JCO Clinical Cancer Informatics

Clinical records preserve patient findings and final recommendations but rarely capture how specialists select evidence, develop alternative interpretations, disagree, and converge on a management strategy. We describe a dataset derived from ten retrospectively selected, completed cases of histologically confirmed colorectal cancer. For each case, an administrator uploaded de-identified textual clinical records to an authenticated web-based annotation platform. A medical oncologist, surgical oncologist, and radiation oncologist, each with more than ten years of clinical experience, independently linked relevant source fragments to patient-specific interpretations and documented clinical problems, differential hypotheses, uncertainty, and diagnostic and treatment strategies. After the three records were locked, a large language model generated candidate questions for review by the tumour-board chair. The four physicians then discussed each case online; recordings were automatically transcribed and verified by two medical residents. The dataset comprises 30 source-linked specialist reasoning records, 10 verified case-level discussion transcripts, 10 final tumour-board decisions, and associated provenance and quality-control metadata. It supports research on evidence-grounded reasoning, cross-specialty disagreement, multidisciplinary decision-making, and multi-agent clinical decision support.

Purpose: Digital-twin frameworks can potentially offer a pathway toward individualized treatment but remain largely untested in radiation oncology. We present a causal digital-twin model for head-and-neck (H&N) cancer that estimates patient-specific survival benefit from chemoradiotherapy (CRT) versus radiotherapy (RT) alone, stratified by HPV status. Methods: Using the public RADCURE cohort (n = 3,346), we developed a causal digital-twin framework in which patients were divided into HPV-positive and HPV-negative groups and assigned to training (RT start year [&le;] 2007) and test sets (RT start year > 2007). Bayesian-network structure learning identified each cohort's Markov Blanket around the 2-year survival endpoint (SVy2), yielding the minimal causally relevant feature set. A causal-forest model (double machine learning, DML framework) estimated the individualized treatment effect {tau}=P(SVy_2|CRT,X)-P(SVy_2|RT,X), with {tau} denoting the model-estimated effect. Overlap restrictions were applied to mitigate extrapolation. Predictive discrimination (AUC, C-index), average treatment effects (ATE), individualized effects (ITE), and decision-curve analyses were computed. Results: For HPV-negative disease, internal evaluation demonstrated strong discriminative performance (test AUC = 0.72; C-index = 0.81, p < 0.01). The estimated test-set ATE favored CRT with a modest but positive improvement in SVy2 (mean {tau} = 0.13). Patterns in {tau} were clinically coherent, including increased benefit in node-positive and good-performance patients and reduced benefit in high-burden or poor-performance presentations. Decision-curve analysis showed that a {tau} -guided selective-CRT strategy provided higher net benefit than uniform treatment policies. In HPV-positive disease, treatment-effect heterogeneity was attenuated, consistent with its more uniformly favorable prognosis. Conclusion: This study demonstrates a proof-of-concept causal digital-twin framework for interpretable, patient-specific survival simulation in H&N cancer. The model captures clinically plausible treatment-effect heterogeneity and establishes a foundation for multi-cohort validation and future integration into individualized treatment decision support.

Cancer registries are essential for population-level surveillance, epidemiologic research, and public health planning, but their accuracy depends on human review of key variables documented in unstructured clinical text. This is time-consuming, and current quality assurance (QA) typically relies on reviewing fewer than 10% of records at random, which can miss facility- or time-specific errors. We developed a large language model (LLM)-based framework for targeted cancer registry QA. The framework uses LLMs to extract biomarker values from clinical text, compares them with registry entries, and flags disagreements for expert review. Six open-source LLMs were benchmarked locally in a privacy-preserving environment, and the best model per site and report type was applied to breast and prostate cancer, two of the most common cancers in the United States. For the four breast cancer biomarkers ER, PR, HER2, and Ki-67, the disagreement rate against quality-assured registry values in 1,852 breast cancer cases was lowest for ER at 1.67% and highest for Ki-67 at 9.18%. In 920 prostate cancer cases, the disagreement rate with the deployed models stayed below 8% for every categorical Gleason variable, while the exact-value PSA metric reached 17.28%. PSA therefore generated the largest volume of candidate discrepancies, though individual PSA flags require more corroboration than categorical flags. Applied to a full year of records (11,740 breast and 6,792 prostate cases), the framework flagged disagreements and assigned each a probable cause, grounded in NAACCR coding rules, to prioritize expert review. Extracted values are linked to their supporting text, and model, prompt, and pipeline versions are logged for reproducibility. By directing review toward records most likely to contain discrepancies, the framework offers an efficient, auditable alternative to random QA sampling for cancer registries.

