Artificial Intelligence in Medicine

Objective. To determine which model family a hospital should use to extract structured data (sections, medication records, laboratory tuples) from Russian/Kazakh stroke discharge summaries, bounded against measured human agreement. Materials and Methods. Three families were compared under identical inputs: a multilingual BERT-style encoder with CRF heads, a locally fine-tuned Qwen3-4B, and zero-shot GPT-5.5, on three tasks of increasing granularity: section segmentation (4,327 annotated documents), seven-field medication records (708 documents), and dated lab tuples (1,888 documents). All arms shared a block-ID representation (extraction as tagging over a fixed segmentation). Blind re-annotation of 50 documents per task provided the human-agreement ceiling; model-vs-ceiling claims use paired same-document tests with Holm correction. Results. On section segmentation all arms fall within {+/-}0.03 of the ceiling (0.905; models 0.919-0.932). On medication detection the cloud model is above it (0.966 vs 0.959); on full-record assembly, re-scored symmetrically against a blind annotator as common reference, both the fine-tuned LLM and the cloud model reach 0.344 agreement versus the gold annotator's 0.183. On lab assembly the fine-tuned LLM matches the full-tuple ceiling (0.898 vs 0.902) while the encoder is far below (0.476). Residual errors concentrate in annotation conventions the annotators themselves could not agree on. Discussion. Detection is solved at the human ceiling even for non-English clinical text, and a minutes-scale fine-tune reaches it with no data egress. Assembly separates the families and is convention-bound: benchmark consumers should read "above human agreement" against what the guidelines fixed. Conclusion. Annotation guidelines and task representation, not model capacity, are the binding constraints for this task family.

Background: General purpose language models such as ChatGPT are increasingly used by physicians and triage nurses during emergency triage. A recent study reported 51.6% undertriage of emergencies when patients queried ChatGPT directly (Ramaswamy et al., 2026). DR. INFO is an agentic AI based clinical assistant that retrieves over a curated clinical knowledge base, and an MTS specific retrieval configuration is available in which the system also retrieves the Manchester Triage System (MTS) textbook at inference time. The safety of these systems as a triage adjunct against a structured framework has not been characterised. Methods: We adapted the clinical scenarios published by Ramaswamy et al. and mapped them to the Manchester Triage System, yielding 39 emergency cases covering all five MTS priority levels. Each case was evaluated in two variants, one without and one with the objective clinical data block (vital signs, examination findings, and laboratory results), and permuted across two genders, giving 156 prompts per condition. Three systems were tested with and without a misleading GP referral statement prepended as an anchoring statement, giving 312 prompts per system: DR. INFO Baseline, DR. INFO with MTS retrieval, and OpenAI GPT-5.1. The primary outcome was the undertriage rate on the ordered MTS scale, tested with Fisher's exact test. Results: GPT-5.1 undertriaged 44.2% of cases (69/156; 95% CI 36.7 to 52.1), including 75.0% of Red and 73.4% of Orange presentations. Both DR. INFO configurations undertriaged 11.5% of cases (18/156; 95% CI 7.4 to 17.5; Fisher's exact p = 1.0 x 10^-10 versus GPT-5.1). GPT-5.1 produced 6 dangerous misses (3.8%), and both DR. INFO configurations produced none (p = 0.030). When the anchoring statement was prepended, GPT-5.1 undertriaged 8 of 8 Red cases, while both DR. INFO configurations continued to undertriage none. Adding objective clinical data to the input reduced undertriage in DR. INFO with MTS retrieval from 19.2% to 3.8% (p = 0.005). DR. INFO Baseline and GPT-5.1 showed no comparable change. There was no significant effect of gender. Conclusion: On this benchmark, replacing a general purpose language model with an agentic retrieval augmented system over a curated clinical knowledge base substantially reduced the undertriage and dangerous miss rates. Adding retrieval of the Manchester Triage System textbook to the agentic system was further associated with a reduced susceptibility to the anchoring statement and with an appropriate change in the assigned MTS priority when objective clinical data became available. Of the three configurations evaluated here, only DR. INFO with MTS retrieval combined a clinically conservative assignment at first contact with appropriate updating as additional clinical information arrived.

