Comparative Analysis of Supervised Machine and Deep Learning Algorithms for Kyphosis Disease Detection

doi:10.21203/rs.3.rs-2091003/v1

Download PDF

Research Article

Comparative Analysis of Supervised Machine and Deep Learning Algorithms for Kyphosis Disease Detection

https://doi.org/10.21203/rs.3.rs-2091003/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

The spinal curve that represents an abnormally rounded back is known as kyphosis, and it can be caused by stress, infectious disease, developmental abnormalities, genetic disease, and occasionally iatrogenic disease. Due to the fact that it provides technical explanations and a significant standard, the emergence of machine learning and deep learning has demonstrated that they can better characterize data. Machine Learning and Deep Learning are promising approaches that help in the prediction, diagnosing of sickness. Compared to conventional computing algorithms, deep learning algorithms are far more effective in disease detection and diagnosing. In this study, Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM) and Deep Neural Network (DNN) models are applied to detect the presence of kyphosis in youngsters supported medicine information and compared to point out the rise in the potency of predictions associated with kyphosis. Each classification algorithm that makes use of the hyperparameter tuning and control feature enhances the algorithms' overall prediction performance, which enhances the algorithms' overall performance. The potency of the planned models were enforced and checked over medical dataset utilized from Kaggle. Overall, the DNN model performed best among the wide range of different models and achieved maximum accuracy and other performance metrics scores derived from the stratified K-Fold cross-validation. It is advised that after a patient has completed a clinical procedure, the DNN model be trained to identify and forecast kyphosis disease.

Machine Learning

Deep Learning

Random Forest

K-Nearest Neighbors

Support Vector Machine

Deep Neural Network

Making predictions using information presented to the computer is a technique known as "machine learning” [1]. Machine Learning (ML) is the conception through which machines will be trained and adapted to new knowledge without human intervention. Deep Learning (DL) is a sort of artificial intelligence and machine learning that, when applied to specific types of data, simulates human behavior. Artificial Intelligence (AI) is a broader umbrella under which ML and DL falls as revealed in Figure 1.

Machine Learning (ML) algorithms arrange the info, gain from it, assemble bits of data and make forecasts that upheld the knowledge it investigated while not the prerequisite for added express programming. ML algorithms are comprehensively arranged into supervised, unsupervised, and reinforcement learning. During this paper, supervised learning was targeted.

One machine learning technique called deep learning allows computers to analyse vast volumes of data and produce intermediate results, which can then be used to support any response [2]. Deep Learning (DL) is another area of artificial intelligence that deals with algorithms that are influenced by how the brain develops and functions (ANN). DL employs ANN with a second hidden layer. Neural networks are set up in a cascading pattern to process non-linear data sets. As demonstrated in Figure 2, deep learning is a technique that uses several hidden layers to discover the key inside the data. Other diseases can also be diagnosed using deep learning. [3].

Thanks to developments in data technologies like storage capacity, processing power, and data transmission rates, machine learning has been widely implemented in a variety of industries, including healthcare. Deep learning enables the healthcare industry to analyze data at lightning-fast speeds while maintaining excellent accuracy [4].

Kyphosis might be a deformity of the backbone (spine) as given away in Figure 3. A toddler is born with humpback or he or she is going to be ready to develop it thanks to alternative health issues. Extreme humpback will cause torment and distorting. The spinal curve that represents an abnormally rounded back is known as kyphosis, and it can be caused by stress, infectious disease, developmental abnormalities, genetic disease, and occasionally iatrogenic disease [5]. Age, as well as the origin and consequences of the curve, all affect how kyphosis is treated. A more thorough diagnosis of kyphosis disease in children at an early age helps prevent uncommon spinal vertebral difficulties [6].

Automating medical diagnosis would be tremendously beneficial because it is a crucial yet challenging task that needs to be completed correctly and consistently [7]. Many individuals have kyphosis operations, and even after the procedure, the disease may still exist. Therefore, the task at hand is to forecast, based on the patient's many characteristics, whether or not the patient would still have kyphosis after the procedure. Since the disorder of kyphosis may be classified, supervised learning systems are the focus. In order to gain insight into which model performs the best, this study compares the performance of supervised ML/DL models in kyphosis diagnosis across datasets.

Alternative medical professionals have already done a decent job of diagnosing and predicting disease by comparing and applying a number of machine learning algorithms.

In [5], Kyphosis must be addressed since some cases are congenital and some are iatrogenic, depending on the patient's needs and current state. Depending on the severity of the patterns, patients who develop kyphosis at a young age may be treated effectively with counseling and exercises; in other cases they may require surgery.

