The Lancet Digital Health

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

Background Histological diagnosis of early-stage mycosis fungoides (MF) is hindered by profound overlap with benign inflammatory dermatoses (BIDs), leading to diagnostic delays and extensive ancillary testing. We developed MIMIC (Multiple Instance-learning for Identification of Mycosis fungoides In Cutaneous biopsies), a weakly supervised deep learning model designed as a triage tool at initial H&E whole slide image (WSI) review to distinguish classic patch and plaque stage MF from BIDs. We externally validated the model and evaluated its clinical utility. Methods In this retrospective multicentre study, we trained a base model using weakly supervised attention based multiple instance learning on 3,339 WSIs from two Dutch centres. Crucially, all MF training labels were derived from a deeply phenotyped national cohort featuring strict multidisciplinary expert panel consensus diagnoses (the clinical gold standard). Transportability was evaluated on 371 WSIs from four independent European centres. A blinded reader study on 171 WSIs compared morphology only performance of MIMIC with 11 (dermato-)pathologists. We then retrained an updated model on all retrospective multicentre data and assessed clinical utility in a strictly held out, consecutive Utrecht cohort (2022-2023; 486 accessions, 863 WSIs). Primary analysis focused on classic MF versus BIDs (453 accessions). Decision curve analysis, using Platt scaled probabilities to correct for spectrum bias, evaluated net benefit at a prespecified, safety oriented threshold of 0.04. Findings The base model showed good multicentre transportability (mean centre specific AUROC 0.91; pooled AUROC 0.84). In the reader study, MIMIC achieved an AUROC of 0.87, exceeding the mean pathologist AUROC (0.79) and the best individual reader (0.83). In the consecutive MF versus BID cohort, the updated model achieved an AUROC of 0.87 (95% CI 0.81-0.92). At the 0.04 threshold, sensitivity was 97.8% (44/45 MF cases) and specificity 50.2%, reducing unnecessary ancillary workups by 39.9 per 100 screening cases versus a test all strategy. Interpretation By identifying nearly half of BIDs as low risk while preserving near complete sensitivity for classic early stage MF in a European digital pathology workflow, this unimodal H&E approach offers a scalable digital solution to reduce defensive ancillary testing and accelerate the diagnostic journey for patients with MF. Further validation is needed in non European centres and in populations with darker skin phototypes.

Background: GPs are often the first clinicians to assess acute chest pain in gatekeeper healthcare systems, but usually with limited diagnostic support. They must distinguish conditions that can be managed safely in primary care from those requiring immediate hospital evaluation, particularly acute coronary syndrome (ACS), while avoiding unnecessary over-referral to already pressured emergency departments. We evaluated HEART-GP, a novel GP-led strategy combining structured clinical assessment, fingerstick point-of-care high-sensitivity troponin I testing, and ECG when available. Methods: We performed a prospective, multicentre, paired diagnostic accuracy study in four Dutch out-of-hours primary care centres implementing the HEART-GP strategy. Patients with acute non-traumatic chest pain and possible ACS according to the treating GP were enrolled from March 2023 to August 2025. Before troponin results were available, GPs recorded intended management, allowing paired comparison with unaided clinical judgement. The HEART-GP strategy recommended urgent referral for high clinical suspicion, troponin of >=4 ng/L, or ischaemic ECG. The primary outcome was adjudicated 6-week major adverse cardiovascular events (MACE). The study was registered as ISRCTN11954040. Findings: Of 1104 patients who underwent HEART-GP, 917 were included. MACE occurred in 62 (6.8%) and myocardial infarction in 52 (5.7%). HEART-GP classified 478 (52.1%) as low risk. For MACE, sensitivity was 95.2% (95% CI 86.5-99.0) and negative predictive value was 99.4% (98.1-99.8); three MACE occurred in the low-risk group, with no cardiac deaths. For myocardial infarction, sensitivity was 98.1% (89.7-100) and negative predictive value was 99.8% (98.6-100). Compared with unaided GP judgement, HEART-GP improved sensitivity for MACE (95.2% vs 67.7%; p<0.001) and myocardial infarction (98.1% vs 67.3%; p<0.001), and reduced urgent referrals (439 [47.9%] vs 483 [52.7%]; p=0.042). Interpretation: HEART-GP showed high rule-out safety and may support GP diagnostic certainty and efficient referrals in patients with possible ACS. Implementation requires training, governance, safety-netting, and local pathways.