Depression in Alzheimer's disease (AD) is clinically challenging to characterise because affective and cognitive manifestations frequently overlap with the underlying neurodegenerative condition. This study investigates depression-related speech patterns in AD using an explainable multimodal framework that integrates linguistic and acoustic features from elicited picture-description speech. To address the absence of directly observed depression labels in the target AD cohort, we employ a source-informed cross-dataset transfer strategy. Unsupervised clustering identified two distinct speech sub-phenotypes within the AD cohort. By comparing these unsupervised representations to a depression-labelled source corpus, one cluster demonstrated mathematical proximity (measured via cosine similarity in a shared latent space) to the source depression centroid. We explicitly acknowledge that this proximity reflects geometric similarity in a shared latent space, which may encapsulate overlapping domain characteristics alongside affective variance. To validate the structural integrity and separability of these discovered clusters, a proxy supervised classification task was employed. The retained multimodal artificial neural network (ANN) integrated information from both modalities and achieved 92.73\% cluster separability accuracy. This demonstrates that the representation-derived two-cluster partition was highly learnable on the independent internal hold-out test set; it does not establish clinical validity, longitudinal stability, or diagnostic validity of the identified sub-phenotypes. For interpretability, Shapley Additive exPlanations (SHAP) analysis was utilised to extract feature attributions and organise influential acoustic and linguistic features into a multimodal taxonomy. Furthermore, a Multimodal SHAP Attribution Magnitude (MSAM) score was derived as an exploratory model-based attribution score rather than a clinically validated severity measure. Ultimately, this study provides an interpretable computational baseline for characterising depression-related speech variance in AD, establishing a structural foundation for future prospective validation using independently labelled clinical cohorts.

Evidence-synthesis teams increasingly depend on machine-learning tools to automate risk-of-bias assessment, but these tools frequently rely on unrunnable, deprecated software stacks. Refactoring legacy tools into modern environments is essential for maintenance, yet introduces a critical risk: silently invalidating published performance metrics. Without rigorous validation, systematic reviewers cannot trust that modernized tools retain their predecessors' behavior. We present a five-tier validation framework applied to RCT-Reviewer, an independent ground-up modernization of the widely used RobotReviewer system. To prove preservation, we engineered a compatibility shim to execute the original 2017 code alongside the modernized tool using byte-identical weight files. The framework evaluated predictive validity on a human-labelled benchmark, quantified the cost of dropping obsolete ensemble components, proved bit-exact fidelity, tested parser robustness at scale, and included an external validation arm evaluating agreement on 313 trials against a recent reference standard. A control experiment showing 100% agreement between legacy and modernized implementations on identical text proved that variance versus published values is driven by PDF provenance (open-access versus publisher typeset) rather than codebase refactoring. The modernized tool perfectly reproduced original risk-of-bias judgements in 6,018 comparisons. While the screening classifier experienced a bounded performance reduction (F1 0.969 to 0.925) from removing obsolete components, the model's core mathematical validity is preserved. These results establish that reviewers can deploy a maintainable modernization with a clearly bounded performance trade-off, offering a reusable validation framework for clinical NLP modernizations. Future work will explore retraining obsolete neural network components to close the performance gap.

Background Hepatocellular carcinoma (HCC) treatment selection demands nuanced integration of heterogeneous patient data, yet prevailing predictive models rely on restricted data modalities and oversimplified therapeutic frameworks, compromising clinical translation. Objective We developed and validated a multimodal artificial intelligence framework to guide optimal treatment strategy selection across the full spectrum of HCC interventions. Methods This retrospective study comprised 1,043 HCC patients (development cohort, January 2017-December 2023) and 55 external validation patients (2023) from Wuxi Peoples Hospital. We engineered Embedding-Augmented Extra Trees (ET-Emb), a novel model fusing structured clinical variables with contextual text embeddings derived from medical histories and radiology reports. ET-Emb quantifies probabilities for five primary treatments: open/laparoscopic resection, transarterial chemoembolization, radiofrequency ablation (RFA), and chemotherapy. Model performance was rigorously assessed via 10-fold cross-validation and external validation using ROC-AUC and PR-AUC metrics. Results ET-Emb demonstrated robust performance in the development cohort (ROC-AUC: 0.84 {+/-} 0.04; PR-AUC: 0.55 {+/-} 0.06), significantly outperforming established benchmarks. This generalizability was preserved in external validation (ROC-AUC: 0.77 {+/-} 0.02; PR-AUC: 0.47 {+/-} 0.03). SHAP analysis identified textual clinical narratives and socioeconomic determinants as critical predictive drivers. Conclusions By unifying structured and unstructured data modalities, ET-Emb delivers accurate, multi-treatment strategy prediction for HCC. Its clinical validity and the demonstrated significance of textual features establish multimodal AI as an essential paradigm for simulating complex oncological decision-making, positioning ET-Emb as a transformative tool for precision HCC management.

