Deep Learning-Based Approach for Thyroid Dysfunction Prediction

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 04 | Apr 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 224
Deep Learning-Based Approach for Thyroid Dysfunction Prediction
Tushar Bhatia
Student, Department of Computer Science and Engineering, HMR Institute of Technology and Management, Delhi,
India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Globally, thyroid dysfunction is a major health
concern caused due to irregular hormone production by the
thyroid gland. Millions of populations are getting affected by
this disease regularly. Accurate diagnosis of thyroid
dysfunction is crucial for effectivetreatmentandmanagement
of the disease, but this is challenging given the condition’s
complex and varied symptoms. In this paper, a deep learning-
based neural network algorithm for generating predictions is
constructed based on a dataset of approximately3772patient
records with 28 features. The ArtificialNeuralNetwork(ANN)
model was trained and evaluated using standard machine
learning techniques and achievedhigh-levelaccuracy(98.8%)
in identifying instances of thyroid dysfunction. The findings
demonstrate that the proposed ANN model can be a reliable
and effective tool for early diagnosis of thyroid dysfunction.
The suggested model has several advantages, including its
ability to handle a large number of input parameters and its
ability to learn intricate relationships between input and
output variables. However, further research is required to
assess if the suggested approach can apply to more extensive
and diverse patient populations. Overall, the results of this
study lay out the potential of machine learning and ANN
models in the diagnosis of thyroid dysfunction and may aid in
creating more precise and effective diagnostic equipment for
this prevalent endocrine illness.
Key Words: Thyroid Dysfunction, Deep Learning, Neural
Network, Artificial Neural Network, Machine Learning,
accuracy, endocrine illness.
1.INTRODUCTION
The thyroid gland is a tiny, butterfly-shaped organ situated
in the front of the neck, surrounding the windpipe. Ourbody
contains glands, which produce and releasecompoundsthat
help the body to perform a specific function. The thyroid
gland produces hormones, namely levothyroxine(T4) and
triiodothyronine(T3), which assistinregulatingmetabolism,
heart rate, body temperature, and other essential processes.
When the thyroid gland is overactive or inactive, it can lead
to various health problems.
Thyroid dysfunction is a widespread endocrine disorder
affecting millions worldwide, irrespectiveofage,gender, and
ethnicity. It occurs when the thyroid gland either produces
excess or insufficient hormones, which can result in several
health issues. Hypothyroidism, characterizedbylowthyroid
hormone levels, and hyperthyroidism, characterizedbyhigh
thyroid hormone levels, are the most common thyroid
disorders. It can affect bodily functions like energy
production, weight management, and mood regulation.
Symptoms of thyroid dysfunction can vary widely and
include fatigue, weight gain, depression, and anxiety. Early
detection and treatment of thyroid disorders are essential
for managing the condition and avoiding severe
complications. Diagnosing thyroid dysfunction requires a
combination of clinical evaluation, biochemical tests, and
imaging techniques. However, traditional diagnostic
methods are time-consuming, expensive, and require
specialized tools and expertise. Therefore,thereisa need for
a methodical and accurate approach to the identification of
thyroid disorder.
Deep Learning-based model architecture has emerged as a
convincing technique for improving the efficiency ofthyroid
dysfunction prediction. This paperpresentsa DeepLearning
Artificial Neural Network (ANN) model for making a
prediction using clinical and biochemical parameters.
Fig -1: Thyroid gland
1.1 Deep Learning
Deep Learning lies withinthestrataofmachinelearning(ML)
and artificial intelligence (AI). Its methodology is influenced
by the human brain's structure and function. It involves
training artificial neural networks, which are complex
mathematical models that can learn to recognize patterns in
data.
Deep Learning has riseninprominenceinrecentyears,owing
to the abundance of extensive amounts of data and powerful
computing resources. It has enabled significant advances in
several fields like natural language processing, computer
vison, speech recognition, and medical science.

