Limpiti, Tulaya

Preferred name

Limpiti, Tulaya

Alternative Name

Limpiti, T.

Main Affiliation

School of Engineering

tulaya.li@kmitl.ac.th

Now showing 1 - 5 of 5

Some of the metrics are blocked by your
consent settings
BAMS: Binary Sequence-Augmented Spectrogram with Self-Attention Deep Learning for Human Activity Recognition
(2024-01-01)
Sricom, Natchaya
;
Charakorn, Rujikorn
;
Manoonpong, Poramate
;
Limpiti, Tulaya
Human Activity Recognition (HAR) has rapidly gained interest over the years due to its wide range of applications in AI-based systems, particularly healthcare monitoring. HAR methods typically involve extracting relevant features from data provided by wearable sensors, smartphone sensors, cameras, or their combinations to classify different activities. Nevertheless, a major challenge lies in achieving high classification accuracy with limited data samples, particularly when distinguishing between activities with similar signal attributes. To address this challenge, we propose a novel HAR method called BinAry sequence-augmented spectrograM with Self-attention deep learning (BAMS). Our proposed method leverages only basic wearable sensor data. It utilizes short-time Fourier transform spectrograms to extract spatio-temporal sensor information. The spectrogram is integrated with a binary sequence that captures movement direction. We integrate a scaled dot-product self-attention mechanism into the model to prioritize data from wearable sensors, thereby enhancing the model's performance. The proposed method is evaluated on a public dataset using leave-one-subject-out cross-validation for efficacy and robustness. The method is found to achieve significant improvement over other state-of-the-art methods with the classification accuracy percentage and weighted F-1 scores of 88.06±5.11 and 87.36±5.96, respectively, for a twelve-activity classification.
Some of the metrics are blocked by your
consent settings
Non-invasive Optical Blood Glucose Measuring System using Regression Models
(2023-01-01)
Limpiti, Tulaya
;
Seemavijai, Mayravee
;
Pongmee, Narutchai
;
Kanayat, Pannita
;
Koseeyaporn, Jeerasuda
Diabetes is one of the chronic diseases with an increasing number of patients every year. Therefore, blood glucose level (BGL) is an essential health information for diagnosis and management of diabetes. Commonly, the BGL is usually measured invasively by taking blood samples, which can cause pain or discomfort to the patients. Our paper aims to design and develop a low-cost, low-computational non-invasive blood glucose measuring system using optical technique. The proposed system consists of a hardware for measuring photoplethysmogram (PPG). The PPG signal is subsequently analyzed to estimate the corresponding BGL, which is sent wirelessly to be stored in a database and displayed on an Android application. The application recommends the appropriate dosage of insulin injection based on the type of users and their diets. The report of historical blood glucose levels in the database can also be generated for further medical examination.Three PPG features, namely peak-to-peak amplitude, maximum amplitude, and standard deviation of signal amplitude, have the strongest correlations to the BGL. We investigate three regression models for estimating the BGL from these PPG parameters. The robustness of the system is assessed using cross validations. The average root mean squares error (RMSE) for three-fold cross-validations are 17.38, 16.76, and 16.39 mg/dL for simple linear regression, principal component regression, and partial least squares regression, respectively. The RMSEs from leave-one-out cross validation are approximately 12 mg/dL, with partial least squares regression model having the best accuracy. Furthermore, the results from Clarke Error Grid Analysis indicate that the system can be implemented with any of the three models for practical usage.
Some of the metrics are blocked by your
consent settings
EEG-BBNet: A Hybrid Framework for Brain Biometric Using Graph Connectivity
(2025-01-01)
Lakhan, Payongkit
;
Banluesombatkul, Nannapas
;
Sricom, Natchaya
;
Sawangjai, Phattarapong
;
Sangnark, Soravitt
