IEEE i-COSTE 2025 – Page 2 – NPS Australia Submission System

XAI-Enhanced Hybrid Models for Effective Chronic Kidney Disease Prediction

Our key contribution lies in developing a hybrid predictive framework that combines machine learning, deep learning, and ensemble methods for highly accurate and interpretable Chronic Kidney Disease (CKD) detection. We addressed data imbalance through SMOTE, SMOTE-Tomek, and SMOTE-ENN, followed by SHAP-based feature selection via XGBoost, where top-ranked features were aggregated across all resampling strategies. Complementing this, statistical analysis with SPSS reinforced the identification of clinically significant features. Multiple machine learning classifiers, ensemble approaches, and advanced deep learning models—including Attention Autoencoder with XGBoost, TabNet, TabPFN, LightCNN, MLP, and DeepCrossNet—were systematically evaluated. DeepCrossNet achieved 97.38% accuracy, while the stacking ensemble attained 97.50% and Random Forest reached 97.71%. Furthermore, SHAP and LIME explanations emphasized GFR and serum creatinine as critical predictors, enhancing clinical trust. Visualization with t-SNE and UMAP confirmed class separability and detected ambiguous cases. Together, these contributions highlight the novelty of integrating SHAP-aggregated features, advanced resampling, and hybrid ML-DL ensembles to advance both accuracy and interpretability in CKD prediction.

Sustainable Development in Transition: Empirical Insights on Green Economy and Carbon Emissions from India

This study contributes by providing empirical evidence on the link between the green economy and carbon emissions in India, a country undergoing rapid economic transition. It fills a critical gap by integrating sustainability and emission dynamics in a single framework, offering insights that extend existing literature. The findings have strong policy relevance, supporting strategies for low-carbon growth and sustainable development pathways in emerging economies.

Women-led Social Entrepreneurship in Fostering SDGs: Myth or Reality? A Phenomenological Perspective

Abstract—The study explores women’s social entrepreneurial approaches that align with the United Nations’ Sustainable Development Goals, reflecting overall sustainable development in the context of Bangladesh, targeting achievement by 2030. The study adopted a phenomenological approach to reveal the lived experiences of six women social entrepreneurs’ journey towards social entrepreneurship. Protocol analysis, conducted through in-depth interviews, was analyzed through a phenomenological lens. Following Husserl’s epoche (bracketing) and eidetic reduction, it captures the essence (noema) and the act of experiencing (noesis) of underlying entrepreneurial approaches towards the SDGs. The findings of the study demonstrate nine key approaches and their alignment with the SDGs and sustainable development of Bangladesh. By examining lived experiences and interpreting realities, it seeks to provide policymakers with actionable insights to facilitate supportive arrangements for schemes, mentorship, and networking. Additionally, it aims to empower educators to serve findings as learning modules and to engage both national and international development organizations in actively involving social entrepreneurs as key partners in sustainability efforts. Future research endeavors will shed light on men and other supportive organizations, addressing the SDGs within the context of different countries.

Preventing Local Minima in Lyapunov-Based Control Scheme through Iterative Repulsive Potential Augmentation

This study introduces the Iterative Repulsive Potential Augmentation (IRPA), a motionplanning framework designed to overcome the local minima problem in Artificial Potential Fields. IRPA builds upon the Lyapunov-based Control Scheme (LbCS) by iteratively increasing the repulsive potential in regions where the robot becomes trapped.
This augmentation is repeated across successive iterations until the robot can successfully escape the local minimum and reach its goal. The approach provides a computationally efficient solution to the local minima problem while preserving the inherent efficiency and scalability advantages of the LbCS.

Comparative Analysis of Supervised Machine Learning Anomaly Classification on Active Power in Smart Grids.

The primary contributions of this study are as follows. First, it presents a focused investigation on active power–based anomaly classification in smart grid physical infrastructure by utilizing the physical attack subset of a cyber–physical dataset. Second, it conducts a systematic comparative evaluation of multiple supervised classifiers under physical attack scenarios, employing rigorous multi-class performance metrics such as precision, recall, F1-score, and confusion matrices. Third, it applies Random Forest–based feature-importance methods, including Gini impurity and permutation importance, to identify influential features, reduce dimensionality, and enhance interpretability. Finally, the study advances toward an explainability-driven and practically deployable anomaly classification and detection framework for smart grid monitoring, addressing limitations of prior works constrained by dataset scope, conceptual focus, or narrow evaluation metrics.

