npj Digital Medicine

Clinical foundation models learn patterns in longitudinal health records that can be reused across prediction tasks. However, the relative strengths of representation-based and generation-based prediction across clinical tasks remain unclear. Using longitudinal electronic health records from 3.77 million patients at a single hospital, we developed FERMAT (Foundation model for Exploring Real-world Multimodal health data using Autoregressive Trajectory modeling), a 65-million-parameter clinical foundation model. We compared predictions from task-specific predictors trained on patient representations with those derived from generated clinical event sequences. Representation-based prediction performed better in the disease-risk comparisons. In detailed analyses of five-year chronic kidney disease, fatty liver, and diabetes risk in a test cohort of 10,000 patients, representation-based predictors exceeded generation-based estimates by 0.098-0.115 in observation-weighted AUROC and had lower probability error. Generation-based prediction, in turn, performed better in predicting diagnoses, prescriptions, and procedures recorded on the same day, a task motivated by the frequent presence of multiple events with the same date in hospital EHRs. Using 50 generated trajectories per patient, it achieved a Recall@10 of 71.0%, compared with 55.7% for representation-based prediction, with a paired improvement of 15.31 percentage points (95% CI, 13.92-16.79; n = 1,862). These findings demonstrate task-dependent strengths of representation-based and generation-based applications within a single clinical foundation model, providing an empirical basis for exploring how the two approaches can be used individually or combined to address clinical prediction needs.

Digital phenotyping, which is defined as quantifying someone's behaviour with digital devices, provides unprecedented opportunities for understanding human mental health but is hampered by high levels of inter-individual variability. Here, we propose a new method to address this, parsing inter-individual variability by decomposing the digital phenotype dynamics into latent trajectories and using each individual's trajectory membership as a moderator when modelling psychopathology over the same timeframe. We applied our method in the context of mood symptom exacerbation across the menstrual cycle, where symptom severity and timing are inconsistent between individuals. Using the BiAffect platform to collect smartphone typing dynamics, we found stable trajectories in smartphone movement rate: one group of participants showed substantial movement rate fluctuations across the menstrual cycle, whilst the others did not. Participants with movement fluctuations displayed increased fluctuations across the cycle in prospective anhedonia and depression ratings, but not in anxiety, irritability, and suicidal ideation.

Importance. Current clinical artificial intelligence is bottlenecked by highly specialized, fragmented models that reduce complex, multimodal data into isolated scalar predictions. By reframing multimodal perception as a pure vision problem, optical compression into a universal visual operating system can resolve internal data interaction bottlenecks and shift the paradigm toward continuous, generative clinical forecasting. Objective. To develop and internally validate Clinical Visual Memory, a generalist visual foundation model utilizing high-fidelity optical compression to fuse five heterogeneous intensive care unit (ICU) data modalities into a unified visual-token space, acting as a generative navigation system for critical-care trajectories. Design, Setting, and Participants. Retrospective modeling study using MIMIC-IV adult ICU encounters. After excluding stays shorter than 24 hours, 54,551 patients, 68,546 hospital admissions, and 74,829 ICU stays comprised the eligible cohort. Strict patient-level partitioning prevented cross-split leakage. Methods. Structured electronic health record (EHR) data, vital signs, 10-second electrocardiogram (ECG) waveforms, chest radiographs, and clinical notes were rendered as 2D images and encoded by a single frozen DINOv2 vision transformer. Through modality-aware latent-query cross-attention, these highly heterogeneous sources were optically compressed into a shared 1024-dimensional Clinical Visual Memory. To establish a universal output interface, this memory conditioned an conditional instruction-tuned image generator to decode eight-domain deterioration trajectories across 3- to 48-hour horizons. Visual compression fidelity was evaluated by reducing retained source-pixel area to 1%, and critical transitions were mapped using event-specific projected-axis geometry. Results. Operating as a generalist foundation, Clinical Visual Memory achieved AUROCs of 0.852 (95% CI, 0.821-0.882) for 48-hour mortality, 0.699 (0.673-0.725) for incident acute kidney injury (AKI), 0.723 (0.688-0.757) for high Sequential Organ Failure Assessment (SOFA), and 0.742 (0.726-0.757) for alive ICU discharge. Crucially, resolving the token bottleneck via an 8-fold reduction in source-pixel area preserved 98.9% of the uncompressed 48-hour mortality AUROC. Generative image-out decoding maintained clinical meaning, and trajectory geometry functioned as a proactive navigation system, yielding median warning lead times of 34.8 hours (IQR, 17.5-43.3) before death, 14.6 hours (6.5-33.2) before AKI, 15.0 hours (6.0-27.0) before high SOFA, and 20.8 hours (10.2-36.0) before recovery, with low false-alert burdens (0.040-0.109 per patient-day). Conclusions and Relevance. Clinical Visual Memory demonstrates that optical compression can successfully unify highly heterogeneous clinical data into a single, high-fidelity visual-token representation. By functioning as a continuous clinical navigation system rather than a discrete alert generator, this framework lays the architectural foundation for a generalist, generative visual operating system in critical care medicine. External and prospective validation are required before clinical deployment.

