Machine Learning on Dynamic Functional Connectivity: Promise, Pitfalls, and Interpretations.

Exploring LoRA variants to adapt whisper models for robust recognition of children’s speech

Modern myoelectric prosthetic hands continue to face reliability challenges due to the non-stationarities of surface electromyography (sEMG) signals, which are highly sensitive to limb positions, electrode shifts, and grasp forces. While data abundance is a common strategy to mitigate these issues, it significantly increases users’ training burden. Hand gesture recognition, which maps the spatiotemporal patterns of muscle activation to key hand gestures for daily activities, remains the standard control strategy for advanced prosthetic hands. However, while temporal information can be reliably extracted from the sEMG signals, spatial information is highly dependent on electrode placements, which vary significantly between subjects. Previous research in myoelectric hand gesture transfer learning has primarily focused on transferring either spatial information or combined spatiotemporal information, leaving the transfer of temporal information alone largely unexplored. We propose a temporal-spectral cross-subject transfer learning framework using multi-stream convolutional neural networks (CNNs), where each stream processes only a single sEMG channel. Evaluated on the Transradial Amputee sEMG Multi-Contraction Forces Dataset, our framework has achieved training accuracy of 92.73% for medium contraction force and generalization accuracy of 74.53%, outperforming several models for sEMG hand gesture recognition. It also significantly improves recognition accuracy compared to the self-training baseline with the same architecture (repeated measures t-test p ≈ 0.032 ). By excluding spatial knowledge transfer, our approach maintains high robustness even under extreme cases of channel mismatch between source and target subjects. Moreover, this study highlights the importance of CNN architecture design, and spatially agnostic feature extraction for advancing myoelectric control systems.

MMCGR: A multimodal cascaded gated fusion RWKV-based model for emotion recognition in conversations

Bayesian-optimized filtering and hybrid ViT-LSTM model for knee joint abnormality recognition using sEMG

Optimizing 2D bridge engineering drawing digitization: A comparative study of text recognition tools and development of lightweight post-recognition structured information extraction methods

Artificial intelligence at the frontlines: Emerging infectious and parasitic diseases in the digital era.

InGesture: An eight-class inertial sensor dataset for fluid intake and hand-gesture recognition.

An adaptive recognition model for tool wear state across multiple datasets

PADI-Location-AR-EN: A normalized Arabic-English spatial entity dataset for epidemiological surveillance.

The classification of Bangla characters and digits is a fundamental component of Natural Language Processing (NLP) and computer vision applications. However, despite advancements in handwritten character recognition (HCR) for other languages, recognizing Bangla script remains a formidable challenge, primarily due to its extensive character set, intricate compound forms, and significant stylistic variations. While existing datasets have significantly advanced the field of Bangla HCR, they frequently overlook the complexity and variability of primary-level learners' handwriting. This paper introduces the first extensive Bangla Primary Script 2025 (BPS2025) dataset, a novel, balanced and comprehensive, demographically oriented collection of isolated characters and numerals specially focused young primary school students. The dataset was selectively curated from 500 students, aged 7 to 12 and in grades 2 to 5, across four districts in Bangladesh. It comprises 24,420 raw images across 60 balanced classes, including 50 basic characters and 10 digits, which followed five stage pre-processing pipelines to process the final dataset. The dataset addresses a significant gap in existing benchmarks, which is expected to support future research and real-world educational applications by facilitating the development of robust and accurate recognition models for Bangla script.

A silent speech interface with machine learning recognition model using microneedle array electrodes and polymer-based strain sensors

An improved activation function for the recognition of knee osteoarthritis severity.

The increasing number of data breaches and unauthorized data exposures has raised serious concerns regarding personal data security in the digital era. Sensitive user information, including email credentials, personal identifiers, and biometric data, is frequently exposed due to insecure systems, large-scale data breaches, and improper data handling practices. Such exposures often remain unnoticed by users, leading to significant privacy risks and potential misuse of personal information. This paper proposes an AI-based Personal Data Leak Detection System designed to identify and analyze potential data leaks associated with individual users. The system integrates intelligent data processing techniques, biometric analysis, and pattern recognition methods to detect exposed information across structured and unstructured datasets. Facial recognition models are utilized to identify potential biometric leaks, while additional analytical modules evaluate user-related data exposure and associated risks. The system is implemented using a Flutter-based frontend and a FastAPI backend, with Firebase authentication ensuring secure and isolated user access. Experimental observations indicate that the proposed system effectively detects user-related data leaks and enhances user awareness regarding privacy risks. The results demonstrate the potential of AI-driven approaches in strengthening digital privacy and improving cybersecurity mechanisms.

The ever-expanding biomedical literature necessitates an efficient and robust mining platform, with the foundational step being a reliable Biomedical Named Entity Recognition (BioNER) system. Existing approaches, such as multi-task and collaborative learning, have attempted to address dataset heterogeneity but often rely on complex architectures with task-specific layers, limiting scalability. A key research gap is the development of a unified model that optimises across multiple datasets without sacrificing performance or introducing architectural complexity. In this study, we propose a novel Loss-Masking Optimisation framework for BioNER models that enables multi-dataset training via a dataset-aware masking strategy. This approach extends the standard BERT-based NER pipeline by introducing a tag-masking array that nullifies logits for tags absent in the originating dataset, thereby reducing cross-dataset interference. Using this methodology, we trained a single BioNER model across all 16 biomedical NER datasets, achieving higher precision and overall F1 scores than conventional multi-dataset training. While some datasets showed performance gains, others stayed near baseline, and a few declined, underscoring the nuanced impact of dataset interactions. To the best of our knowledge, this is among the first studies to apply a dataset-aware loss-masking mechanism to unified multi-dataset BioNER training, offering a scalable alternative to multi-task architectures.