Objectives: To test whether radiology report evaluation metrics distinguish clinically meaningful errors from textual changes and align with radiologist-assessed error burden. Methods: Cross-dataset evaluation used ReXErr-v1 (2,708 report pairs; 5,724 paired error sentences) and 100 RadEvalX report pairs with expert error counts. BLEU-4, ROUGE-L and METEOR were assessed in ReXErr-v1; RadEvalX analyses included these plus BERTScore, CheXbert, RadGraph F1 and RadCliQ. Outcomes were ReXErr-v1 pairwise win rate and AUROC for clinical-content versus linguistic errors, and RadEvalX Spearman correlation with clinically significant error count and AUROC for any significant error. Confidence intervals used 10,000 clustered percentile bootstrap resamples; Holm adjustment-controlled multiplicity. Results: ReXErr-v1 paired-sentence win rates were 0.986 for BLEU-4, 0.999 for ROUGE-L and 0.998 for METEOR, but discrimination of clinical-content from linguistic errors was modest (AUROC 0.609-0.620). Penalty magnitude was strongly associated with textual change after adjustment for error type (normalised character edit distance coefficient 0.746; 95% CI 0.705-0.788; P<0.001). In RadEvalX, CheXbert showed the highest correlation with clinically significant errors (rho=0.413; 95% CI 0.223-0.578) and highest AUROC (0.742; 95% CI 0.638-0.836). Conclusions: Near-ceiling sensitivity to textual corruption did not imply sensitivity to clinical significance. CheXbert showed the highest alignment with expert error assessment, although pairwise superiority was not demonstrated over all comparators and performance remained moderate.

Background: Clinical registries and institutional quality control (QC) in pathology heavily depend on complex histopathological data. However, critical tumor characteristics are mostly locked in unstructured free-text reports. We hypothesize that locally deployed, resource-efficient language models operating within a secure institutional network can reliably extract relevant variables from German reports of hepatocellular carcinoma (HCC), thereby enabling automated data curation without transferring sensitive patient data. Methods: We developed a local, privacy-preserving open-weight Large Language Model (LLM) pipeline consisting of two consecutive stages: Stage 1 utilizes a lightweight gatekeeper model to isolate the target cohort of confirmed Hepatocellular Carcinoma cases from a screening pool of N = 2,864 reports. Stage 2 deploys a high-capacity model to extract 15 predefined histopathological variables, including tumor-node-metastasis staging, histological grading, vascular invasion, and selected immunohistochemical features into structured JSON format. For both Stages, pipeline performance was evaluated using Accuracy, Macro F1-scores, and Cohen's kappa compared to expert-curated reference annotations. Results: The top-performing Stage 1 gatekeeper isolated the target patient cohort with zero classification errors, effectively reaching complete agreement with the reference standard (ground truth) and correcting inaccurate administrative metadata. For Stage 2 feature extraction, current-generation models demonstrated robust structural reliability, reducing the burden of manual post-processing. The 31B model configuration (Gemma 4) virtually eliminated objective factual errors, achieving near-perfect agreement with the reference standard (Macro F1 > 0.99). Conclusion: Local two-stage pipelines offer a highly accurate, resource-efficient, and secure solution for clinical text processing. This framework serves as a privacy-preserving engine for institutional quality control and the scalable curation of pathology registries.

