Thyroid Disease Prediction using AI models

Abstract:

Thyroid disorders represent a significant health concern globally, necessitating

accurate and timely diagnostic tools for effective management. This study
explores the application of Gated Recurrent Unit (GRU) neural networks in
predicting thyroid conditions based on relevant patient data.The proposed model
leverages the temporal dependencies inherent in medical data, utilizing the
GRU architecture's ability to capture and learn sequential patterns. We employ a
dataset comprising a diverse set of patient records, including clinical variables
such as thyroid hormone levels, patient demographics, and historical medical
history.Our experimental results demonstrate the efficacy of the GRU-based
model in predicting thyroid conditions with high accuracy and sensitivity. The
dynamic nature of the GRU units enables the model to effectively capture long-
term dependencies within the sequential input data, providing a robust
foundation for thyroid prediction.Furthermore, we conduct comparative
analyses with traditional machine learning models to highlight the superior
performance of the GRU-based approach. The model's interpretability is also
assessed to enhance its applicability in clinical settings.This work contributes to
the ongoing efforts in developing advanced machine learning techniques for
medical diagnosis, specifically focusing on thyroid disorders. The proposed
GRU-based model showcases promising results, offering a potential avenue for
enhancing the accuracy and efficiency of thyroid condition predictions,
ultimately aiding clinicians in early diagnosis and intervention.
 CHAPTER 1

INTRODUCTION

1.1 THE THYROID GLAND

1.1.1 Anatomy and Histology

The thyroid is one of the largest endocrine glands which is located in the
neck, just below the larynx. It is made of two lateral lobes, right and left,
connected by a central isthmus, with occasionally an accessory pyramidal lobe
discernible as a finger-like projection directed upward from the isthmus. The
parathyroid glands are located behind the thyroid gland which involved in
calcium metabolism (Melmed et al. 2012).

Figure 1.1 Thyroid anatomy, Panel A: thyroid in the neck. Panel

B: details of thyroid gland

The thyroid gland is formed by closely packed follicles, lumen spherical

units containing proteinaceous colloid surrounded by a single layer of thyroid
epithelial cells, named thyrocytes or follicular cells shown in the Figure 1.1.
Even within a single gland The diameter of the follicles varies considerably, but
averages about 200 nm. The follicular cells vary in height with the degree of
glandular stimulation, becomes columnar when active and cuboidal when
inactive. The epithelium rests on the basement membrane which is rich in
glycoproteins and separates the follicular cells from the surrounding capillaries.
Almost 20 to 40 follicles are demarcated by connective tissue septa in order to
form a lobule supplied by a single artery.

The functional activity of a given lobule may differ from that of its
neighbours. The thyroid gland also contains calcitonin-producing parafollicular
cells (C cells), usually localized along with the basement membrane of the
thyroid epithelium (Melmed et al. 2012).

1.1.2 Physiology of Thyroid Gland

The thyrocytes main function is to produce thyroid hormones thyroxine

(T4) and triiodothyronine (T3), homodimer peptides derived from the iodination
and coupling of the amino acid L-tyrosine. Thyroxine is a multifunctional
hormone, which helps in the metabolism of carbohydrate, proteins and lipids. It
maintains the Body Mass Ratio (BMR) and also the heat regulation of the body
(Luo et al.. 2009). It plays a vital role in the growth and development of various
organs. The hypothalamo, hypophyseal-thyroid axis mainly regulates the
thyroxin secretion. The major diseases associated with thyroid gland, hyper and
hypothyroidism is due to increased or decreased secretion of the hormones
respectively.

1.2 HYPOTHALAMIC - PITUITARY - THYROID AXIS

The thyroidal secretion of T3 and T4 is controlled by the

thyroidstimulating hormone or thyrotropin (TSH), a peptide produced by
endocrine cells in the anterior pituitary named thyrotrophs. TSH production is
regulated by two interacting systems. They are,

1) Open-loop neural control by hypothalamic hypophysiotropic factors

2) Negative feedback by thyroid hormones

The neural control is mainly exerted by the thyrotropin-releasing

hormone (TRH), a peptide produced by the paraventricular neurons of the
hypothalamus which binds to specific receptors on the plasma membrane of the
thyrotrophs, that stimulates the production of TSH. TSH secretion is also
influenced by other hormones including oestrogens, glucocorticoids and
possibly Growth Hormone (GH), while it is inhibited by cytokines in the
pituitary and hypothalamus. Thyroid hormones, both T3 and T4, exert their
negative feedback control on both the pituitary and the hypothalamus ,
determining a reduction in TRH and TSH release. Circulating T3 and T4 are
transported inside the brain, where T4 is converted into T3 by type II deiodinase
and then T3 interacts with thyroid hormone receptors. TSH binds to its specific
receptor (TSHR) located on the basolateral embrane of thyrocytes and exerts
several functions acting via different intracellular signaling cascades. TSH main
actions include: 1) induction of thyroid hormone synthesis 2) stimulation of
endocytosis on the apical membrane, causing internalization and digestion of
colloid with consequent thyroid hormone release into the blood from the
basolateral membrane 3) maintenance of trophic thyroid cell integrity and
tropism. In particular, two main aspects of thyroid hormone synthesis induction
are: A) Tg/TPO/NIS genes transcription induction via adenylate cyclase (AC),
an enzyme that catalyzes the conversion of adenosine

triphosphate (ATP) to 3', 5'-cyclic adenosine monophosphate (cAMP) and

pyrophosphate B) H2O2 generation via Gq–phospholipase C cascade leading to
increased intracellular levels of diacylglycerol and calcium ion (Ca2+) (Dumont
et al. 2000). TSH is a glycoprotein hormone composed by two subunits: α and
β. The β subunit is unique to TSH and therefore determines its receptor
specificity. The α subunit contains the effect or region and it is nearly identical
to other hormones, such as the human chorionic gonadotropin (hCG),
luteinizing hormone (LH) and follicle-stimulating hormone (FSH).

1.3 THYROID HORMONE REGULATIONS

Thyroid Hormone activity is regulated at multiple steps; however, the

primary point of regulation is synthesis of thyroid hormones by the thyroid
gland. Here we discuss the basic circuitry which governs the release of thyroid
hormone by the thyroid gland. Overview: Thyroid Hormone release is regulated
by an extended version of the circuit that governs Basic Hypothalamic-Pituitary
Coordination.

Hypothalamic Step: The hypothalamus releases Thyrotropin-releasing

Hormone (TRH) which acts on the anterior pituitary. It should be pointed out

that TRH is also involved in the regulated release of prolactin.

Pituitary Step: Being stimulated by TRH, the anterior pituitary synthesizes

and secretes Thyroid Stimulating Hormone (TSH) which acts directly on the

thyroid gland.

Thyroid Step: The thyroid gland releases thyroid hormones by the

stimulation of TSH.

