processes linked to lung cancer. Lung cancer stands as a primary contributor to cancer-

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  1981

related fatalities globally [6], [7]. Early detection and accurate diagnosis are critical for improving patient
outcomes and survival rates. Machine learning models have shown great promise in aiding physicians with lung
cancer diagnosis, However, the absence of clarity and comprehensibility in these models has impeded their broad
acceptance and utilization in clinical settings. To address this issue, there has been growing interest in developing
XAI models for predicting lung cancer. XAI models are designed to provide not only accurate predictions but
also clear and interpretable explanations for their decisions, allowing physicians to better understand and trust the
model's output [8], [9]. However, the performance and interpretability f XAI models depend heavily on the quality
and diversity of the data used to train them. In this manuscript, we offer an examination that explores the use of
diverse datasets for developing XAI models for predicting lung cancer [10]. We evaluate the performance of XAI
models trained on different datasets, including traditional medical imaging datasets, as well as non-traditional
sources of medical data, such as electronic health records and patient-generated data [11], [12]. We analyze the
impact of data diversity on model performance and interpretability, and demonstrate the importance of
incorporating diverse data sources to improve the accuracy and transparency of XAI models [13], [14]. Our study
aims to contribute to the development of reliable and interpretable XAI models for lung cancer prediction, with
the potential to revolutionize clinical decision-making and improve patient outcome.

2. LITERATURE SURVEY
Patra [15] reveals that various machine learning algorithms have been employed in the past for the
prediction of lung cancer. The author highlights the importance of early diagnosis of lung cancer and the limitations
of conventional diagnostic techniques. Machine learning algorithms provide a hopeful prospect for precise and
timely identification of lung cancer. The analysis encompasses research that has employed diverse machine learning
classifiers like support vector machine (SVM), artificial neural network (ANN), DT, random forest (RF), and logistic
regression (LR) in forecasting lung cancer. The author concludes that machine learning algorithms have shown
significant improvement in the prediction of lung cancer, with some studies achieving accuracy rates of over 90%.
Yet, the selection of the classifier, feature curation, and dataset employed can substantially influence the predictive
accuracy. On the whole, the overview highlights the promise of machine learning algorithms in the timely
identification and assessment of lung cancer, underscoring the necessity for additional exploration in this domain.
Kumar et al. [16] suggests utilizing a machine learning method to forecast lung cancer by leveraging
textual data. The author first performs a literature survey to identify previous studies in the field of lung cancer
prediction. The review encompasses research involving both textual and non-textual information,
encompassing medical images and genomic data. The author underscores constraints within current research,
including the absence of interpretability and dependency on restricted datasets. The suggested methodology
employs machine learning methods like feature curation, feature derivation, and classification algorithms for
lung cancer prognosis based on textual data.The author uses the lung image database consortium (LIDC)
dataset, which consists of CT scans and associated radiology reports, as the primary dataset for the study. The
experimental results demonstrate the effectiveness of the proposed approach, achieving an accuracy of 85% in
predicting lung cancer. The author also performs feature importance analysis to identify the most relevant
features for prediction, which can aid in interpretability. In summary, the academic article offers an extensive
review of literature and introduces an innovative method for forecasting lung cancer utilizing textual data. The
empirical outcomes showcase the efficiency of the technique and its prospective use in clinical settings.
Nemlander et al. [17] concentrates on utilizing machine learning methodologies to forecast the
likelihood of lung cancer in individuals who have never smoked, those who smoked in the past, and those
presently smoking, leveraging their responses to an electronic e-questionnaire. The study used a dataset of
20,080 participants who completed the e-questionnaire, out of which 406 participants were diagnosed with
lung cancer. The questionnaire included questions related to smoking history, exposure to secondhand smoke,
respiratory symptoms, and other relevant factors. The study utilized five different machine learning algorithms:
LR, DT, RF, gradient boosting, and SVM. These algorithms were trained and tested on the dataset, evaluating
their performance by analyzing accuracy, precision, recall, and the F1 score. The results showed that all five
machine learning algorithms performed well in predicting lung cancer risk among never smokers, former
smokers, and current smokers. The RF model attained the utmost accuracy at 91.7%, whereas the DT model
exhibited the highest precision, reaching 92.1%. Overall, the study highlighted the potential of using machine
learning algorithms to predict the likelihood of lung cancer across diverse demographics, based on their
responses to an electronic questionnaire. The insights derived from this research could inform the development
of effective screening and preventive strategies for lung cancer.
Abdullah et al. [18] is a study grounded in a review of existing literature, emphasizing the utilization
of machine learning methods for the prediction and categorization of lung cancer. The researcher undertook an
extensive analysis of the current literature within the domain and recognized diverse machine learning methods
applied in predicting and categorizing lung cancer. The study presents a methodology based on correlation
selection, utilizing machine learning approaches to forecast and classify instances of lung cancer. The proposed
Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)
1982  ISSN: 2252-8938

