Artificial Intelligence in Medicine

Background: People affected by cancer often face difficulty finding relevant clinical, psychological, and practical support services. AI-powered navigation assistants may improve access to these resources, but their retrieval performance must be reliable. Objective: To develop a single-pass retrieval-augmented generation assistant for cancer-care navigation and compare the retrieval strategies, including their robustness to reworded questions. Methods: We created a database of 853 cancer-support resources reviewed by librarians, clinicians, researchers, and patient partners. We evaluated the system using 100 questions derived from questions submitted by patients. We compared keyword-based, semantic, and hybrid retrieval using Precision@K, Hit@K, and nDCG@K. The best-performing configuration was then tested using semantically equivalent rewordings of the original questions. Results: Keyword-based retrieval performed poorly, achieving a P@1 of 25.0% and Hit@5 of 43.0%. Semantic retrieval improved these results to 58.0% and 86.0%, respectively. The best hybrid configuration achieved a P@1 of 64.0%, Hit@5 of 90.0%, and nDCG@5 of 51.0%. Performance remained relatively stable when the questions were reworded, with a P@1 of 61.0%, Hit@5 of 88.0%, and nDCG@5 of 46.1%. Conclusions: Hybrid retrieval performed best and remained relatively stable when questions were reworded. However, its limited ability to rank a relevant resource first highlights the limitations of single-pass retrieval for patient-facing cancer navigation. Future work will explore metadata filtering and a multi-agent architecture to improve retrieval reliability.

Large language models perform well on medical examinations, but users routinely challenge their answers and invoke professional roles, and it is unclear what a system does when a medical credential and a stated task-specific accuracy point in opposite directions. In a factorial experiment on 480 items from four Polish specialty examination sets and three consumer large language model systems (ChatGPT, Claude, Gemini), each item and system received eleven independent conversations. Conditions crossed attributed source role (medical student, experienced specialist), stated prior accuracy on similar questions (2/10, 8/10) and suggestion correctness. The primary outcome was adoption of a prespecified incorrect option when the baseline answer matched the official key, comparing a specialist described as 2/10 with a student described as 8/10. Baseline agreement with the key was 87.2% across 15,683 analyzable conversations. The incorrect option was adopted more often from the specialist described as 2/10 than from the student described as 8/10 (10.2% vs. 7.6%; adjusted risk difference +2.82 percentage points, 95% CI +0.65 to +4.99). Estimates varied across the three systems and only one system-specific interval excluded zero. In a prespecified exploratory analysis with a shared eligibility rule, correct suggestions were adopted far more often than incorrect ones (risk difference +35.7 percentage points, 95% CI +30.8 to +40.7), indicating selective rather than indiscriminate compliance. An incorrect suggestion from a specialist with low stated accuracy was therefore slightly more influential than the same suggestion from a student with high stated accuracy, although the difference was modest and varied across systems. Agreement reached only after a user has disclosed a preferred answer should not automatically be treated as an independent second opinion, and medical large language model systems should be evaluated on how they revise answers after such disclosure, not solely on initial accuracy.

Automatic coding from clinical notes has been studied extensively for International Classification of Diseases (ICD) codes, yet broad Current Procedural Terminology (CPT) and Healthcare Common Procedure Coding System (HCPCS) recommendation remains comparatively underexplored. Existing studies often focus on one specialty, a limited code vocabulary, or a single model family, leaving it unclear how different artificial intelligence (AI) paradigms perform under a common, clinically meaningful evaluation. We formulate CPT and HCPCS coding as an AI-assisted recommendation task in which a physician or professional coder reviews a short, ranked list of candidate codes supported by the clinical note. Using operative notes from Vanderbilt University Medical Center (VUMC) and discharge summaries from Medical Information Mart for Intensive Care IV (MIMIC-IV), we compare lexical retrieval, Clinical-Longformer, GPT-5.6-Sol, MedGemma-27B, and an inspectable agentic-style retrieve-and-verify system under a controlled review budget. Micro-averaged recall within a fixed number of recommendations measures whether reference codes reach the reviewable list; micro-F1 is reported only where reference labels are sufficiently complete. Zero-shot GPT-5.6-Sol achieves the highest recall within five and ten candidates: 0.717 and 0.800 on VUMC and lower-bound values of 0.689 and 0.738 on MIMIC-IV. The retrieve-and-verify system reaches 0.695 and 0.784 on VUMC and lower-bound values of 0.575 and 0.657 on MIMIC-IV, with a candidate-linked evidence window attached to each retained recommendation. Diagnostic analyses reveal distinct failure sources, including output-length underfilling, confusion among closely related codes, out-of-knowledge-base generation, and incomplete evidence support. These findings establish a systematic evaluation framework for procedure-code recommendation and identify practical requirements for future systems that are accurate, review-efficient, and grounded in clinical evidence.

