Artificial Intelligence in Medicine

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.

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.

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.

De-identifying temporal protected health information (PHI) is challenging because temporal expressions appear in diverse formats and must be transformed while preserving clinically meaningful timelines. This study proposes and evaluates an LLM-based framework for temporal PHI de-identification in real-world clinical notes. Using 1,148 notes from 30 sarcoma patients (28,431 annotated temporal entities), we evaluated multiple modern LLMs for temporal entity extraction and surrogate generation. Proprietary models achieved strong extraction performance, with GPT-4o achieving the best overall results. Most models preserved temporal formatting (>99%), but surrogate generation remained challenging. GPT-5.4 achieved the highest shift correctness (90.2% on a shared-entity subset) and the lowest order violations. Error analysis revealed systematic shift deviations ({+/-}1, {+/-}30/31, {+/-}365 days), highlighting persistent limitations in LLM temporal reasoning. These findings suggest that reliable temporal de-identification will require hybrid or multi-agent approaches beyond standalone LLMs.

Background: Identification of in-hospital cardiac arrest (IHCA) through manual chart abstraction is the gold standard, but its time-consuming and resource-intensive, which naturally limits its practicality. Diagnostic codes provide an accessible automated alternative, but prior work has shown this method of extraction has poor sensitivity and poor positive predictive value. More accurate and scalable automated methods are needed to support IHCA surveillance and research. Methods: We developed large language model (LLM)-based systems that incorporated physician-defined clinical knowledge through guideline-enriched prompting, supervised fine-tuning, and a self-correction procedure. The systems identified IHCA events and their locations from electronic health record notes. Performance was evaluated at the encounter level against physician-adjudicated chart review and compared with ICD-code-based identification. Results: Compared with an ICD-code-based algorithm (precision 0.68, recall 0.91, F1 score 0.78), the supervised fine-tuned LLM offered substantially better performance for IHCA identification on a curated evaluation corpus, reaching a precision of 0.88, a recall of 1.00, and an F1 score of 0.94. When tested on a larger, unselected validation cohort of 45,525 encounters, a separate LLM configuration - relying on guideline-enriched prompting and a self-correction procedure rather than fine-tuning - achieved a precision of 0.79 for encounter-level IHCA identification, meaning most encounters flagged by the model were genuine IHCA cases. The fine-tuned LLM underperformed on this cohort due to its inability to self-correct its decisions, an issue that can be mitigated by having a separate model perform the correction step. Conclusion: Explicit integration of clinical knowledge through prompting and supervised fine-tuning enabled accurate identification of IHCA events and locations from routine clinical documentation. In a two-pass workflow, these systems could prioritize encounters for clinician adjudication, reduce manual screening requirements, and support more scalable IHCA surveillance and registry development.

Objective: To develop and evaluate a framework for characterizing clinician editing of Artificial Intelligence (AI)-generated[KS1.1] documentation and assess its feasibility for health system-level monitoring of AI scribes. Materials and Methods: We analyzed outpatient encounters in which an AI scribe was used at a single academic health system, examining the History of Present Illness (HPI) and Assessment and Plan (A&P) sections of notes. We characterized edits on three dimensions[KS2.1][AG2.2]: lexical edit intensity based on Levenshtein distance, embedding edit intensity BERT[KS3.1]Score, and clinical edit intensity based on removed and added UMLS[KS4.1][AG4.2] concepts.[KS5.1][AG5.2] We analyzed the Positive Predictive Value (PPV) of clinical edit intensity as a measure of clinically meaningful editing using clinicians as the gold standard, examined correlations among dimensions, and designed exponentially weighted moving-average control charts to monitor longitudinal changes in clinician editing behavior.[KS6.1][AG6.2] Results: 268,379 encounters were included (267,654 HPI, 267,594 A&P). Clinical edit intensity [&ge;]1 had an 88.9% PPV for clinically meaningful editing. Lexical and embedding edit intensity were highly correlated (Spearman {rho} 0.95), while clinical edit intensity was less strongly correlated with both ({rho} 0.77-0.81). Longitudinal monitoring detected changes coinciding with system-wide rollout.[KS7.1][AG7.2] Discussion:[KS8.1][AG8.2] Clinicians edited A&Ps more heavily than HPIs, potentially reflecting greater attention to content involving clinical decision-making. Over one-third of sections involved clinical concept changes, and clinical edit intensity identified clinically meaningful edits while providing information complementary to lexical editing measures. Conclusion: Clinician editing can be characterized at scale using complementary editing dimensions, providing a scalable signal for post-deployment surveillance of the human-AI documentation process.