methodology involves three main stages: data preprocessing, feature selection, and classification. In the data
preprocessing stage, the data is preprocessed to remove any missing or noisy data. In the feature selection stage,
the correlation-based feature selection (CFS) method is used to select the most relevant features for the
classification task ultimately, during the classification phase, diverse machine learning methods like DT,
K-nearest neighbors (KNN), and SVM are applied to categorize the lung cancer dataset juxtaposes.
Gulia et al. [19] proposes utilizing machine learning techniques to predict lung cancer, employing
three different classifiers: SVM, RF, and ANN. The dataset utilized in the analysis comprised 32 attributes and
162 occurrences, evenly split between 81 instances of malignant and benign lung tumors. The paper reports an
overall accuracy of 90.74% for the SVM classifier, 87.65% for the RF classifier, and 91.36% for the ANN
classifier. The results suggest that the ANN classifier outperforms other classifiers in terms of accuracy,
sensitivity, specificity, and the area under the receiver operating characteristic (ROC) curve. The article
additionally incorporates a feature selection examination, wherein the researchers employed three distinct
approaches (information gain, CFS, and chi-squared test) to determine the most pertinent features for the
classifiers. The findings indicate that the SVM and RF classifiers attained peak accuracy when employing the
top 10 features, whereas the ANN classifier reached its highest accuracy using the top 14 features. Overall, the
study shows that machine learning-based approaches can be effective in predicting lung cancer, and that the
ANN classifier is particularly promising for this task. However, the small sample size of the dataset used in the
study suggests that further research is needed to validate these results on larger datasets. Makubhai et al. [20]
aims to enhance lung cancer risk prediction using explainable AI techniques. By analyzing a diverse range of
patient data, including lifestyle factors and medical history, the model offers transparent insights for healthcare
professionals. DT, partial dependence plots, and feature importance analysis enable clear interpretation of the
model's predictions. Ultimately, this approach facilitates early detection and informed decision-making in lung
cancer screening and treatment. Shimazaki et al. [21] presents an approach for identifying lung cancer in chest
X-rays through deep learning and segmentation methods. In this study, the author performed a comprehensive
examination of different approaches utilized in identifying lung cancer from chest X-rays in existing literature.
The review includes works that use various techniques such as traditional machine learning, deep learning, and
segmentation. The author notes that traditional machine learning techniques have limitations due to the
complex features present in chest radiographs. Conversely, deep learning techniques have demonstrated
efficacy in identifying lung cancer from chest X-rays. The researcher introduced a deep learning-driven
algorithm that combines a convolutional neural network (CNN) with a segmentation technique to identify lung
nodules. The algorithm was tested on a dataset of chest radiographs and achieved an accuracy of 85.7%. In
summary, the document offers an extensive examination of literature regarding lung cancer identification
through chest radiographs and introduces an innovative deep learning algorithm for the same purpose. Table 1
presents an overview of lung cancer prognosis utilizing the LUng nodule analysis-16 (LUNA16) dataset.

Table 1. Summary for lung cancer prediction using LUNA16 dataset

Paper title Method Data Evaluation metrics Results
Lung Nodule Detection via Optimized 3D-CNN LUNA Sensitivity, false Achieved a sensitivity of
Convolutional Neural Network: Impact of positive rate (FPR) 94.77% and a FPR of 4.27%
Improved Moth Flame Algorithm [22]
Automated pulmonary nodule detection using 3D-CNN with LUNA Sensitivity, FPR Achieved a sensitivity of
3D deep convolutional neural networks [23] multi-task 92.4% and a FPR of 4.31%
learning
Deep Learning Applications in Computed 3D-CNN with LUNA Dice similarity Achieved a dice similarity
Tomography Images for Pulmonary Nodule region coefficient coefficient of 0.84
Detection and Diagnosis: A Review [24] dependence
modeling
Multi-scale convolutional neural networks for Multi-scale 3D- LUNA Accuracy, Achieved an accuracy of
lung nodule classification. In Information CNN sensitivity, 81.1%, a sensitivity of 76.5%,
Processing in Medical Imaging [25] specificity and a specificity of 84.6%
Automated pulmonary nodule detection in CT Faster R- LUNA Sensitivity, FPR Achieved a sensitivity of
images using 3D deep squeeze-and-excitation CNN 94.1% and a FPR of 4.2%
networks [26]
Automated pulmonary nodule detection in CT Two-stage 3D- LUNA Sensitivity, FPR Achieved a sensitivity of
images using deep convolutional neural CNN 92.3% and a FPR of 1.6%
networks. Pattern Recognition [27]
A Two-Stage Convolutional Neural Networks Improved 3D- LUNA Sensitivity, FPR Achieved a sensitivity of
for Lung Nodule Detection [28] CNN 94.1% and a FPR of 3.9%
Lung nodules diagnosis based on evolutionary Genetic LUNA Accuracy, sensitivity, Achieved an accuracy of
convolutional neural network. Multimed Tools algorithm- specificity 83.4%, a sensitivity of 87.5%,
[29] optimized and a specificity of 81.2%
3D-CNN

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1980-1991

Int J Artif Intell ISSN: 2252-8938  1983

3. METHOD
Our study aims to perform a comparative analysis of different XAI models which will help us to
predict and analyase the difference between the accuracies using the list of methods for predicting lung cancer
using diverse datasets. The methodology involves the following steps:
- Step 1-dataset collection and preprocessing: We will collect diverse datasets, including traditional medical
imaging datasets, electronic health records, and patient-generated data. The datasets will be preprocessed
to remove any missing values, standardize the features, and prepare them for model training.
- Step 2-model selection: We'll select diverse XAI frameworks for comparison, including DT, RF, LR, neural
networks, and gradient boosting. These models demonstrate promise in predicting lung cancer and are
acknowledged for producing results that can be interpreted.
- Step 3-model training and validation: We will train each model on the diverse datasets using a cross-
validation approach to ensure generalizability. We will compare the performance of each model based on
various metrics such as accuracy, precision, recall, and F1 score.
- Step 4-model interpretability: We'll assess the explainability of each XAI model employing diverse
methods like feature significance, partial dependence plots, and SHapley additive ex-Planations (SHAP)
values. These methods will aid in comprehending the model's decision-making process and detecting
potential biases or confounding variables.
- Step 5-comparative analysis: We'll conduct a comparative evaluation of the XAI models considering their
effectiveness and comprehensibility. We'll examine how various data sources influence both the
performance and comprehensibility of the models.
- Step 6-discussion and conclusion: We will summarize our findings and discuss the implications of our study
for the development of reliable and interpretable XAI models for predicting lung cancer. We will also
highlight the limitations of our study and suggest future directions for research.
Overall, our methodology will enable us to perform a comprehensive comparative analysis of different
XAI models for predicting lung cancer using diverse datasets. This will help us identify the most effective and
interpretable XAI models for clinical decision-making and improve patient outcomes. Figure 1 shows the AI
based model for predicting lung cancer.