Machine learning methods such Random Forest, Support Vector Machines & Artificial Neural Networks were used in [6] to identify and forecast the kyphosis disease. Artificial Neural Networks were discovered to be more accurate than the other two techniques. SVM-Linear model accuracies of 79.01 percent and 77.76 percent, SVM-RBF model accuracies of 82.72 percent and 85.19 percent, and SVM-Poly model accuracies of 82.72 percent and 85.19 percent were obtained by the authors using the SVM, RF & ANN models. They discovered that the ANN (3-6-6-1) version performed the best after using grid search techniques, with 86.42 and 85.19 percent supporting 10-fold & 5-fold cross validations, respectively.

The authors in [8] looked at the connection between a front postural stability, round shoulders, and progressive kyphosis illness. These three issues could appear separately or together. There was a significant correlation between cervical lordosis values and thoracic kyphosis values.

The authors of [9] examined adult kyphosis and covered a variety of surgical and nonsurgical therapies. Physical therapy is part of the non-surgical treatment, which hasn't changed in years but was used to lessen the symptoms. To better the outcomes, however, the surgical procedure has undergone a significant evolution. The three cases that were shown helped to explain these surgical procedures for various ailments.

Artificial intelligence and machine learning have made numerous advances to the field of spine research, according to Fabio Galbusera et al. [10]. It includes various tools and cutting-edge techniques developed for radiological image segmentation, clinical outcome prediction, and picture identification. The author discusses several decision support systems, computer-aided diagnosis, and problems in this review. Additionally, the author turned in a responsible paper on data security and privacy, including all pertinent problems.

In [11], ML performed admirably and has a lot of promise in the spine. ML could aid clinical staff in raising the medical standard, increasing work productivity, and lowering adverse occurrences. However, in order for clinicians to embrace the use of models in actual work, there must be more randomized controlled trials and improvements in interpretability.

Kavitha et al. [12] used thirteen inputs, two hidden nodes, and one output to model an ANN for detecting heart disease. Additionally, ANN was employed in a study by Noura [13] to predict cardiac disease with an accuracy of 88%. Due to its prognosticative power as a model, According to Animesh et al. [14], ANN is frequently used in diagnosis and care systems. Mrudula et al[15] .'s comparison of SVM and ANN for the prediction of heart disease revealed that the ANN model outperformed the SVM model.

Additionally, Tahmooresi et al. [16] evaluated a variety of machine learning techniques for detecting breast cancer, including SVM, ANN, KNN, and DT. They did a thorough literature analysis on numerous machine learning methods to learn more. They came to the conclusion that SVM performed better than all other models, with 99.8% accuracy, in recognizing the illness known as breast cancer.

In their study, Kuo et al. [17] looked at how the NB, SVM, LR, DT, and RF models were presented in predicting the clinical costs of spinal fusion. They came to the conclusion that, when compared to the other procedures, the RF model performed the best in terms of prognostication, with an accuracy of 84.30 percent.

Abdullah et al. [18] used Random Forest & K-Nearest Neighbors models to detect spinal anomalies. They discovered that the K-Nearest Neighbors (KNN) model surpassed the Random Forest (RF) model, which had 79.57 percent accuracy, with an accuracy of 85.32 percent.

The SVM, LR, bagging SVM & bagging LR models, which are all publicly accessible in the Kaggle repository, are used in [19] on a dataset of 310 samples. Several performance metrics, such as training and testing accuracy, memory, and miss rate, were used to assess the performance of the classification of aberrant and normal spinal patients. In addition, curves for precision-recall, kernel parameter optimization, and receiver operating characteristic analysis are employed to assess classifier models. The results show that the observed training accuracies for SVM, LR, bagging SVM & bagging LR are 86.30%, 85.47%, 86.72% & 85.06%, respectively, when 78% of the data are used for training. SVM, LR, bagging SVM & bagging LR, on the other hand, all perform equally well on the test dataset, with an accuracy of 86.96%. SVM, however, stands out from the competition because of its superior recall value and miss rate.

A literature analysis revealed that the art of labour is in decline due to various problems, including a loss of accuracy and a wider reach. Following the review, we attempted to apply the chosen classification algorithms to a dataset on the kyphosis disease that was taken from the Kaggle repository. In this study, we examined how well the most widely used machine learning and deep learning algorithms, RF, SVM, KNN & DNN, performed when tuned hyperparameters and stratified K-fold cross validation were used.

The implementation of the suggested research project is built upon this section. In this section, the pre-processing processes are described, followed by a discussion of the implementation-specific final categorization algorithms. In the current study, RF, KNN, SVM & DNN algorithms are used to create models to identify individuals with kyphosis who have received previous therapy for the condition and have personal records that indicate when they will require skilled surgery. The models' impacts are then evaluated and studied.

