International Journal of Medical Informatics

Objective: We explored the implementation considerations of machine learning (ML) suicide risk identification in a health system that serves primarily American Indian patients in the Southwestern United States. Previous work developed and reported the performance of suicide risk models, but traditional performance metrics do not directly reflect clinical workflows. Materials and Methods: Using EHR data collected for Emergency Department (ED) visits between 10/01/2019 and 10/02/2021, we conducted Decision Curve Analysis (DCA) to identify the standardized net benefit (SNB) of all models previously evaluated and plotted the potential workload implications of the model with the greatest SNB. Finally, we evaluated suicide screening practices during the study period to inform implementation barriers. Results: Among 9,244 patients, 139 (1.5%) had a suicide attempt or death within 90 days of an included visit. The population-specific logistic regression model had the greatest SNB compared to other approaches. When operationalized, the 95th percentile cutoff did not substantially increase clinical workloads, while the 75th percentile cutoff substantially impacted workloads. The secondary analysis on suicide screening depicted variations in screening administration over time. Discussion: While DCA identified advantages of ML suicide risk identification over current standards, the complexities of suicide-related care limited its clinical interpretability. A stepped-care framework may mitigate potential workload burdens. Training initiatives and dynamic implementation strategies will support high-quality implementation of ML suicide risk identification. Conclusion: Our analyses support the feasibility and utility of augmenting existing suicide screening practices with ML suicide risk identification, and demonstrate the importance of nuanced analyses to inform implementation considerations.

Accurate prediction of adverse events in the intensive care unit (ICU) following cardiothoracic surgery is crucial for timely interventions, improving patient outcomes and optimizing value in healthcare. By leveraging advanced machine learning techniques, this study demonstrates the transformative potential of predictive analytics to enhance postoperative care using a dataset from the Society of Thoracic Surgeons' (STS) database and time-series intraoperative data. We developed a multi-modal late fusion approach integrating static patient variables with intraoperative time-series data, utilizing an ensemble of models including residual neural networks for static features, CNN and Bidirectional GRU for time-series processing, XGBoost, logistic regression, SVM and Gradient Boosting. Through five-fold cross-validation, our ensemble model with balanced threshold selection achieved an AUC of 0.87, sensitivity of 0.76, specificity of 0.83, PPV of 0.40, and NPV of 0.96. Preoperative heart failure, IABP insertion, cardiopulmonary bypass (CPB) duration and white blood cell (WBC) count emerged as key predictors. Training the same pipeline on each modality separately showed that the static registry variables alone reached the same AUC of 0.87, and that the intraoperative signals alone reached 0.83, so on this cohort the two modalities act as substitutes rather than complements and fusion changed the sensitivity and precision trade-off rather than the amount of separable signal. A per-signal ablation found that no individual monitoring channel was load bearing, with mean arterial pressure and central venous pressure the least substitutable. This balanced approach optimizes the trade-off between sensitivity and specificity, providing a screening tool with high negative predictive value for monitoring high-risk patients, and the modality comparison indicates that comparable performance is attainable at sites holding registry data alone.

Lines, drains, and airway devices are commonly used in surgical and inpatient care to support treatment and patient monitoring. While these devices are essential, prolonged device retention is associated with complications such as infection, tissue injury, thrombosis, and extended hospital stay. Understanding how long devices remain in place and how duration varies across device types is important for improving patient safety and guiding device management practices. This study analyzes device duration patterns using Patient Lines, Drains, and Airways (LDA) data containing information on device type, placement time, and removal time across multiple patient encounters. Exploratory data analysis and machine learning approaches are used to examine duration distributions, distinguish short and prolonged device retention, and identify structural patterns across device categories. The findings indicate notable variation in device duration patterns across categories, with identifiable differences in device retention behavior across device groups. This study highlights the value of observational electronic health record data for characterizing device exposure patterns and provides a foundation for future data-driven device management and risk assessment.

Background: The Greater Manchester Care Record (GMCR) integrates health and social care data for 2.8 million individuals. While it contains maternity information, identifying individual pregnancies, including timing and outcomes, is complex due to variations in healthcare interactions and recording practices. This study develops open-source code to create a pregnancy register within the GMCR using diagnosis- and procedure-based algorithm established in other databases. Methods: 1,448 maternity SNOMED clinical codes identified pregnancy-related primary care records among women aged 14-49 years between March 2012 and August 2023. Codes were grouped by pregnancy outcome, with a hierarchical ruling approach where multiple outcomes appeared within one episode. Pregnancy start dates were derived from hierarchical coded events or imputed using biologically plausible duration based on outcome type. Internal validation compared GMCR register episodes against Hospital Episode Statistics (HES) hospital maternity admissions; external validation compared outcome rates against published UK estimates. Results: The algorithm was developed across 379,832 women, identifying 880,926 pregnancies. Internal validation showed nearly half of register deliveries had a corresponding HES record, with a positive predictive value (PPV) of 48.1%and a median date difference of 17 days (IQR 1-30). External comparison revealed live birth (927.8 vs. 995.6 per 1,000 deliveries) and stillbirth (2.12 vs. 3.9 per 1,000 births) rates close to national estimates but a notable discrepancy for miscarriage (8.00 vs. 15.3 per 100 pregnancies) and termination of pregnancy (5.36 vs. 25.1 per 1,000 women). Demographic alignment was strong for maternal age (31 vs. 29.8 years) and deprivation (IMD10 median: 3 vs. 2.6), though ethnic minority representation was expectedly higher in the GMCR (37.0% vs. 28.7%). Conclusions: This pregnancy register enriches the GMCR for maternal health research, enabling more comprehensive and regionally tailored studies to inform healthcare delivery and policy. This work represents an important first step in developing a scalable, open-source algorithmic tool to support pregnancy research across diverse populations. By making the code openly available, it can be adapted for use in federated data environments, enhancing reproducibility and collaboration.

