Vicko Bhayyu
Alumni

Title of Final Project: A Deep Learning based for ECG Beat Classification using Auto-encoder and Deep Neural Network

Abstract

Electrocardiograph (ECG) is medical testing for examining heart condition in an electrical signal to provide clinical information about a patient’s heart. With ECG, Cardiologists can diagnose the patient's heart condition either by heartbeats or rhythm. ECG beat classification with a large amount of data has its challenges so that the Deep Learning method that has a high level of abstraction in learning features is highly favored. With the Autoencoder method as feature extraction to learn features and reduce feature dimensions and Deep Neural Network as an ECG beat classifier. The dimension of ECG beat features from raw data is 252 then perform extraction feature by Autoencoder which is then reduced to 32. The results of this extraction feature are then classified by Deep Neural Network with 10 classes. Conducted as many as 15 experimental models with the best model will be tested to another dataset. From the 15 experimental models, the best models are obtained, namely Deep AE - DNN 3 HL, with the results of accuracy, sensitivity, specificity, precision, and F1-Score respectively 99.59%, 91.02%, 99.8%, 93.06%, 91.79%. Then this model was tested back to other datasets SVDB and IncartDB with 99.5% accuracy results, 89.6% sensitivity, 98.39% specificity, 97.62% precision, and F1-Score 93.07% for IncartDB and 97.86% accuracy, sensitivity 87.28%, specificity 94%, precision 91.91%, F1-Score 89.46%.

---------------------------------------------------------------

Work : Metrodata Electronics.

Position: Technical Consultant Analyst

Hanif Habibie Supriansyah
Alumni

Title of Final Project: A Semi-Automated ECG Segmentation  Based on Wavelet and Windowed analysis for Congenital Long QT Syndrome Detection

Abstract

Congenital Long QT Sundrome is a rare heart disease characterized by an extension of the QT interval on the signal. QT interval calculation is obtained by finding the duration from the QRSon start point to the end of the T or Tend wave. To get the location of the electrocardiogram signal features, it requires a segmentation process of each feature on the electrocardiogram signal. This study proposes an algorithm for segmenting electrocardiogram signal features using the Discrete Wavelet Transform and Windowed Analysis methods for the segmentation process of these features. Next, the Bazeet’s Formula method is used to calculate the QT Correction to measure the duration of the QT interval. This study uses the QTDB dataset publicly provided by Physionet. The stages in this research include the pre-processing of data to retrieve and correct the location of the R peak, denoising the signal using Discrete Wavelet Transform, segmenting electrocardiogram signal features with the results of reconstruction each level (Inverse Discrete Wavelet Transform) and Windowed Analysis. Based on these stages, the results of the study have an average value of error (mean error) with Pon -0.001 seconds, Ppeak -0.002 seconds, QRSon -0.007 seconds, QRSoff 0.004 seconds, Tpeak 0.011 seconds, and Tend -0.011 seconds. This study shows that the algorithm designed is able to segment ECG signal features correctly, and the results of QT Correction calculations can be used as a reference for early indications of Long QT Syndrome.

---------------------------------------------------------------

 Work : Blue Power Technology (Computrade Technology International)

Position : Technical Sales Specialist (Pre Sales Engineer)

Sicilia Paledya
Alumni

Title of Final Project: Heartbeat Classification using Artificial Neural Network and Principal Component Analysis 

Abstract

Arrhythmia is one of heart disease characterized by irregular electrical activity or an abnormal heartbeat. Classifying abnormal heartbeats which are symptoms of arrhythmia can be seen from the medical record in the form of an electrocardiogram (ECG). ECG signal records the activity of the heartbeat. The classification of arrhythmia is divided into 6 classes, namely Paced Beat, Atrial Premature, Left Bundle Branch Block Beat, Normal, Right Bundle Branch Block Beat, and Premature Ventricular Contraction. In this study, the classification method used an Artificial Neural Network (ANN) Backpropagation. The result of an ANN Backpropagation method shows an accuracy, F1-score, precision, sensitivity specificity is 99.65%, 99.27%, 96.83%, 97.74%, 99.67%, respectively. This study shows the ANN Backpropagation method can classify heartbeats from arrhythmia through ECG signals.