Connected speech is essential for everyday communication, but clinical constraints and patient fatigue limit detailed evaluation in acute stroke (<1-week post-stroke). Bedside assessments may sample discourse but rarely quantify language impairment (LI) in connected speech, leaving patient communication poorly characterized. We analyzed brief story retellings from 86 patients with left-hemisphere stroke (~4 days post-stroke; 63 classified with LI using composite clinical and naming criteria). From transcripts generated with automatic speech recognition, we derived discrete linguistic features and embeddings with Large Language Models (LLMs). An ensemble of embedding-based classifiers distinguished patients with and without LI with 90% balanced accuracy (79% sensitivity, 100% specificity), outperforming independent embedding and discrete-linguistic-based classifiers, showing distinct LLMs contributed complementary information. Adding the discrete-linguistic-based classifier to the ensemble did not improve balanced accuracy but modestly increased sensitivity at the expense of specificity. We provide proof of concept for a fast, largely automated discourse screener of acute LI.

Background: Large language models (LLMs) may answer cancer questions plausibly while accepting a false presupposition (FP). Prompts that increase FP correction may also challenge questions in which no FP is present (NFP). We compared SkillOpt, an automated skill-optimisation method developed by Microsoft Research, with conventional prompt engineering across models. Methods: Using the physician-verified Cancer-Myth benchmark, GPT-5.4 optimised a written skill for GPT-4o in three independent 12-step executions. Training used 117 FP questions. Rewrites were selected by combined accuracy on a validation set of 58 FP and 30 NFP questions, then tested on 410 unseen FP questions. A separate GPT-4o judged responses. Across seven LLMs, we compared no correction instruction, a deliberately strong seven-example few-shot prompt, three GPT-5.4-generated instruction-only prompts, and full and compact SkillOpt skills. Further tests examined portability, new questions and valid clinical assumptions. Results: SkillOpt increased GPT-4o FP correction from 15.4-17.6% to 82.2-90.5%, gains of 66.3-72.9 percentage points (95% bootstrap confidence intervals 61.7-77.3; all Holm-adjusted p<0.001). Only two or three rewrites were accepted per execution. The first full skill achieved 92.7-96.6% correction on three recipient models but 6.3-16.3% on two weaker models, where few-shot prompting achieved 44.6-54.6%. On Gemini 3.5 Flash and GPT-5.6 Luna, compact SkillOpt retained 90.0-92.9% correction while increasing NFP accuracy from 59.2-79.2% with the full skill to 85.8-93.3%. Both approaches sometimes challenged valid clinical assumptions. Conclusions: SkillOpt produced large, repeatable gains, but it was neither universally portable nor free of over-correction. Automated and conventional prompts should be compared for each intended model using tests of both FP correction and NFP accuracy. Compact SkillOpt may improve this balance on capable models.

Objectives: Searching a reference terminology such as SNOMED CT confronts two retrieval technologies of fundamentally opposite nature. Deterministic lexical matching is precise, order-independent and character-level, and lets a clinician refine a query incrementally, but it is confined to the wording and language of the curated descriptions. Learned semantic matching bridges paraphrase and language, but poorly represents abbreviations and partial, as-typed input. Their requirements and failure modes are so different that combining them naively can let the channel that is wrong for a given query degrade the one that is right. The conventional way to avoid adding a semantic channel at all has been to strengthen the lexical one (curating an additional local interface vocabulary so that more expressions match by surface form), but such vocabularies are labour-intensive to maintain and their quality degrades over time. We ask whether these two opposite paradigms can instead be made to *coexist* in one retrieval flow (each serving the input regime it is strong at, neither harming the other) over SNOMED CT's own curated descriptions, and study the query interpretation, rank fusion and re-ranking design that makes this coexistence safe. Materials and Methods: We implemented one reproducible instance of this architecture over the SNOMED CT International edition. It combines deterministic, order-independent multi-prefix search with three learned components: BioLORD-2023-M embeddings for semantic retrieval, an optional BGE cross-encoder for re-ranking, and a local LLM (gemma 4) for query normalization and structured clinical entity extraction. Lexical and semantic rankings are combined with Reciprocal Rank Fusion, and results can be constrained with a hierarchy (descendant) filter. As an external evaluation we measured search-only linking accuracy on 542 disease mentions from the DisTEMIST corpus (Spanish, zero-shot), resolving gold codes through SNOMED CT historical associations and ablating query pre-processing, re-ranking, and each retrieval channel in isolation; and, on a second corpus of English real-EHR discharge notes (the SNOMED CT Entity Linking Challenge, 12,897 mentions), we estimated field-scoped typeahead performance and tested a context-aware pre-process. Results: Illustrative tests showed complementary behaviour: the lexical and semantic paths resolved different query types, while LLM pre-processing supported cross-lingual and abbreviated input. On a search-only linking evaluation over 542 disease mentions from the DisTEMIST corpus (Spanish, zero-shot), the best and computationally cheapest configuration, semantic search with re-ranking and *no* LLM query pre-processing, placed the exact concept first for 60% of mentions (accuracy@1 = 0.60) and within the top-10 candidates a picker would show for 80% (recall@10 = 0.80). Because these search-only figures are computed on gold mention spans, they isolate normalization from mention detection and are not directly comparable to the DisTEMIST shared task's end-to-end scores. LLM query pre-processing helped dirty, lay input but slightly hurt clean terminology-like queries, and a near-miss analysis showed that many remaining cases retrieved a parent or child of the intended concept. Isolating the channels showed the two coexist without a trade-off: on this clean-normalization corpus the semantic channel carried the accuracy and adding the lexical channel neither helped nor harmed it (accuracy@1 0.60 vs 0.61), because rank fusion followed by re-ranking suppresses whichever channel is off-task; the lexical channel's own strength, partial as-typed input, is a regime this corpus does not test. On the English real-EHR corpus, a field-scoped typeahead (each query constrained to its entry field's domain) placed the concept on the top-10 picker list for 73% of 12,897 mentions across all clinical domains (accuracy@1 0.52; findings 0.54, close to DisTEMIST). Most residual errors were soft (the gold on the list but not first, or a hierarchical neighbour); the hard misses were dominated by ambiguous EHR shorthand, and a single context-aware pre-processing call (normalizing a mention with its surrounding note text) roughly doubled recovery of those hard misses, though, as with LLM normalization on the clean terms of DisTEMIST, applying it to every mention slightly lowered overall accuracy by perturbing the easy majority, so it is best used as a selective rescue. Results are specific to the evaluated configuration, not an architecture-independent estimate. Discussion: The findings support the technical feasibility of a reference architecture that computes additional access routes over SNOMED CT's curated descriptions, and indicate that making the lexical and semantic paradigms *coexist*, rather than choosing between them, is what makes those computed routes safe to add. This shifts rather than eliminates maintenance, from local phrase authoring toward governance and validation of retrieval services. Conclusion: Hybrid concept retrieval lets lexical precision and semantic flexibility coexist over SNOMED CT's maintained descriptions without a trade-off between them, offering an implementation pattern that may reduce additional local lexical curation where direct search is appropriate, while curated interface vocabularies remain useful in selected guided workflows.