Longitudinal clinical cohorts contain heterogeneous progression patterns that are obscured by coarse diagnostic labels. Discovering these latent progression phenotypes requires models that distinguish patients not only by their outcomes, but also by how their clinical trajectories evolve. This paper proposes TRACE, an interpretable temporal phenotyping method that enables the learning of outcome-informed progression phenotypes from longitudinal trajectories. TRACE combines a trajectory- and outcome-aware contrastive objective with a probabilistic predictive phenotype layer, encouraging phenotypes to capture progression structure while remaining predictive of diagnostic state. A dual-attention mechanism further supports interpretation by identifying the visits and clinical variables most associated with each phenotype assignment. TRACE is evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort for cognitive disease modeling, showing improved performance in both diagnostic prediction and phenotype discovery compared with other temporal phenotyping baselines. Finally, by allowing the number of phenotypes to exceed the diagnostic labels, TRACE enables discovering additional progression patterns that are not captured by labels defined a priori.

ICD coding is the backbone of medical-record structuring, insurance payment, and clinical research. Today's best automatic coders are deep models, but they assume tens of thousands of labeled records, GPU servers, and machine-learning engineers. Many hospitals, especially small ones, have none of these. We study Chinese electronic medical records (EMRs) under a hard budget: 2,200 labeled records, CPU only, and fully interpretable decisions. Our method is lightweight and follows the way human coders work. First, codebook knowledge injection marks standard disease and operation names in the text as useful clues. Second, multi-view field weighting gives more weight to the clinical sections that matter most. Third, character n-gram TF-IDF features feed linear classifiers that score each candidate code. Fourth, a small set of clinical rules acts as a final safety check. Our main methodological finding is about this last step: naively applied rules do more harm than good (overall F1 drops from 0.8070 to 0.7995, and main-operation accuracy from 0.9727 to 0.9432). We therefore audit each rule on held-out data by counting how often it fixes a wrong prediction or breaks a correct one, and we remove rules that do more harm than good. This turns rule post-processing into a safety check that never hurts. On the CCL2026-Eval Task 6 dataset, our method reaches macro-F1 of 0.899 (main diagnosis) and 0.966 (main operation) in 5-fold cross-validation. A frequency-stratified analysis shows a clear long-tail limit: codes seen fewer than three times score zero F1. In a head-to-head test, a fine-tuned Chinese pretrained encoder scores lower than our method on all four tasks at this training size, while needing a GPU. Ablation tests show that the gains come mainly from field-weighted text. The framework trains in about 40 minutes on an ordinary CPU, and every prediction can be broken down into interpretable feature contributions.

INTRODUCTION: The digital clock drawing test (dCDT) is a cognitive screening tool capturing high-resolution pen-movement data. Existing approaches for classifying cognitive impairment from dCDTs rely on selected features or image-based methods. We propose representing dCDTs as structural networks to classify cognitive impairment and characterize spatial organization in clock drawings. METHODS: Weighted networks were constructed from 3,362 Framingham Heart Study participants, with and without cognitive impairment. Network representations were generated separately for command and copy conditions, with nodes representing pre-defined clock components (digits, hands, clock face) and edges capturing pairwise spatial relationships between components. Random forest models and graph convolutional networks were trained to predict cognitive impairment from the networks. Models were benchmarked against demographics-only and established summary feature baselines. RESULTS: Network-based models achieved AUC values comparable to the established summary feature baseline across command-only and copy-only conditions. Differences in spatial organization between the cognitively intact and impaired groups were greatest in the upper-right clock region, involving digits 1 and 2, the clock hands, and the surrounding clock face. Differences between the command and copy networks were larger among cognitively impaired participants. DISCUSSION: The proposed network representations provide an objective, scalable, and flexible framework for representing clock drawings for use in classification models. dCDT networks matched the predictive performance of established summary feature-based approaches and identified clock regions where spatial organization differed by cognitive status. After external validation, this approach has the potential to be incorporated within a clinical workflow to support automated and interpretable cognitive screening.