Tissue Step: The thyroid gland mostly releases T4 which does not possess

much endocrine activity. Target tissues possess the enzyme Iodinase that

converts T4 to either T3 or rT3. T3 possesses tremendously more endocrine

activity than T4 whereas, rT3 does not possess any endocrine activity.
Therefore, in reality, the final step of the thyroid hormone regulatory axis can

be occured in the actual peripheral target tissues as shown in the Figure.1.2

Figure 1.2 Major Regulatory steps in TSH release

1.4 CHEMICAL STRUCTURE OF THYROID HORMONES

Thyroid hormones play a major role in the development and functional

maintenance of many organs (Koibuchi and Chin 2000). They are a group of
hormones synthesized and secreted from the thyroid gland, which is located in
front of the trachea, below the thyroid cartilage. In general, two compounds,
thyroxine (3, 5, 3', 5' -tetra-iodo-Lthyronine, T4) and triiodothyronine (3, 5, 3' -
tri-iodo-L-thyronine, T3), are considered thyroid hormones. In addition, another
compound called reverse T3 (3, 3', 5' -triiodoL- thyronine, rT3) is secreted from
the thyroid gland. Their chemical structures are shown in Figure 1.4.

The essential molecular structure comprises two iodinated benzene rings

connected by ether linkage. T4 is the major hormone secreted from the thyroid
gland, whereas the other hormones are mainly generated by the deiodination of
T4 in extrathyroidal tissues. The ratio of the secretion of T4:T3:rT3 from the
thyroid gland is approximately 100:5:2.5. T3, a bioactive thyroid hormone, is
mainly produced by the deiodination of T4 in thyroid hormone target tissues.

1.5 HYPOTHYROIDISM AND HYPERTHYROIDISM

In Primary hypothyroidism TSH will be elevated and fT4 and fT3 are
decreased. Primary hypothyroidism is caused by various conditions namely
Graves‟ disease, multinodular goitre, or toxic nodule in which TSH I usually
very low. Further clinical features help to separate these three conditions.
Similarly, features of hypothyroidism may be seen in patient getting treatment
for hyperthyroidism, patient who had undergone total thyroidectomy along with
patient who had history of radiation. Rare conditions like congenital condition
when thyroid tissue is absent, thyroid disgenesis, iodine transport defect, intake
of goitrogens (broccoli, cabbage) may also lead to hypothyroidism. Other
condition includes systemic amyloidosis, Riedel‟s thyroiditis, thyroglobulin
synthesis defect in which there will be a low thyroglobulin concentration etc.
Secondary hypothyroidism may occur due to various drugs like amiodarone,
lithium etc. Even excess intake of iodine may lead to hypothyroidism. Disease
like pituitary tumor or Sheehan syndrome or any surgery of pituitary may also
lead to hypothyroidism. However, while interpreting thyroid function tests, one
should not forget subclinical hypothyroidism in which TSH is raised while fT4
and fT3 are within normal range. In some condition only fT4 may be decreased
whereas TSH is raised and fT3 may be normal or even supernormal.
 Hyperthyroidism is a pathological syndrome in which tissue is exposed to
excessive amounts of circulating thyroid hormone. In Primary hyperthyroidism
TSH will be decreased and fT4 and fT3 are increased. The most common cause
of this syndrome is Graves' disease, followed by toxic multinodular goitre, and
solitary hyper functioning nodules. Autoimmune postpartum and subacute
thyroiditis (Hasimoto‟s thyroiditis, De Quervain‟s thyroiditis, lymphocytic
thyroiditis) is relatively common diseases with low radioiodine uptake. Other
less common causes include thyroxine ingestion. Secondary hyperthyroidism is
seen in patient with pituitary tumors that secrete thyrotropin, and drug-induced
thyroid dysfunction (amiodarone), are also important causes. Amiodarone may
cause hyperthyroidism up to 10% of cases. Rarely individuals may have
resistance to thyroid hormone syndrome in whom there is marked elevation in
both FT4 and FT3 concentrations in the absence of hyper metabolism are
typical of patients with resistance to thyroid hormone. include thyroxine
ingestion (therapeutic or factitious), ectopic thyroid tissue (including Struma
ovarii), amiodarone therapy, and excess iodine ingestion (typically seen in
multinodular goitre-Jod- Basedow effect). An increase in thyroid tissue favors
the last condition. Some of the patients may present with subclinical
hyperthyroidism in whom TSH is low and normal fT3 and fT4 which is also
seen in patient with thyroxine treatment for hypothyroidism

Data mining

Data mining is a process of discovering patterns, correlations, and

insights from large datasets to extract useful information. It involves various
steps to effectively predict outcomes such as typos, which could be detrimental
in critical data analysis or decision-making processes. Typos prediction,
specifically, is crucial in ensuring data accuracy and integrity, especially in
fields like finance, healthcare, and scientific research. Here's an outline of the
steps involved in typos prediction:
Data Collection:

In the context of typos prediction, acquiring a diverse dataset is crucial. This

dataset should encompass a wide range of text inputs from various sources,
including but not limited to documents, emails, user inputs, and forms. It's
essential to gather data that reflects the potential variability and complexity of
typographical errors.

Data Preprocessing:

Data preprocessing involves several tasks to ensure the dataset's quality and
suitability for analysis. This includes tasks such as:

Cleaning: Removing irrelevant characters, symbols, or formatting artifacts.

Normalization: Standardizing the text by converting it to lowercase, removing

accents, and expanding contractions.

Tokenization: Breaking down the text into individual words or tokens.

Spell Checking: Correcting obvious misspellings or errors that could affect

subsequent analysis.

Noise Reduction: Filtering out noise such as non-textual elements or irrelevant

information.

Feature Selection:

Feature selection is crucial for identifying the most relevant attributes that
contribute to predicting typos. This step involves:

Textual Features: Extracting features such as word frequency, character n-

grams, and word length distributions.
Contextual Features: Considering contextual information such as the position of
the word in a sentence, surrounding words, and grammatical structure.

Typing Patterns: Analyzing typing patterns, including key proximity on a

keyboard and common mistyped sequences.

Language Models: Utilizing pre-trained language models or embeddings to

capture semantic relationships and language nuances.

Feature Engineering:

Feature engineering aims to transform raw data into a format suitable for
machine learning models. This includes:

Vectorization: Converting textual data into numerical representations using

techniques like bag-of-words, TF-IDF (Term Frequency-Inverse Document
Frequency), or word embeddings.

Scaling: Normalizing numerical features to a similar scale to prevent dominance

by certain attributes during model training.

Feature Combination: Creating composite features by combining or interacting

with existing features to capture complex relationships.

Model Selection:

Choosing an appropriate model involves considering factors such as the dataset

size, complexity of the problem, and interpretability of the model. Common
models for typos prediction include:

Supervised Learning Models: Decision trees, random forests, support vector

machines, and neural networks.

Sequence Models: Recurrent neural networks (RNNs) or transformers for

capturing sequential dependencies in text data.
Ensemble Methods: Combining multiple models for improved performance and
robustness.

Model Training:

During model training, the dataset is divided into training and validation sets.
The model learns from the training data to make predictions, while the
validation set helps assess its performance and prevent overfitting.