An explanation of the dataset: Kyphosis dataset was obtained from Kaggle [20]. This kyphosis dataset has address patients’ records that had a remedial kyphosis disorder procedure. This choice of the knowledge represents three inputs (Age, Variety, and Start) and one output (Kyphosis). The credits of the dataset are clarified in Table 1 given beneath:

Table 1. Description of Dataset

Characteristic	Explanation
Kyphosis	Whether or not the kyphosis issue persisted after the procedure
Age	Monthly age of the patient
Number	Involved vertebrae with in procedure
Start	The count of the surgery's first or highest vertebra that was affected

Data Preprocessing: The Scikit-Learn software was used to preprocess the data. Figure 4 demonstrates how the kyphosis region was split into 0s and 1s using the Label Encoder. The data were standardized using the StandardScaler function of the sklearn library to encourage better performances with the ML/DL models.

Classification Learning Algorithms: A computer programme is educated to make new observations or classifications depending on the data it is provided through the process of classification in machine learning and statistics. The following learning algorithms (RF, SVM, KNN and DNN) were used in proposed models for training and prediction of outcomes.

Random Forest (RF): To improve the dataset's predictive accuracy, Random Forest is likely a classifier that consists of a group of selection bushes on various subsets of the given dataset. The random forest uses the forecasts from each tree and the majority votes of predictions to anticipate the final phrase output rather than relying solely on one selection tree [21, 22].

Support Vector Machine (SVM): One of the best classifiers, SVM exhibits some degree of linearity. The SVM is supported by sound mathematical intuition and has the ability to handle some nonlinear scenarios by employing a nonlinear basis function [23]. The goal of the Support Vector Machine rule is to create the sole decision boundary or line that may categorize n-dimensional space, allowing us to quickly assign the novel datum to the correct class over the long term [6, 24].

K-Nearest Neighbor (KNN): The KNN method simply saves the dataset during the preparation stage, and once it receives fresh data, it groups that data into a collection that is plentiful and equivalent towards the incoming information. [25, 26].

Deep Neural Network (DNN): A network with multiple layers and numerous neurons in each layer is known as a DNN. Every layer in a DNN will alternately turn on and off whenever another layer's output serves as that layer's input in the forward direction. An ANN and a deep neural network differ in that the former has 1 input layer, 1 output layer & at most 1 hidden layer. The later, however, should include more hidden layers [27]. Planned work has used 3 hidden layers for the aim of extracting a top quality feature from the dataset. The DNN model was implemented using the Keras model.

Hyperparameter Tuning & Control: Via training a model with existing information, we are competent to suit the model parameters. However, there are other reasonable parameters, called hyperparameters that cannot be directly learned from the regular training method. They're sometimes preset before the actual training method begins. These parameters categorical necessary properties of the model like its complexity or how briskly it needs to learn [28]. During this work, hyperparameter tuning was performed to pick the foremost effective parameters.

Cross validation: As we all know already, K-fold is applied that the model will learn well. It’s typically applied once one doesn't have enough information or uniformly distributed information [29]. K-fold cross validation additionally suffers from sampling. Stratified K-fold got to be hottest over K-fold once handling classification tasks with unbalanced class distributions. As a result, in the current work, to assess the models, stratified K-fold cross validation was employed.

Various Performance Measures:

The subsequent performance indicators are accustomed valuate predictions for stated classification problem using machine learning and deep learning models [30].

Confusion Matrix: An example of a confusion matrix is shown below. It is just a table with the dimensions "Actual" and "Predicted" and the scales "True Positives (TP), "True Negatives (TN)," "False Positives (FP)," and "False Negatives (FN)".

• TP− In this, each actual category & expected category of knowledge purpose is one.

• TN − In this, each actual category & expected category of knowledge purpose is zero.

• FP − In this, actual category of knowledge purpose is zero & expected category of knowledge purpose is one.

• FN − In this, actual category of knowledge purpose is one & expected category of knowledge purpose is zero.

Classification Accuracy: Classification accuracy should be outlined because the range of correct predictions created as a quantitative relation of all predictions created.

Accuracy of classification equals (TP + TN) / (TP + FP + FN + TN)

Precision: Because of the variety of accurate documents that ML/DL models returned, precision, which is used in document retrievals, is also described.

Precision equals TP / (TP + FP)

Sensitivity or Recall: Recall is also discussed since the ML/DL models' range of positive results.

Recall equals TP / (TP + FN)

Specificity: Specificity, in distinction to recall, is also outlined because the range of negatives came back by ML/DL models.

Specificity equals TN / (TN + FP)

F1-Score: The weighted average of recall and precision is used to generate the F1-score. Precision and recall both have an equal impact on the F1 score.

F1-score equals 2 x (recall x precision) / (recall + precision)

AUC- ROC Score: AUC-ROC metric can tell us regarding the competence of model in identifying the categories.

Balanced Accuracy Score: The balanced accuracy in binary and multiclass classification issues are accustomed cope with unbalanced datasets. It’s outlined because the average of recall obtained on every category. The simplest worth is one and also the worst worth is zero.