Most EEG-based biometrics rely on either convolutional neural networks (CNNs) or graph convolutional neural networks (GCNNs) for personal authentication, potentially overlooking the limitations of each approach. To address this, we propose EEG-BBNet, a hybrid network that combines CNNs and GCNNs. EEG-BBNet leverages CNN’s capability for automatic feature extraction and the GCNN’s ability to learn connectivity patterns between EEG electrodes through graph representation. We evaluate its performance against solely CNN-based and graph-based models across three brain–computer interface tasks, focusing on daily motor and sensory activities. The results show that while EEG-BBNet with Rho index functional connectivity metric outperforms graph-based models, it initially lags behind CNN-based models. However, with additional fine-tuning, EEG-BBNet surpasses CNN-based models, achieving a correct recognition rate of approximately 90%. This improvement enables EEG-BBNet to adapt its learning in new sessions and to acquire different domain knowledge across various BCI tasks (e.g., motor imagery to steady-state visually evoked potentials), demonstrating promise for practical authentication.
Some of the metrics are blocked by your
consent settings
Cardiac Arrhythmia Teletriage using Electrocardiogram
(2021-01-01)
Limpiti, Tulaya
;
Chokchaichumnankit, Thunchanok
;
Sanguanchom, Jirayu
;
Soyphan, Natchanon
;
Puttarak, Nattakan
Telehealth has become a favorable method for receiving medical care during the COVID-19 pandemic. It reduces physical contact and also benefits those who live a distance from hospitals. In this paper we present a cardiac arrhythmia teletriage using electrocardiogram (ECG). The system consists of a diagnostic algorithm for arrhythmias and an Android application. The diagnostic algorithm can detect five types of cardiac problems-arrhythmia, bradycardia, tachycardia, bradyarrhythmia, and tachyarrhythmia. The Android application is the main communication channel between patients and healthcare providers. The user uploads their ECG and receives a preliminary diagnosis of their heart health via the application. The system notifies the user if it detects any abnormalities. The user can then make an appointment online for further examination at the hospital. The capability of the proposed system is evaluated using four databases from PhysioNet - MITBIH Normal Sinus Rhythm Database, MIT-BIH Arrhythmia Database, MIT-BIH Atrial Fibrillation Database, and CU Ventricular Tachyarrhythmia Database. It is found that the algorithm is able to detect abnormal ECG signals with an average accuracy of 82.1%.
Some of the metrics are blocked by your
consent settings
Transfer Learning for Classifying Motor Imagery EEG: A Comparative Study
(2021-01-01)
Limpiti, Tulaya
;
Seetanathum, Kornthum
;
Sricom, Natchaya
;
Puttarak, Nattakan
Motor imagery (MI) is the cognitive process when a person imagines performing a specific movement of their body. The corresponding electroencephalographic (EEG) signals can be measured externally to the head using scalp electrodes. Such signals have been applied in healthcare brain-computer-interface (BCI) systems, for example, motor rehabilitation and prosthetics control. These systems convert different MI EEG input signals to directives, so their performances depend on the efficiency of the embedded signal classification algorithm. In this paper we investigate the effectiveness of transfer learning in classifying the MI EEG data. The Continuous Wavelet Transform (CWT) is used to construct the scalograms, which serve as the inputs to the deep learning structure. The efficacies of five pre-trained networksAlexNet, ResNet18, ResNet50, InceptionV3 and ShuffleNet-are evaluated on the BCI competition IV data set 2a. Binary (left hand vs. right hand) and four-class (left hand, right hand, both feet, and tongue) classifications are trained and tested using fivefold cross validation. The result indicates that using the CWT with transfer learning models provides very high classification accuracies. The ResNet18 network achieves the best accuracies in both cases at 95.03±2.95% and 91.86±2.90%, respectively. In addition, we examine the effect of different time-frequency features on the classification performance by comparing the scalogram of the CWT and the spectogram of the Short-Time Fourier Transform (STFT) as the inputs. It is found that the CWT is the preferred choice as it is superior to the STFT.

Limpiti, Tulaya

Filters

Settings

Sort By

Results per page