npj Digital Medicine

Clinical records routinely document medical interventions but often omit the reasoning connecting patient-specific evidence to medical action. This limits the decision-relevant supervision available for training models for personalized care. Here we evaluate a data-augmentation approach that makes selected clinical decision relationships explicit in an 8-billion-parameter Qwen3 model, ChronIQ-DRx2. Using the same training cases and model architecture, expert-guided logic reconstruction was applied across four chronic diseases. On the held-out evaluation, the enhanced model had the highest plotted end-to-end task-completion rate among evaluated systems and higher plotted scores across diagnostic, medication and follow-up measures. The serious clinical error rates shown were 16.7% for the baseline and 12.5% for the enhanced model. In the human-AI evaluation, physician accuracy was 87.40% alone and 88.20% with AI assistance, while mean completion time was 0.959 and 0.780 min, respectively; medical-student accuracy was 62.71% and 74.80%, and mean completion time was 0.210 and 0.195 min. These results are descriptive point estimates; inferential comparisons are not reported here. Structured outputs may make recommendations easier to review, but rationale faithfulness, generalizability and effects on clinical safety require further evaluation.

Objective: Case-finding models for underdiagnosed conditions are usually trained on diagnosis codes, which are missing from the records of many patients with a clinical diagnosis. We asked whether models trained on self-reported diagnoses rank uncoded self-reporters higher than models trained on later codes, and how a score fit only from released summary counts compares with a participant-level model. Methods: In the All of Us Research Program, we followed survey respondents with no qualifying diagnosis code by the survey date for later codes. For chronic fatigue, fibromyalgia, and endometriosis (participants recorded as female at birth), we fit a naive Bayes score with empirical-Bayes shrinkage, from released feature-by-label counts for the self-reported diagnosis and from exact counts inside the data enclave for a later code, and an elastic net on participant-level data for each label. In a random 20% validation split, we measured discrimination of self-report, of uncoded self-reporters (no qualifying code observed), and of later codes. Results: For chronic fatigue (149,860 respondents, including 7,770 self-reporters, 95% with no code observed), the summary-count score trained on self-report ranked uncoded self-reporters with an area under the receiver operating characteristic curve (AUROC) of 0.74 (95% CI, 0.72-0.75), against 0.67 (95% CI, 0.65-0.69) when trained on later codes. For later codes, the two scores were similar (0.69 vs 0.70). Fibromyalgia showed the same pattern (0.73 vs 0.64 and 0.74 vs 0.72). Elastic nets trained on self-report ranked uncoded self-reporters higher than the summary-count scores (0.80 vs 0.74 and 0.78 vs 0.73) but discriminated later codes less well than code-trained elastic nets. For endometriosis, the self-report score also ranked uncoded self-reporters above the code-trained score (0.61 vs 0.53), but discrimination was low and it did not discriminate later codes (0.50 vs 0.70). Conclusion: The choice of training label changes which patients a case-finding model ranks highly. For all 3 conditions, self-reported diagnoses were the more useful label for finding uncoded self-reporters. For chronic fatigue and fibromyalgia, the self-report score also discriminated later codes as well as the code-trained score. Independent clinical validation is needed before these models guide case finding.