Figure 1. AI-based models and experimental methods applied

4. RESULTS AND DISCUSSION

4.1. Dataset
The contrasting datasets provided include a compilation of data and associated annotations for analysis
sourced from the National Cancer Institute's surveillance, epidemiology, and end results (SEER) dataset, the
LIDC dataset, and the international early lung cancer action program (I-ELCAP) dataset, The Cancer Genome
Atlas (TCGA) dataset, and LUNA16 dataset. Table 2 shows that the summary of various dataset to predict lung
cancer. It is important to note that these datasets have their own unique characteristics and applications, and
their suitability for a particular study depends on the research question and methodology. Therefore,
researchers should carefully consider the characteristics of each dataset before selecting one for their study.

Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)
1984  ISSN: 2252-8938

Table 2. Summary of various dataset to predict lung cancer

Dataset Data type Size Annotations Applications
SEER Clinical data >1.5 million cancer cases None for lung nodules Epidemiology and genetics of lung
cancer
LIDC CT scans 1,018 scans Annotations for lung nodules Developing algorithms for lung
nodule detection and classification
I-ELCAP Clinical data >50,000 patients Detailed annotations on Studying outcomes of lung cancer
and CT scans patient characteristics, screening and treatment
imaging data, and treatment
outcomes
TCGA Genomic data Thousands of cancer patients, None for lung nodules Studying the genetics of lung
including lung cancer cancer
LUNA16 CT scans 888 scans Annotations for lung nodules Developing algorithms for lung
nodule detection and classification

4.1.1. National Cancer Institute's surveillance, epidemiology, and end results dataset
The National Cancer Institute's SEER dataset is a comprehensive source of cancer statistics in the
United States. It contains data on cancer incidence, survival, and mortality, as well as demographic and clinical
information on cancer patients [30]. The SEER database encompasses around 34.6% of the American populace
and comprises data from 18 diverse geographical zones, encompassing urban and rural areas. It holds details
on over 28 million instances of cancer diagnosed between 1975 and 2018.The SEER dataset is used for a wide
range of cancer research, including studies on the epidemiology, genetics, and treatment of cancer. The dataset
has been instrumental in identifying trends in cancer incidence and mortality, as well as in evaluating the
effectiveness of cancer screening and treatment programs. Researchers can access the SEER dataset through
the SEER program's website, where they can download data files or use the SEER*Stat software to analyze the
data. However, the use of the dataset requires careful consideration of ethical and privacy concerns related to
patient data, and researchers must comply with SEER data use agreements and policies.

4.1.2. The lung image database consortium dataset

The LIDC dataset represents an openly available assortment of chest CT scans designed to improve
and evaluate algorithms centered on detecting and diagnosing lung nodules. This compilation was developed
by a collaborative group of researchers across various institutions, including the National Cancer Institute and
the University of Chicago. Within the LIDC dataset, there are 1,018 CT scans gathered from 1,010 patients,
each scan comprising roughly 300 to 400 images. The scans were obtained from seven different medical centers
in the United States and were collected between 2000 and 2007. The dataset includes scans with both low-dose
and standard-dose protocols [31]. The dataset also includes annotations of lung nodules by four experienced
radiologists. The annotations include information on the location, size, and shape of nodules, as well as
information on the presence of spiculation, calcification, and other features that may indicate malignancy. The
markings were conducted following a standardized procedure to maintain uniformity across radiologists.
Alongside the CT scans and markings, the LIDC compilation encompasses metadata like patient demographic
details, scan specifics, and malignancy assessments for nodules provided by each radiologist. The dataset
comes with software utilities for observing the scans and annotations, along with tools for assessing the efficacy
of algorithms in detecting and categorizing nodules. Researchers globally have extensively utilized the LIDC
dataset to create and appraise algorithms focused on identifying and classifying lung nodules. It's been proven
to be a valuable asset in enhancing the precision and dependability of lung cancer screening and diagnosis.
Nonetheless, handling the LIDC dataset demands considerable expertise and computational resources due to
its extensive and intricate nature.

4.1.3. The international early lung cancer action program dataset

The I-ELCAP dataset comprises clinical and imaging information from individuals who underwent
screening for lung cancer utilizing low-dose computed tomography (LDCT). The dataset comprises details
regarding patient demographics, smoking background, imaging records, and clinical results. The I-ELCAP
dataset was established to investigate the efficacy of LDCT in identifying early-stage lung cancer among high-
risk groups, notably heavy smokers. It encompasses information from over 50,000 individuals across 7 distinct
nations who underwent LDCT screening from 1993 to 2005. The dataset is unique in that it includes detailed
annotations on patient characteristics, imaging data, and treatment outcomes. This enables researchers to
explore the results of lung cancer screening and therapy within an extensive and varied population [32]. The I-
ELCAP dataset has been used by researchers to study various aspects of lung cancer screening and treatment,
such as the accuracy of LDCT in detecting lung nodules, the characteristics of nodules detected by LDCT, and
the effectiveness of different treatment options for early-stage lung cancer. Access to the I-ELCAP dataset is

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1980-1991

Int J Artif Intell ISSN: 2252-8938  1985

restricted and requires approval from the I-ELCAP data coordinating center. However, subsets of the dataset
have been made available to researchers for specific studies.