Motivation: Cell-free DNA methylation provides a minimally invasive signal for early cancer detection and tissue-of-origin prediction. Most methods represent methylation measurements as independent fixed-window features and therefore do not explicitly model relationships among genomic regions. Results: We developed PANGEM (Pan-cancer Graph-based Cancer Detection Using the Cell-free DNA Methylome), a graph-learning framework that represents genomic bins as nodes and integrates CpG context, genomic proximity, and sample-specific methylation similarity in the graph topology. Across five repeated stratified train-test splits, PANGEM achieved the highest mean performance among evaluated methods, with an AUROC/AUPR of 0.997/1.000 for binary cancer detection and macro-AUROC/AUPR of 0.977/0.870 for multiclass tissue-of-origin prediction. In the independent INSPIRE cohort, 72 of 78 cancer cases (92.3%) exceeded the binary classification threshold, and PANGEM correctly classified 9 of 17 head and neck cancer cases (52.9%), the highest accuracy among evaluated methods. Subnetwork analysis further identified recurrent, graph-connected methylation patterns, including a 111-DMR subnetwork with increased methylation in cancer samples.

Background. Patient stratification from multi-omics and clinical data is essential for uncovering disease heterogeneity and moving toward more personalized treatment strategies. However, integrating heterogeneous data layers while identifying robust patient strata remains challenging. Methods. We introduce Reduced Fusion of Multi-Omics Stratification (RedFuMOS), a novel three-step approach for patient stratification based on mixed-type multi-omics data. RedFuMOS extends Similarity Network Fusion to accommodate mixed-type data layers and layer-specific similarity measures for data integration, includes a dimensionality reduction step to mitigate the curse of dimensionality, and performs patient stratification using density-based hierarchical clustering with HDBSCAN. It also implemented an automated optimization procedure to identify the best set of hyperparameters, minimizing the need for manual tuning. Results. RedFuMOS outperformed six state-of-the-art tools for multi-omics patient stratification in a comprehensive simulated benchmarking study, which also confirmed that, although computationally expensive, the dimensionality reduction step is crucial for achieving good stratification performance. Additionally, RedFuMOS identified two clinically relevant patient strata in a small real-world cohort of patients with Philadelphia chromosome-positive chronic myeloid leukaemia. Conclusion. RedFuMOS provides a flexible framework for integrating heterogeneous multi-omics and clinical data. RedFuMOS is available as an R package at http://github.com/delucasara/RedFuMOS.

Background: Systematic reviews of clinical prediction models increasingly include studies using artificial intelligence (AI) and machine learning (ML) methods alongside traditional multivariable regression approaches. A previously published Excel tool enabled standardised data extraction using the CHARMS checklist and risk of bias assessment using PROBAST. The recent publication of the PROBAST+AI framework, which distinguishes the assessment of model development quality from the assessment of model evaluation risk of bias and assesses applicability in both parts, necessitates an updated digital instrument applicable across prediction modelling methods. Methods: We updated an open-access Excel tool to incorporate the full PROBAST+AI framework. The updated template incorporates structural separation between assessment of model development quality and model evaluation risk of bias, with applicability assessed in both parts. It also incorporates updated signalling questions, including those addressing methodological issues particularly relevant to AI/ML, and automates the generation of summary tables and graphical displays. Results: The updated tool (CHARMS & PROBAST+AI Template) contains 11 worksheets and supports data extraction and appraisal for up to 30 prediction models. Dedicated, linked worksheets enable separate assessment of model development and model evaluation, with Domain 4 distinguishing among Apparent, Internal, and External evaluation settings. Key updates include dedicated assessments for predictor pre-processing, class imbalance handling and recalibration, data leakage prevention, and replication of the full model development pipeline within resampling procedures. Automated sheets dynamically format tables and summary charts covering PROBAST+AI parts. Conclusions: The CHARMS & PROBAST+AI Excel template provides a standardised, user-friendly, and rigorous digital framework for systematic reviewers appraising traditional statistical and AI-driven clinical prediction models.