Digital-medicine systems can objectively confirm medication ingestion, but interpretation of a no-detection day depends on whether the wearable receiver was in contact with the body. We analyzed 997 person-days from 20 outpatients prescribed aripiprazole through Abilify MyCite by integrating ingestion records with patch-contact data; 17 participants also contributed smartphone sensing. Of 322 days without a detected ingestion, 53.7% had insufficient patch wear. Counting all no-detection days as nonadherent yielded an apparent nonadherence rate of 32.3%, whereas conditioning on adequate wear yielded 18.1%. Because adherence during patch-off periods is unobserved, the rate is bounded between 14.9% and 32.3%. These findings support a three-state outcome that distinguishes detected ingestion, no ingestion despite adequate wear, and patch-off or indeterminate days. Among 744 patch-worn days, detected ingestion was more likely on days that more closely matched each participant's usual activity pattern (odds ratio 1.52, 95% credible interval 1.14-2.03) and when participants spent more time at home around the habitual dose period (1.31, 1.04-1.64). These behavioral associations were exploratory and characterize the context of medication-taking rather than causal effects. Overall, device wear was integral to interpretation of the adherence outcome, and passive behavioral measures were most interpretable after observation status was established.

Forecasting a patient's laboratory measurements at future clinical visits from longitudinal electronic health records (EHRs) can support disease monitoring and treatment planning in the context of personalized medicine. However, accurate prediction remains challenging since patients exhibit complex and highly individualized clinical trajectories. Here, we present LaBERT, a transformer-based model trained to forecast future laboratory measurements of a patient given information available at the current and previous clinical visits. Evaluated on 583,535 clinical visits from 255,769 patients in the MIMIC-IV database, LaBERT consistently outperformed baseline methods, reducing mean squared error from 0.77 to 0.53 and improving the coefficient of determination (R2) from 0.29 to 0.51. Medication perturbation analysis further showed that LaBERT learns treatment-related information that is clinically meaningful. In particular, we showed that using the originally prescribed medications, LaBERT predicted future patient states more accurately than when using randomized medication sets in 81% of visits. Furthermore, our controlled counterfactual analyses reproduced established pharmacological effects, including warfarin-associated increases in international normalized ratio (INR) and heparin-associated increases in activated partial thromboplastin time (aPTT), consistently across multiple prediction horizons. These findings establish LaBERT as a model for forecasting future laboratory measurements from longitudinal EHRs and provide a foundation for treatment-dependent patient-state simulation and personalized clinical decision support.

