Transthoracic Echocardiography (TTE) is a primary imaging modality for the evaluation of cardiac anatomy and function. Standardized TTE protocols rely on specific probe maneuvers to acquire predefined anatomical views; however, image acquisition quality is heavily operator-dependent, and particularly novice users often fail to obtain diagnostically adequate images, since optimal probe positioning is inherently a very challenging task. To address this challenge, we propose a novel deep learning-based framework to support less experienced operators in acquiring correct echocardiographic apical views. The proposed approach integrates two core components: a new navigation map that models the relationships between optimal and suboptimal views, linking common acquisition errors to the corresponding corrective probe motion (e.g., tilting, rotation, translation) and a classification system combining a Variational Autoencoder, for semi-supervised latent feature extraction, with a lightweight Multi-Layer Perceptron, for view classification. The framework was first evaluated on two public datasets (CAMUS and HMC-QU) for binary classification (apical 2 and 4 chamber views), achieving average accuracies up to 0.97 while maintaining good generalization across heterogeneous acquisition conditions. The analysis was then extended to a more challenging 11-class classification task, including four standard apical views (2, 3, 4, 5 chambers) and seven suboptimal variants, using a newly collected in-house dataset (PROUD). In this setting, the proposed VAE+MLP achieved an average accuracy of 0.85, representing a promising step toward AI-assisted TTE guidance systems for novice operators.
Connected health increasingly relies on computing systems that can transform diverse, long-term, and context-rich data into reliable and actionable information. This special issue features seven contributions that cover topics such as digital twins and human pose reconstruction, continual learning, oral health monitoring, multimodal dementia detection, domain-generalized medical image segmentation, passive assessment of depression treatment outcomes, and context-aware stress monitoring. Together, these papers demonstrate how advanced sensing and machine learning can evolve from controlled demonstrations to adaptive, interpretable, and human-centered connected health systems.
Conventional radiomics is hindered by manual feature engineering and poor cross-tumor generalization. We introduce SAFARI (Self-supervised Tumor-Agnostic Framework for Radiomics Representation Learning), a novel, self-supervised model for radiomics that learns disease agnostic tumor characteristics. Built on a transformer architecture with a dual-objective self-supervised strategy, combining masked feature reconstruction and contrastive learning across tumor subvolumes, SAFARI was trained on 1,218 radiomic features from 27,144 tumor subvolumes across 1,086 patients with six major cancer types (lung, liver, colorectal, kidney, pancreatic, and colorectal liver metastases). In transfer learning experiments on independent datasets comprising 762 patients, fixed SAFARI embeddings outperformed our matched conventional radiomics baselines across multiple downstream tasks. Performance gains were substantial for predicting HER2 status in gallbladder cancer (AUC: 0.758 vs. 0.610), NSCLC histopathology (accuracy: 71.3% vs. 62.1%), and EGFR mutation status (AUC: 0.805 vs. 0.690). By demonstrating that a single pretrained model outperforms matched task-specific baselines on the majority of endpoints, SAFARI establishes a new, more robust paradigm for radiomic analysis in oncology.
Continuous ICU EEG review is resource-intensive and limited by inter-rater variability. Clinicians interpret harmful patterns by jointly inspecting raw waveforms and region-wise spectrograms within a bipolar montage, integrating temporal context across multiple scales. Existing automated approaches, however, seldom encode these clinical priors or address the full six-class American Clinical Neurophysiology Society (ACNS) taxonomy. This study introduces BiVISTA , a clinically grounded Bi-View Spectro-Temporal Conv-Transformer. The methodology proceeds in four stages: (1) A Bi-View Super EEG Image is constructed by aligning multi-scale waveforms (10, 20, 50 s) with montage-preserving regional spectrograms, strictly enforcing anatomical topology. (2) A convolutional encoder performs manifold compression, extracting local spectro-temporal motifs into a compact embedding. (3) A six-layer Transformer encoder applies multi-head feature projections to refine and disentangle the compact representation. (4) A linear classifier outputs ACNS probabilities. Evaluated on the large-scale HBAC dataset, BiVISTA achieves a mean test accuracy of 96.0% ( \(\pm\) 0.3), macro-F1 of 0.960, and macro-AUROC of 99.6% across ten independent runs. Seizure sensitivity and specificity reach 95.0% and 99.0%, with a Cohen's \(\kappa\) of 0.952. Interpretability analyses demonstrate reliance on physiologically valid cues like focal lateralization, supporting trustworthy clinical decision-making.
Adverse drug events (ADEs) are a critical aspect of patient safety and pharmacovigilance, with significant implications for patient outcomes and public health monitoring. The increasing availability of electronic health records, social media, and patient forums provides valuable yet challenging unstructured data sources for ADE surveillance. To address these challenges, we introduce CONORM, a framework integrating named entity recognition (NER) and normalization (EN) for ADEs across diverse textual domains. CONORM comprises CONORM-NER and CONORM-EN, featuring a dual-encoder with dynamic context refining (DCR). DCR adaptively combines isolated entity embeddings with contextual representations. Our analyses demonstrate that CONORM adapts to text formality, enhances precision on out-of-distribution concepts, and reduces errors compared to context-agnostic baselines. CONORM was evaluated on tweets, forum posts, and structured product labels, covering 30,000 annotated ADE mentions across the SMM4H, CADEC, and TAC datasets. It achieved end-to-end F1-scores of 63.86%, 72.45%, and 84.99%, surpassing the strongest baseline by an average margin of 20%. These results highlight CONORM's robust adaptability across domains. CONORM offers a scalable solution for pharmacovigilance, with pre-computed target embeddings enhancing inference efficiency. Its generalization establishes it as a robust tool for ADE surveillance. The source code is available at https://github.com/ds4dh/CONORM .