Background: Free-text notes in electronic health records (EHRs) contain fine-grained psychiatric information that is essential for psychiatric research and clinical care, and often absent or under-recorded in structured codes alone. Clinical natural language processing (cNLP) can support extraction of this information from EHR notes, yet Spanish-language cNLP remains under-developed. Moreover, broad evaluations comparing multiple encoder-based language models across extensive, fine-grained psychiatric concept sets remain scarce, and it remains unclear how these models compare with traditional NLP (tNLP) systems and much larger generative large language models (LLMs). In addition, cross-site performance of fine-tuned models is rarely tested, and limited annotated training data remains a major challenge, especially for rare symptoms. Objectives: We aimed to advance scalable, global psychiatric cNLP by fine-tuning multiple encoder-based models with differing architectures and pre-training strategies for detecting fine-grained psychiatric concepts in Spanish EHRs. We further evaluated the impact of augmenting the fine-tuning data with precision-weighted weak labels for less-frequent concepts, and compared the performance of the encoder-based models to that of tNLP and a fine-tuned generative LLM trained on the same data. Finally, we evaluated model cross-site generalizability on an external EHR dataset. Methods: Three encoder-based models (BETO, XLM-RoBERTa-large, and bsc-bio-ehr-es) were fine-tuned on 1,642 clinician-annotated EHR documents from Colombia to detect 110 psychiatric concepts in Spanish text. To address the limited annotated examples available for less-frequent concepts, 12,000 additional documents were weakly-labeled for less-frequent concepts using tNLP, and incorporated into the fine-tuning data with labels weighted by pattern precision. Models were compared with tNLP and a generative LLM, and evaluated on an external EHR dataset from another psychiatric hospital in Colombia. Results: Encoder model performance varied substantially, with macro-F1 ranging from 0.64 to 0.81. BETO achieved the highest macro-F1 (0.81; median F1=0.88 [IQR=0.77-0.96]). Adding precision-weighted weak labels for less-frequent concepts improved BETO's overall macro-F1 to 0.83 and increased mean F1 for the 55 augmented concepts from 0.82 to 0.86. Under matched fine-tuning conditions, fine-tuned BETO and the tNLP method were equivalent in F1, whereas the LLM significantly outperformed BETO in F1. After weak-label augmentation, BETO significantly outperformed tNLP in F1 (PFDR<.001) and narrowed the performance gap with the LLM, although equivalence was not established. Lastly, fine-tuned BETO maintained reasonably strong performance on data from an external hospital not used for model fine-tuning (out-of-domain macro-F1=0.78). Conclusions: General-purpose pre-trained encoders had strong performance for psychiatric concept extraction from Spanish EHRs. Weak-label augmentation improved BETO's performance and strengthened results relative to a tNLP baseline, while reducing, but not eliminating, the performance gap with a much larger fine-tuned generative LLM. These findings highlight the utility of these relatively lightweight models for scalable, accurate and reproducible detection of psychiatric concepts in Spanish-language EHRs.