Model Evaluation:

Model evaluation involves assessing its performance using appropriate metrics

such as accuracy, precision, recall, F1-score, and area under the ROC curve
(AUC). It's essential to consider the trade-offs between different metrics based
on the specific requirements of typos prediction.

Hyperparameter Tuning:

Hyperparameter tuning involves optimizing the model's parameters to improve

its performance. Techniques such as grid search, random search, or Bayesian
optimization can be used to find the optimal set of hyperparameters.

Model Validation:

Validating the trained model involves testing its performance on unseen data or
a holdout dataset. This step helps assess its generalization capability and ensures
that it can effectively predict typos in real-world scenarios.

Deployment:

Deploying the validated model into production environments involves

integrating it into existing systems or workflows. Continuous monitoring and
maintenance are essential to address evolving data patterns and ensure the
model's effectiveness over time.
By meticulously following these steps, data mining techniques can be
effectively applied to predict and mitigate typographical errors, thereby
enhancing data quality and reliability in various applications.

Motivation and Scope of Data Mining-Based Thyroid Prediction

The motivation behind employing data mining techniques for thyroid prediction
stems from the imperative need to enhance early diagnosis, treatment, and
management of thyroid disorders. Thyroid diseases, including hypothyroidism,
hyperthyroidism, and thyroid cancer, affect millions worldwide, posing
significant health risks if left undetected or untreated. Leveraging data mining
methodologies offers a promising avenue to address these challenges by:

Early Detection: Data mining facilitates the identification of patterns and trends
within large datasets, enabling early detection of thyroid abnormalities. By
analyzing diverse patient data, including clinical records, demographic
information, and laboratory test results, data mining algorithms can discern
subtle indicators of thyroid dysfunction, facilitating timely intervention and
treatment.

Personalized Medicine: Thyroid disorders exhibit considerable heterogeneity in

their etiology, symptoms, and response to treatment. Data mining allows for the
development of personalized predictive models tailored to individual patient
profiles. By integrating patient-specific data, such as genetic markers, lifestyle
factors, and medical history, predictive models can offer tailored
recommendations for thyroid management, optimizing therapeutic outcomes
and patient care.

Risk Stratification: Data mining enables the stratification of patients based on

their risk of developing thyroid disorders or experiencing disease progression.
By analyzing population-level data, including epidemiological trends and
environmental factors, data mining algorithms can identify high-risk cohorts
and prioritize targeted interventions, such as screening programs and preventive
measures, to mitigate the burden of thyroid diseases.

Clinical Decision Support: Data mining tools can serve as valuable decision
support systems for healthcare practitioners, aiding in the interpretation of
complex patient data and guiding clinical decision-making. By synthesizing
disparate sources of information, including medical imaging, histopathology
reports, and genomic data, data mining algorithms can assist clinicians in
formulating accurate diagnoses, prognosticating disease outcomes, and devising
tailored treatment strategies for thyroid disorders.

Knowledge Discovery: Data mining fosters knowledge discovery by uncovering

novel associations, biomarkers, and prognostic indicators relevant to thyroid
health. By analyzing multidimensional datasets encompassing diverse data
modalities, such as omics data, electronic health records, and patient-reported
outcomes, data mining techniques can reveal hidden insights into the
pathophysiology of thyroid diseases, informing future research endeavors and
therapeutic innovations.

In summary, the application of data mining techniques holds immense promise

in revolutionizing thyroid prediction and management paradigms. By
harnessing the power of big data analytics, machine learning algorithms, and
predictive modeling approaches, data mining endeavors aim to empower
clinicians, researchers, and patients alike with actionable insights and
personalized solutions for combating thyroid disorders and improving health
outcomes on a global scale.
 CHAPTER 2

SURVEY

A. Tiwari, et al investigated an efficiency of various machine learning

models for predicting thyroid disease. Decision Tree, Logistic Regression,
Random Forest, Support Vector Machine (SVM), k-Nearest Neighbor (kNN),
Naive Bayes, Kmeans, and an ensemble model were among the models used.
The results showed that the Logistic Regression model was 96% accurate, while
the Decision Tree and Random Forest models were both 99% accurate. SVM
had an 87% accuracy rate, kNN had a 95% accuracy rate, and the ensemble
model had a 97% accuracy rate. The accuracy rates of Naive Bayes and K-
Means were lower. The outcomes of this study emphasise the potential of
machine learning algorithms, notably Logistic Regression, decision tree,
random forest, and ensemble models, in accurately diagnosing and predicting
thyroid disease.

V. K. Singh et al presented in the paper is basically for making prediction

of the disease called Thyroid. The paper gives a detailed report on the Medical
Science Problem called Thyroid and various Machine Learning methodologies.
We use pycaret to make a single implementation of 14 ML models giving
prediction on the disease. The objective of which was to go on for full scale
development from a prototype having best results. Kaggle is the repository that
provided the Experimental workbench for the work. Also we made a proper
usage of the AI vehicle python to have the overall implementation and
Deployment.

A. Sutradhar et al., propose a novel machine learning (ML) based disease

prediction system that could potentially predict it by considering three crucial
steps. First, to reduce the dimension of the dataset, three feature selection
techniques were employed, including Feature Importance (FIS), Information
Gain Selections (IGS), and Least Absolute Shrinkage and Selection Operator
(LAS). Moreover, recommended medical references were considered while
developing a feature set having the identical attributes as High-Risk Factors
(HRF). Second, the models, including the Three Stage Hybrid Classifier ( 3SHC
) and the Three Stage Hybrid Artificial Neural Network ( 3SHANN ), are used
as classifiers on the training data set. Third, a Local Interpretable Model-
agnostic Explanations (LIME) to the 3SHC with the HRF samples was applied
to individually explain the predictions. Then, the overall behaviors of both
gender and age categories were explored with the help of a Partial Dependence
Plot (PDP). Finally, the proposed system is validated with extensive
experiments where the 3SHC achieves an accuracy (ACC) of 99.29%, which
can play a crucial role in preventing thyroid disease and alleviating stress in the
healthcare sector.

S. Ouartani et al combines machine learning, specifically decision trees,

with ontology integration for interpretability. The ontology classifier provides
insights into the model's decision-making process. The dataset used comprises
9172 observations, each represented by 31 features, accessible for download on
Kaggle. According to the results The ontology model surpassed Decision tree
model.

A. Saini, et al compares of 4 classification models: Naive Bayes,

Logistic Regression, KStar, and Decision tree (J48) has been done. The results
show that all of the aforementioned classification models have an outcome in
accuracy, while the J48 model has outperformed. For the implementation of the
models, the dataset has been taken from Kaggle to build the model. In this
research, each classification algorithm namely Naive Bayes, Logistic
Regression, KStar, and J48 have produced accuracy values of 82.24%, 93.85%,
95.63%, and 97.80%, respectively. However, J48 has the highest accuracy,
which shows that it is the best-performing model among all models.