Background: Many radiology decisions, such as whether a report describes a finding or how urgently a study should be read, are choices among a fixed set of answers. Decision models return a probability for each permitted answer in one step, without generating text, and an open-weight model can run within a hospital so that patient data stay local. No decision model has yet been specialized for radiology and compared with large language models (LLMs). Methods: We fine-tuned the general-purpose decision models Kev-27B (Qwen3.8-27B backbone) and Kev-9B on 36,109 records from public radiology datasets to obtain RadKev-27B and RadKev-9B, and compared them with the Kev models, the hosted decision models Jev and OpenAI Decisions, and the LLM Qwen3.8-27B, on 14,142 held-out questions in fifteen tasks (report reading, case diagnosis and classification, radiological knowledge). The prespecified primary outcome was task-averaged accuracy relative to Kev-27B. Results: RadKev-27B exceeded Kev-27B by 4.9 percentage points (95% CI 3.8 to 6.2), with the largest gain in report error detection (22.6 points); RadKev-9B exceeded Kev-9B by 6.9 points (5.5 to 8.4) and did not differ detectably from Kev-27B, a model three times its size. RadKev-27B was more accurate than the hosted decision models Jev (3.0 points, 1.7 to 4.4) and OpenAI Decisions (4.8, 3.2 to 6.4), than Qwen3.8-27B without reasoning (4.7, 3.4 to 6.0) and with reasoning (2.8 points, 1.0 to 4.6) on 1,675 questions, at 132-fold lower latency (128 ms vs 16.9 s per question). At an error rate of 5%, RadKev-27B could answer 90.5% of questions and Kev-27B 76.4%. Conclusions: RadKev-27B was more accurate than two hosted decision models and an LLM of the same size, with or without reasoning, at a fraction of the latency of reasoning, and can be deployed within an institution for categorical radiology decisions after validation on local data.

Background: General-purpose large language models (LLMs) are increasingly evaluated in diagnostic pathology, but prior studies have largely emphasized diagnostic accuracy rather than how models fail. We evaluated four LLMs for diagnostic performance, pathology-relevant errors and hallucinations, their burden, and potential clinical impact across multi-organ pathology cases. Design: In this retrospective multicenter study, 153 pathology cases from two institutions spanning 20 organs were evaluated using ChatGPT-5.3 (LLM1), Gemini 3 (LLM2), Grok 4.20 (LLM3), and Claude Opus 4.6 (LLM4). Each LLM received multi-magnification histologic images with clinical context and generated a microscopic description and diagnosis. No data splitting, training, or fine-tuning was performed. Twenty-one pathologists assessed 612 outputs for diagnostic correctness, error and hallucination type and burden, clinical impact (0-4), and overall performance (1-5). Results: Strict diagnostic accuracy was 48.9% overall (60.1% including partially correct diagnoses) and ranged from 36.6% to 58.2% across LLMs. Errors occurred in 82.0% of outputs and hallucinations in 76.6%; 90.2% contained at least one error or hallucination. Misinterpretation was the most frequent error (72.1%), while fabricated histologic features were the dominant hallucination type (74.5%). LLM2 had significantly lower misinterpretation rates than the other three LLMs, while LLM3 had significantly higher fabricated-feature hallucination rates than LLM1 and LLM2. Errors were more frequent in incorrect than correct diagnoses (97.5% vs 66.2%), as were hallucinations (95.1% vs 59.9%; both p < 0.001). Every strictly incorrect diagnosis contained an error and/or hallucination, while 79.9% of strictly correct diagnoses also contained at least one. In multivariable analysis, LLM3 was independently associated with higher odds of an incorrect diagnosis. Diagnostic correctness was the strongest determinant of high clinical impact; each one-point increase in hallucination burden increased the odds by 63% (adjusted OR, 1.63). Conclusion: Diagnostic accuracy alone substantially underestimates the safety limitations of general-purpose LLMs in histopathology. Errors and hallucinations were common even when the final diagnosis was correct, and their burden was independently associated with clinical impact. Evaluation frameworks should therefore assess both diagnostic correctness and the reliability and potential consequences of accompanying generated content.