Background: Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder characterized by persistent symptoms of inattention, hyperactivity, and impulsivity that significantly affect cognitive, academic, and social functioning. Although electroencephalography (EEG) has emerged as a promising non-invasive tool for identifying objective neurophysiological biomarkers of ADHD because of its high temporal resolution and cost-effectiveness, many existing machine learning studies rely on single-domain EEG features, which may not adequately capture the complex functional organization of the brain. Methods: This study proposes a subject-level EEG classification framework that integrates three complementary feature domains: spectral characteristics, weighted Phase Lag Index (wPLI)-based functional connectivity, and topological descriptors derived from persistent homology. EEG recordings from 121 subjects were preprocessed and segmented into overlapping windows for feature extraction. Window-level features were aggregated to generate subject-level representations. Recursive Feature Elimination (RFE) was employed to select the 110 most informative features from an initial set of 1,043 descriptors. Five machine learning classifiers, including Support Vector Machine (SVM), Random Forest, XGBoost, CatBoost, and Multi-Layer Perceptron (MLP), were evaluated using stratified five-fold cross-validation. An ablation study was further conducted to assess the contribution of each feature domain. Model interpretability was investigated using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), two widely adopted Explainable Artificial Intelligence techniques for interpreting machine learning models. Results: Among the evaluated classifiers, the SVM achieved the highest classification performance, yielding an accuracy of 80.17%, an F1-score of 80.95%, a balanced accuracy of 80.14%, a receiver operating characteristic area under the curve (ROC-AUC) of 86.72%, and a precision recall area under the curve (PR-AUC) of 87.34%. The ablation study demonstrated that combining spectral, functional connectivity, and topological features consistently outperformed individual feature domains, highlighting the complementary nature of these representations. Explainable AI analyses further identified functional connectivity features as the primary contributors to classification performance while confirming the additional value of spectral and topological descriptors. Conclusions: The proposed framework demonstrates that integrating spectral analysis, functional connectivity, and topological data analysis provides an effective and interpretable approach for subject-level ADHD classification from EEG recordings. The combination of recursive feature selection, multi-domain feature fusion, ablation analysis, and Explainable Artificial Intelligence establishes a transparent and reproducible pipeline that may facilitate the development of reliable EEG-based clinical decision-support systems for ADHD.

The choice of FHIR-to-text serialisation format significantly impacts clinical LLM quality (Kruskal-Wallis H=163.86, p<10^-33, delta=0.24 on a 5-point scale), yet remains unstudied as a clinical deployment variable. We present FHIRBench-UK, evaluating five large language models across six serialisation formats and three clinical tasks on 100 UK Core FHIR patient bundles (18,000 scored prompts across clean and perturbed cohorts). Our findings converge with independent work on open-weight models (Pator, 2026). The optimal format is context-dependent: raw_json dominates for clinical QA, hybrid_adaptive for clinical reasoning, and structured_markdown for summarisation. In 58% of model-task-complexity scenarios, raw_json is suboptimal. Model capability moderates format sensitivity: Claude Sonnet 4.5 shows 0.10-point sensitivity versus Llama 3.3's 0.39, making adaptive serialisation most valuable for budget-constrained deployments using mid-tier models. All findings replicate under clinically realistic data perturbation. The study additionally confirms a complete ranking inversion between token-level F1 and clinical quality (rho=-0.90), replicating US findings across UK Core profiles. We recommend task-aware serialisation routing as a zero-cost quality intervention for NHS FHIR-based LLM deployments.