C. P. Rani, et al propose a framework of iterative Association Rule

Mining and Correlation Analysis, which integrates associa-tion rule mining
algorithm and subsequent correlation-based analysis of the result set to extract
dominant features in the dataset. Apriori and FP Growth algorithms, the two
most popular association rule mining algorithms, were used to extract the major
attributes which causes thyroid related sicknesses. The association rule mining
is used in dimen-sionality reduction of the dataset. We have used a thyroidi-tis
dataset for this study, which is first preprocessed to pre-pare it for association
rule mining. Preprocessing is a signif-icant step in our method. Next step is the
splitting of the dataset into n data subsets. The Apriori and FP growth
algorithms are applied on these multiple data subsets to create frequent k-
itemsets. The most relevant of the associa-tion rules are selected by a majority
voting from the result set generated after the application of association rule
mining. The result set, comprising of the features which are selected as the
principal attributes are again filtered using correlation analysis. The results
show that gender factor is the most dominant of all features. It shows a
predominance in association of women with Thyroid diseases than men. In the
proposed method, the dataset undergoes a two-step process of refinement and
consequently, association rules generated gives valuable insights into the
thyroid related disease’s causative factors.

E. Mohan et al., To detect disease, a novel Long Short-Term Memory

based Convolution Neural Network (LSTM-CNN) is utilized with occurrence
area Vgg-19. For selecting the feature, the approach of bias field correction is
integrated with the hybrid optimization technique i.e., Black Widow
Optimization as well as Mayfly Optimization Approach (HBWO-MOA), also
for classifying the disease the LSTM as well as Vgg-19 of Deep Learning (DL)
is presented. From DDTI dataset image of ultrasound, the disease of thyroid
prediction as well as classification is efficiency. This analysis shown that the
proposed technology is accurate than the convolutional methodology. When
compared to existing prediction techniques i.e., AlexNet-LSTM, ResNet-LSTM,
Vgg16-LSTM, the proposed approach of Vgg-19-LSTM’s precision, sensitivity,
accuracy, recalls as well as F1_score is effective.

Pallavi et al applied a different machine learning as well as deep learning

methods, including dense neural networks to the data to produce a comparison
analysis that would aid the researchers in improving their ability to predict the
disease to use the dataset's specifications. The results of this study show how
data mining and machine learning techniques can advance the medical field.
This study will support the use of this by its doctors as such an additional
system when following the established protocol. The dataset's accuracy and loss
have been assessed. On training data and test data, dense neural networks were
99.45% and 99.15% accurate, respectively.

R. V. Reddy, apply numerous diverse ML strategies to the dataset and

accomplish a comparative examination of their efficiency so that we can more
perfectly anticipate the disease utilizing the factors recognized by the dataset. In
addition, the dataset has been tweaked to improve classification prediction
accuracy. Both the sampled and non-sampled datasets were classified so that the
two could be contrasted more precisely. With some data tweaking, we were able
to boost the random forest algorithm's accuracy to 99 percent, with 97 percent
specificity.
 S. Mondal, et al deployed initial Sixteen ML models, including six
boosting algorithms, on a dataset of 9172 instances with related features. The
model performances have been judged through various standard performance
metrics. The boosting algorithms showed exceptional results, and Cat Boost
(CB) model produced the best accuracy of 95.75%. The hyperparameter tuning
performed on boosting models by implementing Randomized Search CV
increased the accuracy to 96.19% for CB. The stacking ensemble approach was
applied on top of the six boosting tuned models with the CB classifier as the
meta-learner. At the same time, the other boosting algorithms were kept as a
base learner for the final model prediction. The accuracy of the stack model was
impressive, with 95.32% compared with default models, the ROC-AUC at 0.95,
and the other results were also promising. The model’s standard deviation was
significantly less at 0.57, implying the model’s stability and robustness, and the
False Negative (FN) rate reached 1.8%.

B. Saju, et al used a feature selection approach to eliminate certain

irrelevant characteristics from the thyroid dataset (from the UCI machine
learning repository) and to select optimal features. The dataset has three target
classes named normal, hypothyroid, and hyperthyroid. The subjects were
classified through seven different machine-learning algorithms. Random Forest
classifier achieves the highest accuracy 99.58% which is better than the existing
state-of-the-art methods.

B. Saju et al compare the performance of the models by using the

different combinations of bi-predictors and multi-predictors. Our result shows
that models that use k-fold cross-validation have greater performance, and a
higher number of k does not improve the model performance. For the
multivariate analysis, we found that the number of variable is not the key
element to determine the performance of a model, rather than a suitable
combination of strong predictors. Future studies could explore the effects of
cross-validation and multivariate analysis on other machine learning algorithms.

H. M. Almahshi et al presents an analysis and classification model that takes

into account the various factors involved in the prediction of disease. We
applied a number of machine learning algorithms like support vector machine
(SVM), Nave Bayes, decision trees, and ensemble to get the best prediction.
The decision tree algorithm showed the highest accuracy and lowest cost with a
percentage of 97.6 percent and 90, respectively. The results appeared in the
form of three classes (compensated hypothyroidism, primary hypothyroidism,
and negative). Conclusion: The best results were obtained at the milestone point
in computer-aided diagnosis of hypothyroidism. Future work is considered to
increase the number of cases that can be detected and diagnosed using machine
learnin

E. Dritsas, et al used Machine Learning (ML) to design effective models for

predicting the long-term risk of older participants (over 50 years old) being
diagnosed with hypertension. Our purpose is to train models with high
sensitivity in identifying subjects at risk to avoid the future development and
occurrence of hypertension following the proper interventions. In the context of
the adopted methodology, two different class balancing methods are considered,
under which features ranking is applied, and two ML models (namely, Decision
tree and Naive Bayes) are compared based on Precision, Recall, F-Measure,
Accuracy and Area Under Curve (AUC).

S. Nasim et al proposed a novel feature selection approach based on the

optimized chi-squared (CS-PCOS) mechanism. The ten hyper-parametrized
machine learning models are applied in comparison. Using the novel CS-PCOS
approach, the gaussian naive bayes (GNB) outperformed machine learning
models and state-of-the-art studies. The GNB achieved 100% accuracy,
precision, recall, and f1-scores with minimal time computations of 0.002
seconds. The k-fold cross-validation of GNB achieved a 100% accuracy score.
The proposed GNB model achieved accurate results for critical PCOS
prediction. Our study reveals that the dataset features prolactin (PRL), blood
pressure systolic, blood pressure diastolic, thyroid stimulating hormone (TSH),
relative risk (RR-breaths), and pregnancy are the prominent factors having high
involvement in PCOS prediction. Our research study helps the medical
community overcome the miscarriage rate and provide a cure to women through
the early detection of PCOS.

V. T. Manh et al propose a novel deep learning (DL) framework called multi-

attribute attention network (MAA-Net) that is designed to mimic the clinical
diagnosis process. The proposed model learns to predict nodular attributes and
infer their malignancy based on these clinically-relevant features. A multi-
attention scheme is adopted to generate customized attention to improve each
task and malignancy diagnosis. Furthermore, MAA-Net utilizes nodule
delineations as nodules spatial prior guidance for the training rather than
cropping the nodules with additional models or human interventions to prevent
losing the context information. Validation experiments were performed on a
large and challenging dataset containing 4554 patients. Results show that the
proposed method outperformed other state-of-the-art methods and provides
interpretable predictions that may better suit clinical needs.