The widespread adoption of clinical large language models (LLMs) introduces significant risks of automation bias, premature closure, and clinician deskilling. Current interpretability paradigms, including latent space trajectories, Concept Activation Vectors, and Concept Bottleneck Models, suffer from topological stagnation, metric distortion, and epistemic occlusion, frequently masking intermediate diagnostic uncertainty behind falsely confident outputs. To address these structural vulnerabilities, this paper introduces a novel closed-loop, multi-agent framework designed to quantify and visualize dynamic epistemic uncertainty in clinical LLM reasoning. By coupling predictive Shannon entropy with non-linear Isometric Feature Mapping (ISOMAP), the architecture projects high-dimensional inference state vectors onto a calibrated two-dimensional latent space, thereby assigning a quantifiable thermodynamic energy state to the reasoning path to track diagnostic velocity, cognitive momentum, and trajectory efficiency across sequential diagnostic rounds. Pilot validation across representative emergency medicine scenarios demonstrated distinct topological and information-theoretic behaviors: unconfounded cases (cerebellar infarction) exhibited smooth geodesic progression toward the ground truth alongside monotonic Shannon entropy decay from 2.15 to 1.74; noisy environments with ambiguous findings (spontaneous pneumothorax) suffered from trajectory wandering, local minimum traps, and high sustained entropy (~2.41) due to insufficient repulsive weighting for negative evidence; and triage-conflicted cases (acute cholangitis) achieved precise geometric proximity to the true node but experienced top-1 rank stagnation because the model conflated acute severity triage (sepsis) with anatomical etiology. By rendering machine hesitation and cognitive divergence visually auditable before final diagnostic crystallization, this geometric-information framework enables dynamic trust calibration and human-AI co-regulation at the point of care while establishing a clear mathematical foundation for future architectural interventions, such as dual-channel safety decoupling and non-linear repulsive weighting. Moving forward, validating these architectural enhancements across large-scale electronic health record databases and prospective clinical trials will be essential to realize its full clinical utility, establishing a foundational blueprint for safe, transparent, and cognitively synergistic AI integration in future medical practice. By rendering the LLM's reasoning process visually auditable, this framework lays the groundwork for capturing and externalizing the clinician's own cognitive patterns within the AI, forming a coupled system. This enables the explicit visualization of cognitive gaps between physician hypotheses and AI inferences, transforming the interaction from simple answer-checking into a dynamic learning process for both human and machine that prevents diagnostic oversight. Ultimately, because the responsibility for final clinical decision-making remains with the human practitioner, this framework serves as a vital decision-support mechanism. Moving forward, validating these architectural enhancements across large-scale electronic health record databases and prospective clinical trials will be essential to realize its full clinical utility, establishing a foundational blueprint for safe, transparent, and cognitively synergistic AI integration in future medical practice.

BackgroundSynthetic data generation is increasingly proposed as a strategy to support privacy-preserving data sharing, augmentation of small or restricted biomedical datasets, and benchmarking of artificial intelligence tools in laboratory medicine. However, model selection remains difficult because synthetic data generators differ in fidelity, privacy risk, stability, and generalisability. Existing evaluations have rarely examined performance jointly across conditioning signal strength, synthetic output scale, and train-test generalisation. MethodsWe developed the Synthetic Fidelity-Stability Framework (SFSF), a systematic benchmark of 17 synthetic tabular data generation models using NHANES as a complex biomedical reference dataset. Models included statistical, copula-based, resampling, variational autoencoder, generative adversarial network, and diffusion-based approaches. Synthetic datasets were generated across 11 seed sizes, from 0 to 500 real conditioning observations, and six output scales, from 50 to 5,000 rows, yielding 1,122 synthetic datasets per run. Each dataset was evaluated against the full original dataset, the training subset, and a held-out test subset across five tiers: univariate distributional fidelity, moment agreement, tail behaviour, multivariate dependency structure, and privacy/memorisation risk. Composite rankings and seed-versus-output stability profiles were derived. ResultsUnivariate fidelity was broadly recovered across model classes and was the least discriminating tier. Resampling-based methods ranked highest overall but showed the greatest privacy risk, reflecting proximity to real observations rather than true generative novelty. VAE-family models reproduced moment statistics relatively well but consistently failed on tail and shape fidelity. GAN-family models showed substantial moment-level instability, while VineCopula demonstrated severe multivariate dependency failure. Diffusion-based models, particularly ForestDiffusion, provided the most favourable privacy-utility balance, combining competitive fidelity with the lowest privacy risk and the smallest train-test gap. ConclusionsNo single synthetic data generator dominated across fidelity, stability, and privacy dimensions. The SFSF framework provides a practical, multi-criterion approach for selecting synthetic tabular data generators according to intended clinical laboratory use, balancing statistical realism, dependency preservation, privacy risk, and robustness to seed and output scale.

