Thuy Pham earned her Ph.D degree in Electrical and Information Engineering at The University of Sydney, NSW, Australia. Her passion for signal processing started at The University of Arizona, USA with her MSc in Electrical and Computer Engineering in 2010, where she built her background on Neuroscience, Robotics, Digital Signal and Image Processing. She was awarded the UA Meritorious Awards (2010-2012) and the Research Grant, Neuroscience Dept., for her thesis on a real-time neural signal processing system for dragonflies towards hybrid bio-robots, at The Higgins Laboratory, Neuromorphic Vision and Robotic Systems, The University of Arizona. She received the Australian Prime Minister Award (Endeavour, 2013-2017) and Norman I. Price Scholarship for her PhD study.
Before joining UTS, she has been held visiting scholar appointments and collaborated with world-leading medical and physiology groups: The Fuglevand Laboratory on Motor Control Neurophysiology, USA; Brain & Mind Research Institute, Sydney Medical School; Woolcock Institute of Medical Research, NSW; Garvan Institute of Medical Research, Darlinghurst NSW Australia.
Distinctions and awards
- Springer Nature Theses: Recognizing Outstanding Ph.D. Research Award 2018
- Australian Prime Minister Award (The Endeavour Scholarships and Fellowships, AUD$320,000) 2013–2017
- Norman I. Price Scholarship. 2013–2017.
- The UA Meritorious Awards, University of Arizona, USA. 2010–2011
- Progressing Research Grant, Neuroscience Dept., University of Arizona, USA. 2010.
Membership and extracurricular activities
- Knowledge Discovery and Data Mining journalism
- IEEE Student Member
- IEEE Engineering in Medicine and Biology Society Membership
- Life Sciences Community, IEEE
- Cloud Computing Community, IEEE
- IEEE Communications Society Membership
- Sponsored Member of Science Program for Excellence in Science (AAAS-American Association for the Advancement of Science)
Can supervise: YES
Bio-medical engineering; Wearable device design; Applied machine learning; Predictive medicine and healthcare applications. Fashion: Big data; Internet of things; Data science.
Fundamentals of biomedical engineering. Biomedical Engineering & Technologies
Pham, TT 2018, Applying Machine Learning for Automated Classification of Biomedical Data in Subject-independent Settings, Springer.
Pham, TT, Thamrin, C, Robinson, PD, McEwan, A & Leong, PHW 2018, 'Respiratory Artefact Removal in Forced Oscillation Measurements: A Machine Learning Approach', IEEE Transactions on Biomedical Engineering, vol. 64, no. 8, pp. 1679-1687.View/Download from: Publisher's site
Goal: Respiratory artefact removal for the
forced oscillation technique can be treated as an anomaly
detection problem. Manual removal is currently considered
the gold standard, but this approach is laborious and subjective.
Most existing automated techniques used simple
statistics and/or rejected anomalous data points. Unfortunately,
simple statistics are insensitive to numerous artefacts,
leading to low reproducibility of results. Furthermore,
rejecting anomalous data points causes an imbalance between
the inspiratory and expiratory contributions. Methods:
From a machine learning perspective, such methods
are unsupervised and can be considered simple feature extraction.
We hypothesize that supervised techniques can
be used to find improved features that are more discriminative
and more highly correlated with the desired output.
Features thus found are then used for anomaly detection
by applying quartile thresholding, which rejects complete
breaths if one of its features is out of range. The thresholds
are determined by both saliency and performance
metrics rather than qualitative assumptions as in previous
works. Results: Feature ranking indicates that our new landmark
features are among the highest scoring candidates
regardless of age across saliency criteria. F1-scores, receiver
operating characteristic, and variability of the mean
resistance metrics show that the proposed scheme outperforms
previous simple feature extraction approaches. Our
subject-independent detector, 1IQR-SU, demonstrated approval
rates of 80.6% for adults and 98% for children, higher
than existing methods. Conclusion: Our new features are
more relevant. Our removal is objective and comparable to
the manual method. Significance: This is a critical work to
automate forced oscillation technique quality control.
Sadr, N, Jayawardhana, M, Pham, TT, Tang, R, Balaei, AT & de Chazal, P 2018, 'A low-complexity algorithm for detection of atrial fibrillation using an ECG.', Physiological measurement, vol. 39, no. 6.View/Download from: Publisher's site
We present a method for automatic processing of single-lead electrocardiogram (ECG) with duration of up to 60 s for the detection of atrial fibrillation (AF). The method categorises an ECG recording into one of four categories: normal, AF, other and noisy rhythm. For training the classification model, 8528 scored ECG signals were used; for independent performance assessment, 3658 scored ECG signals.Our method was based on features derived from RR interbeat intervals. The features included time domain, frequency domain and distribution features. We assessed the performance of three different classifiers (linear and quadratic discriminant analysis, and quadratic neural network (QNN)) on the training set using 100-fold cross-validation. The QNN was selected as the highest performing classifier, and a further performance assessment on the test data made.On the test set, our method achieved an F1 score for the normal, AF, other and noisy classes of 0.90, 0.75, 0.68 and 0.32, respectively. The overall F1 score was 0.78.The computational cost of our algorithm is low as all features are derived from RR intervals and are processed by a single hidden layer neural network. This makes it potentially suitable for low-power devices.