One of the critical strengths of deep learning is its ability to
extract featuresfromrawdataautomatically.Thismeansthat
it can solve problems where traditional machine learning
approaches require hand-craftedfeaturesordomain-specific
knowledge.
Deep learning models generally feature layers of
interconnected nodes, or neurons, that perform a simple
mathematical operation on inputs. The output of one layer is
fed into the next, and each layer learns to recognize more
complex data features. Model training involves adjusting the
weights of the connections between the neurons with an
objective to minimize a cost function, which evaluates the
difference between the predicted values and actual values.
This is often achieved using an algorithm called stochastic
gradientdescent,whichiterativelyupdatestheweightsbased
on the gradient of the cost function concerning the weights.
1.2 Artificial Neural Networks
Artificial Neural Networks (ANNs) are deep learning-based
models designed to emulate the structure and function of
biological neurons in the brain. ANNs are constructed up of
layers of interconnected nodes, including an inputlayer,one
or more hidden layers, and an output layer. Theyareutilized
for analysing data patterns and makingpredictionsbased on
information.
Each neuron in an ANN receives inputs from neurons in the
preceding layer, which are merged and processed with the
help of an activation function to generate an output. During
training, the weights of the connections between neurons
are altered to minimize the cost.
There are numerous types of ANNs, each having its own set
of advantages and disadvantages. Feedforward neural
networks are the simplest type, with layers that process
unidirectional flow of information from the input to the
output layer. Recurrent Neural Networks (RNN)aresuitable
for tasks involving data sequences due to their cyclic
connections, allowing information to flow in cycles.
Convolutional Neural Networks (CNN) are specialized for
processing images and consist of layers that applya seriesof
convolutional filters to the input image, allowing the
network to learn to recognize patterns at different scales.
Regardless of their success, ANNshavesomelimitations, like
the requirement for substantial amounts oftrainingdata and
the complexity of interpreting the inner workings of the
models. However, they continue to be an active area of
research and development and are likely to play a pivotal
role in the future of artificial intelligence.
2. LITERATURE REVIEW
[1] This study utilizes a range of classification models to
diagnose thyroid disorders based on parameters including
TSH(ThyroidStimulatingHormone),T4U,andgoitre.Various
classification techniques, including K-nearest neighbor
(KNN), were employed to support the study's findings.Naïve
Bayes and Support Vector Machine algorithms are also
implemented. The test was carried out with the help of a
Rapid Miner instrument. The results revealed that the KNN
was more accurate than Naïve-Bayes in detecting thyroid
disorder, with a 93.44% accuracy. The suggested KNN
technique enhanced classification accuracy and contributed
to better results. KNN exhibited superior performance
compared to other methods, since the factors were
independent of each other.
[2] In this research paper, the authors developed a machine
learning algorithm to predict the mosteffectivetreatmentfor
thyroid disease based on patient characteristics and medical
history The data was collected from 282 patients with
thyroidillnessandperformancewasevaluatedusingmultiple
ML algorithms. The findings indicated that the Random
Forest algorithm performed the best, getting an accuracy of
77.83% in predicting the most effective treatment. The
authors noticed that the model could support clinical
decision-making in treating thyroid disease, potentially
improving patient outcomes.
[3] In this study, the authors proposed an ensemble method
for classifying thyroid disease that involves optimization of
features. They obtained data from patients diagnosed with
thyroid disease and extracted a set of parameters related to
the disease. They then used an ensemble classifier that
combined several machine learning methods to predict the
type of thyroid disease based on extracted features. The
results showed that the proposed ensemble approach
outperformed individual machine learning algorithms
regarding accuracy. The study demonstrates the potential of
an ensemble approach for enhancing theefficiencyofthyroid
disease classification.
[4] The authors of this research constructed a deep-learning
model for predicting thyroid disorders by incorporating
clinical data fromover20,000Indianpatients.Themodelwas
based on a CNN architecture and achieved an accuracy of
92.6% and a specificity of 96.3% in predicting
hypothyroidism andanaccuracyof91.5%,andaspecificityof
95% in predicting hyperthyroidism. The study highlightsthe
potential of deep learning models for diagnosing and
managing thyroid disease in India.
[5] This paper proposes an ANN model for the automated
prediction of thyroid disease. The authors collected thyroid
samples and trained an ANN model using an 80:20 ratio split
of data for training and testing. The model achieved an
average accuracy of 85% during training and 82% during
testing. The study concludes that ANNs are a flexible and
robust technique for thyroid disease diagnosis, with high
reliability in different sampling situations.

3. METHODOLOGY
The proposed ANN-based approachconsistsof4stages:data
collection, data preprocessing, model training, and model
evaluation. This section laysoutanoverviewofvarioussteps
involved in the prediction process. The figure below gives a
representation of the workflow involved.
Fig -2: Workflow diagram
(I) Data collection
To conduct this research, a dataset comprising clinical and
biochemical parametersofpatients withandwithoutthyroid
disease was obtained from the UCI Machine Learning
Repository. The dataset consists of 3772 instances, each
containing 28 attributes, including age, sex, thyroxine and
antithyroid medication details, thyroid surgery, pregnancy,
sickness, hyperthyroid and hypothyroid queries, tumor and
psych information, TSH, T3, and T4 levels, and various other
chemical and biochemical parameters that are commonly
used in diagnosing thyroid dysfunction. Table-1andTable-2
show the numerical and categorical attributes respectively.
Table-1: Numerical Attributes
S.No. Attribute Name Data Type
1 age object
2. TSH object
3. T3 object
4. TT4 object
5. T4U object
6. FTI object
Table-2: Categorical Attributes
.No. Attribute Name label
1 sex F = female, M=male
2. on thyroxine f= false, t =true
3. query on thyroxine f= false, t =true
4. On antithyroid medication f= false, t =true
5. sick f= false, t =true
6. pregnant f= false, t =true
7. thyroid surgery f= false, t =true
8. I131 treatment f= false, t =true
9. query hypothyroid f= false, t =true
10. query hyperthyroid f= false, t =true
11. lithium f= false, t =true
12. goitre f= false, t =true
13. tumor f= false, t =true
14. hypopituitary f= false, t =true
15. psych f= false, t =true
16. TSH measured f= false, t =true
17. T3 measured f= false, t =true
18. TT4 measured f= false, t =true
19. FTI measured f= false, t =true
20. TBG measured f= false, t =true
21. referral source other, SVHC, SVI
22. Binary Class P = positive,
N= negative