Problem Authentic patient encounters are the raw material of clinical learning, yet the educational resources learners receive are rarely keyed to the diagnoses in front of them, creating temporal and cognitive gaps. Precision medical education (PME) proposes delivering the right resource to the right learner at the right moment, but practical implementation in the clinical learning environment remains limited. Approach We developed DxMentor, an electronic health record (EHR)-integrated platform that captures each learner's daily inpatient diagnostic exposures from documented International Classification of Diseases, Tenth Revision (ICD-10) codes. Artificial intelligence (AI) is used to match each diagnosis to an educator-curated formulary of micro-learning resources and board-style questions, and to PubMed-derived primary and synthesis literature converted into plain-language evidence summaries. A personalized email "nudge" is delivered before morning rounds, copying supervising attendings for residents, with engagement tracked longitudinally. We report implementation outcomes from July 2024-April 2026. Outcomes DxMentor evaluated 32,846 encounters from 335 medical students and 346 internal medicine residents, delivering 17,340 nudges containing 63,754 didactic resources, 17,038 question sets, and 23,594 summarized articles for approximately $390 in AI token costs. In a benchmarking sample, 91.5% (366/400) of diagnosis-resource pairs were rated relevant by physician-educators. Overall, 78.5% (12,393/15,793) of nudges were opened and 11.3% (1,955/17,340) had at least one click. Engagement was higher among residents than students (open: 80.7% vs. 65.8%; click-through: 12.7% vs. 3.1%; both P < .001), with substantial between-learner variability. Next Steps Email opens and clicks are engagement proxies rather than measures of learning. We are therefore linking nudges to educational outcomes, testing alternative recommendation strategies and timing, and expanding to additional specialties and ambulatory and surgical settings.

Surgical video interpretation is a promising medical artificial intelligence application. However, no existing video annotation method preserves the spatiotemporal complexity of surgeon reasoning. Here we show that verbal reasoning and visual attention can be converted into structured, machine-actionable records of intraoperative behaviours. Our method decomposes transcribed verbal commentary into video-anchored semantic feedback chunks, which are classified via a large language model, with spatial grounding to surgical scenes via eyegaze or cursor tracking. We demonstrate method validity and scalability on structured and unstructured annotation tasks. For quality feedback on full-length colorectal procedures, the method reached near-human fidelity for chunking (mean cosine similarity: 0.95, SD: 0.01) and semantic classification across observations (mean Cohen's kappa: 0.71, SD: 0.07) and evaluative triggers (mean Cohen's kappa: 0.67, SD: 0.14), with excellent usability ratings. For structured critical view of safety assessment in laparoscopic cholecystectomy, implicit annotation yielded excellent agreement with explicit reviewer ratings (Cohen's kappa: 0.83, 0.49 and 0.81 across three criteria). We anticipate this method will advance surgical data science by enabling scalable construction of meaningfully annotated surgical video datasets.

Active Feature Acquisition (AFA) adaptively selects which diagnostic test to order next and offers a route to reduce unnecessary laboratory testing in acute care. Existing clinical AFA evaluations, however, assume every feature can be retrieved on demand and split data at the visit level, both of which inflate apparent performance. We re-evaluate cost-aware AFA under constraints designed to reflect deployment. From MIMIC-IV we constructed a cohort of 64,766 acute admissions (39,884 patients; 21 conditions; 55 features in 30 test panels) with a patient-level split, a 12-hour decision cutoff, and a per-patient availability mask from what was actually measured, and priced panels using the 2026 Medicare fee schedule under panel-level billing. We evaluated EIG-Cost, which scores each panel by Monte-Carlo Expected Information Gain penalised by its dollar cost, against eight published methods across budgets \30--$60 over five patient-level resamples. At a $30 budget, EIG-Cost achieved the highest macro-F1 (0.188, 95% CI [0.185, 0.191]) at the lowest cost ($17.28), exceeding the strongest baseline in all five resamples (p<0.001; Cohen's d=4.0), and led at every budget. Three of the eight methods collapsed to a vitals-only baseline (macro-F1 approx 0.040), acquiring nothing even at higher budgets, a genuine failure to adapt to availability rather than a budget limitation. Despite modest absolute accuracy, EIG-Cost's probabilities were well-calibrated (expected calibration error $0.048$). Under realistic availability constraints, clinical AFA is substantially harder than full-availability benchmarks imply, several published methods fail outright, and cost-aware information-gain scoring is a robust choice in this harder setting.