I. Ahmed, et al propose a novel multimodal MRI-based computer-aided

diagnosis (CAD) system that detects cancerous thyroid nodules using a deep-
learning architecture. Particularly, our system is built with a multi-input
convolutional neural network (CNN) to perform fusion of two MRI modalities:
the diffusion weighted image (DWI) and apparent diffusion coefficient (ADC)
map. The main contribution of our system is three-folded. Namely, (1) it is the
first system to fuse thyroid DWI and ADC using CNN for classification
purpose; (2) it enables independent convolutions process for each of DWI and
ADC images, which can increase the likelihood of detecting deep texture
patterns in thyroid nodules; and (3) it enables adding extra channels in each
input with the possibility to integrate with additional MRI modalities and other
imaging technologies. We compared our system to other fusion methods and
also to other machine learning (ML) frameworks that use hand-crafted features.
Our system achieved the highest performance among them with diagnostic
accuracy of 0.88, precision of 0.82, and recall of 0.82.

M. Riajuliislam, et al attempted to predict hypothyroid in the primary stage. To

do so, we have mainly used three feature selection techniques along with
diverse classification techniques. Feature selection techniques used by us are
Recursive Feature Selection(RFE), Univariate Feature Selection(UFS) and
Principal Component Analysis(PCA) along with classification algorithms
named Support Vector Machine(SVM), Decision Tree(DT), Random
Forest(RF), Logistic Regression(LR) and Naive Bayes(NB). By observing the
results, we could extrapolate that the RFE feature selection technique helps us
to provide constant 99.35% accuracy for all four classification algorithms. Thus
it's deduced from our research that RFE helps each classifier to attain better
accuracy than all the other feature selection methods used.

T. Khan et al used a filter-based feature selection algorithms, specifically

mutual information in conjunction with a two-class Neural Network (NN)
classifier, with Azure Machine Learning tools to construct a predictive model.
Our proposed two-class NN model build using selected features in association
with SMOTE performed better than other recent ML models with an F1Score
(0.982), precision (0.968), recall (0.995), and accuracy (0.981), respectively.
 Chapter 3

Existing system

Data mining often begins with the search for frequent itemsets, and the
Apriori algorithm stands as a cornerstone in this pursuit. Its operations unfold in
two distinct phases, facilitating the discovery of associations within
transactional databases. Firstly, during the Support Counting phase, the
algorithm meticulously scans the dataset to ascertain the frequency of each
itemset, thereby identifying those that occur with notable regularity. This initial
pass enables the algorithm to discern the prevalence of various combinations of
items, laying the groundwork for subsequent analysis. Subsequently, the
Candidate Generation phase comes into play, wherein the algorithm exploits the
insights gleaned from the prior phase to construct candidate itemsets. Through a
process of joining frequent itemsets and subsequent pruning of infrequent
candidates, Apriori refines its search, ultimately revealing significant
associations among items of interest. This iterative approach efficiently
uncovers patterns that hold practical value in diverse domains, including retail
market analysis and recommendation systems.

K-Means Clustering:

Clustering algorithms, such as K-Means, offer indispensable tools for

partitioning datasets into coherent groups based on inherent similarities. K-
Means unfolds through a series of distinct operations, each contributing to the
creation of meaningful clusters. The initialization step marks the outset of the
algorithm, where K initial centroids are randomly selected from the dataset.
Subsequently, during the Assignment phase, each data point is assigned to the
nearest centroid based on a specified distance metric, typically Euclidean
distance. This assignment step forms the foundation of cluster formation, as data
points gravitate towards centroids that best represent their characteristics.
Following assignment, the algorithm proceeds to Update Centroids,
recalculating the centroids' positions based on the mean of the data points within
each cluster. This iterative refinement continues until a convergence criterion is
met, ensuring stable cluster centroids. K-Means' operations culminate in the
formation of cohesive clusters, facilitating insights into data patterns and
enabling applications ranging from customer segmentation to anomaly
detection.

Decision Tree Algorithm:

Decision tree algorithms, renowned for their interpretability and efficacy

in classification tasks, operate through a series of recursive steps aimed at
partitioning data based on attribute values. At the heart of the algorithm lies a
hierarchical structure that unfolds through attribute selection, splitting, and
recursive building. Attribute selection serves as the algorithm's guiding
principle, as it determines the optimal attribute for partitioning the dataset at
each node. This decision-making process relies on criteria such as information
gain or Gini impurity, which quantify the attribute's discriminatory power. Upon
attribute selection, the dataset undergoes partitioning, with subsequent nodes
representing distinct subsets based on attribute values. This recursive
partitioning continues until a stopping criterion is met, such as reaching a
predefined tree depth or a minimum number of instances per node. Decision
trees' innate interpretability makes them invaluable in domains like healthcare
diagnosis, where clinicians seek transparent decision-making models.

FP-Growth Algorithm:

FP-Growth emerges as a potent algorithm in association rule mining,

offering efficiency and scalability in discovering frequent itemsets within
transaction databases. The algorithm's operations revolve around the
construction of a compact data structure called the FP-Tree, followed by the
mining of frequent itemsets and the generation of association rules. The initial
step entails constructing the FP-Tree, wherein transactional data is encoded into
a hierarchical structure that succinctly captures itemset frequencies and
relationships. This compact representation sets the stage for efficient mining of
frequent itemsets, achieved through a depth-first traversal of the FP-Tree.
During this traversal, the algorithm systematically identifies frequent itemsets
by leveraging conditional pattern bases and conditional FP-Trees. Finally,
armed with the knowledge of frequent itemsets, FP-Growth proceeds to
generate association rules, applying support and confidence thresholds to filter
significant associations. This streamlined approach positions FP-Growth as a
cornerstone in market basket analysis and recommendation systems, offering
superior performance on large-scale datasets.

Linear Regression:

Linear regression, a foundational algorithm in predictive modeling,

operates through a series of operations aimed at modeling the relationship
between independent and dependent variables. Beginning with model
initialization, wherein parameters are set to initial values, the algorithm
proceeds to predict the target variable based on input features. This prediction
forms the basis for evaluating the model's performance, as it enables
computation of the loss between predicted and actual values. Subsequently,
through gradient descent, the algorithm iteratively updates model parameters to
minimize the loss, ultimately converging towards an optimal solution. Linear
regression's simplicity and interpretability make it a versatile tool in various
domains, from finance to healthcare, where predictive modeling is paramount.

Support Vector Machines (SVM):

Support Vector Machines (SVM) are powerful supervised learning

algorithms used for classification and regression tasks. SVM operates by
finding the optimal hyperplane that separates different classes in the feature
space. Its operations involve transforming the input data into a higher-
dimensional space using a kernel function, where a hyperplane with the
maximum margin is identified to best separate the classes. Additionally, SVM
can handle non-linear separable data by using kernel tricks to map data into
higher dimensions. This flexibility and ability to handle complex data
distributions make SVM a popular choice in various applications such as image
classification, text categorization, and bioinformatics.

