This study explores the acceptance and perceived effectiveness of AI-powered Augmentative and Alternative Communication (ACC) systems among individuals with speech impairments. Despite increasing advances in inclusive technology, limited research has addressed how people with communication disabilities evaluate such tools regarding usability, privacy, and behavioral intention. An interdisciplinary project involving information technology, rehabilitation medicine, and social work was conducted with 39 participants in Taiwan. Each participant completed pre-assessments, voice recordings to build personal language models, and post-assessments using the Technology Acceptance Model (TAM) framework. Results showed high overall satisfaction and willingness to use the system, particularly among younger and male participants. Although no statistically significant differences in TAM variables were found between subprojects or education levels, male users demonstrated higher behavioral intention, and younger participants exhibited more favorable attitudes toward AI tools. Quality of life (QOL) scores revealed the highest satisfaction in the environmental domain and the lowest in health and life satisfaction. These findings provide a foundational understanding of user experience and acceptance, guiding the future design and deployment of AI-based ACC technologies.
Embedded systems often accumulate complex, monolithic code, making identifying and reusing functionality across projects difficult. This paper presents an automated approach for extracting reusable software features from embedded C code using six pre-trained large language models (CodeBERT, GraphCodeBERT, CodeT5, CodeGPT-small, PLBART, and UniXcoder). The system embeds code-level features (e.g., functions and global variables) and textual artifacts (e.g. comments, control-flow conditions, and README entries) into a shared semantic vector space. Cosine similarity matches each feature with its most relevant textual artifact, and a coherence score is assigned to capture the strongest alignment across models. High-confidence matches are then clustered using unsupervised learning to reveal logical subsystems within the codebase. We evaluate the method on a heating system controller by manually curating 14 feature-requirement description pairs. Each feature is a functional unit selected from the code, and each description summarizes its intended behavior. These pairs serve as the benchmark for assessing extraction performance. Our method successfully traced features with 86% recall, using conservative thresholding that accepts only semantically strong matches. Clustering based on coherence scores produced interpretable groupings corresponding to logical subsystems such as heating control, safety monitoring, and user interface. Compared to single-model baselines (with recall as low as 22%), the multimodel ensemble method significantly improved coverage by leveraging the semantic diversity of different LLM models. These results demonstrate the feasibility of combining diverse pretrained embeddings with multiple textual sources to support feature extraction and modular reuse in embedded software.
Scientific literature contains a vast quantity of validated but fragmented knowledge. Literature-based discovery (LBD) aims to connect this dispersed information to generate new, valuable hypotheses. This paper presents a method for generating new scientific hypotheses by performing symbolic inference over a network of semantic predications-machine-readable scientific claims that are enriched with provenance and contextual metadata. The method constructs inference chains by linking multiple semantic predications from independently published sources. It accounts for specific properties of the predications to improve reliability, including a status of knowledge metric that we define and compute. The approach favors more robust and informative connections when generating hypotheses. To evaluate the quality of the hypotheses produced, we apply a temporal slicing methodology that assesses whether predicted relationships correspond to scientific claims published in the future. This approach enables direct comparison with established LBD methods. In a biomedical use case, our system achieves a Mean Average Precision (MAP) of 0.27 and a Precision@15 of 0.45, indicating the feasibility of interpretable, reusable scientific discovery through symbolic reasoning over structured knowledge.
Large Language Models (LLMs) are increasingly used for decision support in high-stakes domains such as healthcare and law, yet their deployment is limited by output instability and lack of transparency, especially when labeled data is scarce in real-world scenarios. To address these challenges, we propose Med-TAMARA, a trust-aware, reflective multi-agent framework for classifying medical, ethical, and legal risks in patient-doctor conversations. Our system combines self-consistency filtering and few-shot chain-of-thought (CoT) prompting to eliminate unstable predictions, agentic peer review to refine model judgments, trust-weighted aggregation to account for agent reliability, and a confidence scoring mechanism to flag uncertain cases for human oversight. We conducted a proof-of-concept evaluation on the English subset of the NTCIR-18 MedNLP Chat dataset, suggesting that Med-TAMARA may outperform strong baselines in preliminary experiments across standard evaluation metrics. Although our dataset is limited in size, these initial results indicate the potential of the framework as a foundation for a more reliable LLM-assisted risk assessment. The source code for MedTAMARA will be released publicly after further optimization and testing.