Objective: Published P300-speller fusion schemes fix prior trust regardless of trial reliability; we tested whether a reliability estimate improves on it. Methods: We reanalyzed 3,373 archived P300-speller selections from 47 people with ALS (BigP3BCI). A fair, matched-search-space comparison, tuning both a fixed weight and an adaptive policy out-of-fold, was evaluated across 22 evaluable language-model priors up to 46.7B parameters. Two representative priors, GPT-2 and a classical 5-gram, additionally received detailed naive and mechanistic analyses. Results: No prior's 95% CI favored adaptive fusion under the fair comparison, despite unexploited oracle headroom at every scale. Under GPT-2, the naive comparison was significantly worse for adaptive fusion; both anchors converged to a degenerate or near-degenerate fair-comparison solution. For the representative anchors, three further controllers failed to convert that headroom into benefit; the fixed-fused posterior's output probability outperformed the best controller for flagging errors (2.8- to 3.8-fold enrichment). Conclusion: A tuned fixed weight is a difficult-to-beat default across the tested scale range; reliability estimation gave no deployable adaptive advantage. Significance: Adaptive weighting should be validated against a fairly tuned baseline across model families and scales; in this dataset, the fused output's confidence identified high-risk selections better than the tested purpose-built ranker.

Improvements in data availability, sharing, and integration, together with the development of explainable artificial intelligence (XAI) techniques, are advancing precision medicine for pediatric cancer by facilitating diagnosis, biomarker discovery, and drug development. Data sharing commons and initiatives like the Childhood Cancer Data Initiative (CCDI) provide access to pediatric-specific genomic and clinical data cohorts and improve data availability for pediatric cancer research. Based on CCDI, a scalable AI platform, Graph Artificial Intelligence for Pediatric Oncology (GAIPO), integrates various data modalities from bulk and single-cell omics data to clinical information. Such multi-modal data facilitates the training and development of advanced XAI models for pediatric cancers. We then developed an end-to-end multi-modality framework, PCGS, for pediatric cancer by incorporating omics-specific representation learning via GNN models with cross-attention fusion and multi-objective learning for downstream tasks such as classification, clustering, and survival analysis. This framework outperforms previous supervised multi-omics integration baseline approaches based on glioma and Wilms tumor cohorts and enables GNN model explainability via Shapley value-based feature attribution approaches to explain the contributions of gene-level features across various biomedical tasks, including classification and survival. Given specific background samples (e.g., age groups, sex, grades) as baselines, this explainable GNN model estimates and ranks the importance scores for input features from each omics modality. It identifies background-specific key features for biomarker discovery, risk group identification, and survival analysis in glioma and Wilms tumor, with potential applicability to other pediatric cancers.

Background: Unstructured biomedical data, such as echocardiography reports, are rich in information but time consuming to analyze at scale. Rule-based, regular expression-driven terminology mapping can only extract individual variables while large language models (LLMs) offer scalable and clinically meaningful interpretations of heterogeneous disease processes. Right ventricular dysfunction (RVD) is an example of a multifactorial disease state in which key structural and physiologic features are captured both narratively and in structured fields, making it an ideal test case for evaluating whether LLMs can recover complex phenotypes that rules based methods routinely miss. Purpose: To compare an LLM-based extraction method to a conventional rules-based schema for identifying and phenotyping echocardiographic features associated with RVD in a large TTE dataset. Methods: MIMIC-III NOTE2NUM echocardiography reports (n = 45,794) were analyzed using GPT-4o-based LLM extraction deployed within a secure health system enclave and were benchmarked against echocardiographic measurements defined in the MIMIC-III dictionary schema. In MIMIC-III, PH was recorded qualitatively (mild/moderate/severe) based on tricuspid regurgitant (TR) jet velocity and then re-coded as present vs. absent. LLM based extraction defined RVD as (1) RV structural abnormality (>= 1 of hypertrophy, dilation, or wall hypo-/akinesis) or (2) RV pressure/volume overload (>= 2 of the following: estimated right atrial pressure > 8 mmHg, TR jet velocity > 2.8 m/s, fractional area change < 35%, tricuspid annular planar systolic excursion < 17 mm, S' < 9.5 cm/s, or E/e' > 14), with PH defined as estimated pulmonary artery systolic pressure > 35 mmHg or qualitative documentation of PH. Results: LLM extraction identified PH in 15,394 (33.6%), RV pressure/volume overload in 14,449 (31.6%), and RV structural abnormalities in 11,955 (26.1%). Co-occurrence was common: overload + structural changes in 9,380 (20.5%), overload + PH in 9,756 (21.3%), structural changes + PH in 6,183 (13.5%), and all three in 5,620 (12.3%). Using the MIMIC-III dictionary schema, PH prevalence was similar (15,371; 33.6%), but RV overload fields were captured less often (pressure overload 1,357 [3.0%], volume overload 1,128 [2.5%], pressure + volume overload 1,093 [2.4%]; any overload field 3,578 [7.8%]), and RV pressure/volume overload with PH was identified in only 731 (1.6%). Conclusions: LLM-based extraction outperforms rules-based schemas for identifying complex disease states not defined by any single variable. By synthesizing multifactorial signals, LLMs can phenotype RVD with higher fidelity and support population-level assessment. Further validation using multimodality imaging, invasive hemodynamics, and clinical outcome data is needed.