Background: Biomedical studies increasingly collect repeated measurements, such as clinical ratings, speech measures, gait summaries, and wearable-sensor features, at irregular subject-specific follow-up times. These data are often sparse and unaligned, making them difficult to use with standard classifiers that require fixed-length input vectors. Parkinson's disease provides a motivating example because progression is longitudinal and heterogeneous, but clinical follow-up is rarely observed on a common schedule. Objective: This study develops and evaluates a multiclass likelihood contrast classifier for sparse longitudinal trajectory recovery. The goal is to classify subjects using their observed repeated measurements while avoiding aggressive reduction of trajectories to averages or simple slopes. Methods: The proposed method fits one class-specific mixed-effects model per outcome class and assigns each held-out subject to the class with the largest marginal likelihood score. The main specification uses a quadratic mean trajectory with subject-specific random intercept and random slope terms. Performance is evaluated in a three-class simulation with imbalanced class sizes and in a Parkinson's motor-UPDRS sparse trajectory recovery experiment. In the Parkinson's experiment, dense histories are first used to define mathematical trajectory phenotypes; most observations are then hidden, and models are asked to recover the phenotype from the sparse record. Baselines include a linear mixed-likelihood classifier, functional k-nearest neighbors, random forest, and support vector machine. Results: In the Parkinson's sparse recovery experiment, after approximately 90% of dense observations were hidden, the proposed model achieved accuracy 0.8095, macro-F1 0.8160, MCC 0.7107, and macro-AUC 0.9268. A train-validation-test diagnostic gave similar held-out behavior, with validation macro-F1 0.7980 and test macro-F1 0.8160 for the proposed model. Conclusion: Multiclass likelihood contrasts offer an interpretable and competitive approach for sparse, unaligned longitudinal classification. The Parkinson's analysis should be read as an algorithmic sparse trajectory recovery study rather than clinical diagnostic validation. Future work should evaluate the method with externally assigned clinical outcomes, multivariate longitudinal biomarkers, prospective cohorts, and ordinal generalized mixed-model likelihoods that better match bounded clinical rating scales.

Longitudinal electronic health records (EHRs) contain rich information about the evolution of a patient's clinical state, but temporal measurements and events are often represented using summary statistics or complex learned representations that can be difficult to interpret clinically. We propose an interpretable trajectory-based feature representation that transforms heterogeneous longitudinal EHR data into compact, human-readable predictive features. Numerical temporal variables are represented by trajectory states such as INCREASING, DECREASING, and STABLE, while categorical clinical events are represented by interpretable states describing repeated occurrence, stability, or change. These trajectory features are combined with static patient and encounter characteristics and evaluated using L1-regularized logistic regression, which provides sparse feature selection and direct identification of positive and negative predictive features. The approach is evaluated using MIMIC-IV in two clinically distinct prediction tasks: in- hospital mortality following ICU admission and hospital readmission within 30 days after discharge. For ICU mortality, multiple observation windows ranging from 1 to 48 hours are investigated to examine how the availability of longitudinal information affects prediction and the relative contribution of trajectory features. The highest AUROC of 0.8751 is obtained with a 32-hour observation window. For 30-day readmission, trajectories from the final 72 hours of the index hospitalization are evaluated using both three-state and five-state categorical representations. The two representations produce essentially identical discriminative performance (AUROC 0.6875 and 0.6876, respectively), indicating that increasing trajectory granularity provides little predictive benefit in this setting. The results demonstrate that heterogeneous longitudinal EHR observations can be transformed into compact and directly interpretable temporal features while retaining useful predictive information. The proposed framework therefore provides a simple and model-compatible approach for identifying clinically interpretable patterns of patient evolution and can be extended with more detailed trajectory definitions or applied with other predictive models.

Self-repairs, in-utterance revisions in which a speaker abandons and reformulates their speech, are a promising interpretable marker for speech-based cognitive screening. Detecting them automatically is difficult because a self-repair is defined by its relationship to surrounding speech rather than by fixed lexical cues. On the DementiaBank ADReSS corpus, we compared the capability of generative LLMs under a five-condition prompt ablation against a fine-tuned DistilBERT token classifier at detecting self-repairs. GPT-5 performed best (test F1 = 0.73) and was largely insensitive to prompt design, whereas the LLaMA (open-weight alternative) was both weaker and far more prompt-sensitive (test F1 = 0.47). DistilBERT, nearly 100 times smaller, matched the open-weight LLM at a fraction of the computational cost. These results suggest that a locally deployable encoder, given sufficient in-domain annotation, is a more plausible route to clinical self-repair detection than scaling model size or prompt complexity.

Background Naturalistic audiovisual recordings of caregiver-child interactions contain rich developmental signals. However, extracting interpretable clinical measures requires resource-intensive manual coding. To address this bottleneck, we evaluated natural-language queries for retrieving specific behavioral moments from these recordings, applying multimodal embeddings as an automated evidence-selection layer. Methods We compared three embedding models (Jina Embeddings v5 Omni, LanguageBind, and Wave7B) for natural-language retrieval directly from audio and video streams, bypassing transcript text. We assessed performance across 27 behavioral targets in 277 caregiver-child recordings (14, 24, and 36 months of age) from the Early Head Start Talkbank corpus, yielding 7,479 recording-target queries. Results Jina Embeddings v5 Omni achieved the highest top-10 retrieval success (text-to-audio 38.3%; text-to-video 36.4%), ahead of LanguageBind (37.0%; 34.5%) and Wave7B (36.1%; 35.0%). Across models, retrieval was substantially more successful for common targets than for rare vocal and gestural behaviors, such as pointing and babbling. By analyzing the spoken words within the retrieved audio clips, we found that Jina accurately ranked the children by their relative vocabulary size at each age (Spearman = 0.68, 0.82, and 0.90 at 14, 24, and 36 months). However, the model severely underestimated the total number of unique words each child used throughout the full session. Conclusion Multimodal embeddings can successfully pinpoint important developmental behaviors and speech patterns within lengthy caregiver-child recordings. However, these systems still struggle to locate rare events. Additionally, while they can accurately rank children by relative vocabulary size, they fail to measure a child's complete vocabulary. We conclude that these models are currently best suited for automated evidence-selection to prioritize relevant segments for expert interpretation rather than acting as an independent replacement for manual behavioral coding or language assessment. Improving the detection of infrequent behaviors and validating these models across external datasets are essential next steps before real-world clinical deployment.