Proposed System

We demonstrate how the suggested predictive model's working design interacts with other significant deep learning and machine learning applications, tuning and control of hyperparameters, and model testing using k-fold cross validation. The proposed system has been used to classify individuals with, kyphosis illness and healthy individuals. The outcomes of varied ML/DL prognosticative models for kyphosis illness were tested. The favored machine learning models (RF, SVM, KNN), and deep learning models (DNN) were utilized within the system. The framework of the system is demonstrated in Figure 4.

We must select a dataset for Figure 6. that is related to the patients' kyphosis illness. Any and all discrepancies that might have existed in the dataset at the time the data was gathered must be removed during the preparation stage. The next step asks us to select the testing mode and the classification techniques that will be applied during implementation. With the help of the algorithms for hyperparameter tuning and control, it is now possible to put the classification algorithms that were previously explained into work.

We will ultimately need to compare and contrast each of the developed algorithms to see which one is best and offers the highest level of accuracy when predicting the outcome.

Implementation of Proposed System

The models were implemented via Google Colab notebook using the Python programing language version 3.6.9. With the help of the Keras model, the DNN model was implemented. The deep multilayer perceptron architecture with regularization and dropout employed in the developed DNN learning model is based on deep learning [31]. In this part, the aftereffects of the examination effort has been discussed by playing out an analytical investigation on the chosen dataset and ML/DL models with RF, SVM, KNN, and DNN algorithms supporting stratified K-Fold (K=5 and 10) cross validations. The results of the machine learning and deep learning models were given for model evaluation and compared after performing hyperparameter tuning with stratified K-Fold cross validation.

5.1. Exploratory Analysis: Exploratory examination showed that the absence of kyphosis in 79% of patients while the presence of kyphosis in 21% of patients as appeared with Figure 7. In Figure 8, it was possible to see a link between kyphosis and the number addressing 0.36. Figure 9 contains frequent examples of the kyphosis sickness, either present or missing, based on the characteristics of the individuals.

5.2. Random Forest (RF) Model Based Outcome: Utilizing stratified 5-fold cross validation, the RF model was developed. The achieved accuracy using numerous K-folds is shown in Table 2. Figure 10 illustrates how precisely Fold-3 has discriminated. The stratified 5-fold cross validation was validated by the RF model with 85.22% as a mean accuracy & 0.0485 as a std. deviation. After hyperparameter tuning using 5-fold stratified cross validation, the optimal parameter was selected at n_estimators=400 and criterion=gini.

Table 2. RF Model’s accomplished classification accuracies with stratified 5-fold cross validation

K-Folds	Accuracy
1	82.35%
2	81.25%
3	93.75%
4	81.25%
5	87.50%
Mean: 85.22% std. deviation: 0.0485

The stratified 10-fold cross validation had been carried out following extra preparation on the RF model. Table 3 provides the achieved accuracy with the multiple K-folds. Fold-5 had distinguished with foremost extreme accuracy which is envisioned in Figure 11. It is found that, the RF Model at that time accomplished 83.89% as a mean accuracy and 0.0982 as a std. deviation supported the 10-Fold stratified cross validation. Given hyperparameter tuning through stratified 10-fold cross validation, the finest parameter was selected at n_estimators=150 and criterion=gini.

Table 3. RF Model’s accomplished classification accuracies with stratified 10-fold cross validation

K-Folds	Accuracy
1	88.89%
2	87.50%
3	62.50%
4	75.00%
5	100.00%
6	87.50%
7	87.50%
8	75.00%
9	87.50%
10	87.50%
Mean: 83.89% std. deviation: 0.0982

In addition to classification accuracy, Table 4 provides other effective performance measures for the RF model based on stratified 5-fold & 10-fold cross validation.

Table 4. RF Model’s accomplished other performance measures with stratified 5-fold & 10-fold cross validation

K- Fold stratified cross validation	Recall/ Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1 Score	AUC-ROC Score
5-fold (K=5)	0.48	0.95	0.72	0.70	0.51	0.72
10-fold (K=10)	0.35	0.96	0.65	0.60	0.43	0.65

5.3. Support Vector Machine (SVM) Model Based Outcome: The SVM model was created using stratified 5-fold cross validation. Table 5 provides the achieved accuracy with the multiple K-folds. Fold-3 had discriminated with the utmost precision, as seen in Figure 12. It is discovered that the SVM model supported the stratified 5-fold cross validation by achieving 85.22% as a mean accuracy and 0.0839 as a standard deviation. The best parameters were selected after hyperparameter tuning using a stratified 5-fold cross validation, and they comprised kernel=rbf, gamma=0.1, and C=3.

Table 5. SVM Model’s accomplished classification accuracies with stratified 5-fold cross validation

K-Folds	Accuracy
1	82.35%
2	75.00%
3	100.00%
4	81.25%
5	87.50%
Mean: 85.22% std. deviation: 0.0839

Additional modeling work and stratified 10-fold cross validation were done on the SVM model. Table 6 provides the achieved accuracy with the multiple K-folds. Fold-6 had been separated with the utmost precision, as seen in Figure 13. Support Vector Machine (SVM) Model at that time was found to have supported the stratified 10-Fold cross validation with 85.14% as a mean accuracy & 0.0940 as a standard deviation. The best hyperparameter values were chosen to be kernel=rbf, gamma=0.1, and C=8 using stratified 10-fold cross validation.

