Paper Nazek Hassouneh

Predicting Survivability in Leukemia Patients using Deep Learning 1st Nazek Hassouneh 2nd Loai Alnemer 3rd Jamal Alsakran Computer Information Science Computer Information Science Computer and Information Science The University of Jordan The University of Jordan Higher Colleges of Technology Amman, Jordan Amman, Jordan Fujairah, UAE [email protected] [email protected] [email protected] 4th Ali Rodan Computer and Information Science Higher Colleges of Technology Al Ain, UAE [email protected] Abstract—Leukemia is the most commonly diagnosed type of algorithms are used in survivability prediction. Many ML cancer in children in the United States. The prediction of its algorithms have been applied in cancer research for the de- survivability is of great importance for patients as it gives them velopment of the prediction model which provide an effective hope and improves their psychological health. Additionally, it predictive model. Deep Learning is a new trend in machine is crucial for physicians in prescribing the proper arrangement learning. This research aims to predict the survivability of of treatment. This research aims to predict the survivability of Leukemia patients using Deep Neural Network (DNN) al- Leukemia patients using Deep Neural Network (DNN) algorithm. gorithm. The “Deep Neural Network- DNN” is an example The prediction model of the DNN algorithm is experimentally of deep learning techniques. The Leukemia dataset is ob- built using two different methods; cross validation method and tained from The Surveillance, Epidemiology, and End Results ensemble method. The performance evaluation is measured using (SEER) program. SEER is suitable to be used in the survival accuracy, specificity, and sensitivity metrics. For comparison predictive model because it contains information related to purposes, we use traditional Machine learning (ML) algorithms: cancer including cancer grade, type of treatments, and the Decision Tree, Multilayer Perceptron, Support Vector Machine. cause of death.The data of the SEER program was collected The proposed DNN algorithm outperforms the other three ML from 1973 − 2014 [5]. In this research, the usage of DNN is algorithms in predicting the survivability of a Leukemia patient investigated in comparison with the traditional ML algorithms, in both cross validation and ensemble methods. and many parameters are examined in order to select the final structure of the DNN. For building the predictive model the Index Terms—component, formatting, style, styling, insert final structure of DNN is trained and tested on the SEER dataset using two methods: cross validation and ensemble I. INTRODUCTION methods. The performance of the predictive model is evaluated by calculating the accuracy, specificity and sensitivity metrics. Leukemia is the seventh leading cause of cancer death in the United States as reported by the National Cancer II. RELATED WORK Institute (NCI) [1]. Also, Leukemia is the most common type of cancer diagnosed in children in the United States [2]. Machine learning algorithms are used in survivability pre- Every year, Leukemia is diagnosed in about 29, 000 adults diction. Many researchers use a Decision Tree (DT) algorithm and 2, 000 children in the United States [3]. The prediction on SEER data. The authors in [6] have used DT, as well, of survivability is one of the most challenging and difficult for Breast Cancer. The DT algorithm has shown the best tasks for physicians. Due to the importance of this topic, accuracy of the sample data with a limited period time. In there are many techniques that have been implemented for [7], the authors used the DT algorithm for Breast Cancer the predicting survivability such as machine learning (ML) on SEER data to predict ten years survival. Also, the DT algorithms [4]. The survival rate is defined by the National has proved to be the best classifier for ten year survival. Cancer Institute (NCI) as the percentage of people in a study The authors in [8] applied ML algorithms on Lung Cancer who are still alive for a certain period of time after they have on SEER data. The DT have the best accuracy over variant been diagnosed with or started treatment for cancer [3]. There time survival, then come NN and SVM, respectively. After are many factors affecting the survival rate such as: the type that, they use a voting and ensemble methods. The results of cancer, sex, age and grade. The treatment and the patient’s showed more enhancement. The authors in [9] used Breast response to treatment can be different from one patient to Cancer dataset from SEER program and apply many ML another. Therefore, researchers have provided new methods to pre- dict a survivability for cancer patients. Machine learning 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 191 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.