Objective To evaluate the information quality, reliability, readability, and content deficits of ChatGPT, DeepSeek, and Doubao in answering frequently asked questions about bee stings. Methods Twenty-five high-frequency patient questions were selected based on Google Trends, domestic health platforms, and allergy guidelines, and submitted to the three LLMs, yielding 75 responses. Two emergency physicians independently assessed quality using GQS, mDISCERN, QAMAI, EQIP, and PEMAT-U/A. Readability was evaluated using six indices (ARI, FRES, GFI, FKGL, CLI, SMOG). Content deficits were classified qualitatively. Twenty emergency department patients evaluated the comprehensibility and credibility of AI responses. Results Significant differences were found among the three models on all quality metrics (all P < 0.01). DeepSeek achieved the highest mDISCERN score (median 21.00); ChatGPT led on QAMAI (25.00) and EQIP (75.00%); Doubao ranked lowest on all quality indicators but highest on PEMAT-U (73.00%). No model met the sixth-grade reading level standard. Eleven content deficits were identified: outdated medical advice (45.45%), omissions (36.36%), and hallucinations (18.18%). Doubao received the highest patient preference (45.0%). Conclusion The three LLMs exhibited distinct strengths: DeepSeek in source reliability, ChatGPT in overall quality, and Doubao in understandability. All models had excessively high readability levels, and outdated medical advice was the predominant content deficit. Objective To evaluate the information quality, reliability, readability, and content deficits of ChatGPT, DeepSeek, and Doubao in answering frequently asked questions about bee stings. Methods Twenty-five high-frequency patient questions were selected based on Google Trends, domestic health platforms, and allergy guidelines, and submitted to the three LLMs, yielding 75 responses. Two emergency physicians independently assessed quality using GQS, mDISCERN, QAMAI, EQIP, and PEMAT-U/A. Readability was evaluated using six indices (ARI, FRES, GFI, FKGL, CLI, SMOG). Content deficits were classified qualitatively. Twenty emergency department patients evaluated the comprehensibility and credibility of AI responses. Results Significant differences were found among the three models on all quality metrics (all P < 0.01). DeepSeek achieved the highest mDISCERN score (median 21.00); ChatGPT led on QAMAI (25.00) and EQIP (75.00%); Doubao ranked lowest on all quality indicators but highest on PEMAT-U (73.00%). No model met the sixth-grade reading level standard. Eleven content deficits were identified: outdated medical advice (45.45%), omissions (36.36%), and hallucinations (18.18%). Doubao received the highest patient preference (45.0%). Conclusion The three LLMs exhibited distinct strengths: DeepSeek in source reliability, ChatGPT in overall quality, and Doubao in understandability. All models had excessively high readability levels, and outdated medical advice was the predominant content deficit.

Standard uncertainty-informed rejection can unexpectedly trigger severe performance collapse, exposing localized vulnerabilities that common machine learning metrics typically do not show. We systematically diagnose this failure dynamic using Levodopa-Induced Dyskinesia prediction in Parkinson's Disease as a proof-of-concept. By training a heterogeneous ML ensemble, decomposing Aleatoric and Epistemic uncertainty and applying unsupervised subgroup discovery, we isolated the precise drivers of these atypical errors. Stratified error analysis revealed two divergent predictive regimes previously hidden by a global evaluation. While the models successfully extracted a predictive signal for one subgroup, the baseline features of a second subgroup lacked discriminative capacity, resulting in a high rate of confident misclassifications. Operating entirely below rejection thresholds, this single subgroup flatlined predictive metrics, driving the collapse of the global rejection curve. Ultimately, we demonstrate that atypical rejection failures stem from subgroup-specific data ambiguity rather than algorithmic deficiencies, making localized uncertainty-aware evaluation a critical methodological requirement prior to real-world deployment.