In hypertension screening, previously diagnosed cases provide reliable positive labels, whereas individuals without a prior diagnosis comprise both undiagnosed hypertensive and non-hypertensive individuals, creating a canonical positive-unlabeled (PU) learning setting. We propose PU-Boost, a two-stage reliable-sample reconstruction framework that uses only readily obtainable demographic, behavioral, medical-history, and anthropometric variables. During training, only previously diagnosed hypertension cases are treated as labeled positives (P), while all remaining individuals are treated as unlabeled (U); no additional reference labels are supplied to the two-stage reconstruction procedure. Elastic Net first provides a global ranking of P versus U, after which a random forest refines the uncertain region and highly ambiguous observations are rejected. XGBoost is then trained on the reconstructed class-balanced pseudo-labeled sample. In the independent test set, PU-Boost achieved a sensitivity of 0.817 and a balanced accuracy of 0.682, the highest observed values among the compared models. It produced 729 false negatives, 41.1% fewer than standard XGBoost, although specificity decreased to 0.547. SHAP analysis indicated that established hypertension-related features, including age, body mass index, waist circumference, and family history, contributed strongly to model output. PU-Boost therefore provides a statistically motivated framework for reducing missed cases when screening for undiagnosed hypertension under incomplete diagnostic labeling.

Background: Accurate patient understanding of lung cancer screening (LCS) results is critical for engagement and follow-up adherence, given the currently low screening uptake. While artificial intelligence (AI) systems show promise in facilitating patient communication, its effective use in clinical settings requires appropriate invocation of imaging tools and self-regulation by deferring certain questions to physicians. To this end, we developed and evaluated a planner-executor style multimodal agentic system to answer simulated patient questions about LCS CT results. Methods: This retrospective study utilized 116 LCS CT reports and images collected from a tertiary academic hospital. The system employed a multi-agent architecture (planner and executor) to mimic clinical reasoning. We developed a large language model (LLM)-based question generation framework to prepare a comprehensive set of 699 simulated patient questions, categorized as answerable and defer-to-doctor to assess self-regulation. Performance was evaluated on tool-calling, self-regulation accuracy (ability to correctly defer out-of-scope questions to a human provider) assessed by LLM-as-a-judge, and clinical quality assessed by a reader performance study with four physicians evaluating a subset of 100 responses. Results: The system demonstrated high overall tool-calling accuracy of 93.1% (646/694; 95% CI: 91.2%, 95.0%) and self-regulation accuracy of 92.6% (462/499; 90.2%, 94.8%). For answerable questions, the percentage of responses receiving perfect 5.0 scores across all readers included 78.9% (180/228) for clinical accuracy (inter-rater agreement: Gwet's AC2=0.92), and 73.2% (167/228) for patient understandability (0.86). For defer-to-doctor questions, the percentage of responses included 77.3% (133/172) for clinical accuracy (0.80) and 70.3% (121/172) for patient understandability (0.78). Conclusions: The planner-executor style multi-modal agentic system reliably answers patient-specific questions about LCS. Its high tool-calling and self-regulation performance demonstrates its potential as a safe and effective digital communication facilitator in LCS programs.

Accurate classification of lung adenocarcinoma growth patterns from hematoxylin and eosin (H\&E)-stained whole-slide images is essential for treatment planning and prognosis. We develop and evaluate an exponentially weighted ensemble deep learning approach that combines five architectures (EfficientNet-B3, DeiT3, Swin Transformer, Vision Transformer, ConvNeXt) with weights determined by a softmax transformation of per-model Quadratic Weighted Kappa (QWK) scores. Using 25,545 image patches at 40x magnification from the Dartmouth-Hitchcock dataset, we evaluated performance with stratified 5-fold cross-validation. Individual model performance ranged from Cohen's kappa kappa = 0.5563 (ConvNeXt) to kappa = 0.9561 (EfficientNet-B3). The ensemble achieved 96.52% accuracy and kappa = 0.9648, exceeding the best individual model by delta kappa = +0.0087 with statistical significance (McNemar's test, chi^2 = 50.84, p < 10^-12; bootstrap 95% CI on delta kappa [+0.0056, +0.0117]. Pairwise disagreement rates of 16.6--43.2% between architectures indicate that the models capture different aspects of tissue morphology, and the framework automatically down-weights unreliable components, as illustrated by ConvNeXt receiving only 5.7% ensemble weight despite a single-fold training failure. The approach offers a clinically grounded method for ensemble aggregation in digital pathology.