Background: Deep-learning systems for Gleason grading are developed almost entirely on high-end clinical scanners and on cohorts from a small number of Western institutions, yet deployment increasingly involves other devices and other populations. These two distribution shifts, device and population, are rarely tested together on the same physical slides. The PAR dataset, from Erbil, Iraq, digitizes each biopsy on three scanners and provides three distinct pathologist grades, so it permits both tests at once on a Middle Eastern cohort. A concurrent study by the dataset originators validated a task-specific model and two foundation models on PAR; we complement it by testing an inde-pendently developed detect-then-grade pipeline and by separating scanner effects on detection from scanner effects on grading. Methods: We applied one fixed model de-veloped on North American and European material to all 1017 whole-slide images (339 slides from 185 patients, three scanners; 49.6% clinically significant cancer) with no scanner-specific or population-specific tuning. We measured cancer detection (area under the ROC curve of the predicted cancer-tissue fraction), all-slide ISUP agreement of the deployed detect-then-grade pipeline (quadratic-weighted kappa, QWK), and grading agreement on pathologist-confirmed cancers, at the slide level and, because a case carries up to two slides, at the patient level. The reference reader was S.A.; thresholds and operating points were cross-validated leave-one-out; scanners were compared by paired within-biopsy bootstrap and confidence intervals confirmed by patient-cluster bootstrap. Results: Detection was statistically equivalent across scanners (AUC 0.987 to 0.991; paired differences at most 0.003) and transferred to this non-Western cohort with no per-population tuning. At a 95% sensitivity operating point the deployed pipeline reached cross-validated all-slide QWK of 0.86, 0.81, and 0.86 (Grundium, Hamamatsu, Leica), matching the inter-pathologist ceiling of 0.81, against 0.23 to 0.62 for the ungated model. Grading of confirmed cancers was scanner dependent: the compact Grundium (0.63) did not differ from the clinical Leica (0.67; paired difference 0.04, 95% CI -0.03 to 0.11), while both exceeded Hamamatsu (0.44). Results held at the patient level, with grading somewhat lower for two scanners; the two slides of a case disagreed in grade in 43% of cases, and patient clustering did not widen the intervals. Conclusions: Can-cer-tissue fraction is a triage signal robust across scanner and transferable to an un-derrepresented population for detection, while grading is the scanner-sensitive step. Prostate grading models should be deployed as a detect-then-grade pipeline, with grading validated per device and confirmed on the local population.

ICU trauma patients are clinically heterogeneous, and early mortality risk stratification may support monitoring and resource allocation. We developed machine learning models for 30-day mortality prediction using information recorded during the first 24 hours after ICU admission. In MIMIC-III, six feature configurations were trained using 3,411 patients and compared in a patient-level configuration-selection hold-out subset of 853 patients. The selected XGBoost configuration yielded an area under the precision-recall curve (AUPRC) of 0.556 and an area under the receiver operating characteristic curve (AUROC) of 0.863. For cross-database evaluation, a 228-predictor harmonized XGBoost model was refitted on the complete MIMIC-III cohort and evaluated in 13,747 MIMIC-IV ICU stays without using MIMIC-IV outcomes for model development or recalibration. It achieved an AUPRC of 0.495, an AUROC of 0.825, and a Brier score of 0.109. Calibration was monotonic but showed increasing overprediction at higher predicted risks. First-24-hour clinical information retained predictive value across MIMIC database versions, although internal configuration selection, model differences, same-center provenance, and incomplete feature-mapping documentation limit generalizability and deployment readiness.

On the public Regensburg Pediatric Appendicitis Dataset, several recent studies report near-ceiling diagnostic performance using radiologist-interpreted ultrasound findings as tabular model inputs, without processing the underlying images. This raises a question worth answering carefully: how much of that performance reflects genuine diagnostic information, and how much reflects the way this dataset's outcome labels were constructed? We defined four modeling arms, an image-only arm (a backbone-matched BiomedCLIP pipeline as the primary comparator, with an originally-planned ResNet18 model retained as a secondary baseline), a clinical/laboratory-only model (XGBoost), a radiologist-findings model (Random Forest, treated as a reference built from expert-recorded features rather than a live expert comparator), and a multimodal model fusing frozen BiomedCLIP image embeddings with clinical/laboratory data, and compared them using paired DeLong tests, bootstrap confidence intervals, TOST equivalence testing, and pooled repeated cross-validation. The multimodal model (pooled AUC 0.894) showed significantly higher discrimination than the backbone-matched image-only (0.789, p less than 0.0001) and clinical-only (0.855, p less than 0.0001) models. Against the radiologist-findings model (0.934), it initially fell short (p = 0.034). We then found that the dataset's own rule for labeling non-surgically-managed patients uses appendix diameter, one of the radiologist-findings predictors, and, separately, the Alvarado and Pediatric Appendicitis Scores, both retained as clinical/multimodal predictors. Removing appendix diameter dropped the radiologist-findings AUC to 0.790; removing the two scores from the multimodal model changed its AUC only marginally, from 0.894 to 0.890, a difference not statistically distinguishable from no change (p = 0.25). Comparing these two corrected models directly, the multimodal model remained significantly higher (0.890 vs. 0.790, p = 6.2 times 10 to the power of minus 6). A supplementary check found that 137 of 578 patients (24 percent) have ultrasound images containing burned-in caliper measurements; on a matched comparison holding the patient set fixed, excluding these images significantly reduced both the image-only and multimodal models' AUC, indicating a second, image-level information pathway that our design cannot fully attribute to on-screen numeric values versus other differences between marked and unmarked views. An exploratory check that combined both corrections, restricting the doubly-corrected multimodal model to annotation-free images on the smaller, matched patient subset, found a smaller, non-significant difference against the diameter-removed radiologist-findings model (p = 0.176), so our main doubly-corrected comparison should not be read as controlling for this second pathway. The multimodal model showed higher discrimination than the unimodal automated baselines on this benchmark, significantly so in pooled repeated cross-validation; its apparent shortfall against radiologist-derived features is highly sensitive to the dataset's outcome-definition procedure, although the exact fraction of that shortfall attributable to label-construction dependence, as opposed to genuine diagnostic information, cannot be determined from these data alone.

