Scholarly Indexed Publications

Permanent URI for this collectionhttps://hdl.handle.net/10361/28605

Browse

Now showing 1 - 20 of 701

Metadata Only
An efficient metaheuristic approach for finding motifs from DNA sequences
(Institute of Electrical and Electronics Engineers Inc., 2021-01-01) Alam, Syed Md. Shamsul; Kowser, Ibna; Islam, Md. Al-Junaed; Zaman, Shurid Shahriar; Kabir, Tahseen Tayeb; Bin Ashraf, Faisal; Department of Computer Science and Engineering
Finding patterns of the short sequences in DNA, RNA protein sequence has immense biological significance. The characterization and recognition of motifs is therefore an important method for a more in-depth understanding of genes or proteins in their structure, function and relations of evolution. This is one of the classical problems in the field of computational biology and which is an NP Hard problem. In this paper, we have proposed an evolutionary approach to get the motifs from DNA sequence by searching candidate motifs using heuristic way from the data. We have included various mutation techniques in an evolutionary approach and found an efficient way to calculate the fitness of our candidate motifs. We have evaluated the fitness of found motifs from our approach with benchmark data sets. Our method performs better results in terms of accuracy and specificity.
Metadata Only
Bengali sign language to text conversion using artificial neural network and support vector machine
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Chowdhury, Anika Raisa; Biswas, Akash; Hasan, S.M. Farzana; Rahman, Tanjina Mehnaz; Uddin, Jia; Department of Computer Science and Engineering
This paper presents a novel system that converts Bengali Sign language to text using an optimum system comprising of artificial neural networks and support vector machine (SVM). Microsoft Kinect is used to take the input, which is the hand sign performed in front of the camera. The captured hand sign is eventually recognized, after joint and wrist detection and by assessing the contours. Contour feature is extracted and is run through a SVM for classification of the sign. The contour finding algorithm utilizes the convex hull method, and the features extracted after detection is passed through the support vector model for recognition. To validate the performance of the proposed model, a dataset that consists of both male and female hand gesture images is utilized. Experimental results demonstrate 84.11% classification accuracy for our tested dataset.
Metadata Only
Spinal Cord Injured (SCI) patients' length of stay (LOS) prediction based on hospital admission data
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Mazhar, Tabib Ibne; Suha, Nusrat Jahan; Chaki, DIpankar; Ali, Md. Haider; Department of Computer Science and Engineering
In order to take better care and to ensure better facilities to the inpatients, predicting length of stay serves a great importance. Since, the resources and the doctors are limited in the hospital, especially in a developing country like Bangladesh, it is quite difficult to provide proper healthcare to the inpatients. Not only the hospital resources are limited but also, it is difficult for the inpatients to bear the expense for a long period of time. In addition to that, if doctors can predict length of stay at the early stage of pre-admission, they can map a well instructed way for example, which treatment, which instrument will treat patients best. As a result, the patient can start his treatment with a slight assumption of the expenses. If we can predict accurate length of stay, patients do not have to leave in between the treatment without medical advice. Keeping all these points in mind, we develop a system using machine learning algorithms to predict Spinal Cord Injured (SCI) patient's length of stay. For this purpose, we choose Centre for the Rehabilitation of the Paralyzed (CRP) to collect data. They provide us around 600 inpatients' admission data who have been admitted to the hospital. After that we successfully build a model for predicting length of stay (LOS) with 81.62% accuracy.
Metadata Only
Performance evaluation of seven machine learning classification techniques for movie box office success prediction
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Quader, Nahid; Gani, Md. Osman; Chaki, DIpankar; Department of Computer Science and Engineering
Movie industry is a multi-billion-dollar industry and now there is a huge amount of data available on the internet related to movie industry. Researchers have developed different machine learning methods which can make good classification models. In this paper, various machine learning classification methods are implemented on our own movie dataset for multi class classification. The main goal of this paper is to conduct performance comparison among various machine learning methods. We choose seven machine learning techniques for this comparison such as Support Vector Machine (SVM), Logistic Regression, Multilayer Perceptron Neural Network, Gaussian Naive Bayes, Random Forest, AdaBoost and Stochastic Gradient Descent (SGD). All of these methods predict an approximate net profit value of a movie by analyzing historical data from different sources like IMDb, Rotten Tomatoes, Box Office Mojo and Meta Critic. For all these seven methods, the system predicts a movie box office profit based on some pre-released features and post-released features. This paper analyzes the performance assessment of all these seven machine learning techniques based on our own dataset which contains 755 movies. Among these seven algorithms, Multilayer perceptron Neural Network gives better result.
