Nima Wickramasinghe from the Department of Electronic and Telecommunication Engineering (ENTC), together with his mentor Mohamed Athif from the Department of Biomedical Engineering, Boston University (previously an undergraduate at ENTC) have written a research paper, titled “Multi-label Classification of Reduced-lead ECGs using an Interpretable Deep Convolutional Neural Network” which has been accepted to be published in the journal Physiological Measurement as a special issue paper.
In their work, they propose a novel method to identify the presence of 26 cardiac abnormalities in an ECG recording with reduced leads. Even though most of the previous work relies on 12-lead ECGs, classification using reduced leads remained unexplored. In their research, they trained a deep convolutional neural network to classify the ECG recordings and showed that the reduced-lead model performs comparably to the 12-lead model. In addition to accurately classifying the cardiac abnormalities, they have used SHAP (shapley additive explanations: a game-theoretic approach used to explain the output of any machine learning model) to interpret the deep learning model. The authors identified that the model learns almost the same diagnostic criteria used by cardiologists to classify cardiac abnormalities. By analyzing the model through SHAP, they were able to detect why the model underperforms in some of the classes, which was mainly due to the lack of discriminating features in reduced leads, labeling inconsistencies in the dataset, and low number of samples.
Physiological Measurement is a journal that covers the quantitative measurement and visualization of physiological structure and function in clinical research and practice, with an emphasis on the development of new methods of measurement and their validation. The Sustainable Education Foundation facilitated the collaboration between the 2 authors.
DOI Link to the paper: https://doi.org/10.1088/1361-6579/ac73d5Read More
Mr. Pasan Dissanayake and Dr. Prathapasinghe Dharmawansa of ENTC have Published Two Research Papers in the Prestigious Journal IEEE Transactions on Information Theory
Two research papers written by Mr. Pasan Dissanayake and Dr. Prathapasinghe Dharmawansa of ENTC have been published in the prestigious journal IEEE Transactions on Information Theory. IEEE Transactions on Information Theory is the world’s No. 1 journal in the areas of Information and Communication theory research. This truly exceptional achievement will make the research history of ENTC. This extraordinary academic achievement will be a guiding spirit for the current and future researchers in the entire university system of Sri Lanka. Moreover, it will help place ENTC among the top researchers in information and communication theory in the world. The details of the two papers are as follows.
1. Distribution of the Scaled Condition Number of Single-spiked Complex Wishart Matrices
This paper statistically characterizes the scaled condition number (SCN) of single-spiked complex Wishart matrices by deriving its density function. The statistical characteristics of the SCN and its variants have been instrumental in understanding many physical phenomena across a heterogeneous field of sciences. While numerical analysts and statistical physicists are interested in the behavior of the SCN for white Wishart matrices, the case corresponding to correlated Wishart matrices are of paramount importance in wireless communications and statistics. In particular, the SCN has been used as a performance metric in certain wireless signal processing applications involving multiple-input multiple-output (MIMO) systems, in which the antenna correlation gives rise to the correlated Wishart matrix. Recently, the SCN has been proposed as one of the test statistics for blind spectrum sensing in cognitive radio (CR) systems. The key concept behind CR is to opportunistically utilize the underutilized spectrum in view of improving the spectral efficiency of modern wireless networks. Against this backdrop, this paper leverages powerful random matrix theoretic techniques and the novel density of the SCN to statistically characterize the receiver operating characteristics (i.e., ROC) of the aforementioned detector. Since the modern wireless architectures facilitate the use of large antenna/sensor arrays with comparable observational sample acquisition, the analysis has been extended to the asymptotic regime in which the number of antennas of the detector and the samples diverge at the same rate so that their ratio remains constant. It turns out that, in this asymptotic regime, the statistical power of the SCN based detector can be approximated by the most celebrated Tracy-Widom distribution corresponding to the complex matrices. Moreover, numerical results have revealed that those asymptotic results compare favourably with their not so large dimensional counterparts.
2. The Eigenvectors of Single-spiked Complex Wishart Matrices: Finite and Asymptotic Analyses
This paper investigates the finite dimensional distributions of the eigenvectors corresponding to the extreme eigenvalues (i.e., the minimum and the maximum) of single-spiked complex Wishart matrices. These spikes arise in various practical settings in different scientific disciplines. For instance, they correspond to the first few dominant factors in factor models arising in financial economics, the number of clusters in gene expression data, and the number of signals in detection and estimation theory. In particular, the focus is on the distributions of the squared modulus of the eigen-projectors (i.e., projection of the spiked-vector onto the leading and least eigenvectors) of single-spiked Wishart matrices. This metric is commonly used to infer information about the latent spiked-vector using the eigenvectors of the sample covariance matrix. A concrete example in this respect is the principal component analysis (PCA) in which the eigenvectors of the unknown population covariance matrix is approximated by the eigenvectors of the sample covariance matrix. This metric has further been used in the covariance estimation based on the optimal shrinkage of the eigenvalues of the sample covariance matrix in the high dimensional setting when the unobserved population covariance matrix assumes the spiked structure. This paper leverages the powerful contour integral representation of unitary integrals and orthogonal polynomial techniques to derive closed-form expressions for the densities of the above metrics. A somewhat surprising stochastic convergence result pertaining to the above metrics has also been established. Finally, the same analytical framework has been extended to derive the corresponding destines for real and singular Wishart scenarios; however, with closed-form solutions limited to a few special configurations only.
Team DigitX, a team of ten undergraduates from the Department of Electronic and Telecommunication Engineering (ENTC) under the supervision of Dr. Chamira Edussooriya from ENTC, University of Moratuwa, emerged as the champions of the Students Challenge at the Internal Conference on Autonomous Systems (ICAS) 2021 conference. The conference was held virtually in Montreal, Canada from August 11 to 13, 2021.
ICAS 2021 is the premier international forum organized by the IEEE Signal Processing Society (SPS) through IEEE SPS Autonomous Systems Initiative, for presenting the technological advances and research results in the fields of theoretical, experimental and applied autonomous systems. IEEE ICAS was held for the first time this year and it focused on providing an updated state-of-the-art over advanced signal processing theories and techniques that are relevant for developing autonomous systems.
The ICAS 2021 Students Challenge was on “unsupervised anomaly detection through self-aware autonomous systems”, to detect abnormal events during the navigation of a ground/ aerial vehicle. The team was able to develop an unsupervised learning algorithm with state-of the-art signal processing and machine learning techniques such as auto-encoders, recurrent neural networks for time-series data signal processing.
Based on the inertial measurement unit (IMU) sensor data, video data captured by the cameras of autonomous systems in several camera perspectives and light detection and ranging (LIDAR) sensor data (point clouds), the unsupervised model was trained to identify the normal situations that it saw during the training phase as normal and every other situation as abnormal (as per the challenge requirement a reference normal scenario was defined for each dataset). UoM won the first place while a team from Indian Institute of Technology (IIT), India and a team from CentralSupelec, France won first and second runners up, respectively.
An autonomous system is an artificial system capable of performing a set of tasks with a high degree of autonomy. Developing computing systems with advanced levels of autonomy has been a crucial task for decades in order to manage ever-increasing requirements in complexity. This achievement by UoM places Sri Lanka at a higher position in the world signal processing community.Read More