The prevalence of multiple long-term conditions (MLTCs) is associated with increased healthcare utilisation and an elevated risk of unplanned 30-day hospital readmission. Existing prediction tools predominantly focus on single-disease cohorts and fail to capture the clinical heterogeneity, polypharmacy, and care complexity characteristic of MLTC populations. Using data from 99,207 UK Biobank (UKBB) participants with MLTCs ([&ge;]2 long-term conditions), we developed Self-HR, a two-stage self-supervised learning framework that learns transferable patient representations from longitudinal clinical data encompassing hospital admission diagnoses, primary care prescriptions, long-term condition histories, and demographic factors. Self-HR achieved an AUROC of 0.92 and AUPRC of 0.75 in the UKBB discovery cohort, outperforming all supervised baselines --- including XGBoost, Random Forest, and fully supervised neural networks --- across both overall and minority-class metrics. Performance was sustained upon external validation in 79,224 multimorbid individuals from the Clinical Practice Research Datalink (CPRD; AUROC 0.86, F1 score 0.67 for the readmitted class). Self-HR demonstrated superior robustness to partial outcome labelling and class imbalance, maintaining an F1 score of 0.62 for readmitted patients when trained on only 50\% labelled data, compared with 0.28 for the best supervised comparator. Ablation analyses identified incident admission diagnoses as the strongest predictive feature, followed by primary care prescriptions and long-term condition history. Beyond binary classification, Self-HR generalised to regression tasks --- predicting incident and emergency admission durations --- through fine-tuning alone, achieving the lowest MAE and RMSE across all tasks without repeat pretraining. These findings support Self-HR as a data-efficient and generalisable framework for readmission risk prediction in multimorbid populations, with potential to inform proactive discharge planning and targeted post-discharge care.

Background International Classification of Diseases (ICD) codes are widely used for reimbursement, health-service planning, and disease surveillance. Automated systems can identify diagnoses and supporting evidence across clinical documents. Yet coding is more than finding diagnoses. Decisions based on isolated information may not fully reflect the evolving clinical picture during admission. Provisional, superseded, or conflicting diagnoses can therefore persist in the final code set, complicating review and correction. We therefore developed and evaluated Clinico, a coding agent that organises evidence by time and clinical relationships to adjudicate diagnoses and update their status during admission. Methods We did a retrospective multicentre study with internal and external validation using inpatient records from two Shanghai hospitals. Routine coder-finalised codes served as references. Using 5,000 development stays, we refined Clinico's prompts and workflow without fine-tuning model parameters. We compared Clinico with three prespecified comparators, two trained on 80,340 stays. Clinico maintained an evidence-linked diagnostic ledger across each admission and reconciled related or conflicting claims before finalising the code set and selecting the principal diagnosis. The primary outcome was principal-diagnosis exact-match agreement; complete code-set micro-F1 was the key secondary outcome. Findings Evaluation included 5,000 internal and 4,109 external test stays. Principal-diagnosis exact-match agreement was 83.6% (95% CI 82.6-84.6) internally and 52.0% (50.5-53.5) externally; the next-highest observed estimates were 74.7% and 31.5%. Complete code-set micro-F1 was 90.6% (95% CI 90.2-91.0) internally and 42.6% (42.0-43.2) externally; the next-highest observed estimates were 90.3% and 25.5%. Interpretation In both hospital test sets, Clinico had the highest observed estimates of principal-diagnosis agreement and complete code-set micro-F1 among evaluated methods. Its longitudinal adjudication process links decisions to source evidence, allowing them to be checked against the record. Prospective studies should assess its effects in clinical workflows using independently adjudicated references. Funding None.