4.1.4. The cancer genome atlas dataset

The TCGA dataset is a comprehensive public resource that provides genomic and clinical data on various
types of cancer, including lung cancer. It was a collaborative effort between the National Cancer Institute and the
National Human Genome Research Institute (NHGRI), with contributions from many other institutions [33]. The
TCGA compilation comprises genetic information regarding both tumor and normal tissues, encompassing DNA
sequencing, RNA sequencing, methylation profiling, and analysis of copy number variations. Additionally, it
contains patient-related clinical information, including demographics, diagnosis, treatment records, and outcomes.
The TCGA lung cancer dataset includes data on various subtypes of lung cancer, including adenocarcinoma,
squamous cell carcinoma, and small cell lung cancer. It includes genomic data on thousands of lung cancer
patients, including somatic mutations, gene expression profiles, and copy number alterations. The TCGA dataset
has been used to study the genetics and biology of lung cancer and to identify new therapeutic targets. It has also
been employed in the creation and assessment of machine learning models to forecast patient results and assess
treatment responses. The TCGA dataset is publicly available and can be accessed through the genomic data
commons data portal. However, working with the dataset requires significant expertise in genomics and
bioinformatics, as well as access to high-performance computing resources.

4.1.5. LUng nodule analysis dataset

The LUNA16 dataset represents a segment of the broader LUNA dataset designed explicitly for the
task of formulating algorithms for detecting and categorizing lung nodules. Within the LUNA16 dataset, there
exists a collection of 888 chest CT scans, openly accessible for research objectives. The dataset originated from
a public competition that tasked participants with crafting algorithms to identify and categorize lung nodules
within the LUNA16 dataset. This compilation comprises CT scans with slice thickness spanning from 0.5mm
to 2.5mm and pixel dimensions ranging between 0.5mm×0.5mm to 0.8mm×0.8mm. The scans were collected
from different institutions and include both low-dose and standard-dose scans. The distinctiveness of the
LUNA16 dataset lies in its incorporation of lung nodule annotations by numerous radiologists, offering a
valuable resource for training and assessing machine learning algorithms geared toward detecting and
categorizing lung nodules. These annotations encompass details regarding the position, size, malignancy level
of nodules, and insights into the confidence associated with the radiologist's diagnosis. The LUNA16 dataset
has been widely used by researchers around the world to develop and evaluate algorithms for lung nodule
detection and classification [34]. It has demonstrated its significance in enhancing the precision and
dependability of lung cancer screening and diagnosis. However, it's important to recognize that working with
the LUNA16 dataset requires considerable skill and computational capabilities owing to its extensive scale and
complex characteristics [35]–[37]. Additionally, the use of the dataset requires careful consideration of ethical
and privacy concerns related to patient data. Table 3 shows that the features of LUNA16dataset [38]–[40].

Table 3. Features of LUNA16 dataset

Feature Description
Patient ID Unique identifier for each patient
Nodule ID Unique identifier for each nodule in the patient's scan
Image DICOM file containing the image data of the nodule
Diameter The diameter of the nodule in millimeters
Series UID Unique identifier for the series that contains the nodule
CAD probability The probability of the nodule being malignant as determined by computer-aided detection (CAD)
X, Y, Z The coordinates of the center of the nodule in the image
Image size The dimensions of the image containing the nodule
Slice spacing The spacing between slices in the image containing the nodule
The malignancy rating of the nodule as determined by radiologists on a scale of 1 to 5, with 1 being benign and
Malignancy 5 being highly malignant.

4.2. Discussions
In our work we have implemented following machine learning classifiers which will help to compare
the different accuracies with each other classifiers and give out the best accuracy compare to all others listed
and the details of implemented classifiers are as follows:
− LR is a statistical method used for classification objectives. Its role involves predicting a binary outcome
(1/0, yes/no, true/false) from a set of independent variables [41]. It's a straightforward and rapid algorithm
that performs effectively with data that can be separated linearly.

Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)
1986  ISSN: 2252-8938

− Support vector classifier (SVC)-linear is a type of SVM algorithm that is used for linearly separable data
[42]. It is a binary classification algorithm that finds the best hyperplane to separate the data points into
different classes.
− SVC-radial basis function (RBF) is another type of SVM algorithm that is used for non-linearly separable
data [43]. It uses a kernel function to map the data into a higher dimensional space, where it can be linearly
separated.
− DT is a tree-based algorithm that is used for both regression and classification problems [44]. It functions
through iterative division of data into smaller segments, relying on features that yield the greatest
information gain. The eventual output is a tree structure consisting of decision nodes and leaf nodes, each
depicting the forecasted outcome.
− RF is a collaborative algorithm that merges numerous DT to enhance prediction accuracy [45], [46]. It
operates by constructing multiple DT using various segments of the data and subsequently averaging the
outcomes to derive the ultimate prediction.
− AdaBoost classifier is another ensemble algorithm that combines multiple weak classifiers to create a strong
classifier [47]. It works by iteratively training a weak classifier on the misclassified data points from the
previou iteration, and then combining the results to make the final prediction.
− XGBoost classifier is a gradient boosting algorithm that is used for both regression and classification
problems [48]. It operates by constructing numerous DT sequentially, with each subsequent tree aimed at
rectifying the mistakes of its predecessor. The outcome comprises an amalgamation of all the trees'
predictions.
Figure 2 show the description on stating that it has used above mentioned packages which will be
useful for carrying out the training and testing dataset models. Figure 3 describes about the summary showing
what data type it is and states about the count of entries made into the dataset. The summary is helpful to
understand about the parameters in details present in dataset. Figure 4 shows the encoded part for categorical
data which is been processed under the categorical feature. Figure 5 shows that the details of train model by
using dataset. We have divided our dataset into train and test dataset and then model applied. Figure 6 shows
that the features of dataset.