algorithms. They predict the patients if survive or not from using the input medical image. The RNN is used in [14] to Breast Cancer. ML algorithms are applied and achieve a high predict re-admission for diabetic patients in case they need accuracy (70%) with DT and SVM following NN with (68%). to be re-admitted to the hospital or not. RNN obtained good Then, they applied a conformal prediction algorithm to remove accuracy with 76%.DNN learns by creating a more abstract non reliable predictions. Their results improve in accuracy and representation of the data as the network grows deeper; as in the DT algorithm with 72% accuracy have the best results a result, the model automatically extracts features and yields then NN and SVM. Recently, Elberak et al. [10] conduct a higher accuracy results [16]. comparative study to predict the survivability time in months. III. METHODOLOGY Deep learning has shown good performance in the Health Informatics field. In Medical Imaging, deep learning is used The proposed methodology is based on using the DNN for classifying the tumor of cancer using medical image [11]. model with SEER dataset for predicting survivability of DNN has been used recently in Medical informatics field Leukemia patients. Below are the steps of the proposed DNN for prediction of cancer with electronic health records [12]. model for predicting the survivability. The author in [13] uses deep learning for predicting the survivability of childhood Acute Lymphocytic Leukemia using A. Dataset Preparation two datasets from Children’s Hospital in Australia. The dataset consists of 150 rows with 12000 columns. All parameters Preparing and preprocessing the dataset is the most impor- for the DNN (number of hidden layers, number of neurons, tant step, it includes many substeps as follow as: and the percentage of dropout) were examined. She uses the technique of Out-of-Bag sampling with 17 and 15 votes to 1) Preparing the Dataset: The SEER dataset is obtained determine the final prediction of patients. Accuracy, sensitivity from SEER program in text file format. This file contains and specificity are used to evaluate the performance of the many numbers and characters with no spaces between them. prediction model. The deep neural network showed higher It is not clear where the attributes are located. Therefore, a accuracy with 98% for prediction. program is written to separate those attributes. Each patient’s profile is presented in a single record with 133 attributes. Al-Bahrani et al. [12] use colon cancer from SEER program. This file contains 396,485 records for patients with Lymphoma They take 188,336 records between the years (1988-2009) with Cancer, Myeloma Cancer, and Leukemia Cancer. This file is 133 attributes for each patient. They build a deep neural net- filtered to select just Leukemia patients based on attribute work using Keras library. The building of predicting survival name, Site Recode ICD-O-3/WHO 2008 that recodes the code model for 1, 2, and 5 years. DNN is implemented with five related to leukemia and Lymphoma types. After the filtering hidden layers. Each layer had 812 neurons. The results of DNN process is applied, the dataset contains 131,615 records and outperforms random forests and logistic regression algorithms 133 attributes for leukemia patients. Dataset contains patients in the area under curve (0.86) and sensitivity (86.4%) metrics. who died with Leukemia cancer, who died for other reasons The authors in [11] employ Convolutional Neural Networks but not Leukemia cancer, and who are still alive. (CNN) as a feature extractor. They use the CNN model which contains eight hidden layers to extract a set of features by 2) Cleaning the Dataset: The data in the Leukemia dataset passing skin natural image to CNN. Then, these features were has many problems and there are attributes with values that passed to the classification model using k nearest neighbor are not useful in the prediction model and should be removed. classifier (K-NN) to classify the input image whether it is The problems so stated as follows: benign or malignant. • Redundant attributes: many attributes have the same Chopra et al. [14] develop a Recurrent Neural Network value. In this case, only one of these attributes will be (RNN) model to predict 30 day re-admission for diabetic removed. patients in case they need to be re-admitted to the hospital or not. After preprocessing and normalizing the data, the • Attributes have an unknown value: there are many at- selected attributes are applied to the RNN network with two tributes have the value is 99 that means unknown value. hidden layers and with 500 epochs. The RNN has the highest accuracy with 81.12% compared to the accuracy of classifiers • Attributes have not applicable value like value 88 that SVM, Random Forest, and Simple Neural Networks. The means these attributes are not applied and therefore have SVM has had the lowest accuracy by 64%. Rao et al. [15] been removed. present a model that uses a Convolutional Neural Network (CNNs) to classify tumors seen in Lung Cancer using scan • There are attributes have more than 50% of the missing images as malignant or benign. The dataset contains the scans values. These attributes are been removed because they of 1018 patients. They use a convolutional neural network contain large percent of missing values. The reason for that contains four layers. The results show that the accuracy having a large percentage missing value that there are of CNN is 76% compared with traditional neural networks many attributes are created after the years 2004 and 2010, 72%. The CNN has also been used in many researches [11], so that every patient profile, before these dates have a [15] to classify whether the tumor is benign or malignant missing value. So, these attributes have been removed After removing the records and attributes, the dataset contains 128,946 records and 34 attributes for Leukemia patients. After this step the dataset is cleaned. 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 192 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.