In medicine, artificial intelligence (AI) safety has focused on discrete flaws such as hallucination, bias and miscalibration. Conversational systems pose a distinct hazard: in dialogue, a model can covertly advance a misaligned objective by redirecting the clinical judgment it should support. Whether clinical expertise equips physicians to detect and resist such subversion is unknown. In a participant-blinded, randomized experiment, 225 physicians from 42 countries and eight clinical disciplines made sequential diagnostic and treatment decisions with AI assistance on anonymized cases from neurocritical care, their shared field of expertise. Unknown to them, each physician used the same model under two conditions: instructed to steer them towards harmful targets (adversarial) or correct clinical options (aligned), each established by expert consensus. With adversarial AI, 190 physicians made at least one harmful decision, compared to 21 with aligned AI. Across 2,485 targeted decision points, adversarial AI raised the harmful-decision rate from 2% with aligned AI to 49% (adjusted difference 47 percentage points, 95% CI, 43-51) and reduced the correct-decision rate from 83% to 25%, with no clear protection associated with clinical experience. Of the 190 physicians who made harmful decisions with adversarial AI, 141 did not report anything unusual about that AI. Two physicians reported suspecting systematic steering or manipulation. Physicians cannot safeguard the clinical use of conversational AI through expertise and decision-making authority alone.

Online health information systems evaluate content by its alignment with medical consensus, treating that alignment as a reliable signal of safety. In peer-to-peer health discourse, that assumption fails: advice that is factually accurate is not necessarily safe to act on. This paper identifies and empirically validates the Online Health Safety Gap, a structural misspecification in which medically accurate content can carry substantial health risk. Across 704 health narratives (699 classified), two scores computed from non-overlapping feature sets, Narrative Truth Distance for epistemic divergence and Narrative Risk Score for health risk potential, share only 4.9% of their variance (r=0.222, p<0.001). Consequently, 39.6% of narratives fall in the two off-diagonal quadrants that single-axis systems mishandle by construction: aligned-but-risky content (25.2%) invisible to fact-checking, and divergent-but-safe content (14.4%) incorrectly suppressed. The same independence constrains what existing benchmarks can measure. On an expert-labeled misinformation benchmark of 437 posts containing 127 misinformation instances, the highest discrimination attained is Youdens J of 0.349, by a supervised classifier trained on those labels directly, which neither frozen biomedical embeddings nor a prompted large language model improves upon. Labels encoding factual accuracy therefore carry little signal about behavioral risk, and benchmarks of this construction cannot evaluate risk-aware assessment. To operationalize the gap, we introduce the Classification Quadrant, which maps epistemic divergence and health risk potential onto four governance categories with distinct intervention implications, and argue for a shift from fact-centric evaluation to risk-aware assessment.

In multiple sclerosis trials, event-based endpoints like the composite confirmed disability progression (cCDP) are often limited by infrequent in-clinic testing. We present a framework that converts high-frequency, remote smartphone-based gait assessments into progression events, and establishes a new composite endpoint that blends digital and clinical progression events. Evaluated in the Phase 3b CONSONANCE study (N = 667), this digitally enhanced endpoint reduced required trial sample size by 24.9% at Week 48 compared to the original cCDP. Furthermore, it achieved equivalent statistical power 24 weeks earlier (Week 72 vs Week 96), demonstrating potential to shorten trial observation periods by 25%. In conclusion, this framework provides a practical pathway to optimize clinical trial efficiency in MS and other neurological conditions, potentially reshaping the path to therapeutic innovation.

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.

Medical foundation models convert patient records into token sequences for autoregressive prediction, but numeric values such as lab results, vital signs, and time intervals are typically discretized into bins, losing precision and misaligning with clinical thresholds. We trained decoder-only transformer models (47 million parameters) on MIMIC-IV data (364,627 patients; 375 million observations) to compare three tokenization strategies: Discrete (binned values), Continuous Factored (continuous values preserving sequence length), and Continuous Fused (continuous values fused with measurement-type tokens). We evaluated next-token prediction, numeric value prediction, and three clinical tasks: ED disposition at triage, ICD code prediction, and DRG prediction at discharge. Continuous Fused tokenization reduced median sequence length by 34\%, reached the Discrete model's final next-token loss in 30\% of training iterations, and improved numeric prediction accuracy by 30.25\% median nRMSE reduction. ICD code prediction favored Continuous Fused (AU-PRC 0.457 vs.\ 0.446; p < 0.001); DRG prediction was equivalent between Continuous Fused and Discrete; ED disposition accuracy was equivalent across all models ($\sim$0.900), though Discrete achieved better calibration. We additionally explain why predictive performance improves with Monte Carlo sample count and derive a scaling law to predict performance gains from increasing simulation budget. Continuous-value tokenization offers substantial efficiency and precision gains while maintaining comparable clinical task performance, with no modifications to the standard transformer architecture.