---------------------------------------------------------------

Work : Ibnu Sutowo Hospital, Baturaja

Position: IT Department (Medical Record)

Ahmad Noviar
Alumni

Title of Final Project: A Deep Learning based for ECG Noise Beat Classification using Denoising Auto-Encoder and Deep Neural Network

Abstract

Arrhythmia is a indication or a symtop of heart beat disruption or heart rhythm Early detection of arrhythmia could help patient in handling the abnormalities  quickly. Arrhythmia can be detected using electrocardiogram (ECG), which is heart activity electrical signal recording. This research aims to classify normal heart,  premature ventricular contraction, atrial premature beat, right bundle branch block beat and non-conducted p-wave on the ECG signal. Deep Neural Network  is proposed due to the ability of processing non-linear data like ECG signal. The data used in this research is obtained  from Physionet.org website with imbalanced class ratio and noise-contained. To overcome the noise-contained data, Denoising Autoencoder is proposed to denoise the signal and Autoencoder used to extract the feature of the denoised ECG signal. Both the technique above shows the results performance accuracy, sensitivity, specificity, precision and F1 Score is 99.06%, 93.56%, 99.35%, 89.42% and 91.11% repectively.

Rahmi Khoirani
Alumni

Title of Final Project: Atrial Fibrillation Classification using Recurrent Neural Network

Abstract

Atrial Fibrillation (AF) is one of the most common heart diseases characterized by irregular heartbeat rhythms. AF can be detected via an electrocardiogram (ECG) signal. This study proposes one of the Deep Learning technique, namely the Recurrent Neural Network (RNN) with Long Short Term Memory (LSTM) architecture. This study classifies three classes of ECG signals, namely normal, atrial fibrillation, and others. The 2017 PhysioNet Computing in Cardiology (CinC) Challenge database is used. The pre-processing steps to classify the heart signal are bound normalization, denoising the signal with Discrete Wavelet Transform (DWT), data resampling with Random Oversampling (ROS), and segmentation with window size. Based on these stages, the results of the classification model with a window size of 900, and 16 iteration batch size. The distribution of training and testing data ratios of 90% and 10%, respectively. The results are obtained with accuracy, sensitivity, specificity, precision, and F-measure is 92.12%, 88.36%, 94.31%, 87.99%, and 87.89%, respectively. This research shows that RNN-LSTM can classify AF through ECG signals with the characteristic of sequential data.

Varindo Ockta Keneddi P.
Alumni

Title of Final Project: Author Name Disambiguation

Ferlita Pratiwi Arisanti
Alumni

Title of Final Project: Atrial Fibrillation Classification using Long Short Term Memory (LSTM)

Abstract

Early detection of cardiac disease can extend life through proper treatment. One of the most dangerous cardiac diseases is atrial fibrillation. Atrial fibrillation can be detected using an electrocardiogram (ECG), which is a signal recording of the electrical activity of the heart. This research aims to classify normal heart and atrial fibrillation of the ECG signal. Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM)-based is proposed due to can process sequential data such as ECG signal. This study used Physionet.org/Computing in Cardiology (CinC) Challenge 2017 database, which has a large imbalanced data ratio. To overcome the problems of imbalanced data, Synthetic Minority Oversampling Technique (SMOTE) is proposed. SMOTE technique shows the results performance accuracy, sensitivity, specificity, precision, and F1 score is 94.83%. 94.95%. 94.95%. 94.78%. and 94.82%. respectively.

Andre Herviant Juliano
Alumni

Title of Final Project: Atrial Fibrillation Classification using Convolutional Neural Network and Deep Neural Network

Abstract

Electrocardiography plays a very important role in the medical field because it functions to evaluate electrical activity and conditions in the human heart. The results of the evaluation will be in the form of graphs or signals that represent the human heart rate per unit time or better known as the Electrocardiogram (ECG). Based on research that has been done in the last few years, deep learning has succeeded in classifying Atrial Fibrillation with a high degree of accuracy. The deep learning method proposed in this study is Convolutional Neural Networks (CNN). This is because CNN has the advantage of combining feature extraction based on feature learning and classification in a learning process. In the classification process based on ECG signals, 6 trial scenarios will be tested, each consisting of 7 convolution layers, 10 convolution layers, and 13 convolution layers with fully connected layers of 1000 nodes, 1000 nodes and 1 node for 2700 nodes signal and 18300 nodes signal with window size, and fully connected layer 100 nodes, 100 nodes and 1 node for 18300 nodes signal. From several experimental scenarios, the first experiment scenario using the 13 convolution layer model produced performance values ​​for accuracy, precision, sensitivity, specificity, and f1 score respectively 92.97%, 77.78%, 87.46%, 87.46%, and 81.63%. The low precision, sensitivity and f1 scores of some CNN models are because some ECG signals have the same morphology between the normal signal and AF signal, as well as the amount of normal and unbalanced AF data.

---------------------------------------------------------------

Work : PT Asuransi Simas Jiwa.