Background: Adjuvant therapy decision in early-stage node-negative breast cancer remains one of the most challenging problems in oncology. Although most patients in this population achieve favorable outcomes with endocrine therapy alone, a subset remains at risk of relapse and may benefit from chemotherapy. Accurate identification of high-risk and low-risk patients is therefore essential to prevent both undertreatment and overtreatment. Methods: In this study, machine learning models were assessed using MSK-IMPACT (n=645) and METABRIC (n=503) cohorts comprising node-negative stage I-II breast cancer patients. Three experimental settings were evaluated: training on MSK-IMPACT and testing on METABRIC, the reverse direction, and five-fold cross-validation on the combined cohort. Sensitivity analysis and data harmonization were applied to identify and mitigate cohort-specific bias. Survival analysis, SHapley Additive exPlanations (SHAP) interpretability, and ablation analyses were performed to validate clinical relevance and characterize feature contributions. Results: The combined cohort model achieved an area under the receiver operating characteristic curve (AUC) of 0.788{+/-}0.028 across five-fold cross-validation, while bidirectional cross-cohort validation yielded AUC of 0.741 when testing on MSK-IMPACT and AUC of 0.714 when testing on METABRIC. Survival validation on METABRIC out-of-fold predictions from the combined cohort model showed hazard ratios of 3.01 (95\% CI: 2.28-3.98) for disease-free survival and 2.35 (95\% CI: 1.79-3.09) for overall survival, consistent across Luminal A, Luminal B, and Triple-Negative subtypes. Conclusion: Validated across two geographically independent cohorts, the proposed model enables cost-effective, gene expression-free identification of patients at high-risk of distant relapse and may facilitate adjuvant chemotherapy decision-making in early-stage node-negative breast cancer.

Voxel-level prediction of treatment response from longitudinal fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) can enable spatially adaptive dose planning in advanced non-small cell lung cancer (NSCLC), but point predictions alone lack the uncertainty quantification needed for safe clinical decision-making. Standard conformal prediction (CP) provides distribution-free coverage but yields spatially uniform intervals and assumes exchangeability, an assumption violated by intra-tumoral spatial correlation. We propose a multiscale CP framework built on a re-optimized VoxelForecast generalized least squares model that explicitly captures voxel-level covariance. Variogram sensitivity analysis identified the Stable model as significantly superior to the previously used Matern family. The framework combines hierarchical nested leave-one-patient-out calibration, which restores exchangeability for valid voxel-level coverage, with residual-variance CP (VarCP), which learns a heteroscedastic, location-dependent scale to produce adaptive intervals while preserving finite-sample marginal coverage guarantees. On two prospective NSCLC trial cohorts, FLARE-RT (N=25 locally advanced, chemoradiotherapy; 11,100 voxels) and BRIGHT (N=19 metastatic, chemoimmunotherapy; 26,980 voxels), multiscale feature integration reduced voxel-level prediction error by 35% and 31%, respectively. VarCP maintained nominal coverage while producing 7-18% (FLARE-RT) and 4-8% (BRIGHT) narrower voxel-level intervals than a hierarchically calibrated standard CP baseline (paired Wilcoxon signed-rank, significant at most miscoverage levels), with comparable lesion-level gains (7-18% and 6-10%) via median-based aggregation of voxel residuals. VarCP delivers tighter yet mathematically guaranteed spatial uncertainty maps, supporting both voxel-level dose adjustment and whole-lesion treatment planning in advanced NSCLC.