Discourse analysis can reliably predict real-world communicative success, yet is rarely implemented clinically, due to challenges in manually generating stimulus-specific main concept inventories (MCIs) and scoring patient narratives. We evaluated whether automatically derived macrolinguistic (main concept, sentiment) and microlinguistic (fluency, lexical, syntactic) features extracted from movie clip narrations can augment the clinical assessment of discourse in aphasia. The MCIs generated directly from each clip's video, audio and subtitles using a vision-language model approximated the main concepts healthy controls produced. Group classification (control: n=51; aphasia: n=54) based on the discourse features was excellent (AUC=0.95, 95% CI [0.90, 0.99]), with macrolinguistic features showing the largest group differences (main concept completeness: d=-1.84, 95% CI [-2.30, -1.38]) and strongest associations with aphasia severity (semantic distance to inventory: r=-0.79, 95% CI [-0.87, -0.66]). Findings support the feasibility and scalability of automated criterion-based discourse assessment using ecological stimuli, and highlight its potential for clinical implementation.

In recent years, the issue of reproducibility in scientific studies has been at the forefront of scientific discussion. Many of these issues arise due to poor employment of statistical methods, making it crucial to address these knowledge gaps for frequently used methods. Elastic net regression (ENET) is one such algorithm used frequently in biology and public health due to its ability to select important predictors (also known as "features") from large data sets. However, ENET struggles with consistent feature selection and hyperparameter choice in these applications due to low observation counts and noisy data. This limitation hinders ENET's effectiveness in biological research due to low model reproducibility and predictor interpretability. We propose the Instability Quotient Criterion (IQC) as a metric to quantify the consistency of selected features for feature selection models such as ENET, with a specific emphasis on biological applications. IQC leverages ensemble modeling and Shannon's entropy in order to determine the stability of selected features derived from ENET models. We demonstrate IQC's robustness via simulation studies, showing that IQC is capable of detecting significant decreases in model stability with lower observation counts, higher dimensionality, and higher proportions of noisy signal. We additionally utilize simulation to provide a naive threshold for low versus high stability. Finally, we applied IQC to a case study to illustrate how our method can be integrated into a larger analysis pipeline when looking at microarray gene expression data. IQC indicated stability issues for differentially-expressed genes when classifying acute leukemia subtypes in spite of high classification success. Lastly, we discuss how model instability impedes interpretation of biologically significant genes for discerning cancer subtypes. Overall, we present IQC as a robust evaluation metric that can easily be incorporated into high-dimensional biological analyses.

Tabular synthetic data generation (SDG) can facilitate research and development in the health domain, as access to health data is restricted. To accommodate for its structural complexity, such as multimodal distributions and fragmentation, transformer models provides an alternative to conventional SDG methods since they are better equipped to infer global spatial coherence due to textual context of tabular records. However, methods for optimising their adaption to tabular SDG are still largely unexplored. We contribute to the understanding of transformer-based SDG by evaluating fidelity and privacy of three prompting regimes, zero shot, one shot and few shot training, for tabular health SDG using small transformer models. We evaluated llama 3.2 (1B and 3B parameters) and Phi (1.3B and 2.7B parameters), on a real clinical dataset. Our results showed that zero shot learning was less prone to hallucinations and performed overall better for fidelity, but at the loss of privacy in comparison to the other regimes. Our results showed no differences between model sizes and families. Future studies should extend the analysis to other models and datasets, and explore strategies to reduce hallucinations.

BackgroundObstructive sleep apnea (OSA) is a highly prevalent sleep disorder. Major risk factors include obesity, older age, and male sex. Given the well-established association between OSA and cardiovascular disease, markers of cardiovascular health derived from electrocardiography (ECG) may help identify individuals with undiagnosed OSA. This study evaluated whether 12-lead ECG features improve machine learning-based prediction of OSA beyond established clinical factors. MethodsWe conducted a retrospective cohort study of individuals who underwent diagnostic sleep studies at the University of Kansas Health System Sleep Disorders Center (2004-2025) and had a 12-lead ECG acquired within one year. Standard ECG features (intervals, amplitudes, voltage-time integrals, morphology descriptors) were extracted from the representative beat. Latent representations were derived using two variational autoencoder (VAE) models pretrained on >1 million ECGs: a 30-dimensional representative-beat VAE and a 90-dimensional full 10-second ECG VAE. Three LightGBM models were trained: clinical factors only (age, sex, weight, height, body mass index [BMI], Epworth sleepiness scale [ESS]), ECG features only, and a combined model. Feature selection and hyperparameter tuning were performed separately for each model within a 5-fold cross-validation scheme. The outcome was moderate-to-severe OSA (AHI[&ge;]15 events/h), evaluated by AUROC (bootstrapped 95% CI) in a 10% holdout test set. ResultsThe dataset included 7,699 individuals (52.2% women, mean age 58.0 {+/-} 14.2 years, BMI 33.6{+/-}8.2 kg/m{superscript 2}, ESS 8.8{+/-}5.1; 36.4% with moderate-to-severe OSA) Training AUROC (95% CI) was 0.71 (0.66-0.76) for the combined model, versus 0.65 (0.59-0.70) for the clinical-only model and 0.66 (0.61-0.72) for the ECG-only model. Holdout test AUROC was 0.78 (combined), 0.69 (clinical-only), and 0.71 (ECG-only). Feature importance identified age, BMI, weight, ECG VAE latent variables, height, sex, and P-wave voltage-time integral as relevant contributor to the predictions. ConclusionsLatent ECG features extracted by VAE improve prediction of moderate-to-severe OSA when combined with clinical factors.