3) Normalizing the Dataset: Normalization is a scaling backend, the Theano backend and the CNTK backend. We use technique in preprocessing stage. It can be useful for the the TensorFlow that is the default backend of the Keras. The prediction that enhances the performance of classifiers. There TensorFlow is developed toTensorFlow to enable fast experi- are many normalization techniques such as Min-Max and mentations with different structures on deep neural networks. standardization. The standardization technique [12] is applied such that the attributes preserve the normal distribution. The building blocks of DNN in Keras as follows. Dense layers: the dense layer is a fully connected layer. So, all the 4) Extracting the Class Label: The class label for every neurons in the layer are connected to those in a next layer. patient profile is created in this step. The class label is (1) Fully connected layers are defined using the Dense class in if the patient died from Leukemia cancer, and the class label Keras. Initializing the network weights: the network weights is (0) if the patient is still live or died by other reasons not are initialized to a small random number that are generated Leukemia cancer [9]. To create the class label the following from a uniform distribution that called in Keras is a uniform. steps are described as follows: We have used two attributes The random number in this case is between 0 and 0.05 which in the dataset to extract the class label. The attributes are the is the default uniform weight initialization in Keras [18]. Vital Status Recode (VSR) that indicates the vital status for Activation functions: the role of activation function converts patients (dead or alive), and the Cause of Death (COD) that neural networks into non-linear. There are many activation has different values indicating the reason of death if patients functions that are used such as sigmoid, rectified linear unit died as following: (RLU), and hyperbolic tangent. There are two parameters that are important to adjust the model; the loss function and the • Patients who died because of Leukemia cancer. optimizer function. Loss function which is used to measure • Patients who died because of other reasons (e.g. Heart the difference between the predicted value and the actual value where the robustness of model increases, while the value of diseases or accidents) the loss function decreases. There are many functions for loss function, e.g. Mean Squared Error and Cross Entropy. Also, Cause of Death (COD) has value indicating patients Optimizer function which is used to adjust the weights in back who are still alive. The extracted class label is outlined as propagation stage in a way that the error is minimized. There following: are many optimization functions like Adam, Adagrad and Gradient Descent. Adam is adaptive learning rate optimization If (COD = Leukemia, cancer and VSR =Died) then the class algorithm. label = 1. (Not Survive-Died). 2) Deep Neural networks regularization: Many strategies If VSR = alive then class label = 0. are used in deep neural network to reduce the error and If (COD = died by other reasons and VSR = Died) then increase the accuracy as much possible. These strategies are class label = 0. known as regularization. Dropout is a powerful regularization After the class label is created, the attribute Vital Status strategy aims to reduce the complexity of the model and Recode (VSR) and Cause of Death are removed and replaced prevent overfitting. Dropout activation enables the network to it with the class label. After cleaning the dataset the number learn redundant representations. These redundant representa- of attributes in dataset is 33 attribute the final attribute will be tions make the network robust and avoid over fitting [12]. class label. Overfitting means that the trained model performs well on the training data, but not well on new testing data. Dropout is B. Deep Neural Networks Structure randomly setting percentage of activation during the training of networks. 50%, 25%, 10% percentage are examined for The deep neural network structure has an input layer, many dropout. hidden layers, and one output layer. Input data is passed to the hidden layers by activation function. The output of this 3) DNN Implementation: To select the structure of the stage is called feed forward. Then, the error rate is calculated DNN we need to select the number of hidden neurons in by comparing the actual output with predicted output. Based each hidden layer and the number of hidden layers. The core on error rates, a backward is done to adjust the weights of structure of DNN in Keras is called model. The sequential the network. This error is calculated by loss function, that DNN model is used such that the layers are added using add() will be described in more details in this section. Through an method. We can determine the number of hidden neurons iterative training process, the network weights are adjusted when adding a hidden layer. The number of input neurons to minimize error. The iterative training method uses the in DNN model is (32) which is the number of attributes in the optimizer functions to find the minimum error values. The Leukemia dataset. optimizer function will be described in the next subsection. We use (relu) RLU activation function for each hidden 1) Deep Neural Network building blocks in Keras: There neuron because it is simple and efficient [12]. We use the are many tools that provide deep learning models such as sigmoid function as an output activation function because it is Keras, Caffe, Deep Learning4j. One of the most powerful effective in binary outputs [19]. Then, we compile the model tools is Keras [17]. The Keras is an open source library for using Cross Entropy as the loss function because it is effective neural networks that is implemented in python. The our DNN in binary classification model [20] and Adam as an optimizer is implemented using Keras and TensorFlow. Keras uses a specialized manipulation library to do numerical computations. This library serves as the backend engine of Keras. Keras have three backend implementations available; the TensorFlow 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 193 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.