Living clinical guidelines revise recommendations as evidence emerges, so an AI answer can be faithful to a real guideline and still be out of date. Here we examine this failure using ASCOBench, 288 unique questions in 96 three-turn conversations grounded in versioned American Society of Clinical Oncology (ASCO) breast and prostate cancer guidelines, with oncologist-reviewed reference answers. Oncologists corrected one in four version-sensitive answers drafted by a frontier model, against fewer than one in ten factual ones. Across the guideline corpus, 19 recommendations changed between versions, seven of them reversals, yet only 1 of 14 superseded documents states that it has been replaced. With a strong model, retrieval over several guideline versions made stale answers four times more frequent than answering without retrieval. SentryLine, a verification-first system, lowered incorrect answers to 4.2-5.2% from 9.7-22.9% for baselines across three models. Recognizing change before a guideline declares it remains an open problem.

Background: Clinicopathologic (SARCULATOR) and transcriptomic (CINSARC) models are established prognostication tools in soft tissue sarcoma, but neither directly represent the sequence context of tumor genomic alterations. Evo2 is a genomic foundation model trained to learn the grammar of DNA/RNA sequences. We tested whether Evo2-derived features from tumor short variants, fusions, and rearrangements improve survival prognostication beyond clinicopathologic and transcriptomic risk models. Methods: We analysed formalin-fixed, paraffin-embedded tumor material and clinical data from 102 patients with dedifferentiated liposarcoma (n = 25), high-grade leiomyosarcoma (n = 25), and undifferentiated pleomorphic sarcoma (n = 52). A separate TCGA-SARC cohort of comparable histologies (n = 129) served as external validation. Reconstructable alterations (short variants, fusions and rearrangements) were converted into sequence inputs, embedded with Evo2 and aggregated into patient-level genomic representations. Cox proportional hazards models incorporated genomic features alone (Evo2) or together with clinicopathologic (SARCULATOR) or transcriptomic (CINSARC) predictors across overall survival (OS), progression-free survival (PFS), disease-free survival (DFS), relapse-free survival (RFS), and metastasis-free survival (MFS). Discrimination was assessed using Harrell's concordance index (C-index). Results: Evo2 features were generated from 174 study-cohort sequences and 7298 TCGA-SARC sequences. In the study cohort, adding Evo2 to SARCULATOR increased the reported C-index for OS from 0.620 to 0.770, PFS from 0.625 to 0.755, DFS from 0.612 to 0.714, and RFS from 0.633 to 0.771. In TCGA-SARC, SARCULATOR plus Evo2 increased the OS C-index from 0.595 to 0.716 and the MFS C-index from 0.548 to 0.667. Also in TCGA-SARC, adding Evo2 to CINSARC increased the OS C-index from 0.476 to 0.653 and the MFS C-index from 0.553 to 0.650. These are descriptive point-estimate comparisons. Moreover, in TCGA-SARC Evo2 alone outperformed SARCULATOR alone and CINSARC alone, supporting independent prognostic information within altered sequence-context representations. Conclusions: Evo2-derived genomic representations improved clinicopathologic and transcriptomic prognostication across sarcoma cohorts and survival endpoints. These findings support genomic foundation models as an independent and complementary layer for sarcoma risk stratification through direct modeling of altered tumor sequence biology. Confirmation of model transportability, calibration and clinical utility is required before clinical application.

Objective: To determine how cohort versus nested case-control (NCC) design affects EHR-based prediction of a rare outcome. Materials and Methods: We compared two cohort and five NCC designs for predicting early-onset gastrointestinal cancer using L1-regularized logistic regression (LASSO), XGBoost, and multilayer perceptron (MLP). Models underwent nested five-fold cross-validation with model-specific hyperparameter tuning and temporal evaluation in a held-out prospective cohort representing intended deployment. Model discrimination was assessed using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). Model calibration was assessed using Brier scores and calibration plots before and after recalibration with isotonic regression. Feature-ranking similarity was quantified using rank-biased overlap. Results: The cohort development datasets included 256,854-415,529 patients, with 2-yr cumulative incidence of 0.12%-0.17%, compared with approximately 9% in the 1:10 matched NCC datasets. NCC designs generally produced higher internal discrimination, but temporal performance varied by algorithm and index-date definition. Cohort designs achieved the highest temporal AUPRC for LASSO (0.068), whereas diagnosis-anchored NCC achieved the highest temporal AUPRC for XGBoost (0.086) and MLP (0.070). Temporal AUPRC generally decreased as the required lead time before diagnosis increased. Raw predicted probabilities substantially overestimated observed risks in the temporal cohort. Isotonic recalibration improved probability scaling while largely preserving discrimination, although residual miscalibration remained. Cohort-based feature rankings were more consistent across algorithms, whereas NCC-based rankings depended more strongly on algorithm and index-date anchoring. Conclusion: Study design should be selected and evaluated according to the intended clinical prediction task and deployment setting.