Background: Hospital readmission is a common and costly outcome in patients with chronic obstructive pulmonary disease (COPD). Multimodal deep learning models combining structured electronic health record (EHR) data with physiological, imaging, and textual data have been proposed to improve its prediction, but whether such fusion outperforms EHR data alone, once temporal-leakage safeguards are applied, remains unclear. Methods: Using a stratified MIMIC-IV cohort of COPD patients enriched to about 50% readmission (natural prevalence 18.1%; n=1,279; 895/192/192 training/validation/test), we predicted 30-day, all-cause readmission from four frozen, pre-trained encoders (electrocardiogram, chest radiograph, clinical notes, structured EHR). We audited and corrected temporal leakage in the admission-linkage pipeline of each modality, then compared three fusion architectures (naive concatenation, scalar gated fusion, EHR-anchored residual fusion) against unimodal baselines and an EHR-alone reference, using bootstrap significance testing, a missingness stress test, and calibration analysis with and without post-hoc recalibration. Results: The audit corrected three previously undetected temporal-alignment issues, affecting 5-9% of patients for clinical notes, 29-43% for chest radiographs, and 53-58% for ECG. After correction, every modality except EHR showed near-chance discrimination alone (ROC-AUC 0.46-0.53), confirmed with a non-linear adapter probe. No fusion architecture significantly outperformed the EHR-alone baseline (ROC-AUC 0.63); removing auxiliary modalities improved rather than degraded fusion performance under stress testing. The raw Brier score was lower for all three fusion architectures than for EHR-alone (about 0.24 versus 0.32), but this reflected overconfidence of the EHR-alone probe (calibration slope 0.12); after logistic recalibration fitted on the validation set, Brier scores were indistinguishable (EHR-alone 0.239; fusion 0.240-0.244; all paired bootstrap 95% CIs included zero). Conclusions: Structured EHR data alone matched the discrimination and, after recalibration, the calibration of more complex multimodal fusion for 30-day COPD readmission prediction. These findings highlight the importance of rigorous temporal-leakage auditing in multimodal clinical prediction research and suggest that reformulating this task around temporal proximity to the outcome, rather than refining the fusion mechanism further, may be more productive for future work.

High attrition rates across clinical drug development impose immense clinical and financial burdens, yet prospectively forecasting whether an individual trial will achieve its primary endpoint remains challenging. Here we present a machine-learning system that predicts clinical trial success across phases I-IV, independently modeling studies with efficacy versus safety/toxicity primary endpoints. To prevent lookahead bias and outcome leakage, models are trained on registration-time features, including target and mechanism-of-action precedent, drug modality and novelty, trial design and operational characteristics, and frozen literature-derived target-choice embeddings. Each prediction yields a continuous Predictive Trial Index (PTI) between 0 and 1 reflecting success probability. On a test set of trials initiated in 2021 or later, the efficacy model achieved an area under the receiver operating characteristic curve (AUROC) of 0.828 (95% CI 0.812-0.845; n = 2,347) and the safety/toxicity model reached 0.809 (95% CI 0.787-0.831; n = 1,485). These models are embedded in inClinico Agent, an agentic AI system integrating frontier large language models (LLMs) with deterministic scoring engines, retrieval tools and specialized clinical skills. We also introduce ClinBench, a prospective benchmark of 2,040 ongoing trials evaluated by inClinico Agent and five frontier LLMs, establishing a leakage-free foundation for tracking prospective predictive accuracy as trials reach completion.

Large language models (LLMs) are increasingly evaluated for clinical tasks, but safe use requires reliable uncertainty estimates as well as high accuracy. We tested eight general-purpose LLMs on 100 validated neurology board-style multiple-choice questions and compared three confidence signals: self-reported confidence, critique-implied confidence from an independent model instance, and consistency-based confidence from repeated sampling. Accuracy varied from 45% (Llama 3.1 8B) to 92% (Gemini 3 Flash/Pro), but confidence was generally overestimated, particularly in lower-accuracy models. Consistency-based confidence provided the best discrimination between correct and incorrect answers (pooled AUROC 0.74), outperforming self-report, although calibration differences between methods were not significant. Calibration should therefore be reported alongside accuracy when clinical LLMs are evaluated.