Accurate damage assessment following a natural disaster is crucial for guiding effective response and recovery efforts. In this study, we present a deep learning framework that employs multi-temporal fusion of pre- and post-disaster remote sensing imagery, using a shared-weight ResNet50 architecture as the model backbone. To improve both predictive performance and physical realism, we incorporated a physics-informed loss function that integrates texture energy derived from a GrayLevel Co-Occurrence Matrix (GLCM). This additional loss term penalizes outputs that fail to exhibit expected rubble-like texture patterns, thereby encouraging more physically consistent predictions. Model performance was assessed using several evaluation metrics, including confusion matrices, average accuracy, F1 scores, xView2 scores, and multi-class ROC curves. Experimental results show that the inclusion of the GLCM-based physicsinformed loss resulted in measurable improvements in both damage detection and classification accuracy-achieving a 7% gain in average accuracy and a 8.57% increase in xView2 score compared to baseline models without the physics constraint.
Music recommendation is an important aspect for music streaming service users. Studies have proven that music has effects on human emotions, cognition and health. Vice versa, numerous studies have shown that health factors such as emotional state and physical fatigue affect the music preferences of users. In spite of that, most music recommendation systems offered by large companies focus on music trends and listening history, often neglecting the influence of current emotional or health conditions of the individual. Health data from wearable devices can provide valuable insights into music preferences. This study presents a web application that integrates Spotify music data with Fitbit health metrics to train a machine learning-based scoring model. The system predicts music success scores by analyzing user behavior and health data, adapting scoring model accordingly. A supervised user study assesses the system's usability and effectiveness in adapting to user preferences through realworld interactions. Overall, this study highlights the potential of integrating health data into music recommendation systems to enhance personalization and user experience.
The rise of generative Artificial Intelligence (AI) has created the possibility of presenting novel recipes, i.e., recipes that do not exactly match any known recipe and this has led to the creation of AI-based recipe recommendation systems. AI-based recipe recommendation has the possibility of accommodating a variety of preferences - including a person's current health (e.g., diabetes), health goals (e.g., weight loss), taste preferences, cultural or ethical needs (e.g., vegan diet). However, unlike recipes recommended or created by a human dietitian, recipes created by generative AI do not guarantee accuracy, i.e., the generated recipe may not meet the requirements specified by the user. This work quantitatively evaluates how closely recipes generated by OpenAI's GPT4 large language models, created in response to specific prompts, match known recipes in a collection of human-curated recipes. The prompts also include requests for a health condition, diabetes. The recipes are from the largest online community of home cooks sharing recipes (www.allrecipes.com) and the Mayo Clinic's collection of diabetes meal plan recipes. Recipes from these sources are assumed to be authoritative and thus are used as ground truth for this evaluation. Quantitative evaluation using NLP techniques (Named Entity Recognition (NER) to extract each ingredient from the recipes and cosine similarity metrics) enable computing the quality of the AI results along a continuum. Our results show that the ingredients list in the AI-generated recipe matches 67-88% with the ingredients in the equivalent recipe in the ground truth database. The corresponding cooking directions match 64-86%. Ingredients in recipes generated by AI for diabetics match those in known recipes in our ground truth datasets at widely varying levels: between 26-83%. The quantitative evaluation is used to inform the development of a web-based personalized recipe recommendation system for diabetics that uses OpenAI's GPT4 model for recipe generation.
Interdisciplinary paradigms like autonomous vehicles (AVs) have enabled us to lead smarter lives. However, these systems are under threat from adversarial attacks that aim to compromise the machine learning (ML) perception systems focused on traffic sign recognition (TSR), which can put public and driver safety at risk. In this paper, we develop a two-stage neural network architecture to increase resiliency in ML-enabled AV TSR against adversarial attacks. The methodology proposes using both a Manipulation Detection Model, which examines the traffic sign image and determines whether it has been altered, and a Multi-classification Model, which performs multisign classification on predicted "normal" traffic sign images. The results indicate that the use of the two-stage model provides better performance than single CNN setups with P-mu = 0.88, R-mu = 0.87, and F-mu = 0.85. Additionally, this setup undergoes less deterioration than single setups with P-mu, R-mu, and F-mu deterioration by -0.03,-0.01, and -0.03, respectively.