(II) Data Preprocessing
Data preprocessing is a crucial stageinanymachinelearning
project. The following steps are performed in this stage:
• Data Cleaning: The 'binaryClass' column in the
dataset is converted to numerical values, 't' and 'f'
values are replaced with 1 and O, respectively, and
'?' values are replaced with NaN.
• Feature Engineering: The 'sex' column is converted
to numerical values, and the 'referral source'
column is dropped from the dataset.
• Handling missing values: The missing values are
imputed with the mean value of the respective
column.
• Splitting the dataset: The dataset is divided into
training and testing sets with the 'train_test_split()'
function from sklearn.
• Feature scaling: The training and testing sets are
scaled using the 'StandardScaler()' function to
ensure all the features are on the same scale.
(III) Model Building
The proposed research involves the creation of a deep
learning model based on ANN architecturetopredictthyroid
disease. The model is implemented using the Tensorflow
Keras API. The model's architecturecomprisesa sequenceof
four densely connected layers, where each neuron is linked
to every neuron in the next layer. The input layer has 256
neurons, which is equal to the number of features in the
input dataset and uses the Rectified Linear Unit (ReLU)
activation function.
The dropout layer is then added after the first, second, and
third hidden layers, respectively, with 0.4, 0.3, and 0.2
dropout rates. Dropout is, basically, a regularization
technique used in deep learning models to prevent
overfitting. It randomly drops out some oftheneuronsinthe
hidden layer during training, which reduces the co-
dependence between neurons and improves generalization.
The second hidden layer has 128 neurons, and the third
hidden layer has 63 neurons, both activated using the ReLU
activation function. The final output layer has only one
neuron, which produces the probability output of thebinary
classification problem (0 or 1) using the sigmoid activation
function. Figure 2 visualizes the developed ANN model
architecture.
Next, the model is compiled using binary cross-entropy loss
and the Adam optimizer.
ReduceLROnPlateau, ModelCheckpoint, and EarlyStopping
are the callback functions used to monitor the training
process, adjust the learning rate, save the best model, and
stop the training if the accuracy becomes stable for a given
number of epochs.
The model is then fit using the 'fit()' method with the
training data, for 80 epochs, a batch size of 48, and a
validation split of 0.1.
Fig -3: Proposed ANN model architecture
(IV) Model Evaluation
The trained model is evaluated using the test dataset that
has not been used in the training process. The predictions
are compared to the true labels using a confusion matrix,
which is a valuable tool for evaluating the performance of a
binary classification model. It is a table that shows the true
positives (TP), true negatives (TN), false positives (FP), and
false negatives (FN) predictions of the model. From the
confusion matrix,variousperformancemetricslikeaccuracy,
precision, recall, and F1 score are also calculated.
The accuracy metric measures the proportion of accurate
predictions made by the model and it can be described as
follows:

Sometimes, accuracy can be misleading when the dataset is
uneven, indicating that one class is substantially more
prevalent than the other. In such instances, additional
metrics such as precision, recall, andvF1-score are more
informative. The precision metric measures the percentage
of true positives among the predicted positives andisa good
indication of the model's ability to prevent false positives.
The recall metric determinestheproportionoftruepositives
among the real positives and is a good measure of the
model's capacity to detect all positive cases. The F1 score
metric combines precision and recall to measure the
accuracy of a binary classification model. It is the harmonic
mean of precision and recall. The formulas for calculating
precision, recall, and F1oscore are defined as follows:
Overall, evaluating a trained model using a test dataset and
various performance metrics measures how well the model
performs on unseen data and its ability to predict positive
and negative cases correctly.
4. RESULTS & DISCUSSION
Based on the evaluation metrics, the trained ANN model
effectively recognized thyroid dysfunction. The confusion
matrix indicated that the model predicted 691 true positive
(TP) cases and 55 true negative (TN) cases, with only 3 false
positive (FP) and 6 false negative (FN) predictions. The
overall accuracy score was 0.9888, the precision score was
0.992, the recall score was 0.992, and the F1 score was
0.970. The curve depicted in Figure 4 depicts how the
model's accuracy on both the training and test datasets
evolves throughout multiple epochs and rises over time as
the model learns to match the data better. In summary, the
proposed ANN model demonstrated high accuracy and
balanced performance in identifying thyroid dysfunction.
The strong performance on the test dataset implies that the
model is not overfitting to the training set. These results
suggest that the model can potentially assist in diagnosing
thyroid dysfunction. Nevertheless, some limitations to this
study should be considered, like the fact thatthedataset was
not diverse enough or that there were potential biasesinthe
dataset that may have influenced the performance.
Additionally, the model wastrainedandtestedusingmedical
record data, and its performance can be improved by
incorporating other clinical information, such as patient
history and imaging results.
Fig -4: ANN model accuracy
5. CONCLUSION & FUTURE SCOPE
In conclusion, the developed deep learning model exhibited
high accuracy and specificity, which indicates its potential
usefulness in clinical practice. The model outperformed
traditional machine learning algorithms, emphasizing the
potential of deep Learning based neural network models in
thyroid dysfunction prediction. Future studies should
concentrate on expanding the dataset, incorporating
additional relevant features, and further validating the
model's performance on diverse populations. Moreover,the
model can be integrated into clinical decision support
systems to help physicians in accurate thyroiddiagnosisand
management.
REFERENCES
[1] K. Chandel, S. Arora, S. K. Gupta, and V. K. Panchal, "A
comparative study on thyroid diseasedetectionusingK-
nearest neighbor and naïve bayes classification
techniques," CSI Transactions on ICT, vol. 4, no. 2-4, pp.
313-319, Dec. 2016.
[2] L. Aversano, M. L. Bernardi, M. Cimitile,M.Iammarino,P.
E. Macchia, I. C. Nettore and C. Verdone, "Thyroid
Disease Treatment prediction with machine learning
approaches," Procedia Computer Science, vol. 192, pp.
1031-1040, 2021, doi: 10.1016/j.procs.2021.08.106.
[3] A. Shrivas and P. Ambastha, "An ensemble approach for
classification of thyroid disease with feature
optimization," International Education and Research
Journal, vol. 3, no. 5, pp. 1–4, 2019.
[4] A. Singh, S. Dubey, and S. K. Patil, "Deep Learning Based
Prediction of Thyroid Disorder," in 2021 5th
International Conference on Intelligent Computing and
Control Systems (ICICCS), pp. 1149-1153, IEEE, 2021.

[5] V. V. Hegde, and D. N., "Automated Prediction ofThyroid
Disease using ANN," International Journal of Innovative
Research in Science, Engineering and Technology,vol.5,
no. 5, pp. 268-272, May 2016.
[6] R. Chaganti, F. Rustam, J. L. Mazón, C. L. Rodríguez and I.
Ashraf, "Thyroid Disease Prediction Using Selective
Features and Machine Learning Techniques," Cancers,
vol. 14, no. 16, p. 3914, 2022.
[7] G. Kaur, K. Sidhu and E. Kaur, "Artificial neural networks
for diagnosis of thyroid disease," International Journal
for Technological Research in Engineering, vol. 2, no. 1,
pp. 56-59, Sep. 2014, ISSN: 2347-4718.
[8] A. Shukla, R. Tiwari, P. Kaur and R.R. Janghel, "Diagnosis
of Thyroid Disorders using Artificial Neural Networks,"
2009 IEEE International Advance Computing
Conference, 2009, pp. 1016-1020, doi:
10.1109/IADCC.2009.4809149.
[9] R. Chaganti, F. Rustam, I. De la Torre Díez, J. L. Mazón, C.
Rodríguez and I. Ashraf, "Thyroid Disease Prediction
Using Selective Features and Machine Learning
Techniques," Cancers, vol. 14, p. 3914, 2022. doi:
10.3390/cancers14163914.
[10] A. Banduni and R. Mehra, "Interactive Thyroid Disease
Prediction System Using Machine Learning Technique,"
2019 6th International Conference on Parallel,
Distributed and Grid Computing (PDGC), 2019, pp. 689-
693, doi: 10.1109/PDGC.2018.8745910.
[11] V.Sarasvathi and Dr.A.Santhakumaran, “Towards
Artificial Neural Network Model To Diagnose Thyroid
Problems”, Global Journal of Computer Science and
Technology, Vol. 11, No. 5, pp.53-55, 2011.
[12] https://my.clevelandclinic.org/health/diseases/8541-
thyroid-disease

Deep Learning-Based Approach for Thyroid Dysfunction Prediction

More Related Content

Similar to Deep Learning-Based Approach for Thyroid Dysfunction Prediction (20)

More from IRJET Journal (20)

Recently uploaded (20)

Deep Learning-Based Approach for Thyroid Dysfunction Prediction