Background. Chest radiographs (CXRs) are the most common imaging exam performed, but CXR reports can be nonspecific due to incomplete history, relying on vague terms like opacities to convey diagnostic uncertainty. We prospectively evaluated whether introducing a clinical dashboard with relevant patient information affects CXR report specificity and efficiency, report concordance with discharge diagnoses, and follow-up chest CT utilization. Methods. In this stepped-wedge cluster randomized trial (NCT06785246), a clinical dashboard for radiologists was rolled out with user education sessions across 15 service areas of an integrated health system from June 2024 to June 2025. Emergency department (ED) and inpatient CXRs were included. A large language model (GPT-4o) was validated using a Delphi panel of three experts to classify report specificity and CXR report diagnosis. A subsequent model release (GPT-5.4) graded each CXR report impression as negative, specific abnormal, or nonspecific abnormal. It also judged whether an abnormal impression matched an ED discharge diagnosis. The primary outcome was the proportion of specific reports (negative or specific abnormal). Secondary outcomes were report-discharge diagnosis concordance, follow-up chest CT, antibiotic and diuretic orders, and CXR reporting time. Mixed-effects models adjusted for time, patient age, and radiograph view, with random effects for radiologists, facilities, and service area. Results. The trial included 404,860 ED and inpatient encounter STAT priority CXRs read by 387 radiologists (256,303 control; 148,557 intervention). The adjusted proportion of specific reports was 82.2% in control periods and 82.1% in intervention periods (difference, -0.10 percentage points; 95% CI, -0.53 to 0.34; P=.66). Among abnormal reports, concordance with a discharge diagnosis rose from 40.9% to 42.2% (difference, 1.24 points; 95% CI, 0.20 to 2.28; P=.02). Follow-up chest CT fell from 11.4% to 10.9% (difference, -0.55 points; 95% CI, -0.92 to -0.17; P=.004). Antibiotic orders (24.0% vs 23.9%) and diuretic orders (14.7% vs 14.8%) within 7 days did not change, nor did reporting time (difference, -0.8 seconds; 95% CI, -2.7 to 1.0). Conclusions. Giving radiologists EHR context through a dashboard did not change the specificity of STAT priority CXR reports or prescribing patterns. It was associated with small improvements in diagnostic concordance and follow-up CT use, without slowing reporting. Validated large language models make it feasible to measure report quality across a health system.

Background Artificial intelligence computer-aided diagnosis (AI CAD) algorithms for screening mammography have shown promise, but independent head-to-head comparisons of commercial algorithms on large, diverse cohorts remain limited. Methods 786,124 mammography screening exam performed from January 1 2022 to December 31 2023 were identified from electronic medical records. Three FDA-cleared commercial algorithms (A, B, D) and the open-source academic model Mirai (C), were applied to the four standard screening views. Breast cancer within 12 months was ascertained through cancer registry linkage. Algorithms were compared by AUROC; sensitivity, specificity, and PPV at matched AI-positive rates of 5%, 10%, 15%, 20%, 25%; and the proportion of false negatives classified as AI-positive. Results Of 707,922 examinations with complete data, 4,436 were associated with cancer. Radiologists recalled 6.8% of examinations, with a sensitivity of 67.2%. Algorithm D had the highest AUROC (0.849), followed by algorithm B (0.818), Mirai (0.817), and algorithm A (0.815). Algorithm D also had the highest sensitivity at every AI-positive rate, from 55.3% (95% CI: 54.1, 56.6) at 5% to 77.9% (95% CI: 76.9, 79.0) at 25%, with CIs that did not overlap those of the other algorithms. No algorithm matched radiologist sensitivity at AI-positive rates of 10% or less. Algorithm B flagged a proportion of radiologist false negatives similar to that of algorithm D from 5% through 20% (20.8% vs 20.5% at 5%), despite its lower standalone sensitivity. Conclusions AI CAD algorithms differed meaningfully in performance on the same large, diverse screening examination cohort, and an open-source academic model performed comparably to two of the three commercial products. Independent local validation at clinically relevant operating points should precede adoption of mammography AI, and algorithm choice and threshold should be matched to the intended workflow.