This paper presents a 3D Convolutional Neural Network (CNN) that can classify stages of Alzheimer's disease, which is a progressive neurodegenerative disorder marked by memory loss and cognitive decline, in both binary and multi-class settings with great accuracy. The Magnetic Resonance Imaging (MRI) images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset have been used. Even though many studies report lower accuracy with MRI, our approach demonstrates that a carefully designed preprocessing pipeline and optimized hyperparameter tuning can unlock the true potential of MRI for AD classification outperforming state-of-the-art models. Our final model achieved an average accuracy of 94.37 % in the multiclass classification of subjects with Alzheimer's disease (AD), mild cognitive impairment (MCI) and cognitive normal (CN), 91. 79% accuracy in the binary classification of AD vs. CN cases, and 83. 02% accuracy in Early MCI cases vs. CN. In addition, an interpretability occlusion-based technique has been added to these models to highlight the brain regions that contribute most to the model's predictions. This revealed that the most influential regions were part of the extracted regions of interest which are areas known to be critical in AD.
The integration of drone technology and Artificial Intelligence (AI) is revolutionizing ecological habitat monitoring. Deep learning-based semantic segmentation models have shown strong performance compared to traditional manual surveys [1], significantly improving the efficiency and accuracy of habitat mapping. However, these new AI approaches often overlook critical spatial relationships between adjacent habitat areas and do not incorporate domain-specific knowledge about river habitats. To address these limitations, we propose P-KESS, a Prior-Knowledge Enforced Semantic Segmentation Pipeline that combines prior knowledge with deep learning methods to enhance the drone imagery river habitat recognition. P-KESS leverages spatial contextual relationships and expert domain knowledge to correct false predictions made by deep learning models. We also present a dataset collected from rivers to demonstrate the efficacy of P-KESS, achieving a significant performance improvement from a baseline of 75.4% F1 score and 96.6% accuracy to 95.5% F1 score and 99.3% accuracy. These results highlight the superiority of leveraging domain knowledge over deep learning-only state-of-the-art semantic segmentation models.
This study presents a novel approach for generating high-quality class labels for severely imbalanced healthcare fraud data using unsupervised feature selection and labeling techniques. Our results show it is possible that supervised learners can be trained on labels generated with an unsupervised labeling technique and yield performance better than an unsupervised learner. Large unlabeled datasets pose significant challenges for machine learning, particularly in domains like healthcare fraud detection, where class imbalance is extreme; however, manual labeling can be time-consuming and expensive. We propose a labeling technique that combines SHapley Additive exPlanations (SHAP) for unsupervised feature ranking and selection with an autoencoder-based method for label generation. Experiments are conducted on a dataset derived from publicly available Medicare Part D data. The quality of generated labels is evaluated by training multiple supervised classifiers using both our unsupervised labeling method and an unsupervised baseline and comparing their performance. Results also demonstrate that SHAP-based feature selection consistently improves the quality of generated labels across different classifiers and datasets. This work provides an effective framework for addressing the challenges of unlabeled, severely imbalanced healthcare fraud data without requiring manual annotation. The unsupervised nature of the method preserves data privacy while enabling the application of supervised learning techniques to previously unusable datasets.
Breast cancer is one of the most common cancers found in women, affecting 1 in 8 Canadian women over their lifetime, according to the Public Health Agency of Canada. The early diagnosis of breast cancer can be extremely beneficial for reducing the spread of cancer cells and the risk of death. Machine learning (ML) systems can help healthcare professionals diagnose breast cancer with better accuracy by identifying underlying patterns in the images that radiologists might find difficult to detect early on. While deep learning (DL) models can obtain high accuracies, their black-box nature inhibits their ability to explain how they came to their conclusion, creating distrust between DL systems and radiologists. This research employs two CNN model architectures - AlexNet and Houby & Yassin's CNN - to classify recently obtained mammographic images from the KAU-BCMD dataset into BI-RAD categories (0-6). To resolve the problem of non-explainability, two variants of SHapley Additive exPlanations (SHAP), an explainability approach based on Shapley values, have been used to assess the role that each feature has in the prediction process.