Metadata Only
Improved optimum dynamic time slicing round robin algorithm
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Ullah, Shihab; Department of Computer Science and Engineering
In time-shared systems, selection of the time quantum plays a pivotal role in performance of CPU. In this paper, the static use of dynamic time quantum as CPU Time Slice is reviewed and a new algorithm for CPU scheduling named Improved Optimum Dynamic Time Slicing Round Robin Algorithm (IODTSRR) is proposed for process and thread scheduling. The proposed algorithm is based upon dynamic nature of allocation, calculation of the value of time quantum which varies according to the state of queue along with the capability of executing ready processes arriving at the same or different time. The concept of multi-threading by using Dummy Thread is introduced to hold the added processes in the queue during all arrival time intervals respectively. The performance is compared with Optimum Dynamic Time Slicing Using Round Robin (ODTSRR) and the results revealed that the proposed algorithm is much better specifically in response time and turnaround time. As process gets fully or partially executed while others arrive simultaneously, the context switch rates, waiting time and throughput improves hence resulting in optimized CPU performance.
Metadata Only
Bangla sign language interpretation using bag of features and Support Vector Machine
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Uddin, Jia; Arko, Fahmid Nasif; Tabassum, Nujhat; Trisha, Taposhi Rabeya; Ahmed, Fariha; Department of Computer Science and Engineering
To complete any process, communication is necessary. Deaf and dumb people use special language to communicate which is known as Sign Language. In this paper, we propose an image processing based model for interpretation of Bangla sign language. In the model, initially YCBCR color components are used to detect the skin color of the user and then extract the Bag of features for each input image. Finally extracted features are feed to the Support Vector Machine (SVM) for training and testing. To validate the proposed model, we use our own dataset where both male and female hand gestures are used. Experimental results show that the proposed model exhibited average 86% accuracy for our tested dataset. In addition, the proposed model outperforms than other state-of-art models by exhibiting higher accuracy.
Metadata Only
Abnormal event detection in crowded scenarios
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Mostafa, Tahjid Ashfaque; Uddin, Jia; Ali, Md. Haider; Department of Computer Science and Engineering
This paper proposes an autonomous video surveillance system which analyzes footages of extremely crowded scenes and detects abnormal events in the context of that particular scene. The model analyzes the local spatial-temporal motion pattern and detects abnormal motion variations and sudden changes and it can be divided into two major parts, selecting a set of Points of Interest (POI) from given frames and tracking them across multiple frames and dividing the input video frame in a number of cubes and track the motion patterns in each of the cubes for spatial-temporal statistical deviations. To evaluate the performance of proposed model we utilize several datasets and compare the acquired results of the proposed model with various state-of-the art models. Experimental results demonstrate that the proposed model outperforms the other models by exhibiting an average of 96.12% accuracy using Convolutional Neural Network.
Metadata Only
Development of a self-navigating algorithm for library book finder robot
(Institute of Electrical and Electronics Engineers Inc., 2017-07-02) Rashid, Md. Rezaur; Uzzaman N.; Hossain S.; Das Shuvra, Neel Kanto; Department of Computer Science and Engineering
In this paper a self-navigating library robot is presented that goes to the required book shelf as a human would do and does not need to follow any line or fixed path. To achieve this performance a self-navigating algorithm is developed. The library is considered as a 2D array of book shelves. The robot uses 5 ultrasonic sensors to measure the distances of its surrounding objects. These measured distances are prerequisite of achieving all the necessary performances, such as calculating the current row-column positon of the robot, help the robot to move in a straight path, making an accurate turn etc. The robot works perfectly using the developed algorithm and shows satisfactory outcome which indicates the validity of the developed algorithm.