Table 6. SVM Model’s accomplished classification accuracies with stratified 10-fold cross validation

K-Folds	Accuracy
1	88.89%
2	87.50%
3	62.50%
4	87.50%
5	87.50%
6	100.00%
7	87.50%
8	75.00%
9	87.50%
10	87.50%
Mean: 85.14% std. deviation: 0.0940

Apart from accuracy of classification, accomplished other performance measures for the SVM model based on stratified 5-fold & 10-fold cross validation are given away in Table 7.

Table 7. SVM Model’s accomplished other performance measures with stratified 5-fold & 10-fold cross validation

K- Fold stratified cross validation	Recall/ Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1 Score	AUC-ROC Score
5-fold (K=5)	0.48	0.95	0.72	0.68	0.51	0.72
10-fold (K=10)	0.55	0.94	0.75	0.68	0.54	0.75

5.4. K-Nearest Neighbor (KNN) Model Based Outcome: The KNN Model was created using stratified 5-fold cross validation. Table 8 provides the achieved accuracy with the multiple K-folds. Fold-3 and Fold-5 were recognized in Figure 14 with the highest degree of precision. The K-Nearest Neighbor (KNN) Model was found to have 85.22% a mean accuracy & 0.0740 as a standard deviation that supported the stratified 5-fold cross validation. The best parameters were chosen at n_Neighbor=9 and p=2 after hyperparameter tuning by stratified 5-fold cross validation.

Table 8. KNN Model’s accomplished classification accuracies with stratified 5-fold cross validation

K-Folds	Accuracy
1	82.35%
2	75.00%
3	93.75%
4	81.25%
5	93.75%
Mean: 85.22% std. deviation: 0.0740

The stratified 10-fold cross validation was carried out following additional K-Nearest Neighbor (KNN) Model preparation. Table 9 provides the accuracy results obtained with the numerous K-folds. Figure 15 shows how Fold-2, Fold-4, Fold-5, Fold-6, Fold-7, Fold-9, and Fold-10 were differentiated with the utmost precision. It is found that, K-Nearest Neighbor (KNN) Model at that time accomplished 84.03% as a mean accuracy and 0.0535 as a std. deviation supported the 10-fold stratified cross validation. Given hyperparameter tuning through 10-fold stratified cross validation, the finest parameters had been chosen to be n_Neighbor=11 and p=2.

Table 9. KNN Model’s accomplished classification accuracies with stratified 10-fold cross validation

K-Folds	Accuracy
1	77.78%
2	87.50%
3	75.00%
4	87.50%
5	87.50%
6	87.50%
7	87.50%
8	75.00%
9	87.50%
10	87.50%
Mean: 84.03% std. deviation: 0.0535

Table 10 provides accomplished performance measures for the KNN model based on stratified 5-fold & 10-fold cross validation in addition to classification accuracy.

Table 10. KNN Model’s accomplished other performance measures with stratified 5-fold & 10-fold cross validation

K- Fold stratified cross validation	Recall/ Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1 Score	AUC-ROC Score
5-fold (K=5)	0.53	0.94	0.74	0.63	0.52	0.74
10-fold (K=10)	0.50	0.94	0.72	0.55	0.46	0.72

5.5. Deep Neural Network (DNN) Model Based Outcome: The DNN model was created using stratified 5-fold cross validation. The achieved accuracy using numerous K-folds is shown in Table 11. Figure 16 illustrates how precisely Fold-3 has discriminated. The stratified 5-fold cross validation was found to be validated by the DNN Model, which had 87.72% as a mean accuracy & 0.0666 as a standard deviation. The best parameters for hyperparameter tuning were selected using the following settings: learning rate = 0.0005, epochs = 100, batch size = 10, dropout = 0.2, kernel_initializer = "uniform," inputs = 60, 35, and 25 neurons in 3 hidden layers, severally rectified linear measure activation functions in the final layer, as well as the Adam Optimizer.

Table 11. DNN Model’s accomplished classification accuracies with stratified 5-fold cross validation

K-Folds	Accuracy
1	82.35%
2	81.25%
3	100.00%
4	87.50%
5	87.50%
Mean: 87.72% std. deviation: 0.0666

Additional preprocessing and stratified 10-fold cross validation were done on DNN model. The overall accuracy using different K-Folds is shown in Table 12. Fold-10 has discriminated with the utmost precision, as seen in Figure 17. It was discovered that the DNN model at the time achieved 87.64% as a mean accuracy & 0.0561 as a standard deviation. The stratified 10-fold cross validation was supported by the deviance. The best parameters were chosen after hyperparameter tuning using stratified 10-fold cross validation, having three inputs, 30, 30, and 25 neurons in three hidden layers with severally rectified linear measure activation functions, sigmoid activation so at final layer with one neuron, as well as the Adam Optimizer.