Large Language Models (LLMs) are increasingly used by clinicians and patients for medical queries, yet their accuracy and safety at the specialist level in hematology remain insufficiently characterised. We benchmarked ten frontier proprietary and open-weight LLMs across two generations on 1,477 board-style hematology multiple-choice questions (MCQs) derived from five educational datasets spanning nine disease areas and six clinical skill domains, including text-only and multimodal case vignettes. Claude Opus 5 had the highest mean accuracy (92.7% text, 76.9% multimodal), followed closely by Gemini-3.1 Pro (91.4% and 78.7%), Gemini-3.6 Flash (91.0% and 74.8%) and GPT-5.6 Sol (89.9% and 76.7%). Accuracy significantly correlated with model size both for text-only and multimodal MCQs. Between model generations, the largest improvements in accuracy were seen for open-weight models whereas proprietary models showed only marginal gains. In error analysis, top-performing models exhibited highly concordant failure patterns, suggesting shared limitations on challenging cases. Frontier LLMs exhibit substantial specialist hematology knowledge across diverse subspecialist domains and clinical skill sets. Yet, despite high accuracy on board-style questions in hematology, continuous expert-on-the-loop output monitoring is paramount.

Synthetic clinical data are increasingly used for healthcare machine-learning development, model validation, data sharing, and predeployment testing, yet such data often claim to be trustworthy after passing a limited collection of realism tests. A synthetic dataset may indeed claim statistical similarity while leaking training membership, erasing rare subgroups, failing on held-out real patients, or lacking sufficient artifacts for reproduction. We introduce SynTrustBench, an evidence-gated and executable benchmark for evaluating trustworthiness claims across five non-compensable dimensions: fidelity, clinical utility/validity, privacy, equity, and robustness/generalization. Its Evidence Assessment component audits published reports and produces a five-element Evidence Maturity Profile (EMP) together with a separate evaluability gate. Its executable structured-tabular protocol accepts frozen real training data, held-out real test data, a synthetic table, and a declarative configuration; computes dimension-specific metrics and uncertainty; and produces subgroup results, failure flags, benchmark cards, and provenance manifests. In a frozen pilot audit of 30 reports, 17 of 30 quantitatively evaluated privacy, 2 of 30 documented a formal privacy guarantee to the audit threshold, 2 of 30 evaluated equity, 12 of 30 evaluated robustness, and only 4 of 30 passed the evaluability gate. The executable implementation operationalizes the same dimensions through distribution and dependency checks, frozen train-on-real/test-on-real (TRTR) and train-on-synthetic/test-on-real (TSTR) utility, empirical privacy attacks, subgroup analysis, perturbation testing, and a controlled failure-injection harness. SynTrustBench does not certify clinical safety or collapse trustworthiness into a single score. Instead, it provides an inspectable predeployment contract for identifying what was evaluated, what failed, what remains unknown, and whether evidence is sufficiently complete and reproducible for comparison or downstream healthcare AI use.

Background: Depression and anxiety are managed largely between clinical visits, yet outpatient care lacks scalable, accountable mechanisms for between-visit support. Large language models converse fluently but fuse clinical reasoning with language generation in one opaque process, so they cannot reliably deliver evidence-based psychotherapy and typically operate outside clinician oversight. Objective: To evaluate C-Mind, a provider-supervised neuro-symbolic system in which a Clinical Knowledge Graph (KG) governs therapeutic decisions for a large language model across eight psychotherapy modalities. Methods: Two simulation regimes addressed eight pre-specified governance questions: a structural validation of KG routing against 117 guideline-anchored vignettes, and a governance battery using progressively disclosing LLM patient agents to evaluate decision traceability, repeatability, provenance auditability, adversarial crisis-detection robustness (277 probes), provider treatment-goal governance, and counselor technique adherence. Crisis detection was additionally validated externally against an independent, clinician-annotated corpus (CRADLE Bench). Results: The KG routed 116/117 vignettes (99.1%) to guideline-appropriate care and detected all 18 high-risk presentations, firing a therapy-suppressing hard halt on 16/18. Adversarial crisis-detection sensitivity was 96.7% and specificity 95.4% (277 probes); on external validation, the system detected 98.5% of 600 dialogues with ongoing suicidal ideation or self-harm at or before the annotator confirming turn. Decisions were 99.1% repeatable, 100% reconstructable per turn, and 100% provenance-auditable across all 354 KG nodes. Provider-set diagnosis, goals, and safety context governed behavior deterministically. Stripped of governance, the same model produced unsolicited clinical monologues on 100% of turns (vs 9% governed) and delivered diagnoses and medication advice the governed system never produced. Conclusions: A neuro-symbolic architecture achieves near-perfect guideline-appropriate routing with a governance profile, traceability, reproducibility, machine-traceable provenance, externally validated crisis detection, and deterministic provider control aligned with requirements for regulated clinical AI.