Metadata Only
Simulation and analysis of battery performance of a solar car using a model implemented in Simulink
(Institute of Electrical and Electronics Engineers Inc., 2016-01-25) Al Maliniud, Abdullah; Zaman, Asif; Jafar, Imran Bin; Rahman, Mosaddequr; Department of Computer Science and Engineering
Battery performance of a proposed solar car has been investigated under different road conditions while the car makes a nip in one of the longest routes in Dhaka city, using a model developed in Matlab Simulink. A significant reduction in energy consumption is observed when the car runs at lower speed, which is atthbuted to lower air drag at lower speed. Energy consumption further reduces with the increase in number of stoppages. Thus the solar/electric car will be an ideal mode of transport for a city like Dhaka, where traffic congestion is a daily occurrence.
Metadata Only
Propasal of possible OTEC sites in Bangladesh
(IEEE Computer Society, 2014-01-01) Shakil, Shifur Rahman; Hossain, Md. Safwat; Rouf, Nirjhor Tahmidur; Department of Computer Science and Engineering
Bangladesh, a new name among the fast developing country, has a huge population of more than 160 Million people and a large demand for electricity. Although Bangladesh is still struggling to provide a stable energy source to the growing industrial sectors and general mass, the geographical location of this country is very advantageous for the development of renewable energy. The site of Bangladesh is naturally gifted area to establish alternate energy collection sources like Ocean Thermal Energy Conversion (OTEC). Since Bangladesh lies just beneath the tropic of cancer and on the shore of the Bay of Bengal, the two vital elements: constant sunlight and large littoral areas needed for OTEC can easily be found in this region. With huge water bodies throughout, this zone is very much suitable for developing multiple OTEC plant size.
Metadata Only
Forecasting literacy development in Bangladesh: A comparative study of machine learning and hybrid approaches
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Al Arafat Tanzin; Mashrafi, Md Jisan; Hossain, Muhammad Iqbal; Department of Computer Science and Engineering
In general in Bangladesh there have been some progress in literacy rates during the last couple of decades. Though literacy has persisted into the twenty first century as a major concern on the international level, it is still difficult to attain universal literacy. In this study, the possibility of using machine learning and mixed models to predict literacy in Bangladesh is considered, with a focus on the most accurate prediction of literacy levels by the end of the 2030 period. From a set of socio-economic and demographic indicators - including GDP, population, unemployment rate and life expectancy, we assessed few given predictive models. These were Linear Regression, SVR, Gradient Boosting, Random Forest, XGBoost, ANN, LSTM, CNN, RNN, GRU, LSTM + ARIMA, ANN + ARIMA, CNN + ARIMA, GRU + ARIMA. The hybrid models performed better than others in all parameters such as MAE, MSE, & R2 Score which again validate the ability of the hybrid models to capture temporal and non-linear dynamics inherent in the data. Further, based on the ARIMA-models analysis, literacy rate that was below 60% a decade ago is expected to reach nearly 90% in 2030. It also discusses how the policy-driven hybrid modeling can be useful and promotes data drive approach in perusing the SDG goals.
Metadata Only
Hybrid convolutional neural network and random forest model for predicting water level fluctuations in Kaptai reservoir to enhance water resource management
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Sarkar A.; Shahriyar M.F.; Chy A.M.R.; Al Arafat Tanzin, Mohammed Abdul; Mashrafi, Md Jisan; Fahim, Abrar; Department of Computer Science and Engineering
The Kaptai Reservoir, the country's largest artificial freshwater body plays a vital role in hydroelectric power generation, flood control and agricultural support. However, its water levels are subject to fluctuations influenced by both climatic variations and increasing human consumption. Despite its crucial environmental impact there is a lack of research on the Kaptai Reservoir water body. The escalating effects of climate change present significant challenges to water resource management particularly in vulnerable regions such as Bangladesh. This paper presents a machine learning-based approach utilizing Convolutional Neural Networks (CNN) for feature extraction and Random Forest Regression (RFR) for Predicting water level fluctuations in the Kaptai Reservoir. Historical climate data from 2013 to 2022 including rainfall, temperature, and humidity were used to train the model. The CNN model effectively captured both temporal and spatial relationships in the water level time series while the RFR model demonstrated high prediction accuracy with a Root Mean Square Error (RMSE) of 0.5267 and a Mean Absolute Error (MAE) of 0.3128. These results highlight the model's strong performance and its potential for real-time water management and decision-making aimed at mitigating the impacts of climate change on the reservoir. The findings offer valuable insights for policymakers working to ensure the long-term sustainability and resilience of water resources in Bangladesh.
