IEEE i-COSTE 2025 – Page 7 – NPS Australia Submission System

Correlative Analysis on The Effect of Time and Concentration Variation of Various Enzyme Wash (Acidic, Neutral and Granular) On Denim Garments

Dr. Waziha Farha – Conceptualization, Project Administration, Supervision, Writing – Original Draft, Writing – Review & Editing
Dr. Umme Habiba Rahman – Writing – Review & Editing
Rejaul Karim – Writing – Review & Editing
Md. Alauddin Sany – Data Curation
Md. Limon Hosen Miazi – Data Curation
Sanid Al Rafsan – Data Curation
Tanvir Hossain Ashiq – Data Curation

A Multi-functional Smart Pillbox for Elderly Patients: Integrating Reminders, Health Monitoring, and Emergency Features with AI-based Health Condition Prediction

1) Integrated Hardware and Software Solution: We de-
veloped a comprehensive system combining a smart
medicine reminder box with a mobile application and a
web dashboard, ensuring seamless interaction between
physical medication management and digital health
monitoring.
2) Enhanced Medication Adherence and Emergency Re-
sponse: The system features intelligent reminders
(LEDs, buzzer, app notifications) for timely medication
intake and includes a critical panic button for immediate
emergency alerts via SMS and app notifications to
caregivers
3)

IoT-Based Automated Hydroponic System for Smart Agriculture

This work develops a low-cost, solar-powered IoT hydroponic system integrating multi-sensor monitoring, automated nutrient control, cloud logging, and seamless grid fallback, validated on Nutrient Film Technique (NFT) spinach with stable performance and predictive dataset generation.

Multi-Modal Fusion Tuberculosis Severity Classification: Vision Transformer for Chest X-Ray and Tabular Transformer for Clinical Data

1) A novel multimodal fusion framework for TB severity
classification, integrating chest X-ray images via pre-
trained Vision Transformer (ViT) and clinical data via
hyperparameter-tuned Tabular Transformer, with early
feature concatenation for joint learning.
2) Extensive hyperparameter tuning and regularization
(e.g., dropout, learning rate scheduling) to enhance generalization.
3) Achieved 94.01% accuracy on the MIMIC-CXR dataset
using pneumonia as TB proxy, outperforming unimodal
baselines and state-of-the-art methods.

An Efficient Framework for Suicide Risk Detection Integrating Linguistic and Emotional Features Using Graph Neural Network

A summary of the contributions of our study is as follows:
1. A new graph-based framework leveraging social media data for suicide risk detection is proposed.
2. The integration of semantic (GloVe), syntactic (dependency parsing), and emotional (SenticNet) features in a single graph structure is proposed, correlating different perspectives of the data.
3. Here, we demonstrate the creation of more context-sensitive and interpretable models that would outperform certain recent ML, DL, and LLM approaches.
4. The performance of both GraphSAGE and HGNN was significantly improved when the dataset size was increased from 4K to 8K. GraphSAGE kept its reliability high, while HGNN achieved consistent best scores due to its capacity to scale and hence improve the accuracy, recall, and robustness with larger data.

Unraveling the Genetic Tapestry: Ontological Mapping and Survival Modeling of Gene Expression Biomarkers for Autism Spectrum Disorder detection

The following are the primary contributions of this study:
* Proposed a remarkable 2-step feature extraction technique that blends statistics and ML to identify ASD-genetic features.
* Introduced a novel Gene Pathway Analysis framework to identify ASD biomarkers from high-dimensional genomic data.
* Conducted Ontological Analysis to discover critical gene connections, that aids in the exact diagnosis and an awareness of autism.
* Performed Survival Analysis on gene data, uncovering key time-linked biomarkers associated with autism progression.

Latent Representation and Generative Augmentation with Graph-Based Learning for Imbalanced Leukocyte Cytomorphology Classification

In this work, we propose a hybrid framework that integrates autoencoder-based latent representation, GAN-driven minority-class augmentation, and graph-based classification with GraphSAGE. Our key contributions are as follows:
1. We propose a latent-space augmentation strategy that synthesizes minority-class embeddings within the autoencoder feature manifold, avoiding artifacts common in pixel-level augmentation.
2. We design a graph-based relational learning framework that embeds leukocyte representations into a similarity graph and applies inductive classification to leverage contextual neighborhood information.
3. Through extensive experiments on the AML-Cytomorphology-LMU dataset, our method achieves ~91% accuracy and improved macro-F1 scores, particularly enhancing recall for rare subtypes.
4. We show that latent augmentation improves minority-class sensitivity and reduces cross-validation variability, yielding more robust and clinically deployable models.

Real-Time Neural Network Framework for Sub-5 ms Anomaly Detection in NIDS

Network Intrusion Detection Systems( NIDS) are critical for securing modern network infrastructures; still, being deep knowledge- predicated approaches constantly suffer from high conclusion quiescence, limiting their connection in real time and edge computing surroundings. This paper presents an optimized Convolutional Neural Network( CNN)- predicated NIDS frame designed to achievesub- 5 ms anomaly discovery while maintaining high discovery delicacy. The proposed ar chitecture employs depthwise separable complications to reduce computational exodus, INT8 quantization for conclusion accel eration, and channel community to overlap packet internee, preprocessing, and type tasks. Evaluation was conducted on three standard datasets — CICIDS2017, UNSW- NB15, and NSL- KDD — demonstrating delicacy situations above 96 and increment exceeding 15,000 packets per second on an NVIDIA Jetson Xavier NXedge device. The frame achieved up to 60 × hastily conclusion than LSTM- predicated births and demonstrated strong zero day discovery performance, achieving 94.8 recall on unseen Botnet attacks. These results validate the proposed system’s capability to deliver both high- speed and high- delicacy discov ery, making it suitable for deployment in quiescence-sensitive, resource- constrained surroundings. future advancements will explore confederated knowledge for distributed model updates, bettered inimical robustness, and adaptive explainability features to ensure secure decision- making in dynamic network surrounds.

FashFit: Personalized Outfit Illusion Using Face and Garment Matching

The significant research contribution of this work lies in the development of FashFit, a personalized outfit recommendation and visualization system that goes beyond typical e-commerce virtual try-on tools. Instead of focusing on sales, it acts as a personal fashion advisor by analyzing user-specific features such as face shape, body size, skin tone, and fabric preferences, and then matching them with garment attributes like color, silhouette, and fabric type. Using computer vision and image processing, the system generates a virtual illusion of the user wearing the suggested outfits, allowing individuals to visualize suitability before making clothing choices. This contributes to fashion technology by filling the gap between simple visualization and true personalization, helping users make confident outfit decisions and reducing the trial-and-error process in clothing selection.

Conference Site: IEEE i-COSTE 2025