Debasis Samanta

Imagined speech decoding from electroencephalography (EEG) is a challenging problem in brain-computer interfaces (BCI), due to the poor signal-to-noise ratio in non-invasive systems and the complex nature of language decoding. This study proposes an end-to-end deep learning model, SincTran, which internalizes the decomposition of EEG signals into their physiological frequency bands. The proposed model utilizes parallel sinc convolution layers, each tuned to a specific EEG band, aggregated via a softplus-scored, sum-normalized band attention mechanism, followed by spatio-temporal convolution layers and a transformer encoder. The encoded class tokens are passed to a multi-layer perceptron classification head to produce the final predictions. SincTran was evaluated on two benchmark datasets of imagined speech, achieving competitive classification performance across all tasks while maintaining favorable parameter and computational efficiency relative to comparable classifiers. On the 2020 BCI Competition dataset, the model reached an average accuracy of 65.75%, and on the Arizona State University dataset, it reached 79.35% accuracy on the long-word vs. short-word classification task. Additionally, SincTran supports direct interpretability through its architecture, enabling analysis of EEG band importance, temporal attention distribution, and class token geometry. Gradient-weighted class activation mapping was further employed for post-hoc analysis of channel importance. The findings demonstrate the viability of end-to-end pipelines that internalize spectral decomposition for imagined speech EEG classification.

Sleep EEG analysis increasingly requires expressive temporal encoders, yet EEG datasets face a structural scaling limit due to costly overnight acquisition and expert annotation. This makes data-driven scaling impractical and motivates leveraging sequence foundation models as reusable temporal backbones for sleep stage classification. This work presents NeuroWhisper, a convolution–attention encoder-based framework for automated sleep stage classification. First, each 30-second EEG epoch is transformed into a log-mel spectrogram to obtain a compact time–frequency representation. Second, a convolutional front-end extracts local spectro-temporal features from the spectrogram. Third, stacked self-attention layers derived from a speech encoder model capture temporal dependencies within the epoch. Finally, the resulting representations are temporally aggregated and mapped to sleep stage labels using a lightweight classification head. Under time-wise evaluation on PhysioNet Sleep-EDF, NeuroWhisper attains up to 86.4% accuracy and 0.86 macro-F1, exceeding general-purpose and domain-specific baselines under identical conditions. Under subject-exclusive evaluation, accuracy decreases with age, but NeuroWhisper consistently outperforms competing models, with the largest relative gains in elderly cohorts. Model scaling exhibits diminishing returns, as compact variants achieve subject-level performance comparable to larger models at substantially lower computational cost. These results indicate that convolution–attention architectural priors improve robustness in data-limited EEG sleep staging, while demographic and subject-level variability remain limiting factors.

Continuous electroencephalography (EEG) is routinely used in neurocritical care to monitor seizures and other harmful brain activity, including rhythmic and periodic patterns that are clinically significant. Although deep learning methods have achieved high accuracy in seizure detection, most existing approaches remain seizure-centric, rely on discrete-label supervision, and are primarily evaluated using accuracy-based metrics. A central limitation of current EEG modeling practice is the weak correspondence between learned representations and how EEG findings are interpreted and summarized in clinical workflows. Harmful EEG activity exhibits overlapping patterns, graded expert agreement, and temporal persistence, which are not well captured by classification objectives alone. This work proposes a multimodal EEG representation learning framework that integrates signal-domain modeling with structured clinical language supervision. First, raw EEG is transformed into a longitudinal bipolar montage and time-frequency representations. Second, dual transformer-based encoders model complementary temporal and frequency-centric dependencies and are fused using an adaptive gating mechanism. Third, EEG embeddings are aligned with structured expert consensus descriptions through a contrastive objective. Finally, an EEG-conditioned text reconstruction loss is introduced as a representation-level constraint alongside standard classification loss. Experimental evaluation using a controlled train-validation-test split achieves a six-class test accuracy of 0.9797. Ablation analyses show that removing contrastive alignment reduces cross-modal retrieval performance from Recall@10 of 0.3390 to 0.0045, despite minimal change in classification accuracy. These findings demonstrate that discriminative accuracy does not reliably reflect representation quality for clinically meaningful EEG modeling.

Clinical electroencephalography is routinely used to evaluate patients with diverse and often overlapping neurological conditions, yet interpretation remains manual, time-intensive, and variable across experts. While automated EEG analysis has been widely studied, most existing methods target isolated diagnostic problems, particularly seizure detection, and provide limited support for multi-disorder clinical screening. This study examines automated EEG-based classification across eleven clinically relevant neurological disorder categories, encompassing acute time-critical conditions, chronic neurocognitive and developmental disorders, and disorders with indirect or weak electrophysiological signatures. EEG recordings are processed using a standard longitudinal bipolar montage and represented through a multi-domain feature set capturing temporal statistics, spectral structure, signal complexity, and inter-channel relationships. Disorder-aware machine learning models are trained under severe class imbalance, with decision thresholds explicitly calibrated to prioritize diagnostic sensitivity. Evaluation on a large, heterogeneous clinical EEG dataset demonstrates that sensitivity-oriented modeling achieves recall exceeding 80% for the majority of disorder categories, with several low-prevalence conditions showing absolute recall gains of 15-30% after threshold calibration compared to default operating points. Feature importance analysis reveals physiologically plausible patterns consistent with established clinical EEG markers. These results establish realistic performance baselines for multi-disorder EEG classification and provide quantitative evidence that sensitivity-prioritized automated analysis can support scalable EEG screening and triage in real-world clinical settings.

In secure sensor-based applications, generating unique and secure credentials is crucial for ensuring trust and privacy. Traditionally, pseudo-random number generators have been used for this purpose. However, biometric data, especially fingerprint features, has emerged as a robust alternative. This article proposes a dynamic approach for generating a privacy-preserving unique identity from fingerprint sensor data. The method isolates a region of interest (ROI) from the fingerprint image and extracts two feature sets: minutiae-based data and texture-based information. These features are combined and optimized to produce a hybrid feature vector containing discriminative and relevant attributes. A unique identity (termed a key) is then dynamically generated from this vector, characterized by reliability, revocability, unlinkability, and irreversibility. Extensive evaluations using fingerprint images from sensors of varying quality and resolution have demonstrated the method’s robustness. The generated identities have been statistically validated using NIST and Diehard test suites, confirming strong adherence to randomness requirements. Also, comprehensive security analyses have shown resilience against different adversarial attacks. Notably, the approach avoids storing biometric data or generated identities, enhancing privacy protection. These features make the proposed method ideal for secure sensor-based applications such as authentication, data storage security, and digital signature schemes.