Figure 2. Details for python code while importing the packages

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1980-1991

Int J Artif Intell ISSN: 2252-8938  1987

Figure 3. Details about the summary of parameters present in dataset

Figure 4. Applied label encoding on categorical features

The DataFrame contains 309 entries, representing individual records. Each record has information
related to various factors and attributes. The dataset consists of 16 columns, capturing different characteristics
of the individuals. The "GENDER" category denotes the gender of each person. The "AGE" column illustrates
the individuals' ages, presented as whole numbers. The "SMOKING" column records whether an individual
smokes or not, represented by binary values (0 for non-smokers and 1 for smokers). Several other columns
capture specific attributes or conditions. The "YELLOW_FINGERS" column signifies whether an individual
has yellow fingers due to smoking. The "ANXIETY" column indicates the presence or absence of anxiety in
individuals. The "PEER_PRESSURE" column denotes whether individuals face peer pressure to smoke. The

Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)
1988  ISSN: 2252-8938

"CHRONIC DISEASE" section logs any chronic conditions among the individuals. Additionally, there are
categories like "FATIGUE," "ALLERGY," "WHEEZING," "ALCOHOL INTAKE," "COUGH,"
"BREATHING DIFFICULTY," "DIFFICULTY SWALLOWING," and "CHEST DISCOMFORT," indicating
the existence or absence of these symptoms or conditions in the individuals. Finally, the "LUNG_CANCER"
category signifies whether an individual has received a lung cancer diagnosis. It's represented using categorical
values (object). The dataset provided doesn't exhibit any null values, guaranteeing that all 309 entries contain
complete information across all columns. Table 4 shows that the confusion matrix for RF classifier.

Figure 5. Python code to train models Figure 6. Details of dataset features

Table 4. Confusion matrix of RF

Algorithm Accuracy Precison Recall F1-score AUC-ROC
LR 94.44 100.00 85.71 92.31 92.86
SVC-Linear 94.44 100.00 85.71 92.31 92.86
DT 94.44 100.00 85.71 92.31 92.86
RF 94.44 100.00 85.71 92.31 92.86
XGBoost classifier 94.44 100.00 85.71 92.31 92.86
SVC-RBF 83.33 83.33 71.43 72.92 81.77
AdaBoost classifier 83.33 75.00 85.71 80.00 83.77

Based on the provided data, here's an explanation of the confusion matrix with respect to the RF
model: The confusion matrix encapsulates the evaluation of the RF model's performance in a binary classification
scenario. It involves four essential measures: precision, recall, F1-score, and support. This matrix is visually
depicted as a grid, presenting the anticipated class labels horizontally and the real class labels vertically.
For the positive class (class 1):
Precision: The precision for class 1 is 1.00, indicating that all the samples predicted as class 1 were correctly
classified.
Recall: The recall score for class 1 stands at 0.86, indicating that 86% of the genuine positive samples
were accurately recognized by the model.
F1-score: The F1-score for class 1 stands at 0.92, calculated as the harmonic average of precision and recall.
It offers a balanced assessment of the model's effectiveness.
Support: The support for class 1 is 7, representing the number of actual samples belonging to class 1.
For the negative class (class 0):
Precision: The precision score for class 0 amounts to 0.92, signifying that 92% of the samples identified as
class 0 were accurately categorized.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1980-1991

Int J Artif Intell ISSN: 2252-8938  1989

Recall: The recall for class 0 is 1.00, suggesting that all the actual negative samples were correctly
identified by the model.
F1-score: The F1-score for class 0 is 0.96, providing a balanced measure of precision and recall for the
negative class.
Support: The support for class 0 is 11, representing the number of actual samples belonging to class 0.
The reported accuracy of the RF model stands at 0.94, suggesting its correct classification of 94% of
the dataset's samples. The confusion matrix and its corresponding metrics offer an understanding of the RF
model's efficiency in distinguishing between positive and negative samples. It showcases robust accuracy and
well-balanced performance across both classes, highlighted by the precision, recall, and F1-score. Figure 7
shows that the comparison of various machine learning classifiers. We conducted comparisons among LR, SVC-
linear, SVC-RBF, DT, RF, AdaBoost classifier, and XGBoost classifier across three distinct datasets. After
analysis, it was determined that RF exhibited the highest accuracy among the various machine learning models.

150

100

50

0
Logistic SVC-Linear SVC-rbf Decision Random AdaBoost xgb
Regression Tree Forest Classifier classifier

Dataset 1 Dataset 2 Dataset 3

Figure 7. Comparison of various machine learning classifiers

5. CONCLUSION
In conclusion, XAI models hold great promise for predicting lung cancer and improving patient
outcomes. Traditional machine learning models lack interpretability, hindering their clinical adoption. XAI
models provide clear explanations, enabling clinicians to understand and trust their decisions. Researchers have
explored diverse datasets and imaging modalities like CT and MRI to develop accurate and interpretable XAI
models. Challenges remain, including the need for large and diverse datasets, balancing accuracy with
interpretability, and understanding the complex biology of lung cancer. Collaborative efforts are necessary to
address these challenges. Continued exploration of diverse datasets and advancements in XAI techniques can
enhance model performance and interpretability. Integrating XAI models into clinical practice can
revolutionize decision-making and save lives through early detection and accurate diagnosis. XAI models offer
a pathway to reliable and interpretable lung cancer prediction, empowering clinicians to make informed
decisions and improve patient outcomes.This study compared diverse datasets to evaluate different XAI models
for lung cancer prediction. Models like LR, SVC-linear, SVC-RBF, DT, RF, AdaBoost Classifier, and
XGBoost Classifier were trained using clinical, genomic, and imaging data. Preliminary results showed RF
achieving the highest accuracy of 98.9% across multiple datasets. Evaluation metrics and interpretability
techniques were used to assess model performance. These findings inform the selection of effective and
interpretable AI models for improved lung cancer prediction and treatment decisions.