Vo, A, Nguyen, D, Pham, T, Ha, K & Dutkiewicz, E 2018, 'Subject-Independent ERP-based Brain-Computer Interfaces', IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 26, no. 4, pp. 719-728.
Brain–computer interfaces (BCIs) are desirable for people to express their thoughts, especially those with profound disabilities in communication.
The classification of brain patterns for each different subject requires an extensively time-consuming learning stage specific to that person, in order to reach satisfactory
accuracy performance. The training session could also be
infeasible for disabled patients as they may not fully understand the training instructions. In this paper, we propose a
unified classification scheme based on ensemble classifier,
dynamic stopping, and adaptive learning. We apply
this scheme on the P300-based BCI, with the subjectindependent manner, where no learning session is required for new experimental users. According to our theoretical analysis and empirical results, the harmonized integration of these three methods can significantly boost up the average accuracy from 75.00% to 91.26%, while at the same
time reduce the average spelling time from 12.62 to 6.78 iterations, approximately to two-fold faster. The experiments were conducted on a large public dataset which had been used in other related studies. Direct comparisons between our work with the others' are also reported in details.
Mai, HT, Tran, TS, Ho-Le, TP, Pham, TT, Center, JR, Eisman, JA & Nguyen, TV 2018, 'Low-trauma rib fracture in the elderly: Risk factors and mortality consequence', BONE, vol. 116, pp. 295-300.View/Download from: Publisher's site
Pham, TT, Nguyen, D, Dutkiewicz, E, Center, JR, Eisman, JA & Nguyen, T 2018, 'A Profiling analysis of contributions of cigarette smoking, dietary calcium intakes, and physical activity to fragility fracture in the elderly', Nature Scientific Reports, vol. 8, no. 1.View/Download from: Publisher's site
Pham, TT, Leong, PHW, Robinson, PD, Thamrin, C, Gutzler, T, Jee, AS & King, GG 2017, 'Automated Quality Control of Forced Oscillation Measurements: Respiratory Artefact Detection with Advanced Feature Extraction', Journal of Applied Physiology, vol. 123, no. 4, pp. 781-789.View/Download from: Publisher's site
The forced oscillation technique (FOT) can provide unique and clinically relevant lung function information with little cooperation with subjects. However, FOT has higher variability than spirometry, possibly because strategies for quality control and reducing artifacts in FOT measurements have yet to be standardized or validated. Many quality control procedures rely on either simple statistical filters or subjective evaluation by a human operator. In this study, we propose an automated artifact removal approach based on the resistance against flow profile, applied to complete breaths. We report results obtained from data recorded from children and adults, with and without asthma. Our proposed method has 76% agreement with a human operator for the adult data set and 79% for the pediatric data set. Furthermore, we assessed the variability of respiratory resistance measured by FOT using within-session variation (wCV) and between-session variation (bCV). In the asthmatic adults test data set, our method was again similar to that of the manual operator for wCV (6.5 vs. 6.9%) and significantly improved bCV (8.2 vs. 8.9%). Our combined automated breath removal approach based on advanced feature extraction offers better or equivalent quality control of FOT measurements compared with an expert operator and computationally more intensive methods in terms of accuracy and reducing intrasubject variability.
Pham, TT, Moore, ST, Lewis, SJG, Nguyen, DN, Dutkiewicz, E, Fuglevand, AJ, McEwan, AL & Leong, PHW 2017, 'Freezing of Gait Detection in Parkinson’s Disease: A Subject-Independent Detector Using Anomaly Scores', IEEE Transactions on Biomedical Engineering, vol. 64, no. 11, pp. 2719-2728.View/Download from: Publisher's site
Freezing of gait (FoG) is common in Parkinsonian gait and strongly relates to falls. Current clinical FoG assessments are patients' self-report diaries and experts' manual video analysis. Both are subjective and yield moderate reliability. Existing detection algorithms have been predominantly designed in subject-dependent settings. In this paper, we aim to develop an automated FoG detector for subject independent. After extracting highly relevant features, we apply anomaly detection techniques to detect FoG events. Specifically, feature selection is performed using correlation and clusterability metrics. From a list of 244 feature candidates, 36 candidates were selected using saliency and robustness criteria. We develop an anomaly score detector with adaptive thresholding to identify FoG events. Then, using accuracy metrics, we reduce the feature list to seven candidates. Our novel multichannel freezing index was the most selective across all window sizes, achieving sensitivity (specificity) of 96% (79%). On the other hand, freezing index from the vertical axis was the best choice for a single input, achieving sensitivity (specificity) of 94% (84%) for ankle and 89% (94%) for back sensors. Our subject-independent method is not only significantly more accurate than those previously reported, but also uses a much smaller window (e.g., 3 s versus 7.5 s) and/or lower tolerance (e.g., 0.4 s versus 2 s).
