Biomarker Identification by Reversing the Learning Mechanism of Autoencoder and Recursive Feature Elimination
Authors: Fuad Al Abir, S.M. Shovan, Md. Al Mehedi Hasan, Abu Sayeed, Jungpil Shin
Abstract: RNA-Seq has made significant contributions to various fields, particularly in cancer research. Recent studies on differential gene expression analysis and the discovery of novel cancer biomarkers have used RNA-Seq data extensively. New biomarker identification is essential for moving cancer research forward, and early cancer diagnosis improves patients’ chances of recovery and increases life expectancy. There is an urgency and scope of improvement in both sections. In this paper, we developed an autoencoder-based biomarker identification method successfully applied to the UCI gene expression cancer RNA-Seq dataset consisting of five cancerous tumor types. By reversing the learning mechanism of the trained autoencoders, we devised an explainable post hoc methodology for identifying the influential genes with a high likelihood of becoming biomarkers. We applied recursive feature elimination to shorten the list further and presented a list of 11 potential biomarkers that are 99.79% accurate in identifying cancer types using support vector machine. Furthermore, we have compiled and validated separate lists of potential biomarkers for each cancer type. Our results outperform the state-of-the-art methods and confirm the potentiality of the newly identified biomarkers and the efficacy of the biomarker identification procedure. The source code of this project is available at https://github.com/fuad021/biomarker-identification.
November 2021 | Under Review
Gait Recognition with Wearable Sensors using Modified Residual Block-based Lightweight CNN
Authors: Md. Al Mehedi Hasan*, Fuad Al Abir*, Md. Al Siam, Jungpil Shin
(* equal contribution)
Abstract: Gait recognition with wearable sensors is an effective approach to identifying people by recognizing their distinctive walking patterns. Deep learning-based networks have recently emerged as a promising technique in gait recognition, yielding better performance than template matching and traditional machine learning methods. However, most recent studies have focused on improving gait detection accuracy while neglecting model complexity in the deep learning domain, making them unsuitable for low-power wearable devices. Therefore, inference from these models results in latency due to calculation overhead. This study proposes an efficient network suitable for wearable devices without sacrificing prediction performance. We have modified the residual block and accumulated it in shallow convolutional neural networks with five weighted layers only for gait recognition and proved the efficacy of all the architectural components with extensive experiments over publicly available IMU-based datasets: whuGait and OU-ISIR. Our proposed model outperforms all the state-of-the-art methods regarding recognition accuracy and is more than 85 percent efficient on average in terms of model parameters and memory consumption.
18 April 2022 | Published by IEEE Access | Q1 | Impact Factor: 3.367
Webpage |
Paper
Deep learning based air-writing recognition with the choice of proper interpolation technique
Authors: Fuad Al Abir, Md. Al Siam, Abu Sayeed, Md. Al Mehedi Hasan, Jungpil Shin
Abstract: The act of writing letters or words in a free space with hand or finger movements is known as air-writing. Air-writing recognition is a subset of gesture recognition in which gestures correspond to letters written in the air. Air-writing, unlike general gestures, does not require memorization of special gesture patterns and is sensitive to the subject and language of interest. Wide adoption of smart-bands eliminates the requirement of an extra device in air-writing recognition. Therefore, there is a growing interest in developing recognition models using it. However, the variability of signal duration is a key problem in developing an air-writing recognition model. Inconsistent signal duration is obvious due to the nature of the writing and data recording process. The researchers attempted various strategies to make the signals consistent in length, including padding and truncating, but these procedures result in significant data loss. Interpolation is a statistical technique that can be employed in this case to ensure minimum data loss, although it is often overlooked by the researchers. In this paper, we investigated different interpolation techniques upon the smart-band data extensively and developed a method to recognize air-written letters using a 2D-CNN model. Using bicubic interpolation, we acquired 91.34% and 85.59% prediction accuracy in user-dependent and user-independent settings, respectively which outperforms the state-of-the-art method by a clear margin.
13 December 2021 | Published by Sensors, MDPI | Q2 | Impact Factor: 3.576
Webpage |
Paper
Most Dominant Metabolomic Biomarkers Identification for Lung Cancer
Authors: Utshab Kumar Ghosh, Fuad Al Abir, Nahian Rifaat, S. M. Shovan, Abu Sayeed, Md. Al Mehedi Hasan
Abstract: Metabolomic biomarkers play a vital role in the early identification and prediction of cancer. It is possible to save numerous lives if biomarkers are used to assist medical providers in diagnosing their patients faster. Many researchers have been trying to identify the crucial biomarkers in the early diagnosis of diseases. This paper presents several steps divided into two phases for determining the most important metabolomic biomarkers in the blood for lung cancer prediction using Plasma and Serum samples. We used the Shapiro–Wilk Test, Bartlett’s Test, Levene’s Test, Student’s t-Test, and Kruskal–Wallis Test in the first phase to determine the potential biomarkers. Recursive Feature Elimination with Random Forest was used to identify the final most dominant metabolomic biomarker at the second phase. Lastly, we ended with Ridge Classifier and XGBoost Classifier to assess the consistency of our approaches. Despite the declining number of metabolites up to a greater level, our prediction accuracy was 100% and 90.91% for Plasma and Serum samples, respectively which is higher than the state-of-the-art method. Finally, we made some analysis using the most dominant metabolites that can serve as a source of inspiration for our work.
14 December 2021 | Published by Informatics in Medicine Unlocked, Elsevier | Q3 | Impact Factor: 3.373
Webpage |
Paper
Surface Type Classification for Autonomous Robots Using Temporal, Statistical and Spectral Feature Extraction and Selection
Authors: Md. Al Mehedi Hasan, Fuad Al Abir, Jungpil Shin
Abstract: Real-time surface recognition has become a crucial component in assuring the safe walking of intelligent autonomous robots in a complex human-living interior environment. Numerous studies have been done addressing the problem recently. Still, there is a scope of improvements for accurate classification and inference time. In this paper, we have extracted features from accelerometer and gyroscope data in the temporal, statistical and spectral domain and classified them using a tree-based ensembling classification algorithm. We have achieved 80.81% mean accuracy, classifying 9 different surfaces with 1.0% standard deviation in 10-fold cross-validation and 97.25% average AUC score. Our method acquired state-of-the-art accuracy ensuring minimal inference time which is essential for real-time recognition for the autonomous robots.
21 December 2021 | Presented at 14th International Symposium on Embedded Multicore/Manycore Systems-on-Chip (MCSoC)
Paper |
Presentation
Islamic Geometric Patterns with Catmull-Rom Splines
Authors: Fuad Al Abir
Abstract: We present a simple technique for generating Islamic Geometric Patterns with computer algorithm using Catmull-Rom spline - a class of local interpolating spline. All the previous works of developing the methods of generative Islamic Patters was all about the straight line vectors and repetition of a segment based on symmetry group theory. This paper introduces curves in that scenario and flourishes infinite Islamic Geometric Patterns that can be generated with the algorithm.
Course Instructor: Prof. Dr. AHM Sarowar Sattar & Biprodip Paul (Asst. Prof)
April 2019 | CSE 3112: Technical Writing and presentation
Paper |
Presentation