Background. High-throughput drug screening (HTS) is increasingly used in precision oncology across cell lines, patient-derived models such as organoids (PDO) and xenograft organoids (PDxO), and ex vivo patient samples, with multiple active clinical trials incorporating HTS for individualizing treatment. Yet HTS efforts predominantly rely on ad hoc, spreadsheet-based workflows, with no standardized infrastructure for storing, analyzing, and comparing screens across patients, cohorts, or institutions, limiting reproducibility and slowing clinical data exchange. Existing tools address isolated pieces of the workflow, handling either data management or analysis and visualization, but none integrate standardized storage, reproducible dose-response analysis, individualized drug prioritization, and clinician and molecular tumor board friendly reporting in a single open-source framework. Results. We present DS5, an open-source Python framework that operates on an HDF5-based DS5 file format that unifies standardized storage, quality control, dose-response analysis, cohort-level drug prioritization, and automated reporting in a single lightweight package. DS5 enforces raw-data immutability, integrates RxNorm-based drug name standardization, and scales from small clinical cohorts to large pharmacogenomics datasets on a standard laptop. We validated DS5 against two public benchmarks spanning over 428,000 drug, cell line pairs, recovering published LN IC50 values from GDSC2 (Pearson r = 0.973) and Emax values from CTRP (r = 0.961). Using a published breast cancer PDxO dataset (45 compounds by 16 models), we reproduced GR-based profiles via a custom-metric extension and used DS5's cohort-level prioritization to predict in vivo tumor response with ROC AUC = 0.907 across 30 drug PDX pairs. Finally, we demonstrate the utilities of DS5 in an ongoing pediatric brain tumor precision oncology initiative, where its cohort-normalized drug rankings and automated reports directly support molecular tumor board discussions. Conclusion. DS5 provides the cancer basic and translational research community with a standardized file format and reproducible computational analysis methods to support functional drug screening experiments; and its reporting features facilitate bridging raw drug screening measurements to translational decisions. DS5 is released open-source under the permissible MIT license.

Background: As required by the 21st Century Cures Act, all test results, including final radiology reports, are released to patients as soon as the results are finalized. However, radiology reports are primarily intended for communication between medical professionals and can often be difficult for patients to interpret, particularly in cancer care settings. We developed a tool that leverages a large language model (LLM; ChatGPT-5.2, OpenAI, San Francisco, CA) to draft a patient-friendly plain language version of imaging results (Patient-Friendly Scan Result Generation [PFSG] tool) that oncologists can incorporate into their communication with patients via patient portals. This article describes our mixed-methods study protocol. Objective: Our primary aims are to: 1) assess whether patients who receive patient-friendly versions of scan results report better understanding of their results than patients who do not; 2) assess whether introduction of the PFSG tool affects provider workload; and 3) explain how and under what circumstances the PFSG tool enhances communication and workflows, and affects patient and provider outcomes. Methods: We will use an embedded mixed-methods, prospective cohort design within a quality improvement (QI) initiative. The PFSG tool will be implemented using a modified three-wave staircase trial design, with providers in randomized cohorts of oncology disease groups progressively gaining access to the tool over the course of the study. Approximately 60 providers from 11 oncology disease focus groups will be randomized. Longitudinal provider outcomes will be measured through surveys at one month pre-implementation, one month post-implementation, and three months post-implementation. Patient outcomes will be measured cross-sectionally through surveys at one month pre-implementation (n=350) and at post-implementation result point-of-contact (n=350). Additionally, we will conduct interviews with patients and oncologists to inform the iterative design of the tool and its implementation. Results: This study was funded in September 2025 and data collection began in February 2026. As of September 2026, the PFSG tool has been rolled out to all oncology providers, with Cohort 3 post-implementation data collection beginning in August 2026. A total of 738 patient baseline surveys have been collected and patient post-implementation data collection is in progress. Conclusions: This is the protocol for a study of the implementation of a patient-friendly results tool in oncology provider workflows to facilitate improved communication of radiology results to oncology patients. The results of this study will lend evidence to the utility and acceptability of the study tool, and tools like it, to close an important oncology care gap. If the tool is shown to improve key oncology patient and provider outcomes, it may lead to development of a wider release for more general patient populations.