Background. Research data enclaves such as the All of Us Research Program do not allow participant-level data, or models fitted to them, to leave, which makes it hard to deliver case-finding models, which flag patients with possibly uncoded conditions for clinician review, to patients and clinicians. Objectives. We describe and technically evaluate a system that fits models outside an enclave from released summary counts and delivers them through standard health information technology interfaces. Methods. We exported dissemination-compliant summary counts from the All of Us Researcher Workbench and, outside the enclave, fitted versioned models for Long COVID, chronic fatigue, fibromyalgia, endometriosis, Lyme disease, lupus, migraine, depression, and anxiety. A model contract (a model file and a written specification), a Python reference implementation, and synthetic Fast Healthcare Interoperability Resources (FHIR) test patients define the expected outputs. One JavaScript scorer runs in a patient demonstration, a SMART on FHIR clinician application, and a CDS Hooks service. An atlas displays aggregate results. We evaluated the system on synthetic records with automated tests, cross-implementation output comparison, launches through the public SMART sandbox, and a counterexample review of the scoring and display rules. Results. The deployed Long COVID model had an area under the receiver operating characteristic curve of 0.67 in a random validation split. The eight self-report models reached 0.61 to 0.82 in identifying self-reporters without a code. The JavaScript scorer reproduced all 162 reference outputs exactly, and all 437 automated tests passed. A counterexample review found 5 classes of failure, including scoring ineligible patients and treating unavailable data as absent, all fixed and added as tests. Conclusions. Models fitted from released summary counts can be delivered through FHIR, SMART on FHIR, and CDS Hooks with reproducible scoring on synthetic records. Real-record interoperability, usability, and prospective clinical performance remain to be evaluated.

BACKGROUND Timely cause-of-death (COD) information guides health policy, yet many deaths in low- and middle-income countries are not medically certified. In Africa, fewer than 10% of deaths are. Verbal autopsy (VA) yields a structured symptom questionnaire and a free-text narrative. Mortality surveillance increasingly records the narrative alongside the questionnaire, but the algorithms used to produce VA-based statistics were designed for structured responses only, so narratives inform physician COD assignment but do not contribute to algorithmic COD assignment. We evaluated whether supervised models could use both modalities to improve COD assignment. METHODS We analyzed 4,999 VA records for deaths at ages 12 years or older from a national sample of registered deaths in South Africa (RSA), using physician-assigned CODs as the reference standard. We trained unimodal and multimodal pipelines on the labelled records: pretrained language models (PLMs) fine-tuned on narratives alone, machine learning (ML) models on structured responses alone, and the two combined at the data, feature, and decision levels. We also prompted a locally deployed generative large language model (LLM) on the same unimodal and multimodal inputs, without parameter updates. All were compared with InSilicoVA, the operational comparator representing current automated practice. We evaluated individual-level assignment, population-level agreement, calibration, and generalization across provinces in RSA, reporting these metrics rather than ranking pipelines on any one of them. RESULTS Every supervised pipeline improved on InSilicoVA by 8.5-15.9 percentage points in top-1 accuracy and 3.5-7.1 points in cause-specific mortality fraction (CSMF) accuracy. The narrative and structured pipelines performed similarly in aggregate but differed by cause: 63.9% of deaths received identical predictions, 11.6% were correct only from the narrative pipeline, and 13.6% only from the questionnaire pipeline, with the narrative pipeline more accurate for noncommunicable causes and the questionnaire pipeline for communicable causes. Against the unimodal approaches, the decision-fusion pipelines added 4.2-5.4 percentage points in top-1 accuracy, concentrated in the most common causes, whereas feature fusion with cross-modal attention improved the less common causes, adding 6.6 points (95% CI, 1.7-11.5) in balanced accuracy and 5.5 points (95% CI, 0.6-10.3) in macro F1. Feature fusion reached the highest CSMF accuracy (95.7%). Paired differences from the better unimodal representative ranged from -0.2 to +3.4 points, with all intervals including zero. The LLM exceeded InSilicoVA on individual-level metrics and reached the highest balanced accuracy of any pipeline (60.6%), while the supervised pipelines retained higher CSMF accuracy. Supervised pipelines had lower expected calibration error than InSilicoVA. Supervised gains held across all nine held-out provinces, including every top-1 comparison and nearly every macro F1 and CSMF comparison. CONCLUSIONS PLMs using narratives alone substantially improved on current best practice and matched ML pipelines using structured responses, while multiple multimodal strategies produced further gains in individual COD assignment. The zero-shot LLM showed particular strength on rare causes, whereas supervised pipelines better recovered the population cause distribution. These findings support incorporating narratives as a core input to automated VA systems and evaluating representative fusion strategies for the intended output. External validation across settings and age groups will support broader implementation.