Position: Data Scientist

 
Tio Artha Nugraha
Alumni

Title of Final Project: Klasifikasi Author pada Data Bibliografi menggunakan Deep Neural Network dan Support Vector Machine

Abstract

Author Name Disambiguation (AND) is a problem of name ambiguity to the publication in a Digital Library (DL) database caused by the Homonymity and Synonymity of the author's name. The proposed method is a Deep Neural Network (DNN) and Support Vector Machine (SVM) to classify data. The DBLP Labeled Data dataset by Jinseok Kim, et. al. is used for the classification task. This study concerned with processing data with the techniques of normalization and transformation data to create an effective feature for classification. The performance evaluation of the research conducted is accuracy, precision, and recall. The parameters are important to evaluate the AND classification process, especially the identification of the author. For the result, DNN achieves accuracy, precision, and recall, which is 99.98%, 97.71%, and 97.83%, respectively. In addition, SVM produces accuracy, precision, and recall 99.98%, 95.33%, 95.09%, respectively. From the comparison of the two classification methods, DNN outperformed SVM for data classification and author identification.

---------------------------------------------------------------

Work : KAP Mirawati Sensi Idris - Moore Stephens Indonesia.

Position: Junior IT Auditor

 
Muhammad Irham Rizki Fauzi
Alumni

Title of Final Project : ST Elevation Detection In ECG ST Segment Using Discrete Wavelet Transform And Windowing Algoritm

Abstract

There are various types of heart disease, one of which is Myocardial Ischemia (MI) which is caused by anomalies that occur in the heartbeat. The parameter to measure whether a patient has a dangerous disease such as MI is an increase that occurs in the ST segment. This can be seen on the ECG chart which is commonly used in the medical world. This research will detect ST elevation using ECG signal segmentation results obtained by using discrete wavelet transforms and windowing algorithms whose data are taken from the Physionet, the QT database. The preprocessing stage consists of two stages, namely peak R correction and denoising. The next step to do is feature segmentation consisting of QRS wave detection (Onset and Offset), P (Peak and Onset) and T (Peak and Onset). ST elevation detection is done by calculations that have been tested by previous studies. Comparison of average error detection results for P Onset, P peak, QRS Onset, QRS Offset, T Peak, and T end are -1,231 ms, -1,841 ms, -7,020 ms, 4,061 ms, 10,493 ms, -10,797 ms. Based on these detection results, the ST elevation classification was successfully carried out even though there were still errors due to shifting detection points.

Nyayu Latifah Husni
Alumni

SWARM ROBOT TRAJECTORY OPTIMIZATION AND ODOR CLASSIFICATION BASED ON FUZZY-SVM-PSO

Abstract

The need for searching and determining the type of gas (in this research called odor) has been increased.  The usage of human and animal’s nose will be inefficient if they are implemented in the dangerous area.  Finding and determining the type of odor can be conducted by swarm robots.  By spreading them to work in unpredictable area will give more advantages.   However, the robots that are employed to localize and classify the odor usually faces some difficulties, such as: i. they will itinerate to the wrong direction, ii. they will be trapped in local minima, iii. they will collide among them, and iv. they could not avoid the obstacles.  In some cases, they will point the wrong odor source location and will classify the odor to the incorrect class. 

In the real implementation, the most difficult challenge in treating the swarm robots that work for doing 2 tasks directly (finding and classifying the odor) is how to control the swarm trajectory and how to make sure that the robots has recognized the odor.  The swarm robot should be able to: i. adapt and interact with the environment and the other individual robots, ii. Keep away the collision, iii. avoid the obstacles, and iv. be successful in determining the odor type.

To overcome these problems, this dissertation presents a system that has function to control the trajectory of the robot and to classify the odor at the same time.   A combination of 3 artificial intelligences is proposed in this dissertation, i.e. fuzzy logic, Support Vector Machine (SVM), and Particle Swarm Optimization (PSO).  It is due to they provide some advantages.  SVM is good for pattern recognition, fuzzy logic gives smoother trajectory and PSO promises some advantages for target searching.  The contribution of this research is a new idea in the navigation of the robot in finding and classifying odor.  From the experiments in this research, it was obtained that fuzzy-SVM-PSO give more benefits, i.e. the odor searching can be accomplished in 25 – 61 s, while the odor classifying can be done in 8 - 45 s, and the plume finding needs time 17 – 20 s.

Study Program Of Doctoral Engineering, Electrical Department, Sriwijaya University

Ade Silvia

SWARM ROBOT FORMATION CONTROL DESIGN USING BEHAVIOR BASED STRATEGY

Abstract

Swarm robot formation control research has been implemented in various fields. One of the applications is monitoring and mapping the environmental pollution level by using a remote sensing system.  In localizing pollution sources, swarm robots must have ability to cooperate among them.  In their collaboration, they must be able to communicate and be able to avoid obstacles and collisions.  Therefore, for controlling the swarm robot movement, it is needed to set the formation, so that the performance can be improved.