function because it works well and fast on large dataset [21]. Fig. 1. Evaluation results with different numbers of hidden neurons The final structure of DNN is selected by examining differ- patients. Different evaluation metrics are used to measure ent numbers of hidden layers and different numbers of hidden the performance of the prediction models in predicting the neurons for each hidden layer. The results are obtained when survivability of Leukemia cancer patients. trying different numbers of the hidden neurons and selecting the best performance using evaluation metrics. By analyzing For selecting the structure of the DNN, many experiments the obtained results we select the number of hidden neurons are conducted on different values of the number of hidden that have best performance. Then, we train the model with neurons and number of hidden layers, in order to select the depth between two hidden layers to eight hidden layers. This values of the parameters that produce the best performance leads to select the best network depth based on the evaluation of the DNN. This section presents the evaluation results of metrics we used to evaluate our model. The final step involves the experiments at different numbers of hidden neurons and training the DNN with no dropout, 10% dropout, a 25% different numbers of hidden layer to select the best network dropout, and 50% to select the best results. The adding dropout structure. The final structure of the DNN will be used for is performed with add (dropout) method [12]. building the predictive model. C. Evaluation Metrics Previous research has provided various methods to select the number of hidden neurons. The goal in this step is The evaluation process is concerned with measuring the improving the accuracy and minimizing the error. Therefore, differences between expected results and actual results. The many researchers have fixed the number of hidden neurons evaluation of performance of the model is based on counts of based on trial rule [23]. The chosen numbers of hidden neurons the test records that is correctly and incorrectly predicted by are: 25, 32, 45, and 65, where (32) is the number of attributes the model [22]. The performance metrics used are accuracy, in the leukemia dataset, (25) is the number less than the specificity and sensitivity. number of attributes in the dataset, (45) approximately is the 3/2 of the (32) and (65) is the twice as much (32) [24]. The TP +TN (1) number of hidden layers in each experiment is three layers. Accuracy = Fig. 1 presents the evaluation results of the experiments with TP +TN +FP +FN different numbers of hidden neurons to select the best network structure. It presents the evaluation when using 25 , 32 , 45 and TN (2) 65 hidden neurons in each hidden layer. The best evaluation Specif icity = results are obtained when the number of hidden neurons is 45 in each hidden layer. TN +FP In this step, we train the DNN with depth between two hid- TP (3) den layers to eight hidden layers. We select 45 hidden neurons Sensitivity = in each hidden layer at the previous step. We experiment with two hidden layers and evaluate the performance of the model. TP +FN We further expand the DNN by adding one hidden layer in each run and evaluate the performance of the model; until we where T P is true positive, T N is true negative, F P is false reached eight hidden layers. Fig. 2 shows the evaluation results positive, and F N is false negative. with different numbers of hidden layers. The best evaluation results are obtained when the number of hidden layers was six. The accuracy is the percentage number of correct pre- Furthermore, we train the model with no dropout, dropout with dictions to the total number of predictions. The specificity 10%, 25% and 50%. It is found experimentally that the 25% measures the percentage of negative predicted that are cor- of the activation has the best results. Our final DNN structure rectly classified. The sensitivity measures the percentage of consists: (i) Six hidden layers. (ii) 45 hidden neurons in each positives predicted that are correctly classified. We focus on hidden layer. (iii) After each hidden layer we add a dropout predicting the number of patients that survive Leukemia and actual class label is “survive Leukemia”. Therefore, we are particularly interested in TN which means the number of patients is correctly predicted that they will survive from leukemia cancer and actually survived from Leukemia cancer. That means that we want to obtain the highest specificity percentage. To have high percentage we focus on reducing the FP based on equation 2. FP means the number of patients that is predicted to die from leukemia and the actual the patients survived Leukemia. So, we focus on specificity metric. This will help us to assess the performance of ML algorithms. IV. EXPERIMENTAL RESULTS AND ANALYSIS The experiments are conducted on a personal computer with a 2.4 GHz core i3 CPU and 4GB running Windows 10. Two methods are used to build the prediction models of each ML algorithm: (i) cross validation method (ii) ensemble method. We have used four algorithms (i.e. DNN, DT, MLP, and SVM) to predict the survivability of Leukemia cancer 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 194 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.