Metadata Only
The effectiveness of different deep learning models in detecting hate speech on social media
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Jilan, Tahsin Zaman; Department of Computer Science and Engineering
Recent changes in social media made it harder to control the propagation of hate speech. One potential solution can be use of deep learning models for automated hate speech recognition. In this work, we evaluate how well different deep learning models classify hate speech on social networks. In our experiments, we use a dataset of social media posts with and without hate speech.We examine the results of a number of different models, including attention-based models, convolutional neural networks (CNNs), and long short-term memory (LSTM). We also examine the effects of additional variables, such as the amount of training data and the use of pre-trained word embeddings, on the performance of these models. Our results demonstrate that attention-based models perform better than CNN and LSTM algorithms in identifying hate speech. To sum up, our research offers valuable perspectives on enhancing deep learning models for the identification of hate speech.
Metadata Only
Ret-detect: Deep learning-driven automated detection of retinal diseases
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Rahman, Md Mahfujur; Roy, Amit; Karim, Dewan Ziaul; Department of Computer Science and Engineering
Visible light from the environment is captured by human eyes, and the retina creates an image of it and sends it to the human brain. However, this process gets interrupted when our eyes get affected by diseases. Retinal diseases, including AMD, DR, and Glaucoma, are heavily responsible for vision impairment and blindness. To get rid of them and prevent loss of our vision, we need to diagnose these diseases early. Sometimes it becomes difficult to move to a medical center and manage a trained doctor to diagnose on time. In order to solve this issue, we can take advantage of deep learning-based machine learning. This paper proposes a CNN-based deep learning model named "Ret-Detect"to detect retinal diseases using retinal images with a higher success rate. In this model, we use a total of 39,800 retinal images, and we split them for training and testing. We use 80% of the total image, which is 32000 images to train the model, 20% of the images, totaling 7,800 images, will be used for testing purposes. Our model got a training accuracy of 97.96%, and a validation accuracy of 96.00%. We also made a comparison with some pre-trained models, such as MobileNet, DenseNet121, VGG19, InceptionV3, VGG16, ResNet50 and Ensemble model and we got an accuracy of 86.10%, 75.59%, 92.62%, 94.35%, 23.75%, 73.62% and 94.76%for validation. We observed that our custom model shows a better accuracy to detect retinal diseases, and it can provide a noticeable assistance to reduce the number of blindness all over the world.
Metadata Only
Detecting different stages of Alzheimer's disease from MRI images using deep learning and computer vision techniques
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Sharif, Sadman; Islam, Syed Shams; Karim, Dewan Ziaul; Department of Computer Science and Engineering
The precise and early detection of Alzheimer's disease (AD) is essential for prompt intervention and effective management. Neuroimaging methods, such as magnetic resonance imaging (MRI), offer crucial information about brain structure and function; however, they can be time-consuming and prone to observer variability when interpreted manually. This study proposes a novel deep learning framework that leverages computer vision techniques to automate the classification of AD stages, including early & late Mild Cognitive Impairment (MCI), using MRI images. A custom Convolutional Neural Network (CNN) model was developed and optimized, achieving a classification accuracy of 96.6%, surpassing state-of-the-art architectures. To address challenges like class imbalance and data scarcity, we employed advanced data augmentation techniques and unsupervised pre-training using a denoising convolutional autoencoder. Additionally, model explainability was improved through gradient-weighted class activation mapping (Grad-CAM), highlighting critical parts in MRI scans that influence predictions. The findings show that this method could help enhance diagnostic accuracy and provide a robust tool for neurodegenerative disease research and clinical application.