To improve performance in the durability and efficiency of formation control on a swarm robot, it is necessary to choose the right formation strategy.  The swarm robot must have ability to coordinate and cooperate. The choice of strategy type in controlling the swarm robot formation using basic behavior (behavior-based), depends on some factors, such as: (i) reaction behavior between the robot and the environment; (ii) behavior among robots; and (iii) robot behavior with obstacles. In swarm robot control with basic behavioral strategies, the ability that must be possessed by individual robots is to avoid obstacles between robots and obstacles, and to maintain formation during movement.

Maintaining formation control during avoiding obstacles and avoiding collisions between the robots when they perform the tasks in unknown and uncertain environments are difficult to achieve.  Therefore, a soft computational technique is needed.  The Interval Type-2 Fuzzy Logic System (IT2FLS) algorithm is used. This algorithm can be used to deal with uncertainty problems that arise as long as the robot completes its tasks in the real environment.

The problem that often occurs in the use of IT2FLS algorithm is the need of large memory requirement.  This need will affect the overall performance of the swarm robot. However, in this dissertation, the individual specifications of the swarm robot used are simple robots with limited abilities, however, the swarm robot can control the formation during its movement.  All of this can be achieved by modifying the IT2FLS algorithm that is embedded in individual robots and by choosing the right strategy that can improve the formation performance.

The above statement is the contribution of this dissertation.  In this dissertation, the formation control on the swarm robot is carried out using a behavior based strategy.   the swarm robot used in this study consists of simple individuals with low-cost microcontrollers, and has low power as well.  However, researchers were able to embed the IT2FLS controller algorithm into the individual of robot swarm.  From the experiments, it can be proved that it can maintain swarm robot formation.

The individual mobile robots produced in this study have good performance and endurance capabilities in maintaining the formation.  The  swarm robot is able to maintain the stability of the formation control by determining the direction of rotation and speed of each individual robot. This proves that the swarm robot formation control can run well by using behavior based strategies or basic robot behavior.

The formation control testing in order to see the performance of swarm robots that work based on behavior based strategies, was conducted by comparing the movement of the robot using the T1FLS embedded controller with the movement of the robot using the IT2FLS embedded controller, either in the indoor or outdoor.  The test was carried out for several conditions, such as: (i) indppr environmental conditions with obstacles and without obstacles; (ii) outdoor environmental conditions with obstacles and without obstacles, and (iii) environments with obstacles in motion.  The test results show that the IT2FLS controller produces better performance than T1FLS, i.e., with shorter track length, and more stable formation control.

Study Program Of Doctoral Engineering, Electrical Department, Sriwijaya University

Annisa Septiani
Alumni

Title of Final Project: Atrial Fibrillation Classification in Imbalanced Data using Deep Neural Network 

Meiryka Yuandini
Alumni

Title of Final Project: Heart Failure Classification in Imbalanced Data using Deep Neural Network 

Chintya Caroline
Alumni

Title of Final Project: Food Image Classification By Using CONVOLUTIONAL NEURAL NETWORK (CNN)

Abstract

Social media are popular platforms frequently used to share mementos and as a means of marketing strategy, especially by the food industry. Unfortunately, most food photos in social media are not labelled or properly explained, and can lead to confusion by the user. To combat this problem, a novel solution was developed to detect food photos swiftly and automatically with three Convolutional Neural Network (CNN) architectures such as AlexNet, Inception V3, and Resnet 50. Two image processing techniques were implemented, namely Histogram Equalisation and data augmentation. The Food-101 dataset was used, which incorporated a range of diverse food images from the internet and social media. This study revealed that Inception V3 with data augmentation was the best model. Its accuracy was 99.03%, whereas the precision, recall, and F1-score was 99%. Moreover, it had an error rate of 0.005, false positive rate of 0, and false negative rate of 0.013. In addition, this paper demonstrated that the worst model was AlexNet with Histogram Equalization with an accuracy, precision, recall and F1-score of 23%. Furthermore, the error, false positive, and false negative rates were 0.03, 0.015, and 0.76, respectively.