Deep learning models for electroencephalography (EEG) pathology detection achieve competitive accuracy but offer limited interpretability. In this paper, we introduce concept bottleneck models (CBMs) for EEG pathology detection which use clinical EEG biomarkers computed analytically from the signal as a bottleneck to be used for prediction. Those concepts include band power, spectral ratios, alpha peak frequency and amplitude, hemispheric asymmetry and phase locking values. With the backbone based on shallow convolutional neural network (CNN), we compare a black box baseline with two variants of CBMs. The hybrid variant is a classifier receives a residual projection of the backbone features. The pure variant is a CBM where the classifier receives only the predicted concepts. Models are evaluated on abnormality detection and dementia classification followed by concept prediction quality and leakage analysis. Both CBMs approach the baseline on TUAB (82.7-82.9% vs 83.6%) and are 2-5 percentage points lower on CAUEEG (57.9-60.5% vs 62.7%). The pure CBM exhibits concept fidelity concern where the predicted concepts deviate from their ground truth values. Replacing them with their true values causes severe cost on accuracy by 26 and 22 points on TUAB and CAUEEG. The concept fidelity concern is mitigated by increasing the weight of the concept loss to match the ground truth concepts. the mitigation comes with small accuracy cost on TUAB but a relatively larger cost on CAUEEG. The code of the experiments and full logs are available at https://www.github.com/mhmdrdwn/cbm.

Explicit timing and context features may complement short electrocardiogram waveforms, but aggregate gains on imbalanced benchmarks can conceal errors in individual classes. We therefore compared Logistic Regression, Random Forest, XGBoost, and a one-dimensional convolutional neural network (CNN) in four cumulative input stages. Stage S1 used a 255-sample beat waveform alone, and stages S2 to S4 added 6, 13, and 21 engineered features. Using the inter-patient division of the MIT-BIH Arrhythmia Database, with 50,992 development and 49,686 evaluation beats in four classes, we estimated 95% intervals with a paired bootstrap over the 22 evaluation recordings. Adding six RR interval features increased macro F1 by 0.025 to 0.093, with intervals excluding zero for every model family. Further context helped Random Forest and XGBoost, whereas later CNN gains fell within resampling uncertainty. At the final stage, XGBoost and CNN reached macro F1 of 0.580 and 0.574, and the interval for their difference included zero. Despite opposite operating points, their precision for fusion beats was similar at matched recall, suggesting that neither separated fusion beats better. The most accurate pipeline, Random Forest at S2 with 94.45% accuracy, detected only 1.4% of supraventricular beats, and development and evaluation favored the same stage for only one family. The results for the minority classes were based on single recordings. Record 213 supplied 203 of the 207 correct CNN fusion detections, and elsewhere both final models detected at most 4 of 26 fusion beats. Record 232 contains 75% of the evaluation supraventricular beats, and nearly all were missed. Excluding this record nearly doubled XGBoost supraventricular F1, from 0.199 to 0.383. These findings are limited to retrospective analysis with annotated beat positions and a single training seed.