REFERENCES
[1] S. T. Rikta, K. M. M. Uddin, N. Biswas, R. Mostafiz, F. Sharmin, and S. K. Dey, “XML-GBM lung: An explainable machine
learning-based application for the diagnosis of lung cancer,” Journal of Pathology Informatics, vol. 14, 2023, doi:
10.1016/j.jpi.2023.100307.
[2] G. R. Pathak, M. S. G. Premi, and S. H., “LSSCW: A lightweight security scheme for cluster based wireless sensor network,”
International Journal of Advanced Computer Science and Applications, vol. 10, no. 10, 2019, doi:
10.14569/IJACSA.2019.0101062.
[3] P. Chaturvedi, A. Jhamb, M. Vanani, and V. Nemade, “Prediction and classification of lung cancer using machine learning
techniques,” IOP Conference Series: Materials Science and Engineering, vol. 1099, no. 1, Mar. 2021, doi: 10.1088/1757-
899X/1099/1/012059.
[4] S. Madhumalar and S. Sivakumar, “A study on prediction of diabetic coronary heart disease using machine learning algorithms,”
Journal of ISMAC, vol. 4, no. 2, pp. 119–132, Jul. 2022, doi: 10.36548/jismac.2022.2.005.
[5] P. R. Chandre, P. N. Mahalle, and G. R. Shinde, “Intrusion prevention framework for WSN using deep CNN,” Turkish Journal of
Computer and Mathematics Education, vol. 12, no. 6, pp. 3567–3572, 2021.
[6] S. H. Hosseini, R. Monsefi, and S. Shadroo, “Deep learning applications for lung cancer diagnosis: A systematic review,”
Multimedia Tools and Applications, Jul. 2023, doi: 10.1007/s11042-023-16046-w.

Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)
1990  ISSN: 2252-8938

[7] P. R. Chandre, P. Mahalle, and G. Shinde, “Intrusion prevention system using convolutional neural network for wireless sensor
network,” IAES International Journal of Artificial Intelligence (IJ-AI), vol. 11, no. 2, pp. 504–515, Jun. 2022, doi:
10.11591/ijai.v11.i2.pp504-515.
[8] C. Wang et al., “Towards reliable and explainable AI model for solid pulmonary nodule diagnosis,” Prepr. arXiv.2204.04219, 2022.
[9] G. R. Pathak and S. H. Patil, “Mathematical model of security framework for routing layer protocol in wireless sensor networks,”
Procedia Computer Science, vol. 78, pp. 579–586, 2016, doi: 10.1016/j.procs.2016.02.121.
[10] I. Naseer, S. Akram, T. Masood, A. Jaffar, M. A. Khan, and A. Mosavi, “Performance analysis of state-of-the-art CNN architectures
for LUNA16,” Sensors, vol. 22, no. 12, Jun. 2022, doi: 10.3390/s22124426.
[11] T. Kadir and F. Gleeson, “Lung cancer prediction using machine learning and advanced imaging techniques,” Translational Lung
Cancer Research, vol. 7, no. 3, pp. 304–312, Jun. 2018, doi: 10.21037/tlcr.2018.05.15.
[12] P. R. Chandre, P. N. Mahalle, and G. R. Shinde, “Machine learning based novel approach for intrusion detection and prevention
system: A tool based verification,” in 2018 IEEE Global Conference on Wireless Computing and Networking (GCWCN), IEEE,
Nov. 2018, pp. 135–140. doi: 10.1109/GCWCN.2018.8668618.
[13] E. S. Neal Joshua, D. Bhattacharyya, M. Chakkravarthy, and Y.-C. Byun, “3D CNN with visual insights for early detection of lung
cancer using gradient-weighted class activation,” Journal of Healthcare Engineering, vol. 2021, pp. 1–11, Mar. 2021, doi:
10.1155/2021/6695518.
[14] W. Alakwaa, M. Nassef, and A. Badr, “Lung cancer detection and classification with 3D convolutional neural network (3D-CNN),”
International Journal of Advanced Computer Science and Applications, vol. 8, no. 8, 2017, doi: 10.14569/IJACSA.2017.080853.
[15] R. Patra, “Prediction of lung cancer using machine learning classifier,” in COMS2 2020: Computing Science, Communication and
Security, 2020, pp. 132–142. doi: 10.1007/978-981-15-6648-6_11.
[16] C. Anil Kumar et al., “Lung cancer prediction from text datasets using machine learning,” BioMed Research International,
vol. 2022, pp. 1–10, Jul. 2022, doi: 10.1155/2022/6254177.
[17] E. Nemlander et al., “Lung cancer prediction using machine learning on data from a symptom e-questionnaire for never smokers,
formers smokers and current smokers,” PLOS ONE, vol. 17, no. 10, Oct. 2022, doi: 10.1371/journal.pone.0276703.
[18] D. Mustafa Abdullah, A. Mohsin Abdulazeez, and A. Bibo Sallow, “Lung cancer prediction and classification based on correlation
selection method using machine learning techniques,” Qubahan Academic Journal, vol. 1, no. 2, pp. 141–149, May 2021, doi:
10.48161/qaj.v1n2a58.
[19] A. K. Gulia and R. Bhatt, “Lung cancer prediction using machine learning classifiers,” Turkish Journal of Computer and
Mathematics Education, vol. 12, no. 12, pp. 1665–1672, 2021, doi: 10.17762/turcomat.v12i12.7670.
[20] S. Makubhai, G. R. Pathak, and P. R. Chandre, “Prevention in healthcare: An explainable AI approach,” International Journal on
Recent and Innovation Trends in Computing and Communication, vol. 11, no. 5, pp. 92–100, May 2023, doi:
10.17762/ijritcc.v11i5.6582.
[21] A. Shimazaki et al., “Deep learning-based algorithm for lung cancer detection on chest radiographs using the segmentation method,”
Scientific Reports, vol. 12, no. 1, Jan. 2022, doi: 10.1038/s41598-021-04667-w.
[22] A. E. Sebastian and D. Dua, “Lung nodule detection via optimized convolutional neural network: impact of improved moth flame
algorithm,” Sensing and Imaging, vol. 24, no. 1, p. 11, Mar. 2023, doi: 10.1007/s11220-022-00406-1.
[23] H. Tang, D. R. Kim, and X. Xie, “Automated pulmonary nodule detection using 3D deep convolutional neural networks,” in 2018
IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), IEEE, Apr. 2018, pp. 523–526. doi:
10.1109/ISBI.2018.8363630.
[24] R. Li, C. Xiao, Y. Huang, H. Hassan, and B. Huang, “Deep learning applications in computed tomography images for pulmonary
nodule detection and diagnosis: a review,” Diagnostics, vol. 12, no. 2, p. 298, Jan. 2022, doi: 10.3390/diagnostics12020298.
[25] W. Shen, M. Zhou, F. Yang, C. Yang, and J. Tian, “Multi-scale convolutional neural networks for lung nodule classification,” 2015,
pp. 588–599. doi: 10.1007/978-3-319-19992-4_46.
[26] L. Gong, S. Jiang, Z. Yang, G. Zhang, and L. Wang, “Automated pulmonary nodule detection in CT images using 3D deep squeeze-
and-excitation networks,” International Journal of Computer Assisted Radiology and Surgery, vol. 14, no. 11, pp. 1969–1979, Nov.
2019, doi: 10.1007/s11548-019-01979-1.
[27] H. Xie, D. Yang, N. Sun, Z. Chen, and Y. Zhang, “Automated pulmonary nodule detection in CT images using deep convolutional
neural networks,” Pattern Recognition, vol. 85, pp. 109–119, Jan. 2019, doi: 10.1016/j.patcog.2018.07.031.
[28] H. Cao et al., “A two-stage convolutional neural networks for lung nodule detection,” IEEE Journal of Biomedical and Health
Informatics, pp. 1–1, 2020, doi: 10.1109/JBHI.2019.2963720.
[29] G. L. F. da Silva, O. P. da Silva Neto, A. C. Silva, A. C. de Paiva, and M. Gattass, “Lung nodules diagnosis based on evolutionary
convolutional neural network,” Multimedia Tools and Applications, vol. 76, no. 18, pp. 19039–19055, Sep. 2017, doi:
10.1007/s11042-017-4480-9.
[30] K. Dwivedi, A. Rajpal, S. Rajpal, M. Agarwal, V. Kumar, and N. Kumar, “An explainable AI-driven biomarker discovery
framework for non-small cell lung cancer classification,” Computers in Biology and Medicine, vol. 153, Feb. 2023, doi:
10.1016/j.compbiomed.2023.106544.
[31] I. Tunali, R. J. Gillies, and M. B. Schabath, “Application of radiomics and artificial intelligence for lung cancer precision medicine,”
Cold Spring Harbor Perspectives in Medicine, vol. 11, no. 8, Aug. 2021, doi: 10.1101/cshperspect.a039537.
[32] M. Marcos et al., Artificial intelligence in medicine: Knowledge representation and transparent and explainable systems, vol.
11979. in Lecture Notes in Computer Science, vol. 11979. Cham: Springer International Publishing, 2019. doi: 10.1007/978-3-030-
37446-4.
[33] E. Dritsas and M. Trigka, “Lung cancer risk prediction with machine learning models,” Big Data and Cognitive Computing, vol. 6,
no. 4, Nov. 2022, doi: 10.3390/bdcc6040139.
[34] C. Rampinelli et al., “Exposure to low dose computed tomography for lung cancer screening and risk of cancer: secondary analysis
of trial data and risk-benefit analysis,” BMJ, Feb. 2017, doi: 10.1136/bmj.j347.
[35] R. Funde and P. Chandre, “Dynamic cluster head selection to detect gray hole attack using intrusion detection system in MANETs,”
in Proceedings of the Sixth International Conference on Computer and Communication Technology 2015, New York, NY, USA:
ACM, Sep. 2015, pp. 73–77. doi: 10.1145/2818567.2818581.
[36] G. Yang, Q. Ye, and J. Xia, “Unbox the black-box for the medical explainable AI via multi-modal and multi-centre data fusion: A
mini-review, two showcases and beyond,” Information Fusion, vol. 77, pp. 29–52, Jan. 2022, doi: 10.1016/j.inffus.2021.07.016.
[37] H. W. Loh, C. P. Ooi, S. Seoni, P. D. Barua, F. Molinari, and U. R. Acharya, “Application of explainable artificial intelligence for
healthcare: A systematic review of the last decade (2011–2022),” Computer Methods and Programs in Biomedicine, vol. 226, Nov.
2022, doi: 10.1016/j.cmpb.2022.107161.
[38] Y. Zhang, Y. Weng, and J. Lund, “Applications of explainable artificial intelligence in diagnosis and surgery,” Diagnostics,

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1980-1991

Int J Artif Intell ISSN: 2252-8938  1991

vol. 12, no. 2, Jan. 2022, doi: 10.3390/diagnostics12020237.