Anggy Tias Kurniawan
Alumni

Title of Final Project: Author Name Disambiguation

Jannes Effendi
Alumni

Title of Final Project: Automation Of Electrocardiogram Delineation Using LONG SHORT-TERM Memory Based ON 1-DIMENTIONAL CONVOLUTIONAL NEURAL NETWORK FEATURE Extraction

Abstract

Electrocardiogram (ECG) is electrical records that contains information about human heart. In the medical field, humans heart condition can be diagnosed by analyzing the changes in hearts beat or rhythm that contain p wave, QRS-Complex and T wave. Delineation can be very hard for doctor to do because of human errors. Because of that, automation of ECG delineation by using deep learning is preferred. The deep learnings methodology used in this study is Recurrent Neural Network(RNN) with Long Short-Term Memory(LSTM) combined with Convolutional Neural Network(CNN) as feature extraction. LSTM is an effective method for classifying time series data. LSTM can also overcomes vanishing gradient’s problems that occur in RNN. In this study, delineation is applied to 4 and 7 types of waves. There are 14 models generated with the best learning rate, number of hidden layers and batch size. Every time step in LSTM have 370 nodes for every types of waves. From the 14 experimental models, the best model is obtained by using CNN as feature extraction before using Bi-LSTM in both 4 and 7 types of waves. CNN and Bi-LSTM’s model have the highest evaluation values in 7 types of waves scenarios with performance value of sensitivity, precision, specificity, accuracy and F1-Score respectively 98.82%, 98,86%, 99.9%, 99.83%, and 98.84%

Febby Nurheliza
Alumni

Title of Final Project: QT-Correction Detection System using Long Short-Term Memory

Abstract

The bioelectric activity of the heart produces an electrical signal called an Electrocardiogram signal. This recording of electrical signals helps doctors diagnose abnormalities in the patient's heart. One form of congenital abnormality in the human heartbeat that can cause syncope, sudden cardiac arrest and sudden death is Long-QT syndrome, which is characterized by an extension of the interval between Q and T waves on an electrocardiogram signal. Classification of ECG beats in large amounts of data and sequences has its challenges, so Deep Learning that have the advantage in processing data automatically and are able to learn their own computational features and methods are highly recommended in this research. From the 3 experimental models, the best models were Bi-LSTM Model 3, with results of accuracy, sensitivity, specificity, precision, and F1-Score of 99.52%, 96.23%, 99.72%, 96.53%, 96.37% respectively. Then this model was tested back to the other datasets NSRDB with 99.76% accuracy results, 98.30 % sensitivity, 99.87% specificity, 98.37% precision, and 98.34% F1-Score.

Muhammad Amir Hamzah
Alumni

Title of Final Project: A system for ST-Elevation Syndrome using LSTM

Bima Kurniawan
Alumni

Title of Final Project: Klasifikasi Sudut Pandang Ultrasonography(USG) Pada Jantung Fetal Menggunakan  Histogram Of  Oriented Gradient(HOG) Dan Support Vector Machine(SVM)

Abstract

Humans generally have four inner segments of the heart, among others, the Right Upper Atrium, Left Upper Atrium, Right Lower Ventricle, and Left Lower Ventricle. By using ultrasound, we can see the process of taking images from the fetal heart, in the fetal heart there are 4 classes including the Four Chamber View (FCV), Left Ventricular Outflow Tract (LVOT), Right Ventricular Outflow Tract (RVOT), and Three Vessel of Trachea (TVT). The methods to be used are the Histogram of Oriented Gradient (HOG) and Support Vector Machine (SVM). The dataset used is in the form of a video from the journal Hunter L.E et al, from the Radiopedia website, and from the Intersocietal Accreditation Commission youtube. This study aims to classify the 4 classes, the parameters that determine the 4 classes are classified properly, namely accuracy, precision, recall, and F1-score. The classification resulted from the HOG and SVM methods is 78.75% accuracy, 75.5% precision, 75.25% recall, 74.75% F1-score.

Irawan
Alumni

Title of Final Project: Screening View Classification of Fetal Heart from Ultrasound Media Using Convolutional Neural Network(CNN)

Abstract

Heart is a vital organ in human body, however congengital heart disease is an anomaly in fetal body structure that frequently happens, because of that fetal
echocardiograph is done by pediatric cardiologist to identify the congengital heart disease
on fetal. Convolutional Neural Network (CNN) is the method used by the authors in this
study. Here will be classified 4 screening views of fetal heart using CNN with the pretrain model architecture VGG16, VGG19 and ResNet-50, where the main focus of the
research is the average f1-score of the models produced during training and testing.
CNN's classification yielded average f1-score of 100% for VGG16, and 99% for VGG19
and ResNet-50 for model training. After the trained model are obtained, it is used to
classify the dataset outside the training data, the result is that the average f1-score is 99%
for VGG16, 87% for VGG19, and 69% for ResNet50. VGG16 has the best results for
classifying the used dataset