This paper proposes a data-driven framework for quantifying disaster vulnerability using social media analytics, repurposing a previously collected Twitter dataset originally intended for evacuation behavior analysis. After refining the dataset to isolate signals of distress and need, a category based classification strategy is introduced in which thematic dictionaries guide the grouping of Tweets based on the semantic similarity of their embeddings. Focusing on Hurricane Dorian, a compound disaster during the COVID-19 pandemic characterized by high distress and negative sentiment, a weighted amplification factor is incorporated that prioritizes Tweet categories based on the immediacy of impact on human life, while normalizing by Tweet volume and population density. The resulting Media Impact Index (MII) is calculated at the Census Block Group (CBG) level for the United States. To demonstrate the cross-cultural flexibility of the pipeline, the same methodology is applied to Typhoon Hagibis in Japan, with a comparable vulnerability index generated at the district level. The findings suggest that the proposed framework can provide emergency management agencies with a scalable and adaptable tool for identifying and prioritizing vulnerable regions in diverse types of disasters and sociocultural contexts.
Graphical user interfaces (GUIs) are important software engineering artifacts that commonly connect end users with software apps to fulfill requirements. While previous work extracts metadata to build a GUI's search profile and exploits deep learning to caption a GUI, little is known about how to capture the GUI's essential functionality. We shorten the gap in this paper by presenting a novel approach that leverages fewshot prompting and meta-learning. In particular, we devise four attributes of a GUI in order to instrument the meta-learner. The meta-learner is then integrated with the few-shot learner through an iteration mechanism. We comprehensively evaluate our approach with a GUI description dataset and a GUI tracing dataset. The results show that Qwen is a viable implementation option for our approach, and that our approach outperforms the baseline prompting methods.
Change Impact Analysis (CIA) and requirement change propagation are increasingly critical in complex software systems, particularly in the context of information reuse and integration for data science. Currently, CIA is predominantly conducted manually with the support of domain experts. This approach is often time-consuming and susceptible to human error, especially in overlooking affected requirements. Leveraging data science and machine learning (ML) offers promising opportunities to automate and enhance CIA by enabling intelligent prediction and reuse of previously analyzed change patterns. These predictive models can assist experts by highlighting the likely affected requirements, improving accuracy, and reducing effort. This paper explores the current state of CIA practices, evaluates existing ML-based solutions aimed at enhancing the process, and examines their effectiveness, limitations, and opportunities for improvement. We conducted a systematic literature review (SLR) guided by three primary research questions, analyzing 43 studies published over the past two decades. Our findings highlight a lack of collaboration between academia and industry, as well as the limited credibility of existing tools and methods, largely due to the scarcity of comprehensive and diverse datasets.
Early detection of cognitive impairment is essential for timely intervention and treatment of conditions such as Alzheimer's disease. Supervised machine learning models require labeled data for training, which is often scarce, costly to obtain, and subject to diagnostic uncertainty. In this work, we propose CAE-IF, a fully unsupervised hybrid approach that combines Convolutional Autoencoders (CAE) for feature extraction with Isolation Forest (IF) for anomaly detection. Our method is evaluated on a real-world, imbalanced cognitive dataset derived from the Health and Retirement Study (HRS). CAE-IF consistently outperforms two baseline unsupervised models, Local Outlier Factor (LOF) and Isolation Forest, across key evaluation metrics. For AUPRC, CAE-IF achieves 0.3042 compared to 0.2164 for LOF and 0.2490 for IF. For F1-score, CAE-IF achieves 0.3380, outperforming LOF (0.2089) and IF (0.2640). For MCC, CAE-IF scores 0.1818, higher than LOF (0.0323) and IF (0.0892). These results demonstrate the superior performance of CAE-IF in detecting cognitive impairment under class imbalance. These results indicate that CAE-IF can serve as a useful tool for early screening of cognitive impairment using survey-based data.