Biologically rich clinical trials generate heterogeneous datasets whose relationships and clinical context must be preserved. Here we present the POETIC-AI environment, a trial-centric digital environment for integrating, analysing, and exploring multimodal data from POETIC, a randomised phase III trial of peri-operative aromatase inhibition in hormone receptor-positive breast cancer (NCT02338310). We implemented a graph-based ontology linking patients, treatments, sampling timepoints, specimens, assays, images, spatial regions, and derived measurements; and designed code-based and graphical analytical workflows. Using proliferation as a test case, we reproduced POETIC's established reduction in Ki67% and examined variation underlying sample-level measurements. In response-enriched cohorts, mean Ki67% decreased while relative variation among pathologist-assessed fields increased (paired Wilcoxon, P < 0.0001). Pathologist-assessed Ki67% correlated strongly with an 18-gene transcriptomic proliferation signature (r = 0.834), and both decreased significantly during treatment (P < 0.0001). An annotation-free pixel-wise Ki67 H-score showed weaker correlations and a similar but non-significant reduction (r = 0.327-0.472; P = 0.15), highlighting differences between tumour-focused and whole-section image quantification. Digital pathology and spatial molecular measurements remained linked to their source images and annotated tissue regions, enabling exploration within their morphological and spatial context using bespoke interactive applications. POETIC-AI demonstrates how the structure, provenance, and tissue context of evolving multimodal clinical-trial data can be preserved within a connected research environment, supporting reproducible analyses and the laying the groundwork for future development of biologically interpretable computational biomarkers.

Objectives Contemporary oncology has achieved remarkable advances in biomedical treatment while increasingly emphasizing patient-centered care. Nevertheless, routine oncology consultations must address multiple competing clinical priorities, potentially limiting opportunities for patients' experiences to become visible. To address this gap, we established a Dialogue-Based Outpatient Clinic, a complementary patient-directed consultation model embedded within routine oncology care. This study explored the dimensions of patients' experiences that became visible through this consultation model. Methods We conducted a qualitative study of 44 consecutive patients who attended a Dialogue-Based Outpatient Clinic embedded within routine oncology care between 2021 and 2023. Patients attended the clinic either on their own initiative or following a recommendation from a healthcare professional who recognized that they were struggling. Regardless of how they came to the clinic, patients independently determined the consultation agenda. Results Of the 44 patients, 13 (29.5%) sought biomedical information regarding diagnosis, treatment, or adverse effects, whereas 31 (70.5%) discussed concerns extending beyond biomedical issues. Reflexive thematic analysis of these 31 patients identified four interconnected themes: (1) making sense of a cancer diagnosis within one's own life; (2) negotiating treatment in relation to personal values and life priorities; (3) being treated for disease but not seen as a person; and (4) reconstructing life and identity through the cancer journey. Across all themes, patient-directed dialogue consistently revealed how patients understood cancer within the broader context of their lives rather than as an isolated biomedical condition. Conclusions Patient-directed dialogue within routine oncology care revealed dimensions of patients' experiences that may be less readily recognized in conventional consultations. A complementary consultation model that allows patients to determine the agenda may provide opportunities for these experiences to become clinically recognizable.