Table 12. DNN Model’s accomplished classification accuracies with stratified 10-fold cross validation

K-Folds	Accuracy
1	88.89%
2	87.50%
3	87.50%
4	87.50%
5	87.50%
6	87.50%
7	87.50%
8	75.00%
9	87.50%
10	100.00%
Mean: 87.64% std. deviation: 0.0561

In addition to classification accuracy, Table 13 provides other effective performance measures for the DNN model based on stratified 5-fold & 10-fold cross validation.

Table 13. DNN Model’s accomplished other performance measures with stratified 5-fold & 10-fold cross validation

K- Fold stratified cross validation	Recall/ Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1 Score	AUC-ROC Score
5-fold (K=5)	0.62	0.95	0.79	0.90	0.65	0.79
10-fold (K=10)	0.55	0.97	0.76	0.70	0.57	0.76

We'll look at a comparison study of all the different algorithms that were applied in this. The goal of this research was achieved by developing ML/DL algorithms to identify kyphosis disease. Stratified K-Fold cross validation with hyperparameter tuning has been tested using ML/DL models with learning algorithms. 5-fold & 10-fold cross validations were carried out for each test. 5-fold or 10-fold management of K-fold cross validation is typically most acknowledged [32].

A Random Forest (RF) Model achieved 85.22% accuracy by activity 5-Fold stratified cross validation, followed by Support Vector Machine (SVM) Model at 85.22% accuracy, the K-Nearest Neighbor (KNN) Model at 85.22% accuracy, and Deep Neural Network (DNN) Model at 87.72% accuracy as shown in Table 14 and depicted in Figure 18.

By activity 10-Fold stratified cross validation, the Random Forest (RF) Model accomplished 83.89% accuracy, the Support Vector Machine (SVM) Model accomplished 85.14% accuracy, K-Nearest Neighbor (KNN) Model accomplished 84.03% accuracy, and the Deep Neural Network (DNN) Model accomplished 87.64% accuracy as demonstrated in Table 15 and pictured in Figure 19.

In addition to classification accuracy, comparative analysis of various performance parameters for the ML/DL models based on stratified 5-fold & 10-fold cross validation are provided by Tables 14 & 15, as well as depicted in Figures 20 & 21.

Table 14. ML/DL models’ performance measures comparison accomplished using stratified 5-fold cross validation

ML/DL Models	Classification Accuracy (%)	Recall or Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1-Score	AUC-ROC Score
Random Forest	85.22	0.48	0.95	0.72	0.70	0.51	0.72
Support Vector Machine	85.22	0.48	0.95	0.72	0.68	0.51	0.72
K-Nearest Neighbor	85.22	0.53	0.94	0.74	0.63	0.52	0.74
Deep Neural Network	87.72	0.62	0.95	0.79	0.90	0.65	0.79

Table 15. ML/DL models’ performance measures comparison accomplished using stratified 10-fold cross validation

ML/DL Models	Classification Accuracy (%)	Recall or Sensitivity	Specificity	Balanced Accuracy Score	Precision	F1- Score	AUC-ROC Score
Random Forest	83.89	0.35	0.96	0.65	0.60	0.43	0.65
Support Vector Machine	85.14	0.55	0.94	0.75	0.68	0.54	0.75
K-Nearest Neighbor	84.03	0.50	0.94	0.72	0.55	0.46	0.72
Deep Neural Network	87.64	0.55	0.97	0.76	0.70	0.57	0.76

When stratified 5-fold & 10-fold cross validation with hyperparameter tuning was taken into account, the accuracy of the RF, SVM, KNN, & DNN Models were amazingly 83%. Overall, it was found that the DNN model beat a variety of different ML models in terms of classification accuracy, obtaining accuracy validated by stratified 10-fold & 5-fold cross validation of 87.64% and 87.72%. As shown in Tables 14 & 15, the DNN model performed better than a wide range of other models in terms of classification accuracy, recall or sensitivity, specificity, balanced accuracy score, precision, F1 score, and AUC-ROC score supported stratified 5-fold & 10-fold cross validation. The DNN model also achieved maximum score based on these other performance measures.

