Medicine & Health

The work proposes and implements a hybrid retrieval architecture that lets deterministic lexical matching and learned semantic matching coexist over SNOMED CT's own curated descriptions, combining multi-prefix search, BioLORD-2023-M embeddings, an optional BGE cross-encoder for re-ranking, and a local LLM for query normalization, with Reciprocal Rank Fusion and a hierarchy filter; on a search-only linking evaluation over 542 disease mentions from the DisTEMIST corpus (Spanish, zero-shot), semantic search with re-ranking and no LLM query pre-processing reached accuracy@1 of 0.60 and recall@10 of 0.80, and on 12,897 mentions from English real-EHR discharge notes a field-scoped typeahead placed the concept on the top-10 picker list for 73% of mentions (accuracy@1 0.

This work proposes ICARus, a positive-unlabelled read-out refinement framework that integrates protein-protein interaction information derived from a protein language model into a genome-wide RNA interference screen to boost the discovery of vaccinia virus-host interactions and human genes with potential antiviral function, and releases the raw and refined read-outs of the screen as an agentic-AI-enabled community resource.

This work presents ABCP_finder, a computational framework for predicting anti-breast cancer peptides (ABCPs) that combines pretrained protein language model embeddings (ProtBERT and ESM2) with a multilayer perceptron classifier, uses a homology-aware train-test split via CD-HIT at 30% sequence identity with 80% coverage to reduce data leakage, reports ProtBERT as the stronger model with 93.82% accuracy, 86.88% recall, 90.59% F1-score, 0.8618 MCC, 96.67% AUC and a Brier score of 0.0633, selects a 0.7 probability threshold from calibration analysis for high-confidence ABCPs, and shows through external validation with xDeep-AcPEP that unknown peptides predicted as ABCPs exhibit favourable IC values.

The work continues the self-supervised objective of the general video world model V-JEPA 2 on roughly 62 hours of unlabelled vocal-tract real-time MRI to obtain a frozen Arti-JEPA encoder, and evaluates it on phoneme prediction, fluent-versus-disfluent classification, and pre/post-glossectomy transfer, finding that a temporal video prior clearly beats per-frame image encoders, that latent prediction is at least as strong as pixel reconstruction, that domain adaptation roughly doubles cross-domain Cohen's kappa for phonemes but helps binary stuttering detection only marginally, and that phoneme signal remains partly decodable after surgery.

The work proposes HierSTT, an end-to-end hierarchical Transformer framework that uses a Temporal Fusion Transformer for national-level ED demand and spatio-temporal Transformer encoder-decoders to produce regional and hospital forecasts conditioned on higher-level predictions, with a coherence-aware loss penalizing cross-level inconsistency; it also releases a nationwide Portuguese ED dataset covering 81 hospitals across 5 regional health administrations with heterogeneous level-specific covariates, and reports a 32% average WAPE reduction versus the best non-hierarchical deep learning baseline, outperforming classical hierarchical reconciliation methods while producing near-coherent predictions across levels.

Using sparse-autoencoder (SAE) features in Gemma 3 4B/12B IT and Qwen3-8B on clinical triage vignettes, this study finds that medical features fire on the shared clinical narrative under both formats but are inactive at the multiple-choice decision token; emergency-tier information is linearly decodable from vignette representations (ROC-AUC 0.95–1.00) with no significant format difference yet is attenuated at the decision token, and scaffold-peaking features account for over 91% of unsigned attribution in both Gemma models, placing the strongest correlates of the format effect at answer selection.

The work proposes a generation-oriented 3D-CLIP encoder trained with structured hard negatives constructed exclusively at the text level (Attribute-Aware Negatives, AAN, and Semantic-Aware Negatives, SAN) to strengthen contrastive learning under the small-batch constraints of volumetric encoders, and uses it to condition a fully end-to-end latent diffusion model operating directly in 3D latent space, achieving lower FID, higher pathology-classification AUC and precision, faster inference, and lower GPU memory than competing methods on CT-RATE across 18 pathological conditions.