Decision models like Jev read text and return probabilities, categories, or ordinal levels, with confidence scores for the latter two, at a fraction of large language models' (LLMs) cost. We explored how Jev could be integrated into clinical workflows for all these output types: mortality risk estimation, categorical Choosing-Wisely decisions for patient questions, and ordinal colon cancer staging. With standard LLM-style inputs, Jev trailed five flagship LLMs: mortality AUROC was 0.822 versus 0.837-0.846; Choosing-Wisely accuracy, 87.7% versus 98.4-100%; and staging accuracy, 45.4% versus 94.2-100%. Developer-recommended input alterations improved staging accuracy to 75.8% with medical context, 82.1% with task decomposition, and 85.8% with both. None of Jev's staging errors had full confidence. Using Jev for fully confident staging decisions and LLMs for the rest was non-inferior to each LLM alone and cut flagship LLM costs by 33-44%. Prompted as designed, decision models may efficiently share structured clinical tasks with LLMs.

Background Chemotherapy (CT) de-escalation strategies based on dual HER2 blockade with trastuzumab and pertuzumab (HP) have shown promising efficacy in HER2-positive (HER2+) early breast cancer (EBC), but biomarkers to guide patient selection are lacking. The Ataraxis Breast platform (ATX) uses artificial intelligence (AI) models integrating hematoxylin-eosin (H&E) whole-slide images and clinical variables to predict pathological complete response (pCR) and long-term outcomes. We evaluated ATX in the phase II PHERGain (NCT03161353) and PHERGain-2 (NCT04733118) trials. Methods The ATX NEO model was assessed for pCR prediction in PHERGain (n=306 evaluable patients) and PHERGain-2 (n=383 evaluable patients). The ATX RISK model was evaluated for prognosis in PHERGain (n=309 evaluable patients). Prespecified cutoffs classified patients into low-, medium-, and high-probability groups for pCR and low- or high-risk groups for recurrence. Associations with pCR were assessed using logistic regression. Prognostic performance was evaluated using Cox proportional hazards models, Kaplan-Meier estimates, and Harrell's C-index. Results Overall pCR rates were 40.2% in PHERGain and 60.3% in PHERGain-2. Continuous NEO scores were significantly associated with pCR in both PHERGain (odds ratio [OR], 1.28; 95% CI, 1.11-1.48; P<0.001) and PHERGain-2 (OR, 1.30; 95% CI, 1.13-1.52; P<0.001). Using prespecified cutoffs, pCR rates increased across low-, medium-, and high-probability groups (19.2%, 41.3%, and 48.3% in PHERGain; 29.5%, 61.6%, and 70.1% in PHERGain-2). In PHERGain, the RISK score was significantly associated with event-free survival (hazard ratio per 1-SD increase, 1.56; 95% CI, 1.17-2.08; P=0.002; C-index, 0.66). Five-year event-free survival was 93.9% in the low-risk group and 83.1% in the high-risk group. Conclusions NEO score was consistently associated with pCR across two independent HER2+ EBC trials, while RISK score provided prognostic information for long-term outcomes in PHERGain. Prespecified score categories identified clinically distinct response and risk groups. These findings support further evaluation of AI-based biomarkers to refine patient selection for CT de-escalation strategies in HER2+ EBC.

Fairness audits of chest radiograph artificial intelligence usually score subgroup performance against labels extracted from radiology reports by natural language processing (NLP). We tested whether these labels can make a model look fairer than it is, a failure we call phantom fairness. On 4,376 NIH ChestX-ray14 images with both NLP and radiologist-adjudicated labels, we audited classifiers from three backbones by sex, age, projection and older women. NLP labels missed 54% to 59% of radiologist-confirmed airspace opacity, pneumothorax and nodule or mass, and every fracture. These misses were not random: models scored them lower than NLP-caught positives for every finding and backbone (median rank-biserial correlation 0.20 for the primary backbone and 0.16 for the others; all adjusted p [&le;] 0.038), so the labelling errors hid the models' own mistakes. Subgroup gaps were larger on adjudicated labels in 16 of 24 comparisons, none individually significant. For pneumothorax in women older than 60 years, the NLP audit showed a sensitivity gap of 0.064 against a true gap of 0.184; removing the hardest positives reproduced this concealment (0.124 against 0.120 observed), whereas random label errors produced less (0.077). Fairness audits based on NLP-extracted labels should be treated as provisional until checked against radiologist-adjudicated labels.