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.

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.

Large language models (LLMs) are increasingly consulted for health advice, but they can produce fluent statements that are false or unsupported, known as hallucinations, and this risk is poorly characterised in low-resource languages such as Bangla. We measured hallucinations of locally deployable open-weight LLMs in Bangla diabetes question answering and tested whether retrieval-augmented generation (RAG) grounded in Bangladeshs National Guideline on Diabetes Mellitus reduces them. Forty patient questions spanning six clinical sub-domains were each answered twice, under zero-shot prompting, by Qwen3.5-9B, TigerLLM-9B, Qwen3.6-35B-A3B and Gemma-4-26B-A4B, and by BanglaDiaRAG, which adds hybrid dense (BGE-M3) and sparse (BM25) retrieval from a physician-reviewed Bangla guideline knowledge base to Qwen3.5-9B. Two physicians rated the 400 de-identified responses by consensus for medical accuracy, completeness, safety, readability and reliability on a 0-1 scale, and responses with accuracy below 0.70 were classified as hallucinated. Hallucination rates were 2.5% for Gemma-4-26B-A4B, 3.8% for TigerLLM-9B, 22.5% for Qwen3.6-35B-A3B and 77.5% for Qwen3.5-9B. Guideline grounding lowered the rate of Qwen3.5-9B to 11.3% (difference -66.3 percentage points; 95% CI -76.3 to -55.0) and raised every criterion by 0.24 to 0.38 points (Holm-adjusted p < 0.001). BanglaDiaRAG outscored Qwen3.6-35B-A3B on all five criteria and matched or exceeded TigerLLM-9B, but it remained below Gemma-4-26B-A4B in safety, readability and reliability, and its residual errors were specific, confident and potentially harmful. Guideline grounding can turn a weak general-purpose model into a much safer Bangla diabetes assistant, but the choice of base model mattered as much as retrieval and no configuration was free of hallucination, so clinical oversight remains essential before patient-facing use.

Background: Artificial intelligence (AI) may improve electrocardiographic differentiation of Takotsubo syndrome from acute anterior ST-segment elevation myocardial infarction (STEMI), but its impact on human diagnostic performance, the added value of visual explanations, and differences by reader background and experience remain uncertain. Methods: Using 12-lead electrocardiogram (ECG) images from two hospitals, we developed and externally validated an ensemble of three deep learning models after self-supervised pretraining and fine-tuning. The diagnostic task used ECGs recorded within 14 days after symptom onset, regardless of reperfusion status. We conducted a four-phase reader study. Phases 1-3 compared unaided interpretation, a reference sheet, and additional AI assistance. In Phase 4, readers evaluated a new ECG set after randomization to reference-sheet support alone, with AI predictions, or with AI predictions plus gradient-weighted class activation mapping (Grad-CAM) heatmaps. Diagnostic correctness was analyzed using generalized linear mixed-effects models. Exploratory analyses examined differences by cardiologist status and experience. Results: Development and external-validation cohorts, including healthy controls, comprised 832 and 95 participants, respectively. For Takotsubo-STEMI discrimination, external-validation accuracy was 0.914, with areas under the receiver operating characteristic and precision-recall curves of 0.970 and 0.984. Seventy healthcare providers completed the reader study. The estimated probability of correct diagnosis increased from 0.632 without assistance to 0.864 with AI assistance (odds ratio [OR], 3.71; P<0.001). In randomized Phase 4, AI assistance improved accuracy vs reference-sheet support alone (OR, 4.28; 95% CI, 3.08-5.95; P<0.001), and Grad-CAM provided additional benefit (OR, 1.53; 95% CI, 1.04-2.25; P=0.033). Exploratory analyses showed greater improvement among less experienced readers; with AI assistance, non-cardiologists outperformed cardiologists. AI-assisted readers remained less accurate overall than the model alone. Conclusions: An ECG image-based ensemble differentiated Takotsubo syndrome from anterior STEMI and improved reader accuracy. Visual explanations provided additional benefit, while experience-related differences highlight the importance of effective human-AI interaction.