Pham, TT & Dutkiewicz, E 2019, 'Quantify Physiologic Interactions Using Network Analysis', Computational Science and Its Applications – ICCSA 2019, International Conference on Computational Science and Its Applications, Springer, Saint Petersburg, Russia, pp. 142-151.View/Download from: Publisher's site
© 2019, Springer Nature Switzerland AG. To better understand the neural interactions amongst human organ systems, this work provides a framework of data analysis to quantify forms of neural signalling. We explore network interactions among the human brain and motor controlling. The main objective of this work is to provoke unique challenges in the emerging Network Physiology field. The proposed method applies network analysis techniques including power coherence for connectivity discovering and correlation measurement for profiling relationships. We used a well-designed dataset of 50 subjects over 14 different scenarios for each individual. We found network models for these interactions and observed informative network behaviours. The information can be used to study impaired communications that can lead to dysfunction of organs or the entire system such as sepsis.
Pham, T, Takalkar, M, Xu, M, Hoang, DT, Truong, HA, Dutkiewicz, E & Perry, S 2019, 'Airborne Object Detection Using Hyperspectral Imaging: Deep Learning Review', Computational Science and Its Applications – ICCSA 2019, International Conference on Computational Science and Its Applications, Springer, Saint Petersburg, Russia, pp. 306-321.View/Download from: Publisher's site
Hyperspectral images have been increasingly important in object detection applications especially in remote sensing scenarios. Machine learning algorithms have become emerging tools for hyperspectral image analysis. The high dimensionality of hyperspectral images and the availability of simulated spectral sample libraries make deep learning an appealing approach. This report reviews recent data processing and object detection methods in the area including hand-crafted and automated feature extraction based on deep learning neural networks. The accuracy performances were compared according to existing reports as well as our own experiments (i.e., re-implementing and testing on new datasets). CNN models provided reliable performance of over 97% detection accuracy across a large set of HSI collections. A wide range of data were used: a rural area (Indian Pines data), an urban area (Pavia University), a wetland region (Botswana), an industrial field (Kennedy Space Center), to a farm site (Salinas). Note that, the Botswana set was not reviewed in recent works, thus high accuracy selected methods were newly compared in this work. A plain CNN model was also found to be able to perform comparably to its more complex variants in target detection applications.
Ha, M, Pham, T, Nguyen, D & Dutkiewicz, E 2018, 'Non-Laboratory-Based Risk Factors for Automated Heart Disease Detection', International Symposium on Medical Information and Communication Technology, IEEE, Sydney, Australia.View/Download from: Publisher's site
Developing a heart disease detection model using simple non-laboratory risk factors plays an important role in preventive care, especially for high risk subjects. The model allows physicians/epidemiologists to effectively diagnose a person as having heart disease. In this work, we aim to develop a non-invasive risk prediction model for automated heart disease detection that involves age, gender, rest blood pressure, maximum heart rate, and rest electrocardiography. We examine four public datasets from 1071 participants who were referred for a special X-ray of the heart's arteries (i.e., to see if they are narrowed or blocked). The subjects also undertook a physical examination and three non-invasive tests. To estimate the heart disease status, we apply a generalized linear model with regularization paths via coordinate descent. Even without laboratory-based data (e.g., serum cholesterol, fasting blood sugar), we observed a prediction accuracy as high as 72%, compared with 76% of other comprehensive models. This observation suggests that few non-invasive factors utilizing recent advances in data analytics can replace the current practices of heart disease risk assessment.
Pham, T, Nguyen, D, Dutkiewicz, E, McEwan, A, Leong, P & Fuglevand, A 2018, 'Feature Analysis for Discrimination of Motor Unit Action Potentials', International Symposium on Medical Information and Communication Technology, Sydney, Australia.
