Emotion recognition through electroencephalography (EEG) is an important part of passive brain–computer interfaces (BCIs), yet its practical implementation is often hindered by the challenges of limited data availability. Modern deep learning architectures, while powerful, typically require extensive datasets, leaving a significant research gap regarding the comparative robustness of shallow versus deep learning models in small-data conditions. In particular, existing literature has not systematically examined how model performance changes as the available training data is progressively reduced. This study addresses the gap by investigating the optimisation of learning architectures for EEG-based emotion recognition under scarce training data. Two novel evaluation frameworks are proposed: the variable test data scheme, which dynamically adjusts train–test ratios to assess generalisation, and the static test data scheme, which isolates the impact of training data reduction. Ten widely used shallow (SVM, DT, RF, LR, KNN) and deep (MLP, CNN, LSTM , GRU, KAN) models were evaluated across three benchmark datasets, namely the EEG Brainwave, GAMEEMO, and LUMED datasets. The models are evaluated at training sizes ranging from 95% down to 5% of the available data. Key findings indicate a shallow advantage, where Random Forest and K-Nearest Neighbour consistently outperformed deep models in low-data conditions, maintaining accuracies above 90% even at 5% training capacity. Conversely, Recurrent Neural Networks exhibited extreme instability and performance collapse as data decreased. However, Convolutional Neural Networks demonstrated competitive resilience when training data exceeded 50%. These results provide the first systematic, data volume-based evidence of model performance with the available training data. This has important implications for BCI design, suggesting that model selection must be strategically aligned with data volume, specifically labelled data, rather than being driven solely by model complexity. This research contributes a foundational guide for developing efficient emotion recognition systems. Code is publicly available at https://github.com/Nimishan/small-labelled-data.
Large language models (LLMs) are increasingly used in healthcare chatbots; however, they are prone to hallucination and lack reliable domain knowledge without external grounding. This paper presents an empirical study of knowledge modelling approaches for healthcare chatbots, using a corpus of Australian post-stroke driving guidelines. We evaluate the performance of raw LLMs, retrieval-augmented generation (RAG), and knowledge-graph-based retrieval (GraphRAG). Experimental results show that knowledge augmentation improves accuracy, with GraphRAG achieving the best performance among retrieval-based methods. However, our findings also indicate that neither RAG nor GraphRAG provides a complete solution, highlighting the need for more robust and hybrid strategies to support reliable healthcare chatbot systems.
Parkinson’s disease is a progressive neurological movement disorder which causes significant motor and non-motor symptoms, including mobility dysfunction. People with Parkinson’s (PwP) often experience symptom fluctuations relating to the time of day, medication intake time and other factors including sleep quality. Accurate detection and quantification of symptom fluctuations and longitudinal progression in Parkinson’s are crucial for testing new medications in clinical trials and informing treatment strategies for PwP. The current gold-standard snapshot clinical scales struggle to capture the full profile of fluctuating symptoms in everyday home life. Digital devices such as inertial measurement units (IMU) have the potential to objectively assess movement disorders like Parkinson’s in a free living environment continuously. In this work, we present a wrist-worn wearable device with an IMU and a pressure sensor, and infrastructure that is designed based on specific clinical requirements for tracking mobility-related symptoms of PwP in a home environment. The device is characterised in terms of sensors performance and the current consumption. Characterisation shows sensor bias, noise, and drift within published reference ranges for motion analysis, and a measured battery life of 57 h under continuous at-home operation. In addition, a free-living overnight pilot test was conducted to examine the overall performance of the designed system in terms of its robustness, redundancy and signal coverage. It establishes technical feasibility rather than clinical validity. The system maintained reliable data capture throughout, the proposed system achieves an overall reception rate of 82.53% with a median inter-arrival package jitter of 131 ms.
Breast cancer remains one of the leading causes of death among women, and mammography plays a critical role in early detection. However, analyzing these images is often complicated by low contrast, visual noise, and subtle tissue patterns that are difficult to interpret. To address these challenges, we propose IR-EANetSh, a deep learning and explainable AI based framework that improves classification accuracy while offering model transparency for clinical use.In this work, we propose IR-EANetSh (Inception-ResNet channel attention mechanism and SHAP), a deep learning framework that integrates multi-stage image enhancement, an Inception-ResNetV2 backbone that has been pre-trained, a lightweight channel attention mechanism, and SHAP-based explainability to enhance classification and interpretability. The suggested method not only improves the representation of features but also manages the imbalance between classes effectively and also gives the visual explanations that may be used to aid the clinical decision making.We evaluate our approach on three benchmark datasets: mini-DDSM, DDSM, and CBIS-DDSM. IR-EANetSh achieves up to 98.27% accuracy, with strong recall and AUC values, outperforming existing approaches. These results underscore its potential as a dependable and explainable aid for breast cancer screening.