Background: Early cancer detection is aided by timely GP referrals. However, GP practices face pressures, particularly in deprived areas, which may be associated with diagnostic inequalities. Aim: To examine whether within-practice changes in access, workforce, and population income deprivation were associated with diagnostic care indicators. Design and setting: Retrospective longitudinal study of 6,194 practices between 2015-2019 and 2022-2023. Method: Panel fixed-effects regression examined USC referral rates, USC PPV, USC sensitivity, and emergency presentation rates. Predictors were patient-reported access, GP workforce, and practice population income deprivation. Results: Per one SD increase within practices, (i) USC referral rates were lower with better appointment-making experience (-1.4%) and older average GP age (-3.5%), and higher with female GP share (+1.3%); GP FTE showed a small positive association (+0.7% to +0.8%). (ii) USC PPV increased with older average GP age (+2.2%). (iii) USC sensitivity was lower with older average GP age (-1.1%); emergency presentations showed limited association with access or workforce factors. Practices had higher emergency presentation rates when serving the most deprived rather than least deprived populations (+13.7%). There was little evidence that deprivation modified access and workforce associations. Average GP age was an exception: USC referral associations weakened from Q2 to Q5, USC PPV peaked at Q3, and USC sensitivity was significant only in the most deprived quintile. Conclusion: Practice-level access and workforce factors were associated mainly with USC referral activity. Persistent deprivation gaps in emergency presentation suggest monitoring USC referrals alongside PPV, sensitivity, and emergency presentations.

[Abstract] Objective: Intraventricular haemorrhage (IVH) carries high mortality and morbidity; accurate early prediction of 6-month functional outcome is essential for guiding treatment decisions and optimising prognosis. We developed and externally validated a prognostic model that integrates a novel neuroimaging marker, the Brainstem Dorsal Line (BSDL), with the TabICLv2 algorithm to predict 6-month functional outcomes after IVH. Methods: In this multicentre retrospective study, patients with IVH were enrolled from nine tertiary centres in China between Dec 1, 2020, and Dec 31, 2024; seven centres constituted the derivation cohort (8:2 training-validation split) , and two centres formed the external validation cohort. Feature selection followed a three-step pipeline comprising univariable testing, Spearman correlation analysis, and LASSO regression; eight machine-learning models were trained with five-fold cross-validated hyperparameter tuning. Performance was assessed by the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPRC), calibration curves, and decision curve analysis (DCA). Interpretability was evaluated using SHapley Additive exPlanations (SHAP). Results: In total, 728 eligible patients were enrolled 610 in the derivation cohort (unfavourable outcome, 27.9%) and 118 in the external validation cohort (26.3%). Feature selection identified nine optimal predictors, of which BSDL grade 2 ranked highest in feature importance. TabICLv2 showed the best performance in external validation (AUC 0.9014 [95% CI 0.825-0.985], AUPRC 0.8192 [0.688-0.916], F1 score 0.7742 [0.651-0.877]); calibration was excellent (Hosmer-Lemeshow test, P>0.05) and DCA showed clinical net benefit across threshold probabilities of 0.05-0.97. SHAP confirmed BSDL grade 2 as the strongest predictor of unfavourable outcome. Conclusions: Integrating BSDL with TabICLv2 yielded a robust, interpretable prognostic model for 6-month outcomes after IVH, providing a practical tool for early risk stratification and individualised treatment planning.