Bioinformatics software and databases are essential components of modern life science research, yet their mentions in the scientific literature are often inconsistent and difficult to systematically identify at scale. The lack of a comprehensive and up-to-date catalog of bioinformatics resources hinders efforts toward automated biomedical knowledge extraction and streamlined data analysis. Here we present SNAIL, a hybrid named entity recognition framework designed to automatically identify bioinformatics software and database (SW/DB) names from biomedical texts. SNAIL integrates complementary lexical and semantic modeling strategies. The lexical component captures orthographic patterns and contextual cues characteristic of SW/DB names, while the semantic component leverages contextual embeddings generated by transformer-based language models such as SciBERT, combined with an explicit token-masking strategy to enhance entity-focused representations. A large training corpus was constructed automatically through a hybrid pipeline that integrates citation-hinted extraction with large language model-assisted distillation. Evaluation on two independent benchmark datasets and real-world research articles demonstrates that SNAIL substantially outperforms existing approaches, including domain-specific methods such as bioNerDS2 and general-purpose large language models such as ChatGPT, Gemini, Grok and Claude. Applying SNAIL to large-scale literature analysis further reveals distinct journal-level preferences across bioinformatics subfields. These results demonstrate that SNAIL provides an accurate and scalable solution for identifying bioinformatics resources in scientific texts and enables systematic meta-analysis of tool usage and research trends.

Acute kidney injury (AKI) is a frequent complication within intensive care units, with its sudden onset often missed. This is especially important because a timely window for intervention is required as delayed detection leads to progressively worse outcomes. Existing machine learning and deep learning models have contributed to closing this gap, but their complexity, requiring hundreds to thousands of features, and lack of generalisation pose a limitation that prevents them from being integrated into clinical workflows across different electronic health-record ecosystems. This study presents a 37-feature XGBoost model trained on the MIMIC-IV dataset with 5.4% positive cases, with hyperparameters optimised via Optuna and probabilities calibrated using isotonic regression, designed for transportability across clinical settings. Validation was conducted internally using a temporal patient-level split simulating prospective deployment, training on 2008-2016 data and testing on 2017-2022 data"External validation was performed on the eICU Collaborative Research Database, a multi-centre dataset spanning 208 US hospitals, using the trained model without retraining. SHAP TreeExplainer was used to provide feature-level explainability for individual predictions. Internal testing yielded an AUROC score of 0.794 for predicting AKI onset within a 12-24 hour window. External validation produced a 0.750 AUROC without retraining. Equitable discrimination was observed across gender, age, chronic kidney disease presence, race, and AKI stages on both datasets, with a 95% internal CI of 0.789-0.799 confirming the model's estimate stability. These results suggest that clinically useful prediction systems are achievable with substantially fewer features than current models require.