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.

The explosive growth of ambient AI scribes in healthcare has focused almost exclusively on relieving clinician documentation burdens and enhancing clinician efficiency and reducing burnout. While most research has focused on the clinician experience, these same types of tools, such as a patient held AI Personal Health Assistant, can be put in the hands of patients to facilitate communication with healthcare providers, to provide greater autonomy and activation for patients in managing their own healthcare, and to facilitate the task of receiving and retaining complex new information. However, to ensure these tools serve as more than just administrative aids, we partnered with a large technology company to conceptually codesign how such a tool might be studied in real world settings in a way that reflected outcomes of greatest importance to patients and families. We used the Nominal Group Technique and engaged thirty three patients and caregivers and ten multidisciplinary clinicians in structured virtual sessions to identify the outcomes on which subsequent studies should hinge. Our findings reveal a significant misalignment between current industry trends and patient priorities. While the market emphasizes the scribe function for billing, stakeholders overwhelmingly prioritized the AIPHAs role in longitudinal care management and post visit coordination. The most critical outcome identified was the impact on communication and respect, suggesting that patients view AI as a mechanism to improve the human to human dynamic and ensure a shared understanding of the care plan. Our findings further highlighted that the primary hurdle for the use of patient held AI is not technical utility but clinician trust and the need for formal consent processes. Importantly, as this technology is new and integration into the clinical experience is still being explored, patients cannot assume they will have access to the transcript of their own encounter when that record is created and held by the health system. This gap highlights the potential value of a patient held tool, preserving discussions from their visit and follow up steps in clearer detail than they may receive in their clinical note or traditional visit summary. From a system perspective, this transition requires a strategic shift from measuring workflow focused efficiency to measuring patient side advocacy and engagement. By centering the evaluation of AI on these patient defined metrics and fundamentally designing subsequent AI investigations using a partnership model, health systems can move beyond simple automation toward a model of care delivery that empowers patients to navigate complex clinical journeys with greater autonomy and clarity.

Biomedical artificial intelligence (AI) increasingly relies on heterogeneous data sources, yet differences in data structure, terminology, evidence quality, and provenance complicate their integration into reliable knowledge resources. Here, we develop an evidence-aware framework for processing and integrating heterogeneous pharmacovigilance data from FDA Adverse Event Reporting System (FAERS) reports, biomedical literature, clinical trials, and electronic health records in Mass General Brigham. The framework combines large language model (LLM)-based information extraction, ontology-driven semantic normalization, cross-source evidence assessment, statistical and machine learning-based prioritization, and ingredient-based adverse drug event (ADE) relationship inference to transform heterogeneous evidence into a structured and evidence-based knowledge resource. The constructed knowledge system contains 11,588 drug nodes, 44,843 ingredient nodes, 3,275 ADE nodes, 993 indication nodes, and more than 2.47 million relationships, with each drug-ADE relationship assigned a confidence score. We further developed a drug-level web interface for accessing the knowledge resource and validated its utility in downstream AI tasks. Specifically, incorporating the developed knowledge resource improved the comprehensiveness, relevance, and overall quality of LLM-based drug-specific ADE synthesis and improved the precision and F1 score of clinical ADE relation extraction. These results demonstrate that AI-assisted processing combined with evidence-aware curation can transform heterogeneous biomedical evidence into reusable, AI-ready knowledge with explicit provenance and evidence strength information. The framework provides a general strategy for integrating heterogeneous biomedical data to support reproducible knowledge construction and downstream biomedical AI applications.