Random Forest:

Random Forest is an ensemble learning technique based on decision

trees, known for its robustness and accuracy in classification and regression
tasks. The algorithm operates by constructing multiple decision trees during
training and outputs the mode (for classification) or average (for regression)
prediction of the individual trees. Its operations involve bootstrapping the
training data to create multiple random subsets, building decision trees on each
subset, and aggregating their predictions through voting or averaging. Random
Forest mitigates overfitting and enhances generalization by averaging
predictions from multiple trees, making it resilient to noise and outliers. Its
versatility and performance have led to widespread adoption in various
domains, including finance, healthcare, and remote sensing.

Principal Component Analysis (PCA):

Principal Component Analysis (PCA) is an unsupervised dimensionality

reduction technique used to identify patterns and structure in high-dimensional
data. PCA operates by transforming the original features into a new set of
orthogonal components, called principal components, that capture the maximum
variance in the data. Its operations involve computing the covariance matrix of
the input data, performing eigendecomposition to extract the principal
components, and selecting a subset of components that retain most of the
variance. PCA facilitates data visualization, noise reduction, and feature
extraction, aiding in exploratory data analysis and improving the efficiency of
downstream machine learning algorithms. Its applications span diverse fields,
including image processing, genetics, and signal processing.

Neural Networks (Deep Learning):

Neural Networks, particularly deep learning models, have emerged as

state-of-the-art techniques for various data mining tasks, including
classification, regression, and feature learning. These models operate by
mimicking the structure and function of the human brain, comprising
interconnected layers of neurons that process input data and learn complex
representations. Operations in neural networks involve forward propagation,
where input data is passed through the network to generate predictions, and
backward propagation, where errors are backpropagated through the network to
update model parameters via gradient descent. Deep learning models, such as
Convolutional Neural Networks (CNNs) for image data and Recurrent Neural
Networks (RNNs) for sequential data, have achieved remarkable success in
domains like computer vision, natural language processing, and speech
recognition, fueled by their ability to automatically learn hierarchical features
from raw data.

Gradient Boosting Machines (GBM):

Gradient Boosting Machines (GBM) are ensemble learning algorithms

that sequentially combine weak learners, typically decision trees, to create a
strong predictive model. GBM operates by iteratively fitting new models to the
residuals of the previous model, minimizing the loss function gradient. Its
operations involve training base learners on the negative gradient of the loss
function, calculating the residuals, and updating the model predictions with each
iteration. GBM effectively leverages the strengths of individual weak learners,
iteratively refining predictions and reducing bias and variance. XGBoost,
LightGBM, and CatBoost are popular implementations of GBM known for their
scalability, efficiency, and superior performance in various competitions and
real-world applications, including web search ranking, financial forecasting, and
click-through rate prediction.

These algorithms represent a diverse array of techniques in data mining, each

offering unique capabilities and advantages suited to different problem domains
and data characteristics. Understanding their operations is crucial for selecting
the most appropriate algorithm and optimizing its performance in practical
applications.
 CHAPTER 4

Proposed system

Our proposed system aims to develop a robust thyroid prediction model using
Gated Recurrent Unit (GRU) neural networks. The system is designed to
leverage the inherent temporal dependencies present in medical data, thereby
enhancing the accuracy and efficiency of thyroid condition predictions. The
system consists of several interconnected modules, each serving a specific
purpose in the prediction process:

Data Preprocessing Module:

This module is responsible for preprocessing the input data to ensure its
compatibility with the GRU model. It involves tasks such as data cleaning,
normalization, and feature extraction. Clinical variables such as thyroid
hormone levels, patient demographics, and historical medical history are
carefully curated and formatted for input into the neural network.

GRU-based Prediction Model:

The core of our system is the GRU-based prediction model. This module
comprises multiple GRU units arranged in a sequential manner to capture the
temporal dependencies present in the input data. The GRU architecture's ability
to retain information over long sequences makes it particularly well-suited for
modeling medical data characterized by temporal patterns.

Training Module:

In this module, the GRU-based prediction model is trained using the

preprocessed dataset. The training process involves optimizing the model
parameters to minimize prediction errors. We employ techniques such as
backpropagation through time (BPTT) to update the weights of the GRU units
iteratively.

Evaluation Module:

Once trained, the performance of the GRU-based prediction model is

evaluated using validation data. This module assesses various metrics such as
accuracy, sensitivity, specificity, and area under the receiver operating
characteristic curve (AUC-ROC) to gauge the model's predictive performance.

Comparison Module:

This module conducts comparative analyses between the GRU-based

prediction model and traditional machine learning models commonly used in
thyroid prediction tasks. By comparing performance metrics and computational
efficiency, we highlight the superiority of the GRU-based approach in capturing
complex temporal patterns inherent in medical data.
Interpretability Module:

The interpretability of the GRU-based prediction model is crucial for its

applicability in clinical settings. This module employs techniques such as
attention mechanisms and feature importance analysis to interpret the model's
predictions. Clinicians can gain insights into the factors influencing thyroid
condition predictions, enhancing their confidence in the model's
recommendations.

GRU Model Architecture and Training Module:

The core of the system lies in designing and training the GRU-based
model. This module encompasses the design of the GRU architecture,
configuring layers, embeddings, and optimizing hyperparameters for effective
training. Fine-tuning parameters and managing epochs are critical aspects to
ensure the model converges efficiently.

The Gated Recurrent Unit (GRU) is a type of recurrent neural network (RNN)
architecture, similar to the more commonly known LSTM (Long Short-Term
Memory) networks. It's designed to address the vanishing gradient problem in
traditional RNNs by using gating mechanisms.

The GRU has two gates: the update gate and the reset gate. These gates control
the flow of information within the unit, allowing it to capture dependencies over
longer sequences more effectively than traditional RNNs. The architecture
diagram is shown in Figure 2.
Figure 2 :GRU architecture

Let's break down the equations and components of a GRU:

Update Gate

The update gate zt determines how much of the past information ht−1 should be
passed along to the current time step.

zt=σ(Wz⋅[ht−1,xt]+bz)

Reset Gate

The reset gate rt decides how much of the past information to forget or reset.

)rt=σ(Wr⋅[ht−1,xt]+br)

Current Memory Content

The candidate activation or memory content h~t is computed.

h~t=tanh(W⋅[rt⊙ht−1,xt]+b)

Final Hidden State

The new hidden state ℎ�ht is a combination of the previous hidden state and
the current memory content, controlled by the update gate.

ht=(1−zt)⊙ht−1+zt⊙h~t

Where:

ht is the hidden state at time t.

xt is the input at time t.

Wz,Wr,W are weight matrices specific to the gates and candidate activation.

bz,br,b are bias terms.

σ is the sigmoid activation function.

⊙⊙ denotes element-wise multiplication.