Signet: A Low-Cost Prosthetic Glove For Deaf And Mute Patients

Speech is the key to human existence. These are
not easily communicable in the speech-and-hearing-impaired
people. If you consider traditional alternatives like sign
language or even writing in a notepad, the other person also
has to be familiar with those forms of communication rather
than there being a real-time disconnect. In this paper, a low
cost touch-sensor-based assistive communication glove is
proposed, which differentiates between seven specific touches
across fingers and converts them into text messages displayed
on the LCD screen. The system incorporates capacitive touch
sensors, Arduino Nano microcontroller for computation, and
an LCD with an I2C module for display. The proposed design
is intended to provide a portable, low-cost, and user-friendly
solution improving the independence of the deaf and mute
society. Experimental results indicate that the proposed
system is efficient, easy, and ideal for routine daily
communication.

An Efficient Attention-based Deep Learning Model for Masked and Unmasked Face Recognition

• Proposed the Custom Aug-CNN-5, a CNN-based architecture optimized for recognizing masked and partially occluded faces.
• Integrated attention mechanisms (CBAM) and cosine annealing to improve learning efficiency and enhance discriminative power.
• Conducted a comprehensive evaluation on the larger customized VGGFace2 dataset, demonstrating high accuracy, generalizability, and fairness across varied conditions.

Modeling Stop–and–Go Wave Dissipation under Partial CAV Penetration using a Multi-Class CTM

This paper investigates the dissipation of stop-and
go traffic waves under partial penetration of connected and
autonomous vehicles (CAVs) using a multi-class cell transmission
model. A controlled numerical study was conducted along a
corridor with a localized bottleneck, with experiments varying
CAV penetration levels and smoothing intensities. The validated
results show that even moderate penetration of CAVs substan
tially improves traffic performance: unstable density patterns are
suppressed, shockwave propagation is attenuated, and total delay
decreases monotonically with increasing adoption. Fundamental
diagrams with triangular envelopes confirmed capacity gains
consistent with reduced effective headways, while regression of
the jam front demonstrated qualitative agreement between mea
sured and theoretical shockwave speeds. These findings indicate
that CAVs can deliver network-level benefits well before full
market penetration, supporting their role as a viable strategy
for congestion mitigation. The study contributes both a method
ological framework for evaluating CAV impacts in macroscopic
models and empirical insights to guide deployment and traffic
management policy.

Leveraging Large Language Models to Investigate the Decoy Effect in Route Choice Behavior

Route choice behavior is a cornerstone of transport
research, traditionally modeled under the assumption that trav
elers act as rational agents who maximize utility by trading off
time, cost, and reliability. However, behavioral economics shows
that real-world decision making often departs from rationality
due to systematic biases. One such bias is the decoy effect, which
occurs when the introduction of a dominated alternative increases
the attractiveness of another option. While widely studied in
consumer contexts, its role in transport decision making remains
underexplored. This paper presents a novel methodology that
leverages large language models (LLMs) to investigate the decoy
effect in route choice. Using ChatGPT-4o mini, we generated
textual framings of a dominated route alternative and employed
the model as a synthetic respondent to simulate route selections.
Four experimental conditions were tested: baseline (two routes),
decoy with neutral framing, decoy with positive framing, and
decoy with negative framing. Results demonstrate that the decoy
increased the relative attractiveness of the premium route, partic
ularly under positive framing, while negative framing attenuated
the effect. Synthetic responses closely aligned with predictions
from a multinomial logit model, confirming consistency with
behavioral theory. The findings illustrate that LLMs can serve
as both framing generators and behavioral simulators, offering
a rapid testbed for prototyping hypotheses in transport research.
This work establishes proof-of-concept evidence that LLMs can
complement traditional behavioral models, opening pathways
for future integration of artificial intelligence into the study of
systematic biases in mobility decisions.

Explainable Flight State Recognition: Toward Intelligent Pilot Training Systems

This study presents a novel framework that integrates explainable artificial intelligence (XAI) to enhance classification of flight states of an aircraft in the context of pilot training. By combining probabilistic outputs from Artificial Intelligence (AI) models, the approach identifies high-confidence and low-confidence predictions. Further, Shapley Additive Explanations (SHAP) technique is employed to uncover feature-level insights.
The main contribution of the study is to know how model identifies the ambiguous flight states and to know the most influential features in classifying these states. Moreover, to demonstrate how interpretable AI can support the development of intelligent, data-driven pilot training systems that are grounded in real flight data and model transparency.

Conference Site: IEEE i-COSTE 2025