Open-loop insulin pump therapy depends on clinical expertise for dose initiation and titration. We developed a long short-term memory-based system, using 1738 treatment courses from 1641 patients with type 2 diabetes for initial prescribing and daily basal and bolus titration. Patient-grouped five-fold validation showed closer agreement with recorded prescriptions for subsequent titration than for initiation. Prospective evaluation in 97 patients from the model-directed arm of a randomized trial yielded a median derived time in range (3.9-10.0mmol/L) of 84.5%, with low hypoglycemic exposure. Favorable control extended to previously treated patients. Higher mean glucose during the first 24 hours was associated with poorer glycemic control thereafter. The association between distance from development data and glycemic outcomes was stronger when early glucose responses were included alongside baseline characteristics. These findings support model-directed pump dosing and highlight early treatment response as a clinically relevant signal for assessing applicability and identifying patients who may require additional oversight. Trial registration: ClinicalTrials.gov, NCT07066891.

Background: Conventional performance metrics do not indicate whether artificial intelligence (AI)-platforms and radiologists are right or wrong on the same examinations, nor whether their non-agreement is more often beneficial or harmful. Objective: To develop, illustrate, and apply directional metrics quantifying non-agreement between radiologists and AI-platforms against a common reference standard, and to adapt a fragility-based framework for quantifying the statistical robustness of these metrics. Methods: We defined Positive Non-Agreement (PNA, AI-platform correct, radiologist incorrect), Negative Non-Agreement (NNA, radiologist correct, AI-platform incorrect), and Net Non-Agreement (NENA, their difference). Metrics were illustrated in three hypothetical scenarios of 1,000 examinations each and applied to three previously published AI-radiologist studies on pulmonary embolism and mammography. Bootstrap confidence intervals were calculated, and a reclassification fragility index (FI) quantified the minimum case reclassification needed to remove statistical support for NENA's direction. Results: In illustration, NENA was positive in two scenarios and negative in one. In the three published studies, NENA ranged from -4.18% to +10.74% and decreased consistently with increasing radiologist experience (junior residents, +6.27%; consultants, +1.45%), a gradient driven primarily by declining PNA with modest increase in NNA. The reclassification FI was defined for all three illustrative scenarios (range, 0.10%-1.70%) and for six of the seven comparisons from these studies (range, 0.34%-4.31%), but not for the weakest comparison, whose confidence interval for NENA included zero. Conclusions: PNA, NNA, and NENA provide a directional, case-level complement to conventional performance metrics for AI-radiologist non-agreement. The accompanying reclassification FI offers a statistically grounded, generalizable method for quantifying their robustness.

Neurologic autoantibody cell-based assay images are difficult to interpret when fluorescence is weak, focal or dominated by background signal. We developed a visual artificial intelligence overlay that transfers detected cell boundaries from the structural channel to the antibody-staining image and evaluated standalone model concordance and its association with reader classifications. The analysis used 2 763 records for model development and 499 held-out records spanning 14 antibody targets. Four readers assessed the held-out records first without assistance and then with the boundary overlay and model category available. The selected model agreed with the reader-derived three-class reference in 85.2% of records, but recall was lowest for weak-positive or equivocal records. Reader agreement increased in the assisted session, and almost all revised ratings moved towards the model category. The combined display may support consistent visual review while also creating a risk of model-directed anchoring; its role as a second-read aid requires independent validation.