CHPATER 5

SYSTEM IMPLEMENTATION

USECASE
In software and systems engineering, a use case is a list of actions or event
steps, typically defining the interactions between a role (known in the Unified
Modeling Language as an actor) and a system, to achieve a goal. The actor can
be a human or other external system. In systems engineering, use cases are used
at a higher level than within software engineering, often representing missions
or stakeholder goals. The detailed requirements may then be captured in the
Systems Modeling Language (SysML) or as contractual statements. Use case
analysis is an important and valuable requirement analysis technique that has
been widely used in modern software engineering since its formal introduction
by Ivar Jacobson in 1992. Use case driven development is a key characteristic
of many process models and frameworks such as ICONIX, the Unified Process
(UP), the IBM Rational Unified Process (RUP), and the Oracle Unified Method
(OUM). With its inherent iterative, incremental and evolutionary nature, use
case also fits well for agile development
4.7 SEQUENCE DIAGRAM

A sequence diagram is an interaction diagram that shows how objects operate

with one another and in what order. It is a construct of a message sequence
chart. A sequence diagram shows object interactions arranged in time sequence.
It depicts the objects and classes involved in the scenario and the sequence of
messages exchanged between the objects needed to carry out the functionality
of the scenario. Sequence diagrams are typically associated with use case
realizations in the Logical View of the system under development. Sequence
diagrams are sometimes called event diagrams or event scenarios. A sequence
diagram shows, as parallel vertical lines, different processes or objects that live
simultaneously, and, as horizontal arrows, the messages exchanged between
them, in the order in which they occur. This allows the specification of simple
runtime scenarios in a graphical manner.
4.8 ACTIVITY DIAGRAM

Activity diagram is another important diagram in UML to describe dynamic

aspects of the system. Activity diagram is basically a flow chart to represent the
flow from one activity to another activity. The activity can be described as an
operation of the system. So the control flow is drawn from one operation to
another
4.9 COLLABORATIVE DIAGRAM

A collaboration diagram resembles a flowchart that portrays the roles,

functionality and behavior of individual objects as well as the overall operation
of the system in real time. Objects are shown as rectangles with naming labels
inside. These labels are preceded by colons and may be underlined. The
relationships between the objects are shown as lines connecting the rectangles.
The messages between objects are shown as arrows connecting the relevant
rectangles along with labels that define the message sequencing.

Data input
User Training

Result
Process
Data separation

Feature points
Classification
Process Entire process

Result analysis
4.10 CLASS DIAGRAM

In software engineering, a class diagram in the Unified Modeling Language

(UML) is a type of static structure diagram that describes the structure of a
system by showing the system's classes, their attributes, operations (or
methods), and the relationships among objects.
 CHAPTER 5

REQUIRMENT ANALYSIS

4.1 Software Requirements:

Python 3.7 is a prerequisite for this software setup. It serves as the

foundational programming language for developing machine learning
applications. To ensure compatibility and functionality, it's imperative to have
Python 3.7 installed on the system. The installation files can be obtained from
the official Python website.

The software stack further extends with the integration of key libraries:

TensorFlow is pivotal for constructing and training neural network

models. Leveraging its robust architecture, developers can delve into a myriad
of machine learning tasks ranging from image recognition to natural language
processing.

pandas furnishes a versatile toolkit for data manipulation and analysis.

With its DataFrame and Series structures, analysts can seamlessly navigate
through datasets, perform data cleaning, and execute complex transformations.

NumPy plays a fundamental role in numerical computing and array

manipulation. Its array-oriented computing capabilities are pivotal for executing
intricate mathematical operations with optimal efficiency.

scikit-learn serves as a comprehensive toolbox for machine learning

algorithms and tools. With an extensive array of supervised and unsupervised
learning algorithms, it facilitates model selection, evaluation, and
hyperparameter tuning.

Depending on the specific requirements of the project, additional

packages may be necessary to augment the functionality of the software stack.

4.2 Hardware Requirements:

The operating system of choice for this software setup is Windows 11.
This environment provides a stable platform for executing machine learning
tasks seamlessly.

To accommodate the computational demands of TensorFlow and other machine

learning libraries, a minimum of 8 GB RAM is recommended. This ensures
smooth execution, especially when dealing with voluminous datasets or intricate
models that necessitate substantial memory allocation.

By adhering to these software and hardware requirements, developers can

establish an optimal environment conducive to the development and
deployment of machine learning applications on the Windows 11 platform.

In a traditional feed-forward Convolutional Neural Network (CNN), each

convolutional layer except the first one (which takes in the input), receives the
output of the previous convolutional layer and produces an output feature map
that is then passed on to the next convolutional layer. Therefore, for 'L' layers,
there are 'L' direct connections; one between each layer and the next layer.
 CHAPTER 6

SYSTEM TESTING

Unit Testing:

Unit testing for the GRU-based model would involve testing individual
components or functions within the model to ensure they perform as expected.
This includes testing the GRU cell implementation, input data preprocessing
functions, loss function calculation, and gradient updates. Each unit test should
validate the correctness of the specific functionality and handle edge cases
appropriately. For example, one unit test might verify that the GRU cell outputs
the correct hidden state given a specific input and initial hidden state.

Integration Testing:

Integration testing focuses on testing interactions between different components

or modules of the model to ensure they integrate smoothly and produce the
expected results. This involves testing the interactions between the GRU layers,
input data pipelines, and output layers. Integration tests should cover various
scenarios, such as testing different combinations of input features and assessing
the model's performance when integrating with external libraries or frameworks.

Functional Testing:

Functional testing evaluates the overall functionality of the GRU-based model

by assessing its ability to predict thyroid conditions accurately. This involves
feeding sample input data through the model and verifying that the predictions
align with expected outcomes. Functional tests should cover different scenarios,
including varying input data distributions and assessing the model's robustness
to noise or missing data.
Usability Testing:

Usability testing assesses the ease of use and user-friendliness of the model,
particularly from the perspective of clinicians or healthcare professionals who
would utilize it in a clinical setting. This involves gathering feedback from
potential users through surveys, interviews, or observational studies to identify
any usability issues or areas for improvement. Usability tests should focus on
aspects such as the model's interpretability, ease of input data integration, and
clarity of output predictions.

User Acceptance Testing:

User acceptance testing involves validating whether the GRU-based model

meets the requirements and expectations of its intended users. This typically
involves end-to-end testing scenarios where users interact with the model as
they would in a real-world clinical setting. User acceptance tests should cover
the entire prediction process, from data input to result interpretation, and gather
feedback on the model's performance, accuracy, and overall utility in clinical
practice.
 CHAPTER-6