Pham, TT, Nguyen, DN, Dutkiewicz, E, McEwan, AL & Leong, PHW 2017, 'An Anomaly Detection Technique in Wearable Wireless Monitoring Systems for Studies of Gait Freezing in Parkinson’s Disease', International Conference on Information Networking, International Conference on Information Networking, IEEE, Da Nang, Vietnam, pp. 41-45.View/Download from: Publisher's site
Wearable monitoring systems have been in need for studies of gaits especially freezing of gait detection in patients with Parkinson's disease. The causality of gait freezing is still not fully understood. The histogram of gait freezing is the key assessment of the disease, thus monitoring them in patients' daily life is much appreciated. A real-time signal processing platform for wearable sensors can help record freezing time instances. However, current monitor systems are calibrated with offline training (patient-dependent) that is cumbersome and time-consuming. In this work, by using acceleration data and spectral analysis, we propose an online/real-time detection technique. Periods of low acceleration and low spectral coherence are identified and patient-independent parameters are then extracted. Using this set of new features, we validated our method by comparing it with clinicians' labels. The proposed approach achieved an overall mean (±SD) sensitivity (specificity) of 87 ± 0.3% (94±0.3%). To our best knowledge, this is the best performance for automated subject-independent approaches.
Pham, TT, Nguyen, DN, Dutkiewicz, E, McEwan, AL & Leong, PHW 2017, 'Wearable Healthcare Systems: A Single Channel Accelerometer Based Anomaly Detector for Studies of Gait Freezing in Parkinson’s Disease', Proceedings of the IEEE, IEEE International Conference on Communications (ICC), Institute of Electrical and Electronics Engineers, Paris, France.
Pham, TT, Nguyen, DN, Dutkiewicz, E, McEwan, AL, Thamrin, C, Robinson, PD & Leong, PHW 2016, 'Feature Engineering and Supervised Learning Classifiers for Respiratory Artefact Removal in Lung Function Tests', Proceedings of the 2016 IEEE Global Communications Conference (GLOBECOM), IEEE Global Telecommunications Conference, IEEE, Washington, USA, pp. 1-6.View/Download from: Publisher's site
A critical task in forced oscillation technique (FOT),
a promising lung function test, is to remove respiratory artefacts.
Manual removal by specialists is widely used but time-consuming
and subjective. Most existing automated techniques have involved
simple thresholding methods in an unsupervised manner. Breath
cycles can be classified by a binary classification model (classes:
artefactual and accepted). While attempting to use off-the-shelf
sorting algorithms (e.g., one-class support vector machine, knearest
neighbours, and adaptive boosting ensemble), we noticed
their poor detection performance. This may result from the
dependence of samples as found in physiological studies of the
lung function that challenges the learning process. Specifically,
statistics of breaths that we recorded may change from one to
another patient and even within the same recording of a patient.
We introduce an additional feature engineering step that is an
intermediate module to decorrelate samples, called feature learning
(using Wilcoxon signed rank tests). To that end, we collected
FOT recordings from various groups of patients (paediatric and
adult including healthy and asthmatics). Artefacts in this work
were recorded naturally and processed in a complete-breath
approach. Performance metrics include evaluations on preservation
of “accepted” breaths in the filtered output (including F1-
score, throughput, and approval rate). Our experiment found that
our feature engineering steps significantly improve the artefact
removal performance of all implemented classifiers especially with
feature inputs selected by mutual information criterion.
Pham, TT & Higgins, CM 2014, 'A visual motion detecting module for dragonfly-controlled robots', Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE, International Conference of the IEEE Engineering in Medicine and Biology Society, IEEE, USA, pp. 1666-1669.View/Download from: Publisher's site
Pham, TT, Fuglevand, AJ, McEwan, AL & Leong, PHW 2014, 'Unsupervised discrimination of motor unit action potentials using spectrograms', Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE, International Conference of the IEEE Engineering in Medicine and Biology Society, IEEE, USA, pp. 1-4.View/Download from: Publisher's site
Pham, TT 2011, 'A REAL-TIME NEURAL SIGNAL PROCESSING SYSTEM FOR DRAGONFLIES'.
At The Fuglevand Laboratory on Motor Control Neurophysiology, Arizona, USA, she explored unsupervised spike sorting methods using machine-learning techniques to automate action potential discrimination.
At Brain & Mind Research Institute, Sydney Medical School, University of Sydney, Australia, she developed algorithms towards medical wearable devices, e.g., a real-time monitor for freezing of gait events in patients with advanced Parkinson’s disease.
At Woolcock Institute of Medical Research, Glebe, NSW Australia, Thuy enjoys the dual challenge of applying advanced machine learning and statistics to improve quality control of lung function tests, and embedding a data-first culture.
At Garvan Institute of Medical Research, Darlinghurst NSW Australia, she has involved in applied machine learning for predictive medicine studies such as contribution of lifestyle habits to human health.