Background: Public and clinical attention to postural orthostatic tachycardia syndrome (POTS) has increased, particularly since the COVID-19 pandemic. We quantified changes in United Kingdom Google search interest and examined whether searches increasingly used diagnostic and self-assessment language. Methods: We extracted monthly Google Trends relative search volume (RSV; 0-100) for the Health-category search term 'Pots syndrome' in the United Kingdom from January 2004 through July 2026. Five extraction attempts were made; two returned complete, identical monthly series and were retained. Prespecified eras were summarised and an exploratory interrupted time-series model at March 2020 used ordinary least squares with Newey-West heteroskedasticity and autocorrelation consistent standard errors (12 lags). Comparator searches included conventional orthostatic diagnoses, POTS diagnostic terms, associated conditions and YouTube searches. Results: The primary series comprised 271 complete months. Mean RSV increased from 18.6 during 2015-2019 to 64.8 during 2022-2023 (3.49-fold) and remained 50.6 during January 2024-July 2026 (2.73-fold above baseline). Search interest peaked in October 2022 (RSV 100); July 2026 RSV was 57. The interrupted time-series model estimated an immediate March 2020 level increase of 21.8 points (95% CI 2.8-40.7; p=0.024), while the slope change was not statistically supported (0.069 points/month, 95% CI 0.299 to 0.438; p=0.713). Searches for 'POTS symptoms', 'POTS test' and 'POTS heart rate' increased more steeply than the general term, although low baseline volumes made fold changes unstable. Conclusions: UK Google search interest in POTS rose before 2020, increased sharply after the pandemic began, and remained substantially above its prepandemic baseline. The results demonstrate a sustained change in public attention, not disease incidence or social-media causation. The growth of symptom- and testing-oriented searches is compatible with increased diagnostic self-investigation and warrants linkage to referral, diagnosis and social-media exposure data.

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.

Predicting disease risk from prior diagnoses is fundamental to clinical decision-making, particularly during health emergencies such as the COVID-19 pandemic, when individuals with long-term conditions may be disproportionately vulnerable to adverse outcomes. Despite intense interest in developing models to predict disease risk from prior diagnoses (1-3), most prediction models do not estimate effects of each prior diagnosis on disease risk conditional on other diagnoses, limiting interpretability and clinical utility. We developed the Comorbidity Risk Score (CRS), trained on 13 million individuals (age 40-69) from linked electronic health record (EHR) datasets of the entire population of England, to predict COVID-19 hospitalisation and 87 other disease outcomes. CRS was trained at close to saturated sample size and precisely estimated the effects of 212 prior diagnoses on the 88 disease outcomes, conditional on all other prior diagnoses. Correlations of CRS effect sizes across outcomes (e.g. 0.76 for myocardial infarction vs. hyperlipidaemia) matched the corresponding genetic correlations (e.g. 0.79 for myocardial infarction vs. hyperlipidaemia), confirming that comorbidity architectures capture disease aetiology. On average, CRS identified 5% of the population with 3.4-fold higher disease risk, including myocardial infarction (4.4-fold), lung cancer (6.5-fold), and COVID-19 hospitalisation (6.3-fold). Using prior diagnoses alone, CRS outperformed state-of-the-art clinical COVID-19 models (4). Furthermore, CRS (N=13 million) substantially outperformed state-of-the-art AI (1) (N=0.5 million) and linear (3) (N=0.5 million) models in predicting disease risk, suggesting that training sample size outweighs model complexity. CRS attained near-perfect transferability across self-reported ethnicities (e.g., Black vs. White: AUROC ratio = 97.3%). Finally, CRS distinguished independently predictive comorbidities from indirect associations, e.g., lipid metabolism disorder was a strong predictor of myocardial infarction risk but not ischaemic stroke, after conditioning on other prior diagnoses. In conclusion, CRS provides a comprehensive resource for understanding the impact of comorbidities on COVID-19 and other future diseases, revealing disease aetiology while enabling powerful prediction of disease risk.

Agentic workflows can coordinate modelling, but balancing predictive performance, measurement burden and reproducibility is unclear. We developed DXA Agent, an agentic workflow for dual-energy X-ray absorptiometry (DXA) outcomes integrating planning, feature-model refinement, tools, provenance and hypothesis-generating interpretation. Models were independently developed and tested in UK Biobank (5,318 participants) and the National Health and Nutrition Examination Survey (NHANES; 3,777 participants), using cost-efficient and no-limit strategies. Across 20 UK Biobank and three NHANES bone mineral density sites, cost-efficient models achieved lower RMSE and higher R2 than the best conventional comparator, with median relative RMSE reductions of 10.9% and 9.9%, respectively. Classification was task dependent: UK Biobank osteoporosis averaged AUROC 0.839 and PR-AUC 0.182, whereas NHANES performance was comparable with conventional models. Higher-burden features did not consistently improve prediction. These retrospective, cohort-internal findings position DXA Agent as an inspectable, measurement-burden-aware research workflow requiring independent prospective validation.