Fig. 2. Evaluation results with different numbers of hidden layers Fig. 4. Specificity results of ML algorithms Fig. 3. Accuracy results of ML algorithms Fig. 5. Sensitivity results of ML algorithms activation with 25%. Fig. 5 shows that the sensitivity results of each proposed After selecting the final structure of our DNN. We train algorithm using cross validation and ensemble method. The Ensemble method enhances the sensitivity results compared our DNN using training set with multiple iterations. We use with cross validation method. That is because the Ensemble two methods to build the prediction models of the DNN; 10- method removes some of false negative (FN) cases. The per- fold cross validation and ensemble method. For comparison centage of enhancement of the sensitivity of DNN algorithm purposes, we use three ML algorithms DT, MLP, and SVM. is 2.6%. The prediction models of each ML algorithm are also built using 10-fold cross validation and ensemble methods. V. CONCLUSION The accuracy results of the classification ML algorithms are Deep learning has been widely used in the Health Infor- shown in Fig. 3. It presents the accuracy of each proposed matics field. Previous efforts have motivated us to investigate algorithm using cross validation and ensemble methods. It the performance of DNN in predicting the survivability of shows that the DNN algorithm outperforms the other three ML Leukemia cancer patients on SEER data. The new data of algorithmsin both cross validation and ensemble. In addition, the SEER program that is collected from 1973-2014 enabled the ensemble method enhances the accuracy results of all us to experiment the survivability of the cancer with a new set the ML algorithms. Specifically, after applying the ensemble of attributes and get better predictive model. We employed 10 method on DNN, the percentage of enhancement is (0.28) to fold cross validation and ensemble method experiments using become 74.85% accuracy. The best accuracy result is obtained our DNN algorithm, evaluated the model using the evaluation by the DNN using ensemble method. The SVM has a relatively metrics, and obtained the results. Moreover, for comparison good accuracy, but it did not outperform the DNN. purposes, we designed experiments using three different ML algorithms DT, MLP, and SVM. The proposed DNN algorithm Fig. 4 shows that the specificity results of all ML algorithms. obtained the best performance results of the cross validation The DNN algorithm slightly outperforms the other three ML method. After analyzing all the results, we can conclude that algorithms. In addition, the ensemble method enhances the the DNN outperformed the other three ML algorithms. The specificity results compared with cross validation because experiments conducted on the Leukemia dataset proved that the ensemble method removes some of false positives (FP) the DNN is the most effective in predicting the survivability cases. The reduced number of false positive cases increases of a Leukemia patients. the percentage of specificity. Reducing the number of false positive is essential in predicting survivability. The percentage of enhancement of the specificity for DNN algorithm is 6%. 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 195 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.