In this present work, four distinctive machine learning/deep learning algorithms were applied, which are, the Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM) & Deep Neural Network (DNN) models to detect kyphosis disease. It had been determined that the RF, SVM, KNN, and DNN models can to spot and foresee kyphosis once a patient had skilled a procedure. The general accuracies of the models were found between 83%–85% supported the K-fold stratified cross validations once performing the hyperparameter tuning. It is advised that machine/deep learning need to be utilized to acknowledge and analyze kyphosis disease. In terms of classification accuracy, the Deep Neural Network (DNN) model outflanked all the other ML models; accomplishing 87.64% and 87.72% with the stratified 10-fold & 5-fold cross validation correspondingly. These results surpass those of as of now found examination in differed investigation papers. We consider a thorough analysis of algorithms using different performance measures since it helps us comprehend the algorithms better. Numerous performance criteria were taken into account, including accuracy, recall or sensitivity, specificity, balanced accuracy score, precision, F1 score, and AUC-ROC score. By contrasting several algorithms using various criteria, it is possible to assess an algorithm's predictability. Overall, the DNN model performed best among the wide range of different models and achieved maximum accuracy and other performance metrics scores derived from the stratified K-fold cross-validation. Considering the analysis of performance measures, it is advised that Deep Neural Network (DNN) be used as a deep learning model since it is more likely to be able to recognize and predict kyphosis after a patient had a clinical procedure. Future research frequently tries different deep learning algorithms to improve the classification accuracy and other performance metrics in diagnosing kyphosis disease.