Metadata Only
CDFNet based fusion network for interpretable Alzheimer's Disease prediction using hybrid imaging and clinical meta-ensemble learning
(Institute of Electrical and Electronics Engineers Inc., 2026-01-01) Sakib, Tauhidur Rahman; Shakib M.A.; Shourov, M. M. Nasim Osmani; Badhon S.I.; Shabbir, Lotifur; Hasan, Shayonton; Department of Computer Science and Engineering
This work presents a CNN-DenseNet Fusion Network (CDFNet) for interpretable Alzheimer's disease (AD) stage classification using two complementary data streams: structured clinical attributes and MRI scans. A stacking ensemble on the clinical dataset identified Functional Assessment, ADL, and MMSE as dominant predictors, achieving precision 0.98 and recall 0.97, confirming clinically meaningful feature usage. The MRI pipeline, powered by CDFNet, reached 97.4% accuracy and macro F1 of 0.98, outperforming single-branch CNN and DenseNet121 baselines. t-SNE clustering, ROC curves (AUC ? 0.998), and Grad-CAM maps over hippocampal and ventricular regions validated both discrimination and biological plausibility. Together, these pipelines form a stable, leakage-safe, and clinically aligned framework for multimodal dementia staging.
Open Access
Are remittances a buffer against food insecurity? Lessons from a national survey in Bangladesh
(Public Library of Science, 2025-10-01) Alam, Md. Muhitul; Ria, Faria Rauf; Mansur, Mohaimen; Uddin, Md. Azad; Rayhan, Md. Israt; Department of Mathematics and Natural Sciences
Food insecurity continues to be a major global challenge, affecting many people worldwide. Bangladesh is particularly vulnerable due to its susceptibility to frequent climate shocks and socioeconomic challenges. This study investigates the causal relationship between remittance receipt and food security through a comprehensive analysis. Using data from the Household Income and Expenditure Survey (HIES) 2022, we developed a food security index incorporating calorie intake, dietary diversity, food expenditure, and the Food Insecurity Experience Scale (FIES) score. Advanced statistical methods, including Seemingly Unrelated Regression (SUR), Zero-Inflated Negative Binomial, linear regression model, inverse-probability-weighting (IPW), and doubly robust method were employed to identify the factors associated with food security and assess the causal effect of remittance earning. Our findings reveal a strong and positive causal effect of remittance receipt on food security. The observed causal effect remained robust against model misspecification and unmeasured confounders, as confirmed through sensitivity analysis. Key factors such as wealth index, residence type, regional differences, household head’s education, and number of earners also influenced food security outcomes. However, the significance of variables like land ownership, household head’s age, and sex varied across measures. This study highlights the transformative role of remittances in reducing food insecurity. Policies that support remittance flows, improve rural infrastructure, and promote skill development and financial literacy can further strengthen their impact. © 2025 Alam et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Open Access
Exploring regional air pollution transition dynamics: A multi-state markov model approach
(Public Library of Science, 2025-10-01) Hossain, Md. Ismail; Sarkar, Shuvongkar; Methun, Md. Injamul Haq; Rahman, Azizur; Department of Mathematics and Natural Sciences
Background Air pollution, commonly measured by the Air Quality Index (AQI), is a significant global health risk, yet its transition dynamics remain poorly understood. This study aims to investigate the regional air pollution transition dynamics across different air quality states. Materials and methods We analyzed weekly average Air Quality Index (AQI) data from January to September 2024 for 19 countries across Asia, Africa, and Europe, collected from an open-access air quality monitoring platform. According to international standards, AQI was categorized into three states (Good, Unhealthy, Very Unhealthy). We applied a multi-state Markov model to assess weekly transitions between these states and estimate the average time spent in one state before transition. Results Findings indicate that in Asia and Africa, air quality tends to deteriorate more frequently than it improves, with low transition rates from “Very Unhealthy” to better states. Transitions from Unhealthy to Good were less frequent in Asia (HR: 0.09, 95% CI: 0.04, 0.19) and Africa (HR:0.25, 95% CI: 0.11, 0.55) compared to Europe, where air quality showed more stability and improvement. The Good and Unhealthy states in Asia had similar sojourn times of 6.80 (±1.77) and 6.64 (±1.38) weeks, while the Very Unhealthy state lasted 3.36 (±0.98) weeks. The Very Unhealthy state persisted for 0.95 (±0.48) weeks in Africa. Europe maintained the “Good” state longest at 7.68 (±1.98) weeks, with shorter durations for Unhealthy and Very Unhealthy states. Conclusion The study highlights lengthy pollution incidents in Asia and Africa, while Europe demonstrates effective pollution control. These insights can guide policymakers in formulating strategies to mitigate pollution based on regional AQI transition trends. © 2025 Hossain et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Open Access