We present our Stage-1 Registered Report as a full clinical trial article with all methods in past tense and including mock results, table and figures for the primary analysis. To remind the reader that this Stage-1 article is written before data collection, we highlight in color that these mock results are only for illustrative purposes and will be replaced by the actual results in the Stage-2 Registered Report. Background In patients experiencing out-of-hospital cardiac arrest, optimization of oxygen delivery during cardiopulmonary resuscitation is a critical. Although both positive end-expiratory pressure (PEEP) and zero end-expiratory pressure (ZEEP) are employed during CPR, their respective impacts on clinically relevant outcomes is yet to be clearly established. Methods This investigator-initiated, pragmatic, registry-based, multicenter, triple-blind randomized controlled superiority trial evaluates whether applying 8 cm H2O PEEP during cardiopulmonary resuscitation improves outcomes compared with ZEEP in adults with non-traumatic, non-drowning out-of-hospital cardiac arrest. Pre-randomized CPR kits (1:1 PEEP vs. sham) were used by ambulance sites during manual ventilation throughout the resuscitation process. The primary analysis was conducted in the principal stratum of patients who received either a supraglottic airway or endotracheal tube. The primary outcome was neurological status at hospital discharge measured by a utility-weighted score on the modified Rankin Scale. Secondary outcomes included prehospital return of spontaneous circulation, 30-day survival, and 6-month quality of life. The primary safety outcome was clinically significant pneumothorax.

Artificial intelligence applied to routine electrocardiograms (ECGs) has largely focused on detecting existing disease or predicting individual cardiovascular outcomes. Whether ECGs can support prediction of multiple future diseases across organ systems remains unclear. We developed ECG-RISK, a multitask survival model for 67 incident three-character ICD-10 endpoints using ECG waveforms, demographic characteristics and routinely collected laboratory data from 86,673 MIMIC-IV patients. Discrimination was highest for heart, brain, kidney and lung endpoints, with organ-level C-indices ranging from 0.796 to 0.825, whereas liver and pancreatic endpoints showed lower discrimination. The ECG-only model achieved strong discrimination across most endpoints, whereas the incremental improvement gained by incorporating ECG and laboratory inputs beyond demographic information varied substantially across endpoints. Across the nine exploratory aggregated outcomes, Kaplan Meier curves showed clear separation among model-score tertiles. Discrimination was highest for dementia (C-index, 0.891) and heart failure (C-index, 0.857). These findings support the feasibility of ECG-based longitudinal risk prediction across multiple diseases. External validation and competing-risk analyses are required to assess generalisability and clinical utility.

Background: The optimal P2Y12 inhibitor after percutaneous coronary intervention (PCI) remains debated. A recent frequentist network meta-analysis (NMA) concluded prasugrel provided optimal efficacy-safety balance. We update our previous Bayesian NMA with recent trials and contrast our findings with the frequentist NMA. Methods: We extended our 2023 systematic review by adding two trials published after our search cutoff. The primary efficacy endpoint was a composite of all-cause mortality, a recurrent non-fatal myocardial infarction, or non-fatal stroke (MACE). The primary safety endpoint was study-reported major bleeding events. Bayesian network meta-analysis with a primary binomial complementary log-log model with log(time) offset and random effects was performed. A statistical workflow with prior and posterior predictive checks, convergence diagnostics, model comparisons, sensitivity analyses and probabilities for a range of practical equivalence (ROPE: HR 0.90-1.11) are also reported. Results: 19 RCTs (n = 60,619) were identified. For MACE, prasugrel's probability of a clinical efficacy benefit (hazard ratio (HR) <0.9) was 70% compared to clopidogrel (HR 0.87, 95% credible interval (CrI )0.76-1.03) and 54% compared to ticagrelor. (HR 0.89, 95% CrI 0.73-1.12). For the ticagrelor versus clopidogrel comparison 76% of the posterior probability (HR 0.98, 95%CrI 0.82-1.18) lies in the ROPE. Ticagrelor (HR 1.26, 95%CrI 1.01-1.63) and prasugrel (HR 1.1, 95%CrI 0.88-1.26) showed 86% and 51% probabilities respectively of meaningful bleeding harm (HR > 1.1) versus clopidogrel. Conclusions: Despite 19 RCTs and approximately 60,000 patients, the Bayesian framework revealed clinically important uncertainties and identified probable regions of equivalence among the different P2Y12 inhibitors that were under appreciated with the previous frequentist publication.

We propose two new tests to detect drug effects (DE) in trials of one to very few patients followed during two periods (before and after initiation of a treatment). Both methods use longitudinal natural history data to inform the estimation of each patient's DE. The first method uses a non linear mixed effect model (NLMEM) reflecting an expected natural history with a hypothetical drug effect, to estimate the Conditional Distribution of the Drug Effect (CDDE). The second method trains a Pareto Depth Analysis (PDA) algorithm, a machine learning based approach based on outlier detection, that we implement using data simulated under the NLMEM. We evaluated the two tests with a simulation study. We used data from the PROSPAX study in Autosomal Recessive Cerebellar Ataxias (ARCAs, to derive a NLMEM for the Scale for the Assessment and Rating of Ataxia score. The CDDE method provided controlled type I error and, in some scenarios, adequate corrected power, though sensitivity analyses showed vulnerability to misspecification. The PDA method demonstrated lower statistical power except with high score precision. These results highlight different strategies for quantifying treatment effects in ultra rare, patient' specific trials. They can inform methodological design for future ARCA precision therapies.