Urgent diagnostic imaging and consent: A comparative review of Australian law and policy.

The accurate identification and discrimination of multiple endogenous bioactive ingredients in medicinal and edible honeysuckle are of considerable importance for food quality control and human health. Herein, the integration of a high-entropy nanozyme-based "chemical tongue" with a machine learning algorithm is proposed for the first time to recognize seven phenolic components of honeysuckle. First, we have rationally designed a quinary metal element (RuNiFeMnCo) high-entropy layered double hydroxide (HE-LDH) nanozyme with peroxidase-like (POD-like) performance. The specific activity of quinary HE-LDH with up to 34.85 U/mg is significantly higher than binary NiFe-LDH, ternary NiFeMn-LDH, and quaternary NiFeMnCo-LDH, which is approximately 50 times that of binary NiFe-LDH. The enormously boosted catalytic performance benefits from both metal Ru as the main active site and the "cocktail effect" from multiple transition elements' synergistic interaction. Second, the HE-LDH nanozyme can catalyze three substrates to their corresponding color-oxidized products as three channels of "chemical tongue". The endogenous phenolic compounds of honeysuckle can inhibit the chromogenic process to different extents due to their discrepant antioxidant properties, offering distinct "fingerprints" for each species. Furthermore, the fingerprint decryption displays the excellent recognition of seven phenolic compounds at concentrations as low as 10 nM and the wide linear range of 0.01-50 μM for individual species with the assistance of the machine learning algorithm. Finally, the robust recognition model with simultaneous concentration-independent and matrix-independent performance is first constructed to discriminate seven phenolic compounds with concentrations ranging from 0.01 to 10 μM spiked into four actual honeysuckle samples. The identification accuracy of phenolic compounds against the colorimetric sensor array is increased from 58.84% by linear discrimination analysis (LDA) to 98.04% by the random forest (RF) algorithm. More importantly, this strategy provides a new avenue of high-entropy nanozyme as a fascinating platform for the configuration of the sensor array.

Piezoionic sensors perceive the physical world based on polymer ionogels with advantages of flexibility, lightweight, and high sensitivity, and are suitable for physical signal extraction and virtual space construction in artificial intelligence. However, the electrode-electrolyte interface of conventional sensors presents mechanical modulus mismatch, which is prone to interface cracking under external strain and affects cyclic stability. Here, we engineer a gradient sensor interface based on graphene and ionogel that alleviates modulus mismatch by eliminating the interface of the electrode and electrolyte. The piezoionic sensor displays superior cyclic stability in an air environment with signal retention as high as 97% over 4000 bending cycles. It also delivers millisecond-level rapid response and sensitive strain perception in a complex environment for detecting diverse human joint movements. Meanwhile, we integrate the flexible sensors with a large language model for accurate path recognition and realize feature extraction and correlation analysis of the large number of sensor signals. Our study provides an insight into the interface optimization of electrochemical devices and will shed light on the development of flexible sensors in artificial intelligence.

Visual assessment of membrane motion is essential for managing EXCOR VAD, but accuracy depends on observer experience. We evaluated a high-speed image AI model to support healthcare providers. Patients on EXCOR Pediatric admitted to the University of Tokyo Hospital (May 2022-May 2024) were included. Membrane images were obtained from patients and a manually filled pump at bench. An image recognition model was trained to estimate membrane position. Experienced physicians (N = 11) and inexperienced physicians (N = 11) assessed pump status in a sample dataset (N = 12) with and without AI assistance. A total of 142 movies from five patients were collected (98 training, 45 validation), plus 1100 bench images for training. Model accuracy was 0.91, with AUROCs of 0.99 ("fill") and 0.96 ("empty"). Among experienced physicians, accuracy significantly improved with AI assistance from 0.83 (0.67-0.88) to 0.92 (0.92-1.0) (median (IQR); p = 0.016). Among inexperienced physicians, accuracy also significantly improved with AI assistance from 0.67 (0.5-0.75) to 0.83 (0.75-0.92) (p = 0.049). A high-speed image AI can facilitate the visual assessment of EXCOR VAD by healthcare providers.

Dialects, in their remarkable diversity, serve as repositories of rich historical and cultural heritage. The Guanzhong dialect, native to the Guanzhong region of Shaanxi, is characterized by distinct phonetics, vocabulary, and cultural nuances. With the rapid advancement of speech emotion recognition (SER) technology, research into emotion analysis for dialectal speech has gained traction. However, this field remains challenged by limited dialect-specific datasets, an inadequate representation of emotional features in speech, and subpar model performance. The core objective of research is to achieve an accurate recognition of emotions in dialect environments, overcome the limitations of scarce dialect data, and meet the emotional interaction needs of specific groups of people. This work presents a dual contribution: first, there is the construction of a high-fidelity Guanzhong dialect SER dataset through systematic data collection, noise reduction preprocessing, and the annotation of four emotional categories (joy, sadness, anger, neutral), which establishes a robust foundation for subsequent research. Second, there is the development of FCAL-Net, which is a novel framework integrating multi-dimensional feature fusion. This approach leverages the CAM (convolutional attention module) to enhance CNN's local temporal feature extraction, couples it with Bi-LSTM for global contextual modeling, and demonstrates superior performance. The experimental results validate an accuracy of 77.67% in emotion classification, outperforming traditional CNN (12.31% improvement), Bi-LSTM (7.23%), Wav2vec 2.0 (8.08%), and Conformer (4.61%) baselines-highlighting its efficacy in addressing core challenges in dialectal SER.