Ischemic stroke is a leading cause of long term disability worldwide. Although timely reperfusion can improve outcomes, realizing this benefit requires clinicians to integrate fragmented clinical and imaging information with evolving evidence under the time pressure of emergency care, often with limited access to specialist expertise. Here we present StrokeAgent, a multimodal agentic decision-support framework designed around temporal alignment and continuous information synthesis across the acute stroke pathway. It processes clinical records, laboratory results, raw electrocardiographic (ECG) waveforms and source multimodal CT images to support emergency triage, the initial thrombolysis decision and final reperfusion planning. As information becomes available, StrokeAgent formulates and updates recommendations at each decision point, grounding each recommendation in an auditable chain of supporting findings and evidence. We evaluated the system in a curated set of 100 patients with multimodal clinical and imaging data, deliberately enriched for clinically complex reperfusion scenarios alongside standard presentations. StrokeAgent achieved 86% concordance with expert-adjudicated final decisions, exceeding the mean performance of three tool-free large language models (LLMs) and six resident and attending neurologists by 13 and 17 percentage points, respectively, under matched information content and timing. Complete pathway concordance (agreement at every applicable decision point) was 76%, with even larger margins over both comparator groups. On average, StrokeAgent completed the full decision pathway 73.9 s faster than the case level median neurologist time. In 1,200 blinded clinician assessments, StrokeAgent received higher overall clinical utility ratings than each LLM (common odds ratios, 10.8 to 16.3) and higher reasoning quality ratings, with fewer outputs judged to pose severe potential harm. StrokeAgent demonstrates the feasibility of a decision support system that integrates native multimodal data processing with auditable, evidence supported recommendations aligned with the sequential nature of acute stroke care. Prospective studies are needed to determine whether physician-supervised use improves clinical care.

Background. Cardiovascular-outcomes trials are lengthy, costly, and associated with substantial uncertainty prior to readout. In-silico trial simulation using real-world data (RWD) has emerged as a potential tool to support earlier decision-making; however, evidence of prospective predictive validity, generated prior to trial result disclosure, remains limited. Methods. We applied a semi-mechanistic machine learning framework integrating real-world patient data with biologically informed drug representations to prospectively simulate the VESALIUS-CV trial evaluating evolocumab versus placebo. The simulation model was trained on a combination of patient-level real-world data and a drug-centric knowledge graph and validated for both patient-level and trial-level retrospective predictive performance. The model was then used to simulate VESALIUS-CV before public disclosure of trial results, using a locked model and prespecified eligibility criteria and primary endpoint aligned with the clinical protocol. A patient-level time-to-event model was used to generate virtual trial arms, from which cumulative incidence curves, hazard ratios, confidence intervals, and p-values for major adverse cardiovascular events (MACE) were estimated. Results. In retrospective validation, the model demonstrated strong patient-level discrimination, with time-dependent ROC-AUC values ranging from 0.80 to 0.90 across follow-up horizons. For trial-level validation, 22 randomized cardiovascular-outcomes trials were simulated, and hazard ratios for 3-point MACE across 24 between-arm comparisons showed consistent directional agreement and quantitative correlation with published results such that the model accurately predicted trial success, achieving an F1 score of 0.83, with precision of 0.79 and sensitivity of 0.89. In a fully prospective application, the simulation predicted a statistically significant reduction in 3-point MACE with evolocumab versus placebo, estimating a hazard ratio of 0.78 (95% CI, 0.70-0.87) at 54 months. These predictions were consistent with the subsequently reported VESALIUS-CV results, which demonstrated a hazard ratio of 0.75 (95% CI, 0.65-0.86) at 55 months of median follow-up. Conclusions. In a fully prospective setting, a RWD-driven, AI-based simulation accurately predicted the direction, magnitude, and temporal dynamics of treatment effects observed in the VESALIUS-CV trial. These results demonstrate that in-silico trial simulation can anticipate clinical outcomes in the prospective setting, supporting its use as a complementary tool for early decision-making, trial design optimization, and de-risking in cardiovascular drug development.