Purpose To evaluate the external transportability of source-derived probability calibration for chest radiograph deep learning classifiers under distribution shift and determine its dependence on sample size and disease prevalence. Materials and Methods This retrospective secondary computational analysis of public, de-identified radiographic datasets involved no participant contact or new data collection. Cohort sizes were prespecified by availability without prospective power calculation: 106,686 NIH ChestX-ray14 images (29,323 patients; 1992-2015) for development and internal testing, 20,000 CheXpert images (20,000 patients; October 2002-July 2017), and 14,998 VinBigData images (2018-2020) for external testing. ResNet-18, DenseNet-121, and DeiT-Tiny classifiers for cardiomegaly and pleural effusion were trained across four patient fractions and three seeds (36 fits). Calibrators were fitted exclusively on NIH data. Twelve primary external Brier-score contrasts (temperature-scaled versus raw probabilities) were evaluated using paired cluster-bootstrapping with Holm adjustment. Results Temperature scaling reduced Brier scores in nine of 12 primary contrasts but worsened pleural effusion calibration on VinBigData across all architectures (Brier increases: ResNet-18, 0.001460 [95% CI: 0.001347, 0.001568]; DenseNet-121, 0.001193 [95% CI: 0.001109, 0.001274]; DeiT-Tiny, 0.000498 [95% CI: 0.000417, 0.000577]; all adjusted P=.006). Standardizing class weights to NIH prevalence reversed all CheXpert improvements. Although full versus 10% training improved external AUROC across all comparisons, DeiT-Tiny CheXpert pleural effusion Brier score worsened by 0.007039 despite a 0.032466 AUROC gain. At 80% selective image retention, calibrated ensembles retained only 12.3%-17.2% of VinBigData positive pleural effusion examinations. Conclusion Source-fitted calibration is not uniformly transportable across health systems, and discrimination gains do not guarantee calibrated probabilities. External evaluation must simultaneously assess probability reliability, prevalence shifts, and retained-case composition.

Cervical cytology screening relies on the visual assessment of nuclear and cellular morphology across thousands of cells per slide, a labor-intensive task that could benefit from quantitative computational support. Here we present CytoLens, a modular deep learning system that detects and counts nuclei, delineates nuclei and whole cells, measures the nuclear-to-cell area ratio (N:C), classifies cells from multifocal images and scores candidate abnormality. We trained and tested each module on a public dataset matched to its task, without pooling datasets. CytoLens detected nuclei with an F1 score of 0.931 (95% confidence interval (CI), 0.914-0.946) and counted 75 of 100 fields to within one nucleus. Point-prompted segmentation achieved mean Dice coefficients of 0.901 for nuclei and 0.847 for whole cells and, without retraining, transferred to a second dataset in which measured N:C increased with lesion grade in parallel with expert outlines. Multifocal hierarchical classification reached a mean accuracy of 0.947 across four Bethesda-aligned categories, and abnormality scores ranked annotated abnormal candidates with an area under the receiver operating characteristic curve (AUROC) of 0.865. These results support the evaluation of CytoLens on patient-linked clinical specimens.

Background and Objective: The volume of patient portal messages has grown substantially and artificial intelligence (AI) can support clinical staff and improve the timeliness of patient care. Our objective is to test if an AI tool for managing the electronic inbox In-Basket reduces response times to patient portal messages in neurology. Methods: We conducted an implementation study from January to May 2026 in the neurology department at two clinical sites within a regional medical system. A single AI tool was implemented to auto-categorize and auto-route all patient messages received in the electronic inbox. The tool also generates optional draft replies to messages. Primary process measures were the median weekly time (hours) from message creation to the first action and to closure. Impact was analyzed using a controlled interrupted time-series (ITS) design, comparing message-volume-matched control divisions. Results: Among 24,142 messages analyzed, pre-intervention median weekly time to first action was 11.3 hours [IQR: 5.6, 15.1] (Site A) and 16.1 hours [IQR: 13.8, 19.0] (Site B), falling post-intervention to 1.0 hour [IQR: 0, 3.1] (Site A) and 0.6 hours [IQR: 0.2, 2.4] (Site B). Controlled interrupted time-series analysis showed a significant immediate differential level change in median weekly hours to first action at Site A (-92.5%, 95% CI: -97.5% to -78.0%), and a significant downward trend at Site B (-43.5%, 95% CI: -65.2% to -8.2%). Changes in time to closure were directionally favorable but not statistically significant at either site. There was low use of the generated draft replies, only 0.5% at Site A and 3% at Site B. Discussion: Implementation of AI-assisted message categorization and routing was associated with a substantially reduced initial action time for patient In-Basket messages in Epilepsy. The utilization of AI-generated draft-replies to messages was low. These findings support the feasibility and effectiveness of AI-augmented workflows to improve inbox management.