PURPOSE: To demonstrate an end-to-end research workflow combining a large electronic health record (EHR) resource and a multi-agent artificial intelligence (AI) system governed by Human-Guided Orchestration (HGO), using sequential sacituzumab govitecan (SG) and trastuzumab deruxtecan (T-DXd) orders in breast cancer as the clinical use case. METHODS: A June 2026 EHR release contained 4.95 million patients, including 156,051 with breast cancer codes. Muse Spark supported cohort construction, executable statistical analysis, data-quality checks, literature retrieval, and reporting executed through Symbiotic Agent Teaming (SAT) under investigator direction. The main clinical measure was the second-drug record span, from first to last order. Analyses were descriptive and unadjusted; orders did not confirm administration or progression. RESULTS: The workflow identified 106 patients: 73 had T-DXd ordered first and 33 had SG ordered first. Median second-drug spans were 63 days (interquartile range [IQR], 29-116) for SG after T-DXd and 49 days (IQR, 21-126) for T-DXd after SG; the median difference was -14 days (95% bootstrap CI, -41 to 21; distributional P = 0.42). Reported outputs additionally included cohort definitions, aggregate analyses, evidence context, figures, and a preliminary draft manuscript. CONCLUSION: This study provides a concrete model for integrating large EHR resources, agentic AI, and human expertise across the research cycle. The ADC use case illustrates how computational execution and Expert-in-the-Loop (EITL) adjudication can produce interpretable research outputs and refine their scientific meaning, providing a foundation for extending and evaluating this approach across clinical questions.

Background: Artificial intelligence (AI) models for mammography can achieve high diagnostic performance when training and test data originate from similar populations and imaging environments, but performance often deteriorates under cross- cohort domain shift. Differences in acquisition systems, image characteristics, preprocessing, population composition, and disease prevalence can introduce distributional variation that is poorly captured by conventional within- dataset validation. Recent studies have further shown that Foundation Model representations can retain substantial dataset-specific information. However, it remains unclear how image representation influences diagnostic transfer to completely unseen cohorts and whether domain-adversarial learning provides comparable benefits across different representation types. Methods: We systematically evaluated cross-cohort mammography classification using seven independent datasets spanning diverse geographic, institutional, scanner, and imaging environments. Three image representations were compared: handcrafted radiomic features, EfficientNet-B0-derived features, and embeddings from the medical vision-language model MedSigLIP-448. Each was evaluated under three preprocessing conditions: unprocessed images, Otsu-based breast-region extraction, and Otsu extraction combined with Contrast Limited Adaptive Histogram Equalization (CLAHE). Patient-level classification was performed using a mean-pooled multilayer perceptron (MLP) baseline and an attention-based Domain-Adversarial Neural Network (DANN). Generalization was evaluated using a multi-source Leave-One-Domain-Out (LODO) framework in which six datasets were used for model development and the seventh was completely withheld as an unseen test cohort, iteratively across all seven datasets. Performance was primarily assessed by area under the receiver operating characteristic curve (AUC-ROC). Results: Cross-cohort performance depended strongly on image representation and model architecture. MedSigLIP consistently provided the highest performance on unseen cohorts. With MedSigLIP features, mean LODO AUC-ROC increased from 0.598 with MLP to 0.685 with DANN for unprocessed images, from 0.607 to 0.678 after Otsu preprocessing, and from 0.603 to 0.659 after CLAHE+Otsu preprocessing. When DANN-MLP differences were evaluated across held-out datasets and preprocessing conditions, a significant gain was observed for MedSigLIP after correction for multiple comparisons, whereas corresponding gains were not significant for Radiomics or EfficientNet-B0. A mixed-effects analysis similarly showed a positive MedSigLIP X DANN interaction, although this narrowly missed conventional statistical significance. Unprocessed images produced the highest numerical mean performance for MedSigLIP-DANN, but preprocessing itself had no significant overall effect. Notably, linear probing showed strong dataset discriminability in MedSigLIP and EfficientNet representations but little linearly recoverable dataset information in Radiomics, indicating that low dataset separability alone did not predict better diagnostic generalization. Conclusion: Cross-cohort mammography generalization depends strongly on the choice of image representation and its interaction with the classification strategy. MedSigLIP provided the most transferable representation in our multi-source unseen-domain evaluation, and the performance gain associated with DANN was concentrated in MedSigLIP rather than consistently observed across representation types. These findings suggest that the effectiveness of domain-adversarial learning is representation-dependent and that neither conventional preprocessing nor low dataset discriminability alone ensures robust transfer to unseen clinical cohorts.