Introduction Aging populations are placing increasing pressure on Emergency Medical Services (EMS) providers to address the potentially suboptimal use of scarce resources. EMS utilization that falls outside the established deployment framework, referred to here as non-indicated use, may constitute an important source of such suboptimal resource use. Frequent users of EMS represent a key target group in efforts to reduce non-indicated utilization. The focus on frequent users is driven by the substantial workload and costs associated with their use of EMS, as well as by the need to address potential underlying quality-of-care issues reflected in high, persistent, or intensive EMS utilization. Addressing possible non-indicated EMS use among frequent users requires the development of monitoring systems that can be applied in both clinical practice and research to identify such use and support appropriate action once it has been detected. This study aims to explore the relationships between the drivers of frequent EMS use, potential remedies for non-indicated use, and subsequent reductions in utilization. Based on these insights, the study seeks to develop, through prototyping, a framework to guide the design of monitoring systems for research and clinical practice. Methods and analysis We will employ a mixed-methods observational study design. A systematic literature review and interviews with stakeholders who play key roles in the EMS referral process will be conducted to identify the drivers of frequent use and non-indicated EMS use and their potential remedies. Pseudonymized routinely collected data covering 2013-2026 from three Dutch EMS providers will be analyzed using machine learning (ML) and artificial Intelligence (AI) techniques to develop a typology of drivers and trustworthy, interpretable models for identifying frequent use and non-indicated use. A prototype monitoring system will be developed to demonstrate the online and the offline functionalities of the proposed framework for use in both clinical practice and research. The prototype usability, trustworthiness, and interpretability will be evaluated in collaboration with relevant stakeholders involved in the EMS referral process. Ethics and dissemination The data is routinely collected for administrative purposes only. Accordingly, the data will be completely pseudonymized, and as such this study falls outside the scope of the Dutch Medical Research Involving Human Subjects Act. After assessment we obtained a full waiver for using pseudonymized data from the EMS services from the Medical Ethics Review Board of the University Medical Center Groningen (reference number METc 2022/053). Patients and/or the public are not involved in the design, conduct, reporting, or dissemination plans of this research. Interviewees and participants in evaluating prototype monitoring systems (e.g. EMS clinicians, medical directors, ambulance nurses, emergency physicians, general practitioners (GPs), policy makers and police officers) will provide informed verbal consent prior to their participation in the study. The study findings will be disseminated through peer-reviewed publications, presentations at conferences, and in social media. Also, summary reports will be made available to participating institutions and relevant stakeholders.

Patient-reported and clinically documented narratives often contain informal, fragmented and linguistically heterogeneous expressions, complicating medical concept normalization (MCN) and frequently necessitating pre-processing before terminology mapping. Despite rapid advances in biomedical language modeling, the comparative utility of representation models for MCN and their integration with instruction-tuned LLMs in automated normalization workflows remain underexplored. In this study, we systematically benchmarked 15 general, biomedical and clinical transformer-based representation models together with 12 instruction-tuned LLMs across 12,713 instances from five established datasets: TAC2017_ADR, TwADR-L, TwiMed, CADEC and SMM4H2017. Representation models were evaluated using embedding-based semantic retrieval, whereas instruction-tuned LLMs were assessed as upstream text-correction modules. SapBERT achieved the highest Top-5 accuracy among representation models, reaching 63.8% for SNOMED CT and 58.0% for MedDRA without correction. Qwen 2 Instruct was selected as the preferred corrector on the basis of its favorable balance between Top-1 normalization performance and computational efficiency relative to substantially larger models, including Llama 3.1 Instruct (70B). Incorporation of Qwen 2 Instruct increased Top-5 accuracy to 69.4% for SNOMED CT and 63.2% for MedDRA. The resulting framework accepts heterogeneous short medical expressions without manual input pre-processing and automatically performs text refinement, semantic retrieval and terminology mapping to standardized concepts and vocabulary codes. These findings establish a benchmark-guided, scalable framework for automated medical terminology standardization.

Logistic regression remains a widely used classification method due to its interpretability and computational efficiency, but its direct application to high-dimensional biomedical data is limited when the number of features greatly exceeds the number of samples. In this paper, we propose a reformulated logistic regression framework designed for feature selection and classification in complex high-dimensional settings. The method is evaluated on three biomedical datasets, including scenarios with tens of thousands of attributes and substantially fewer samples. Across these datasets, the proposed approach achieved clear separation between control and disease groups while selecting a compact set of features. Several selected features were consistent with previously reported disease-associated markers, supporting the biological plausibility of the model, while additional selected features suggest potential novel candidates for further investigation. These results indicate that the proposed framework may provide an interpretable and computationally efficient alternative for feature selection in high-dimensional computational biology applications.