Deepfakes are AI-generated manipulated videos that closely mimic real individuals, posing significant risks to privacy, digital security, and information authenticity. To address these threats, we propose DFD-V, a novel deepfake detection framework that combines spatial, frequency, and temporal analysis to capture diverse forgery artifacts. The model extracts per-frame features using an EfficientNet-B0 backbone for spatial cues and a custom CNN operating on 2D FFT spectra to detect frequency-domain anomalies. A Motion-Aware Frequency Attention (MAF) module further emphasizes temporal inconsistencies in spectral patterns. These dual features are fused via Selective Cross-Domain Attention Fusion (SCDAF), aligning complementary evidence. The fused representations are then modeled over time using a Bidirectional LSTM, enabling detection of frame-level inconsistencies such as flicker or jitter. We preprocess videos by extracting aligned facial regions and apply standard augmentations to improve robustness. The model is trained using binary cross-entropy loss, with optional adversarial domain adaptation to enhance cross-dataset generalization. Experimental results demonstrate that DFD-V achieves strong performance on multiple benchmarks, offering a reliable and interpretable framework for robust deepfake video detection.
Detecting objects in aerial images with complex backgrounds remains a significant challenge. This difficulty arises not only from the requirement of precise regression of bounding boxes but also from the presence of confusing backgrounds that share visual similarities with target objects. Existing object detectors primarily focus on improving convolutional representations for objects, with limited attention given to background proposal sampling. This paper introduces the Self-Rectification Faster R-CNN, a novel object detector designed to address the hard background proposal sampling problem for most of the object detector. The proposed method incorporates four key innovations: a balanced batch sampler, a balanced minibatch sampler, a hard background image sampler, and the SRNMS (Self-Rectification Non-Maximum Suppression) technique. Experiments demonstrate that the Self-Rectification Faster R-CNN achieves at least 7% improvement in F1-score compared to three state-of-the-art object detectors evaluated on our waterfowl detection dataset.
Legal document retrieval poses unique challenges due to the structural complexity and domain-specific language of statutes and case law. This paper presents MultiRAG, a multi-layered Retrieval-Augmented Generation (RAG) framework designed for intelligent legal information systems. The framework segments legal texts-specifically from Title 18 of the U.S. Code-into sections, pages, and chapters, and encodes them using two embedding models: a general-purpose model (Google-text-embedding-005) and a legal-domain model (VoyageLaw-2). A fuzzy logic-based query analysis module evaluates linguistic and semantic features of the input query to compute a legal intensity score, which guides dynamic weighting between the embedding outputs. Retrieved content across granularities is fused using a weight-aware deduplication and scoring strategy, and passed to a generative model (Gemini-2.0-flash) for structured and contextually faithful response generation. Evaluation on 2,000 queries from four user groups-ranging from legal experts to the general public-demonstrates a 97.75% success rate in generating relevant responses, with average response times under six seconds. These results show that MultiRAG offers a scalable, interpretable, and adaptable solution for legal document analysis and generation.
Zero shot time series forecasting is the challenge of forecasting future values of a time dependent sequence without having access to any historical data from the target series during model training. This setting differs from the traditional domain of time series forecasting, where models are typically trained using large volumes of historical data, from the same distribution. Zero shot time series forecasting models are designed to generalize to unseen time series by leverage their knowledge learned from other, similar series during training. This work proposes two architectures designed for zero shot time series forecasting: zSiFT and zSHiFT. Both architectures use transformer models arranged in a Siamese network configuration. The zSHiFT architecture differs from the zSiFT by the introduction of a hierarchical transformer component to the Siamese network. These architectures are evaluated on vehicular traffic data in California available from the Caltrans Performance Measurement System (PeMS). The models were trained with traffic flow data collected in one region of California and then are evaluated by forecasting traffic in other regions. Forecast accuracy was evaluated at different time horizons (4 to 48 hours). The zSiFT model achieves a Mean Absolute Error (MAE) that is 8.3% lower than the baseline LSTM with attention mechanism model. The zSiFT model achieves an MAE which is 6.6% lower than zSHiFT's MAE.