Objective Specialty referral packets often arrive by fax and clinicians must manually read and organize a large amount of disparate information to affect continuity of care. Commercial large language models (LLMs) can organize packet information, but protected health information (PHI) must be secured. We evaluated whether an open-weight pipeline running entirely on one consumer graphics card can accurately extract clinical information from faxed referral packets. Materials and Methods We generated 200 synthetic colorectal cancer referral packets from simulated patient records, rendered them as faxes without a text layer. A local pipeline transcribed each packet with an optical character recognition model (OvisOCR2) and extracted structured data with an open-weight LLM (Qwen3.8 27B). Twenty outcomes spanning patient identifiers, laboratory results, dates, diagnoses, procedures, medications, imaging, and event chronology were scored against ground truth on 140 validation packets. Results Laboratory values were extracted exactly in 98.9% of 7879 instances, dates in 99.7%, diagnoses in 100%, and procedures in 92.9%. Timeline recall was 96.7% and precision 99.4%. Weighted accuracy was 0.990 (95% CI 0.987 - 0.992). None of 11,993 extracted facts were fabricated. Median processing time was 90 seconds per packet. Conclusion An open-weight pipeline on consumer hardware accurately extracted multi-domain clinical information from synthetic faxed referrals.

Cross-cohort evaluation of multi-omics prognostic models can fail because molecular features are not assayable, fitted models do not transport, feature selection is unstable, or molecular data add little beyond clinical predictors. We used a dual-track evaluation design with TCGA-BRCA as the source cohort and METABRIC as the independent target cohort. Track A tested outcome-blind transport of fixed historical RNA and copy-number models without target-cohort refitting, whereas Track B reconstructed the dependency-aware selection procedure within METABRIC using leakage-controlled repeated cross-validation. Literal transport provided little or negative incremental overall-survival discrimination: RNA reduced Harrell's C-index relative to clinical prediction (Delta C = -0.0136), while copy number was essentially neutral. Track B showed no reliable overall-survival gain for RNA, copy number, or mutation; methylation and the reconstructed multimodal model performed worse than matched clinical comparators. In a protocol-locked post-hoc extension, the same fixed features were retained but model parameters were re-estimated within METABRIC. Incremental C-index increased by 0.0162 for RNA, 0.0091 for copy number, and 0.0201 for the combined panel. Random-panel benchmarks showed that the RNA gain was common among alternative panels, whereas the historical copy-number panel ranked above all 200 panels sampled from its recovered assayable candidate space. The recurrence-free-survival sensitivity analysis showed a positive RNA contrast (Delta C = 0.0146). These results separate failure of fixed-model transport from loss of information in the underlying feature set and show that successful local redevelopment does not by itself establish feature-set specificity.

Doubly robust survival estimators require nuisance models that the available information can support. We examine design and estimator dependence in TCGA breast cancer data. Treatment indicators disagree for 337 of 1076 patients; 67.4% have unobserved five-year binary outcomes; and receptor imputation expands the triple-negative subgroup from 156 directly measured patients to 226. We compare chemotherapy initiation by day 90 with non-initiation, using adjusted restricted mean survival time (RMST) over the next 1095 days. Weighting-only estimates are 55 days (95% bootstrap interval: -24, 133) in triple- negative disease (16 deaths) and 38 days (12, 72) in the whole cohort (72 deaths). A Cox outcome-model benchmark yields 108 and 35 days, respectively, demonstrating material model dependence in the smaller population. A parameter-to-information ratio summarizes the burden on complete-outcome regressions within cross-fitting. Its simulation-calibrated warning level is 0.31 (0.28 - 0.35) for the original regime; additional simulations examine covariate-dependent censoring, delayed benefit and nonlinear misspecification. The ratio does not certify estimator validity or replace weight diagnostics. The contribution is a reproducible workflow that separates design, nuisance-model support and estimator stability before interpreting observational survival contrasts.