CONCLUSION
In conclusion, this study has presented a comprehensive approach to
thyroid prediction using Gated Recurrent Unit (GRU) neural networks, aiming
to address the pressing need for accurate and timely diagnostic tools in
managing thyroid disorders. Through the exploration of temporal dependencies
inherent in medical data, our proposed system offers a promising avenue for
improving the accuracy and efficiency of thyroid condition predictions.By
leveraging a diverse set of patient records, including clinical variables such as
thyroid hormone levels, patient demographics, and historical medical history,
our GRU-based prediction model demonstrates notable efficacy in capturing
sequential patterns and making accurate predictions. The dynamic nature of
GRU units enables the model to effectively capture long-term dependencies
within the input data, providing a robust foundation for thyroid prediction.Our
experimental results showcase the superior performance of the GRU-based
approach compared to traditional machine learning models, highlighting its
ability to capture complex temporal patterns inherent in medical data. Moreover,
the interpretability of the model enhances its applicability in clinical settings,
allowing clinicians to gain insights into the factors influencing thyroid condition
predictions and make informed decisions.This work contributes to ongoing
efforts in developing advanced machine learning techniques for medical
diagnosis, specifically focusing on thyroid disorders. The proposed GRU-based
model offers a potential avenue for enhancing the accuracy and efficiency of
thyroid condition predictions, ultimately aiding clinicians in early diagnosis and
intervention.Moving forward, further research could explore the integration of
additional data sources, such as genetic information or environmental factors, to
improve the predictive performance of the model. Additionally, validation
studies in diverse clinical settings and patient populations would be valuable for
assessing the generalizability and scalability of the proposed approach.Overall,
our study underscores the potential of GRU neural networks in revolutionizing
thyroid prediction and management paradigms, ultimately contributing to
improved patient outcomes and healthcare delivery.
References

1. M. Riajuliislam, K. Z. Rahim and A. Mahmud, "Prediction of Thyroid

Disease(Hypothyroid) in Early Stage Using Feature Selection and
Classification Techniques," 2021 International Conference on
Information and Communication Technology for Sustainable
Development (ICICT4SD), Dhaka, Bangladesh, 2021, pp. 60-64,
I. Ahmed, R. Mohiuddin, M. A. Muqeet, J. A. Kumar and A.
Thaniserikaran, "Thyroid Cancer Detection using Deep Neural
Network," 2022 International Conference on Applied Artificial
Intelligence and Computing (ICAAIC), Salem, India, 2022, pp.
166-169,
2. V. T. Manh et al., "Multi-Attribute Attention Network for Interpretable
Diagnosis of Thyroid Nodules in Ultrasound Images," in IEEE
Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol.
69, no. 9, pp. 2611-2620, Sept. 2022, doi:
10.1109/TUFFC.2022.3190012.
3. S. Nasim, M. S. Almutairi, K. Munir, A. Raza and F. Younas, "A Novel
Approach for Polycystic Ovary Syndrome Prediction Using Machine
Learning in Bioinformatics," in IEEE Access, vol. 10, pp. 97610-97624,
2022, doi: 10.1109/ACCESS.2022.3205587.
4. E. Dritsas, S. Alexiou and K. Moustakas, "Efficient Data-driven Machine
Learning Models for Hypertension Risk Prediction," 2022 International
Conference on INnovations in Intelligent SysTems and Applications
(INISTA), Biarritz, France, 2022, pp. 1-6,
5. H. M. Almahshi, E. A. Almasri, H. Alquran, W. A. Mustafa and A.
Alkhayyat, "Hypothyroidism Prediction and Detection Using Machine
Learning," 2022 5th International Conference on Engineering Technology
 and its Applications (IICETA), Al-Najaf, Iraq, 2022, pp. 159-163, doi:
10.1109/IICETA54559.2022.9888736.
6. B. Saju, V. Asha, A. Prasad, V. A, A. S and S. P. Sreeja, "Prediction
Analysis of Hypothyroidism by Association," 2023 Third International
Conference on Advances in Electrical, Computing, Communication and
Sustainable Technologies (ICAECT), Bhilai, India, 2023, pp. 1-6, doi:
10.1109/ICAECT57570.2023.10117641.
7. S. Mondal, S. Dutta, S. Ghosh, S. Gupta, D. Kakati and A. Nag, "Thyroid
Disease Prediction Model on Boosting-based Stacking Ensemble
Approach," 2023 IEEE 8th International Conference for Convergence in
Technology (I2CT), Lonavla, India, 2023, pp. 1-6, doi:
10.1109/I2CT57861.2023.10126389.
8. R. V. Reddy, M. G. S, A. M. Ali, P. Yamini Devi, M. M. V and Y.
Sucharitha, "Empirical Method for Thyroid Disease Prediction Using a
Deep Learning Techniques," 2023 IEEE 8th International Conference for
Convergence in Technology (I2CT), Lonavla, India, 2023, pp. 1-5
9. Pallavi and H. Singh, "Prediction of Thyroid Disease using Deep
Learning Techniques," 2023 Second International Conference on
Electrical, Electronics, Information and Communication Technologies
(ICEEICT), Trichirappalli, India, 2023, pp. 1-7,
10.E. Mohan et al., "Thyroid Detection and Classification Using DNN Based
on Hybrid Meta-Heuristic and LSTM Technique," in IEEE Access, vol.
11, pp. 68127-68138, 2023, doi: 10.1109/ACCESS.2023.3289511.
11.C. P. Rani, T. Nagaraju, N. S. H. Vardhan, P. N. Teja and P. Charishma,
"Machine Learning Model for Accurate Prediction of Thyroid Disease,"
2023 International Conference on Advances in Computing,
Communication and Applied Informatics (ACCAI), Chennai, India, 2023,
pp. 1-7, doi: 10.1109/ACCAI58221.2023.10199375.
 A. Saini, K. Guleria and S. Sharma, "Machine Learning
Approaches for Early Identification of Thyroid Disease,"
2023 World Conference on Communication & Computing
(WCONF), RAIPUR, India, 2023, pp. 1-6, doi:
10.1109/WCONF58270.2023.10235086.
12.S. Ouartani, N. Taleb and B. Ayoub, "Integrating Ontology and Machine
Learning for Enhanced Decision Support in Thyroid Disease Prediction,"
2023 International Conference on Decision Aid Sciences and
Applications (DASA), Annaba, Algeria, 2023, pp. 208-212, doi:

A. Sutradhar et al., "An Intelligent Thyroid Diagnosis System

Utilising Multiple Ensemble and Explainable Algorithms
with Medical Supported Attributes," in IEEE Transactions
on Artificial Intelligence, doi: 10.1109/TAI.2023.3327981.

13.V. K. Singh, K. K. Nagwanshi, D. K. Singh, A. Lal Mahto, S. Mishra and

N. Behar, "A Collection of ML Algorithms for Detection of Thyroid : A
Computer Based Approach to Handle the Critical Medical Issue," 2023
IEEE 5th International Conference on Cybernetics, Cognition and
Machine Learning Applications (ICCCMLA), Hamburg, Germany, 2023,
pp. 534-539

A. Tiwari, G. Singh, S. Singh and T. Jain, "Thyroid Disease

Detection Using Supervised Machine Learning
Techniques," 2023 3rd International Conference on
 Innovative Sustainable Computational Technologies
(CISCT), Dehradun, India, 2023, pp. 1-6, doi:
10.1109/CISCT57197.2023.10351305.

14.T. Khan, "Application of Two-Class Neural Network-based classification

model to predict the onset of Thyroid Disease," 2021 11th International
Conference on Cloud Computing, Data Science & Engineering
(Confluence), Noida, India, 2021, pp. 114-118, doi:
10.1109/Confluence51648.2021.9377097