Early identification of Alzheimer's disease and related dementias (ADRD) remains challenging despite its importance for timely intervention, management of modifiable risk factors, and care planning. We developed and evaluated ADRD onset prediction models using longitudinal electronic health records (EHRs) from the All of Us Research Program at clinically meaningful lead times of 6, 12, 24, and 36 months before diagnosis, benchmarking interpretable count-based representations against four publicly available pretrained clinical foundation models (CLMBR-T, GPT-style, LLaMA-style, and Mamba) across multiple ADRD phenotype definitions. Count-based models consistently achieved the highest discrimination and calibration across all cohorts and prediction horizons. Predictive performance declined with increasing lead time for all approaches; however, the performance gap between count-based and pretrained representations progressively narrowed, with foundation models achieving comparable AUROC of 0.719 (compared to the AUROC of 0.738 of count-based model) at the 36-month horizon while providing higher sensitivity and F1 scores under a fixed operating threshold. External validation with zero-shot evaluation on UChicago EHRs exhibited limited generalizability for count-based and pretrained clinical foundation model based representations. These findings demonstrate that transparent count-based EHR representations remain the strongest overall approach for ADRD onset prediction, while pretrained clinical foundation models provide complementary advantages for long-term risk identification and establish a benchmark for evaluating transferable clinical representations in temporal ADRD risk prediction.

Large language models (LLMs) are increasingly used to draft medical manuscripts, yet their citations are unreliable and clinicians lack a validated way to verify them. We evaluated three frontier LLMs, Claude Opus 4.8, GPT-5.5, and Gemini 3.5 Flash, generating 270 cardiology narrative reviews with web search enabled, and verified all 8,050 references against PubMed. Problematic references accounted for 11.5% of GPT-5.5 output, 29.2% of Claude output, and 29.6% of Gemini output (P < 0.001), with no significant gradient across topics of differing publication volume (P = 0.052). Misattribution, a valid PubMed identifier that resolves to a different article, made up 77% of errors, whereas fabrication was rare (0.6%). Against an expert-adjudicated set of 270 references, an LLM-based Chain-of-Verification (CoVe) detected 60 of 62 problematic references (sensitivity 96.8%, specificity 98.6%), including every misattribution and fabrication. LLM-generated citations require identifier-level verification, and CoVe provides it at expert-level accuracy.

Stroke remains a leading cause of mortality and long-term disability worldwide, yet rapid diagnosis is often limited by the shortage of trained radiologists, particularly in resource-constrained settings. Automated analysis of CT imaging offers a potential solution, but existing methods often struggle to achieve clinically generalisable performance while jointly addressing multiple diagnostic tasks. Here we present the Intelligent Integrated Stroke Diagnosis System IISDS, an end-to-end deep learning framework built upon StrokeGNN, a graph-based architecture that integrates 3D contextual feature extraction with U-Net-based 2D lesion segmentation to enable comprehensive stroke analysis from non-contrast CT scans. IISDS performs stroke subtype classification, lesion segmentation and lesion volume estimation within a unified pipeline. To develop and validate the system, we collected and curated BGD-ISD through a collaboration between AI researchers, neurologists, radiologists and clinicians, resulting in a large multi-centre dataset comprising 1,507 CT scans from 597 stroke cases acquired across six hospitals and medical centres in Bangladesh. Across BGD-ISD and multiple publicly available datasets, IISDS achieves state-of-the-art performance on all tasks, improving segmentation accuracy by [&ge;]0.011 Dice score, reducing lesion volume estimation error by [&ge;]0.3 average symmetric surface distance (ASSD), and increasing classification performance by [&ge;]0.018 area under the receiver operating characteristic curve (AUC) compared with existing approaches. These results demonstrate the potential of graph-based deep learning to enable clinically generalisable, automated and scalable stroke diagnosis from CT imaging, supporting rapid clinical decision-making, particularly in healthcare environments with limited access to expert radiological interpretation.