REFERENCES [1] cancer Stat Facts: Leukemia,(2018), (n.d.). Retrieved March 07, 2018, from https://seer.cancer.gov/statfacts/html/leuks.html. [2] cancer in Children and Adolescents, National cancer Institute,(2018), Retrieved April 06, 2018, from http :// www.cancer.gov/types/childhood- cancers/child-adolescent-cancers-fact-sheet. [3] National cancer Institute, SEER Training Mod- ule ,(2018),(n.d.). Retrieved April 07, 2018, from https://training.seer.cancer.gov/leukemia/intro/. [4] J. A. Cruz, and D. S. Wishart, Applications of Machine Learning in cancer Prediction and Prognosis. cancer Informatics. 2006. [5] Surveillance, Epidemiology, and End Results Program,(2018), (n.d.). Retrieved March 1, 2017, from http://www.seer.cancer.gov/ [6] D. Delen, G. Walker, and A. Kadam, Predicting breast cancer survivabil- ity: A comparison of three data mining methods. Artificial Intelligence in Medicine,34(2), pages 113-115. 2005. [7] D. Solti, and H. Zhai, Predicting Breast cancer Patient Survival Us- ing Machine Learning. Proceedings of the International Conference on Bioinformatics, Computational Biology and Biomedical Informat- ics.2007. [8] A. Agrawal, M. Sanchit, n. Ramanathan, P. Lalith and C. Alok, A Lung cancer Outcome Calculator Using Ensemble Data Mining on SEER Data., Proceedings of the Tenth International Workshop on Data Mining in Bioinformatics. [9] L. Alnimer et al., Conformal Prediction Technique to Predict Breast cancer Survivability, International Journal of Advanced Science and Technology, vol. 96, 2016, pp. 1–10. 2016. [10] O. Elberak, L. Alnemer, M. Sawalha, J. Alsakran, Predicting Cancer Sur- vivability: A Comparative Study, International Conference on Emerging Internetworking, Data & Web Technologies, 2019. [11] V. Pomponiu, H. Nejati, N.-M. Cheung, DEEPMOLE: DEEP NEURAL NETWORKS FOR SKIN MOLE LESION CLASSIFICATION, IEEE, 2016. [12] R. Al-Bahrani, A. Agrawal, and A. Choudhary, Survivability prediction of colon cancer patients using neural networks. Health Informatics Journal, 2017. [13] R. S. Muthalaly, Using Deep Learning To Predict The Mortality Of LEUKEMIA Patients , Thesis , Queen’s University , Kingston, Ontario, Canada, 2017. [14] C. Chopra, S. Sinha, and S. Jaroli, Recurrent Neural Networks with Non- Sequential Data to Predict Hospital Readmission of Diabetic Patients ,ICCBB 2017, October 18–20, 2017, Newark, NJ, USA , ACM, 2017. [15] P. Rao, N. A. Pereira, and R. Srinivasan, Convolutional Neural Networks for Lung cancer Screening in Computed Tomography (CT) Scans, IEEE, 2016 [16] D. Rav, C. Wong, F. Deligianni, M. Berthelot, J Andreu-Perez, B. Lo,and G. Yang, Deep Learning for Health Informatics , IEEE : Journal of Biomedical and Health Informatics, VOL. 21, NO. 1, JANUARY 2017 [17] Keras: The Python Deep Learning library. (n.d.). Retrieved March 06, 2018, from https://keras.io/ [18] Initializers- Keras Documentaion. (n.d.). Retrieved February 07, 2018, from https://keras.io/initializers/ [19] J. Han, and C. Morag, The influence of the sigmoid function parameters on the speed of backpropagation learning, From Natural to Artificial Neural Computation. Pp. 195–201, 1995. [20] N. Michael, Neural Networks and Deep Learning, Determination Press, 2017. [21] D. Kingma, and J Ba., Adam: A method for stochastic optimization, 2014. [22] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning Book, MIT Press Publisher. 2016. [23] K. G. Sheela, and S. N. Deepa, Review on Methods to Fix Number of Hidden Neurons in Neural Networks, India ,Mathematical Problems in Engineering Volume, Article ID 425740, 11-15. 2013. [24] P. Gaurang, A. Ganatra, Y. P. Kosta and D. Panchal, Behavior Analysis of Multilayer Perceptrons with Multiple Hidden Neurons and Hidden Layers, International Journal of Computer Theory and Engineering, Vol. 3, No. 2, pp. 332–337, 2011. 978-1-7281-5061-1/19/$31.00 ©2019 IEEE 196 Authorized licensed use limited to: University of Glasgow. Downloaded on June 02,2020 at 16:07:00 UTC from IEEE Xplore. Restrictions apply.