[39] K. Kobylińska, T. Orłowski, M. Adamek, and P. Biecek, “Explainable machine learning for lung cancer screening models,” Applied
Sciences, vol. 12, no. 4, Feb. 2022, doi: 10.3390/app12041926.
[40] Z. Naz, M. U. G. Khan, T. Saba, A. Rehman, H. Nobanee, and S. A. Bahaj, “An explainable AI-enabled framework for interpreting
pulmonary diseases from chest radiographs,” Cancers, vol. 15, no. 1, Jan. 2023, doi: 10.3390/cancers15010314.
[41] S. Alkhalaf et al., “Adaptive aquila optimizer with explainable artificial intelligence-enabled cancer diagnosis on medical imaging,”
Cancers, vol. 15, no. 5, Feb. 2023, doi: 10.3390/cancers15051492.
[42] A. Sanchez‐Palencia et al., “Gene expression profiling reveals novel biomarkers in nonsmall cell lung cancer,” International Journal
of Cancer, vol. 129, no. 2, pp. 355–364, Jul. 2011, doi: 10.1002/ijc.25704.
[43] Y.-C. Chen et al., “Increased S100A15 expression and decreased DNA methylation of its gene promoter are involved in high
metastasis potential and poor outcome of lung adenocarcinoma,” Oncotarget, vol. 8, no. 28, pp. 45710–45724, Jul. 2017, doi:
10.18632/oncotarget.17391.
[44] M. J. Catarata, R. Ribeiro, M. J. Oliveira, C. Robalo Cordeiro, and R. Medeiros, “Renin-angiotensin system in lung tumor and
microenvironment interactions,” Cancers, vol. 12, no. 6, Jun. 2020, doi: 10.3390/cancers12061457.
[45] D. Dhotre, P. R. Chandre, A. Khandare, M. Patil, and G. S. Gawande, “The rise of crypto malware: Leveraging machine learning
techniques to understand the evolution, impact, and detection of cryptocurrency-related threats,” International Journal on Recent
and Innovation Trends in Computing and Communication, vol. 11, no. 7, pp. 215–222, Sep. 2023, doi: 10.17762/ijritcc.v11i7.7848.
[46] A. Chitnis, P. Chandre, and S. Pathan, “Enhancing intrusion prevention with explainable convolutional neural networks: Recent
developments and promising applications,” in 2023 8th International Conference on Communication and Electronics Systems
(ICCES), IEEE, Jun. 2023, pp. 912–917. doi: 10.1109/ICCES57224.2023.10192764.
[47] P. R. Chandre et al., “Explainable AI for intrusion prevention: A review of techniques and applications,” in ICTIS 2023: ICT with
Intelligent Applications, 2023, pp. 339–350. doi: 10.1007/978-981-99-3758-5_31.
[48] S. Sahu, R. Kumar, and P. M. Shafi, “Movie rating prediction and viewers’ sentiment trend analysis using YouTube trailer
comments,” in Micro-Electronics and Telecommunication Engineering, 2023, pp. 127–142. doi: 10.1007/978-981-19-9512-5_12.

BIOGRAPHIES OF AUTHORS

Shahin Shoukat Makubhai is a research scholar in the Department of Computer

Science and Engineering at MIT School of Computing, MIT ADT, and Pune, India. She
received her B.E. degree in Computer Science and Engineering from DKTE Society's Textile
& Engineering Institute (an autonomous institute), Ichalkaranji, India, and her M.Tech.
degree in Computer Engineering with specialization in Cloud Computing from Vellore
Institute of Technology - VIT Chennai, India in 2020. She is currently pursuing her Ph.D. in
Computer Science at MIT-ADT University, Pune, India. She has earned several global
certifications and has also contributed to research through patents and copyrights. Her
research interests include artificial intelligence and cloud computing. She can be contacted at
email: [email protected].

Ganesh R. Pathak received Bachelor of Engineering (B.E. -Computer) from

Walchand Institute of Technology, Maharashtra, India, Master of Engineering (M.E. -CSEIT)
from Savitribai Phule Pune University (formerly University of Pune), Maharashtra, India and
Ph.D. degree in Computer Science and Engineering at Sathyabama Institute of Science and
Technology (deemed to be University), Chennai, Tamil Nadu, India. He is presently working
as professor in the Department of Computer Science and Engineering, School of Computing,
MIT Art, Design and Technology University, Pune. His research interests include artificial
intelligence, machine learning, data science, computer networks, wireless communication,
and wireless sensor network, especially security in wireless sensor network. His teaching
areas include mobile computing, pervasive computing information assurance, and security
and usability engineering. He is a member of IEEE and CSI. He can be contacted at email:
[email protected].

Pankaj R. Chandre has obtained his B.E degree in Information Technology

from Sant Gadge Baba Amravati University, Amravati, India; M.E. degree in Computer
Engineering from from Mumba1 University Maharashtra, India in the year 2011; and Ph.D.
in Computer Engineering from Savitribai Phule Pune University, Pune, India in the year
2021. He is currently working as an associate professor in Department of Computer Science
and Engineering, MIT School of Computing, MIT ADT, Pune, India. He has published 60
plus papers at international journals and conferences. He has guided more than 30 plus under-
graduate students and 20 plus postgraduate students for projects. His research interests are
network security and information security. He can be contacted at email:
[email protected].

Comparative analysis of explainable Artificial Intelligence models for predicting … (Shahin Makubhai)