Deep learning methods for glucose prediction in type 1 diabetes (T1D) are often constrained by data availability, restricted access to data, and substantial heterogeneity across treatment settings and institutions. Federated Learning offers a privacy-preserving alternative to centralized training, but its performance under clinically realistic, non-independent and non-identically distributed (non-IID) data remains challenging. This work proposes a hybrid federated-personalized framework that combines cross-silo training with patient-level personalization for multi-horizon glucose prediction. Clinically defined silos based on treatment modality and insulin type are used to explicitly model real-world heterogeneity. A Temporal Fusion Transformer is trained using federated optimization (FedProx), incorporating clinician-prescribed basal insulin profiles as known future inputs. The resulting global model is subsequently adapted at the patient level, enabling personalization. The proposed approach is evaluated on the T1-DEXI dataset and externally validated on the OhioT1DM cohort.Centralized and federated personalization are compared to quantify performance differences between in-distribution and external evaluation and to assess patient-specific adaptation. Results show that the federated global models reduce prediction errors at longer horizons compared to centralized training on unseen data distributions, whereas centralized training achieves lower in-distribution errors. After patient-level personalization, centralized models show slightly lower absolute errors than federated models. These findings are consistent with a balance between population-level behavior and individual adaptation under the evaluated configurations. They support the feasibility of combining cross-silo training with patient-level personalization for privacy-preserving glucose prediction in heterogeneous clinical environments. Furthermore, incorporating scheduled basal insulin profiles enhances in-distribution prediction accuracy, underscoring the importance of clinically meaningful inputs in model design.
Heart failure (HF) is a major cause of morbidity, mortality, and healthcare utilization worldwide, yet early risk stratification remains challenging because prognostic information is distributed across heterogeneous clinical narratives and subtle electrophysiological abnormalities. We propose AGraphHF, a multimodal framework for incident HF risk prediction that integrates 12-lead electrocardiograms (ECGs) with demographic-aware clinical text under a survival objective. The prediction task is temporally anchored at an index ECG and estimates the relative risk of subsequent HF onset under a longitudinal time-to-event formulation. The ECG encoder models individual leads as graph nodes and learns both inter-lead relationships and temporal dependencies using dual graph attention mechanisms. To enhance sensitivity to weak but prognostically relevant signal deviations, the model employs a two-phase adversarial training strategy with dual decoders, where stable reconstruction of typical ECG morphology is followed by deviation-focused refinement conditioned on reconstruction residuals. In parallel, clinical notes are encoded using a demographic-aware language model that extends BioClinicalBERT through age- and sex-conditioned contrastive pre-training. ECG and text representations are aligned in a shared latent space and fused hierarchically to produce patient-level risk estimates. We evaluate AGraphHF on high-risk cohorts derived from the MIMIC-IV database, including patients with hypertension and myocardial infarction. Experimental results demonstrate improved survival risk discrimination compared with established baselines on the hypertension and myocardial infarction cohorts, while also revealing cohort-dependent differences in the contribution of ECG and clinical text. These findings support the use of coordinated graph-based ECG modelling and demographic-aware multimodal integration for HF risk stratification, while emphasizing the need for cohort-specific interpretation of model performance.
Map2VidSeg uses classifier-derived localization maps to generate prompts for video segmentation without segmentation training data. We revisit the pipeline under a fixed 224 × 224 visual-prompt regime and test whether newer foundation-model components improve performance. Starting from the CBMS 2025 system of DINOv2 attention, YOLOE, and SAM-2 on SUN-SEG polyp video, we replace the components along three axes: DINOv3 for map generation, YOLOE-26, for prompt-based segmentation, and SAM-3 for temporal refinement. A completed 2 × 2 prompt-source × YOLOE-stage factorial with SAM-3 fixed shows that the full DINOv3 + YOLOE-26 + SAM-3 upgrade reaches 0.6040 IoU and 0.6684 Dice on SUN-SEG Unseen, whereas the DINOv2 + YOLOE-11 + SAM-3 control reaches 0.6750 IoU and 0.7426 Dice. The prompt-source swap is near neutral with YOLOE-11, whereas YOLOE-26 is weaker at this prompt resolution and the full upgrade underperforms the additive expectation. In a matched comparison using identical cached YOLOE masks, SAM-3 tracker double-pass produced a modest paired gain over SAM-2 double-pass (＋0.0275 IoU; 95% BCa CI [+0.0122,+0.0458]; p=0.0012). A second-task diagnostic study on CholecSeg8k binary instruments shows the same pattern: integration changes improve the YOLOE stage, whereas SAM-3 reduced performance on multi-tool clips under shared-identity tracking, and class-aware tracking eliminated this loss. The results support pipeline-level evaluation and identify class factoring as a third design axis alongside prompt quality and prompt-model compatibility.
Autism spectrum disorder (ASD) is a neurodevelopmental condition whose timely identification can support earlier intervention. Conventional assessment relies on specialist observation and caregiver reports, motivating research into computer-aided screening using facial and behavioral data. This systematic review examines 171 studies published between 2016 and May 2026. It synthesizes machine learning and deep learning approaches to ASD assessment using static facial images, facial video, eye tracking, and multimodal data containing a facial signal. It also covers large language models, vision-language models, and multimodal large language models applied to ASD assessment, which were not addressed by the previous reviews included in the structured comparison, together with facial-data foundation-model studies in other clinical conditions that offer transferable methods. The review compares datasets, modeling strategies, and evaluation practices, and assesses methodological quality and risk of bias. Reported performance was often promising, but the strength of evidence varied with dataset provenance, diagnostic confirmation, participant independence, and validation design. Static-image studies relied heavily on reused public datasets and image-level partitions, whereas facial-video, eye-tracking, and multimodal studies more often used participant-level cohorts but still provided limited external and prospective validation. Differences in populations, acquisition protocols, prediction targets, and evaluation units prevent direct ranking of reported metrics across studies. Foundation models broaden the field toward behavioral description, diagnostic reasoning, and clinician-readable reporting, although their evidence remains preliminary. Facial-related data therefore show potential for computer-aided ASD screening, but clinical translation depends on clinically verified datasets, participant-independent evaluation, and testing in independent cohorts.
Objective: Tabular data learning is crucial in medical data analysis, but it has consistently overlooked a vital component within tables: column names. We aim to leverage column names and pertinent clinical knowledge when training medical tabular data, bridging the gap between the decision-making process of medical experts and conventional models. To protect privacy, all pre-processing is done locally. Materials and methods: We introduce a novel algorithm, named Interpretable Context-Aware Medical Concept Embedding strategy (ICAME), designed to emulate human experts’ data understanding. ICAME is among the first methods that explicitly utilize clinical knowledge in data analysis, applying LLMs for context-aware embedding generation. To ensure the patient data privacy, ICAME stores the embedding of predefined pseudo-values in the local database. This keeps private data local and minimizes the cost of online LLM API usage. Results: ICAME consistently improves performance across classification, regression, and prediction tasks when applied to the MIMIC-III public benchmark (average gains of 1.6%p and 2.9%p on AUROC and AUPRC, respectively). The performance improvement of ICAME is notable in a data-scarce scenario (average gains of 10.5%p and 6.3%p on AUROC and AUPRC, respectively). ICAME also provides interpretability of embeddings, confirming the incorporation of cell values and clinical knowledge. Discussion and Conclusion: ICAME effectively integrates explicit clinical knowledge into tabular data learning for the first time, boosting performance. Additionally, it showcases high scalability, covering numerical, categorical, and binary data types. Through extensive experiments, we demonstrate the potential of ICAME as a promising tool for future medical tabular data analysis.