Optimizing network bandwidth slicing identification: NADAM-enhanced CNN and VAE data preprocessing for enhanced interpretability
(Public Library of Science, 2025-10-01) Ul Islam, Md. Fahim; Hossain, Shahriar; Alam, Md. Golam Rabiul; Mansoor, Nafees; Chakrabarty, Amitabha; Department of Computer Science and Engineering
Communication networks of the future will rely heavily on network slicing (NS), a technology that enables the creation of distinct virtual networks within a shared physical infrastructure. This capability is critical for meeting the diverse quality of service (QoS) requirements of various applications, from ultra-reliable low-latency communications to massive IoT deployments. To achieve efficient network slicing, intelligent algorithms are essential for optimizing network resources and ensuring QoS. Artificial Intelligence (AI) models, particularly deep learning techniques, have emerged as powerful tools for automating and enhancing network slicing processes. These models are increasingly applied in next-generation mobile and wireless networks, including 5G, IoT infrastructure, and software-defined networking (SDN), to allocate resources and manage network slices dynamically. In this paper, we propose an Interpretable Network Bandwidth Slicing Identification (INBSI) system that leverages a modified Convolutional Neural Network (CNN) architecture with Nesterov-accelerated Adaptive Moment Estimation (NADAM) optimization. Additionally, we use a Variational Autoencoder (VAE) for preprocessing initial data, along with reconstructed data for data validity assessment. The model we propose outperforms other alternatives and reaches an accuracy peak of (84%) in the system environment. A range of accuracy was achieved by (k-nearest neighbors algorithm) KNN (76%), Random Forest (69%), BaggingClassifier (70%), and Gaussian Naive Bayes (GaussianNB) (55%). The accuracy of additional methods varies, including Decision Trees, AdaBoost, Deep Neural Forest (DNF), and Multilayer Perceptrons (MLPs). We utilize two eXplainable Artificial Intelligence (XAI) approaches, Shapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), to provide insight into the impact of certain input characteristics on the network slicing process. Our work highlights the potential of AI-driven solutions in network slicing, offering insights for operators to optimize resource allocation and enhance future network management. © 2025 Islam et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Metadata Only
Predicting super-class drug mechanisms from large-scale IC50 cell-line sensitivity profiles
(Institute of Electrical and Electronics Engineers Inc., 2026-01-01) Sakib, Tauhidur Rahman; Shakib, Mostofa Adib; Munir, Saad Ibn; Badhon, Saiful Islam; Sakif, Md. Sadman Yeasir; Roy, Mim Deb; Department of Computer Science and Engineering; School of Pharmacy
Predicting mechanisms of action of drug compounds at the super-class level is indispensable for early drug discovery and functional assessment of compounds. The aim of this study is to perform Super-Class-level mechanism-of-action prediction from phenotypic IC50 cell-line sensitivity data under strict leakage control. In this work, we present a leakage-safe machine learning pipeline for Super-Class Mechanism-of-Action (Super-Class MoA) prediction using drug-cell line sensitivity profiles prepared at large scale from the GDSC dataset. The standardized features derived from IC50 were encoded into 32-dimensional latent embeddings using an autoencoder, while SMOTE was applied solely to the training embeddings of each outer fold to address class imbalance. The resulting embeddings were used to train a stacked neural-boosted ensemble consisting of a neural network and XGBoost, with validation embeddings kept completely untouched to ensure unbiased evaluation. The framework achieved a mean accuracy of 0.9182, a mean macro-F1 score of 0.9118, and a mean macro-AUROC of 0.9568 across five outer folds, demonstrating consistent performance across the three Super-Class MoA categories. These results indicate that latent representations derived from phenotypic drug-response profiles capture biologically meaningful variation, supporting robust Super-Class MoA prediction in early-stage screening settings.

Browse

Recent Submissions