Rastogi, V., Satija, S., Dr, P.K., Sharma, S.S., “MACHINE LEARNING ALGORITHMS: OVERVIEW”, International Journal of Advanced Research in Engineering and Technology (IJARET) Volume 11. ISSN Online. Issue 9, 512–517 (September 2020). https://doi.org/10.34218/IJARET.11.9.2020.052
Sinha, A., Gupta, S.K., Tiwari, A., Chaturvedi, A.: Deep Learning: An Overview and Innovative Approach in Machine Learning. In B. Pandey, & A. Khamparia (Ed.), Hidden Link Prediction in Stochastic Social Networks (pp. 108–134), 2019, IGI Global. https://doi.org/10.4018/978-1-5225-9096-5.ch007
Schmidhuber, J.: Deep Learning in neural networks: An overview. Neural Netw. Issue 61, 85–117 (Jan 2015)
Khan, R., Srivastava, A.K., Gupta, M., Kumari, P., Kumar, S.: "Medicolite-Machine Learning-Based Patient Care Model", Computational Intelligence and Neuroscience, vol. Article ID 8109147, 12 pages, 2022. (2022). https://doi.org/10.1155/2022/8109147
Chatter, P., Ramana, D., Suzain, S., Latha, P.: Prediction of Kyphosis Disease Using Machine Learning. Issue (2021). https://doi.org/10.1007/978-981-15-9712-1_30
Stephen Dankwa, Zheng, W.: Special Issue on Using Machine Learning Algorithms in the Prediction of Kyphosis Disease: A Comparative Study. Appl. Sci. Issue (2019). https://doi.org/10.3390/app9163322
Singh, S.K., Khamparia, A., Sinha, A.: Explainable Machine Learning Model for Diagnosis of Parkinson Disorder. In: Khamparia, A., Gupta, D., Khanna, A., Balas, V.E. (eds.) Biomedical Data Analysis and Processing Using Explainable (XAI) and Responsive Artificial Intelligence (RAI). Intelligent Systems Reference Library, vol. 222. Springer, Singapore (2022). https://doi.org/10.1007/978-981-19-1476-8_3
Singla, D., Veqar, Z.: Association between forward head, rounded shoulders, and increased thoracic kyphosis: a review of the literature. J Chiropr. Med. Issue 16(3), 220–229 (2017)
Zhang P, PengW,Wang X, Luo C, Xu Z, Zeng H,… Ge L, Minimum 5-year follow-up outcomes for single-stage transpedicular debridement, posterior instrumentation and fusion in the management of thoracic and thoracolumbar spinal tuberculosis in adults. Br J Neurosurg 30(6):666–671, 2016
Galbusera, F., Casaroli, G., Bassani, T.: "Artificial Intell. Mach. Learn. spine Res. " JOR spine. Issue 2(1), e1044 (2019)
GuanRui Ren, K., Yu, M.S., ZhiYang Xie, P.Y., Wang, W., Zhang, Y., Huang, Y.T., Wang: XiaoTao Wu, Current Applications of Machine Learning in Spine: From Clinical View. Global Spine Journal. Issue 1–14 (2021). https://doi.org/10.1177%2F21925682211035363
Kavitha, K.S., Ramakrishnan, K.V., Manoj, K.S.: Modelling and Design of Evolutionary Neural Network for Heart Disease Detection. Int. J. Comput. Sci. Issues. Issue 7, 272–283 (2010)
Noura, A.: Heart Diseases Diagnoses Using Artificial Neural Network. Netw. Complex. Syst. Issue 5, 7–11 (2015)
Animesh, H., Subrata, K.M., Amit, G., Arkomita, M., Asmita, M.: Adv. Comput. Sci. Technol. Issue 10, 2137–2159 (2017). Heart Disease Diagnosis and Prediction Using Machine Learning and Data Mining Techniques: A Review
Mrudula, G., Kapil, W., Snehlata, D.: Decision Support System for Heart Disease Based on Support Vector Machine and Artificial Neural Network. In Proceedings of the 2010 International Conference on Computer and Communication Technology, Bradford, UK, 29 June–1 July ; pp. 17–19. (2010)
Tahmooresi, M., Afshar, A., Bashari, R., Babak, B.N., Bamiah, K.M.: Early Detection of Breast Cancer Using Machine Learning Techniques. J. Telecommun Electron. Comput. Eng. Issue 10, 21–27 (2018)
Kuo, C.Y., Yu, L.C., Chen, H.C., Chan, C.L.: Comparison of Models for the Prediction of Medical Costs of Spinal Fusion in Taiwan Diagnosis-Related Groups by Machine Learning Algorithms. Healthc Inf. Res. Issue 24, 29–37 (2018)
Abdullah, A., Yaakob, A., Ibrahim, Z.: Prediction of Spinal Abnormalities Using Machine Learning Techniques. In Proceedings of the 2018 International Conference on Computational Approach in Smart Systems Design and Applications, Kuching, Serawak, Malaysia, 15–17 August ; pp. 1–6. (2018)
Raihan-Al-Masud, M., Mondal, M.R.H.: Data-driven diagnosis of spinal abnormalities using feature selection and machine learning algorithms. PLoS ONE. Issue 15(2), e0228422 (2020). https://doi.org/10.1371/journal.pone.0228422
Kyphosis Dataset from Kaggle:. Avaliable online: https://www.kaggle.com/abbasit/kyphosis-dataset
Breiman, L.: Random Forests. Mach. Learn. Issue 45, 5–32 (2001)
Tyagi, H., Agarwal, A., Gupta, A., Goel, K., Srivastava, A.K., Srivastava, A.K., “Prediction and diagnosis of diabetes using machine learning classifiers”,International Journal of Forensic Software Engineering 20221:4,335–347, DOI: 10.1504/IJFSE.2021.10044205
Singh, Y.V., Singh, P., Khan, S., Singh, R.S., “A Machine Learning Model for Early Prediction and Detection of Sepsis in Intensive Care Unit Patients Journal of Healthcare Engineering, vol. Article ID 9263391, 11 pages, 2022. (2022). https://doi.org/10.1155/2022/9263391
Goyal, N., Trivedi, M.C., “Breast cancer classification and identification using machine learning approaches”, Materials Today: Proceedings, ISSN 2214–7853, (2020). https://doi.org/10.1016/j.matpr.2020.10.666
Ayeldeen, H., Elfattah, M.A., Shaker, O., Hassanien, A.E., Kim, T.H. Case-Based Retrieval Approach of Clinical Breast Cancer Patients. In Proceedings of the 2015 3rd International Conference on Computer, Information and Application, Qingdao, China, 28–29 November ; pp. 38–41. (2015)
Singh, S.K., Sinha, A., Yadav, S., "Performance Analysis of Machine Learning Algorithms for Erythemato-Squamous Diseases Classification," 2022 IEEE International Conference on Distributed Computing and Circuits, E., Electronics (ICDCECE), pp. 1–6, doi: (2022). 10.1109/ICDCECE53908.2022.9793000
Palaniappan, S., Awang, R., “Intelligent heart disease prediction system using data mining techniques”, 2008 IEEE/ACS Int. Conf.Comput. Syst. Appl., pp.108–115, (2008)
Kathleen, H., Miao, Julia, H., Miao, Coronary Heart Disease Diagnosis using Deep Neural Networks,International Journal of Advanced Computer Science and Applications, Vol. 9, No. 10, (2018)
K-Fold Cross-Validating Neural Networks:. Available online: https://chrisalbon.com/deep_learning/keras/kfold_cross-validating_neural_networks/
Model Evaluation: : Metrics and scoring. Avaliable online: https://scikit-learn.org/stable/modules/model_evaluation.html
Chauhan, Dr.A.S., “Kyphosis Disease Prediction Using Deep Neural Networks”, International Journal of Engineering Research & Technology (IJERT) ICRADL – 2021 (Volume 09 – Issue 05), 2021, p.p. 126–131, ISSN: 2278 – 0181. DOI: 10.17577/IJERTCONV9IS05026
Ayeldeen, H., Elfattah, M.A., Shaker, O., Hassanien, A.E., Kim, T.H. Case-Based Retrieval Approach of Clinical Breast Cancer Patients. In Proceedings of the 2015 3rd International Conference on Computer, Information and Application, Qingdao, China, 28–29 November ; pp. 38–41. (2015)

No competing interests reported.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Comparative Analysis of Supervised Machine and Deep Learning Algorithms for Kyphosis Disease Detection

Status:

Version 1

Abstract

Figures

Introduction

Literature Review

Materials And Methods

Results And Discussion

Conclusions

References

Additional Declarations

Status:

Version 1