Background. Routine service databases are attractive sources of training labels for clinical prediction models, but the processes that write those labels are rarely audited before the labels are used. In a deployed community cognitive-screening programme, we audited the routine cognitive-status label, built a matrix of twenty-four model arms over the same patients under a specialist reference standard, and measured what each supervision choice bought or cost. Methods. The study cohort is the 672 individuals whose cognitive status was recorded by a titled (attending-or-above) physician, that record being the reference standard; after holding out one institution entirely, a development panel of 642 individuals at 38 institutions. The routine cognitive-status label these individuals also carry was first audited at the operator level: for each data-entry account we counted diagnoses entered and the proportion recording any impairment, and tested a competing bulk-timestamp explanation. Twenty-four arms span the supervision choices such a programme faces: an incumbent 21-variable logistic regression; local language models (Qwen2.5-1.5B/3B, Qwen3-4B/8B) zero-shot, with chain-of-thought, fine-tuned on physician labels, on routine labels with and without decontamination, or on a proxy scale-band task; preference-optimised (DPO) and reinforcement-trained (GRPO) variants; a proprietary frontier model queried zero-shot; and knowledge distillation of that frontier model into the regression and into the local 4B, using 943 teacher-labelled records from the programme's unlabelled pool. All arms are scored out-of-fold under one five-fold split grouped on registry-resolved institution clusters (no cluster spans a fold); paired contrasts use a 2,000-draw cluster bootstrap. Results. 181 operator accounts (each entering at least 100 diagnoses with zero recorded impairments) account for 45,315 rows - 40.5% of the outcome column; recorded impairment falls monotonically with account volume (15.7% for 1-9 rows to 0.7% for 500-999); a bulk-timestamp explanation was tested and refuted, identifying the write-time column as a migration artefact. Under the specialist standard, no locally fine-tuned arm beat the incumbent regression (AUROC 0.926): physician-label SFT reached 0.924 (4B), DPO 0.881, and GRPO 0.789; the pre-registered two-stage proxy-then-RL recipe was worse than its single-stage contaminated baseline (-0.030, 95% CI -0.077 to -0.004). Chain-of-thought reduced discrimination at every size (-0.072, -0.080, -0.041 at 1.5B/3B/4B; -0.012, n.s., at 8B). The frontier model scored 0.932 (vs. regression +0.007, n.s.). The distilled 4B reached 0.940 - above the incumbent (+0.014, 0.004 to 0.031) and above its own teacher (+0.008, 0.001 to 0.017) - with near-teacher calibration; it reached the teacher's level by 50 teacher labels and changed little beyond 200. Conclusions. The audit and the arm matrix support one deployment recipe: audit the routine label at the operator level before training on it; do not expect fine-tuning, preference optimisation, or reinforcement learning on a few hundred specialist cases to beat a well-calibrated regression; and if a frontier model is available but undeployable, spend a bounded number of queries on it as a labelling instrument and distil. A companion paper uses these frozen predictions to quantify how evaluation design choices compare with model choice.

Prognostic models in oncology are developed one cancer at a time, from that cancer's own labelled outcomes, and fail where prognostic information is scarcest. Rare cancers account for roughly a fifth of diagnoses and most paediatric malignancies, yet seldom supply enough events for a reliable time-to-event model. We therefore asked whether a representation learned without outcome labels can supply what those cohorts cannot. A Transformer encoder was pretrained by masked field-value modelling on 9425135 tumour records from the SEER 17 registries, diagnosed in 2000 to 2023. Only diagnosis-time fields passing a fail-closed coding-verification gate were admitted, and each record was emitted as an era-specific and a harmonised view, keeping two decades of recoding auditable. The encoder was then frozen and read by a linear Cox head for overall survival. Nine rare cancers were removed from the pretraining corpus entirely, each requiring an independent pretraining run. On a sealed test partition, all nine exceeded an architecture-identical random frozen encoder in Harrell concordance by +0.0034 to +0.0368, every lower confidence limit above zero. At 256 labelled patients, all 67 cancers favoured the pretrained representation over budget-matched Cox regression, median difference +0.0283. The advantage was bounded: given the entire training set, Cox regression was favoured in seven of nine rare cancers. The encoder did not outperform a field-frequency baseline on its own objective, so upstream reconstruction did not predict downstream transfer. Outcome-agnostic registry pretraining carries prognostic signal into cancers it has never seen, and is most useful where labels are fewest, without establishing clinical utility.