Background: Ambient artificial-intelligence (AI) scribes are being deployed beyond medicine into allied-health practice, but comparative evidence mainly concerns physician documentation. Objective: To compare quality and transcript-grounded clinical-risk error burden of AI-generated and allied health professional notes from identical simulated consultations. Methods: Paired simulation across dietetics, physiotherapy, occupational therapy and speech and language therapy (five consultations each; 20 pairs; 40 notes). Heidi generated an unedited first-draft note and a professional documented each encounter. Professionals rated PDQI-9 quality. Professional and automated review identified unsupported statements and omissions; two language models graded clinical risk. Co-primary outcomes were PDQI-9 score and Critical+High error density. Results: Mean PDQI-9 scores were 38.45 and 35.05 for AI and professional notes (paired difference +3.40, 95% CI +0.01 to +6.79; Wilcoxon p=0.042; dz=0.47). The Bayesian estimate was +3.47 (94% equal-tailed credible interval +0.66 to +6.30; P[AI>professional]=0.989). Automated-reviewer all-error density was 3.35 vs 8.55 per note (professional-minus-AI difference 5.20, 95% CI 3.85 to 6.55); omission density was 1.45 vs 5.20 (difference 3.75, 2.80 to 4.65). Critical+High density was 0.35 vs 0.75 (difference 0.40, -0.10 to 1.00). The pooled quality result was driven principally by dietetics. Conclusions: AI-generated notes achieved modestly higher pooled quality and fewer transcript-grounded errors, chiefly omissions; serious-error evidence was inconclusive. Small discipline samples and incomplete ground truth preclude safety or discipline-wide superiority claims.

ObjectivesTo determine whether single-layer evaluation characterises clinical risk in the final note of a multilingual ambient AI scribe, and where serious errors arise. DesignTwo-arm evaluation on one common clinical-risk scale, using a frozen, reference-aligned synthetic corpus. SettingScripted consultations spanning 77 languages and five complexity levels, from simple general practice to expert multidisciplinary-team handover. Main outcome measuresPer-session incidence of at least one serious (HIGH or CRITICAL) note-layer error in intrinsic (reference script to note, n=385) and end-to-end (automatic speech recognition (ASR) transcript to note, n=2,302) generation; three LLM raters classified discrepancies using a four-mode taxonomy. ResultsIntrinsic note generation had a 3.4% serious-error rate, with no detectable gradient across complexity (1% to 6%; Cochran-Armitage p=0.42) or language resource (high 2%, medium 5%, low 4%). End-to-end serious errors rose steeply with complexity (L1 1% to L4/L5 24%) and language scarcity (high 7%, medium 10%, low 15%; OR 1.54 per tier, 95% CI 1.15-2.07; p=0.004, GEE clustered on language). Of serious in-note errors, 87% were ASR-derived and 7% note-originated; amplification exceeded correction threefold (fate entropy 1.19 bits). Repeated generation showed moderate reproducibility (Fleiss kappa 0.42); word error rate explained 24% of cross-language variance versus [~]1% for transcription risk density. ConclusionsLow serious-error rates in individual layers did not preclude higher end-to-end note risk. Evaluation should therefore include the clinician-facing note, because component-level metrics cannot fully capture risk created or transformed at the transcription-to-generation interface. What is already known on this topicO_LIAmbient AI scribes are being deployed globally, and safety evaluation has characterised the transcription layer in isolation, most commonly with word error rate C_LIO_LIFrequency-based transcription metrics such as word error rate do not directly encode clinical consequence, and component-level performance may not represent the risk of the final generated note C_LIO_LILarge language models produce clinically relevant output stochastically, so identical prompts can yield materially different content C_LI What this study addsO_LINote generation from a perfect transcript had a low observed serious-error rate and no detectable complexity or language-resource gradient, whereas the end-to-end system showed steep gradients on both axes. C_LIO_LISerious note-layer errors were predominantly ASR-derived, but note generation transformed upstream errors variably, amplifying serious errors about three times more often than it corrected them C_LIO_LITranscription-layer metrics had limited predictive value for note-layer risk after accounting for complexity and language-level dependence; word error rate explained approximately one-quarter of cross-language variance C_LI How this study might affect research, practice or policyO_LIEvaluation and procurement of ambient scribes should include end-to-end assessment of the clinician-facing note, with results stratified by consultation complexity and language where relevant C_LIO_LIA three-surface root-cause alignment using a four-mode root-cause taxonomy (note-originated, propagated-ASR, amplified-ASR and silent-correction) can attribute documentation errors to their point of origin or transformation C_LIO_LILanguage-related performance should remain a surveillance target because disparities can emerge after components are composed, even when corresponding gradients are not detectable in an individual layer C_LI