Background: Artificial intelligence (AI) is increasingly used in healthcare, but data on its real-world use and perceptions among obstetricians and gynecologists remain limited. Objectives: To assess familiarity, use, perceptions, and expectations regarding AI among physicians practicing obstetrics and gynecology in France. Study Design: We conducted a cross-sectional survey between June 2 and July 30, 2025, mainly distributed through professional organizations or associations and social media. Participation was voluntary and self-selected. AI use was defined as any use of an AI tool in routine clinical practice, regardless of type or frequency. Comparisons were made between AI users and non-users and between junior and senior physicians. Results: Among 469 respondents, representing approximately 5% of the target population, 199 of 466 (42.7%) reported AI use in practice; three respondents were excluded from this analysis because of discordant questionnaire responses. AI use was associated with male gender and professional status, with residents more represented among users. Junior physicians reported greater AI use than senior physicians and expressed greater concern regarding medical confidentiality. Current applications primarily involved administrative tasks, medical writing, literature retrieval, and translation; anticipated future uses centered on clinical decision support and medical data interpretation. Perceptions of several AI applications differed between groups. Conclusions: AI was reported to be used by a substantial proportion of respondents, primarily for supportive and informational tasks. These findings support further evaluation of training, validation, governance, and safe implementation of AI in obstetrics and gynecology.

Background: Passively collected smartphone and wearable data, combined with genomic information, may enable scalable early identification of youth who are at risk for substance use disorder. However, heterogeneous data collected in naturalistic settings limit researcher control over data quality, requiring substantial preprocessing. Data cleaning and feature engineering decisions can meaningfully influence analytic outcomes yet are often minimally described or omitted in published work. Objectives: We provide rationales and recommendations for common preprocessing challenges in passive sensing research, including establishing inclusion criteria when "outliers" may reflect true behavioral variability, analyzing differences across smartphone operating systems, handling observations collected outside protocol windows, and engineering features for machine learning applications. Methods: We illustrate principled approaches to these challenges using smartphone and Fitbit data from the Adolescent Brain Cognitive Development Study (ABCD). The sample was drawn from the available half of the year-4 follow-up (data release 5.1; N = 4,754; ages 13-14). Results: Extreme outlying observations were concentrated among low-engagement participants with few observed days. Smartphone usage among Apple users appeared under-recorded relative to Android users when compared with self-report measures. Fitbit-derived behavior differed systematically between observations collected within versus outside protocol windows. We also identified potential errors in the study's default preprocessing pipeline. Conclusion: We recommend prioritizing person-level inclusion criteria that preserve meaningful behavioral variability, examining temporal distributions rather than assuming observations are exchangeable, and thoroughly investigating and reporting data quality issues. Although illustrated using the ABCD Study, these principles broadly apply to research using smartphones, wearables, and other digital data sources.

Background: Digital systems are central to regulating access to care in Brazil's Unified Health System (SUS), but fragmented data can hinder coordination. Objective: This study examines the database structure of SUSfacilMG and identifies modernization priorities, using CORE Saude MG as a comparative reference. Methods: We assessed metadata from 1,586 SUSfacilMG tables, focusing on naming patterns, relationships, and repeated attributes. We then compared architectural features described for CORE Saude MG. Results: Approximately 40% of the assessed tables repeated descriptive attributes, including patient identifiers, suggesting substantial duplication. Metadata alone cannot establish operational failures, security incidents, or effects on care. Conclusion: A phased review of data models, interfaces, and governance is warranted. Standardized exchange, referential integrity, and controlled access should be evaluated alongside service continuity and data quality.

Social media language and digital lifestyle behaviours are promoted as scalable signals of population mental health, but predictive models are often evaluated in ways that overstate accuracy. We analysed 20,000 tweets from 72 users, labelled by whether the account belonged to a user identified as depressed, and two synthetic datasets (100,000 and 30,000 records) linking screen time, sleep and social media use with self-reported stress. Under a conventional tweet-level split, a term-weighted logistic regression appeared accurate (area under the receiver operating characteristic curve 0.835), but when whole users were held out its performance fell to near chance (0.595; 95% confidence interval 0.505 to 0.686), because the model largely recognised users and their topics. Depression-associated tweets were more often negative (31.7% versus 24.2%), yet sentiment alone could not identify unseen users (0.539). The two behavioural datasets gave opposite answers: strong discrimination in one (0.926), driven by short sleep, and chance in the other (0.506), so synthetic results reflect how the data were generated. At a population prevalence of 10%, positive predictive values would be 0.10 to 0.30. Digital distress signals should be validated across individuals, on real linked data and at realistic prevalence before informing public health surveillance or prevention.

Background: Clinical decision support systems (CDSS) are increasingly embedded in national telemedicine services, however how different stakeholders who use them experience these tools, particularly in low- and middle-income countries, is less explored. In the assisted model of India's eSanjeevani service, spoke health workers enter structured symptoms and hub clinicians review a symptom summary and an artificial intelligence (AI)-generated differential diagnosis. Objective: To describe role-specific experiences of eSanjeevani's CDSS functions and to estimate how acceptance coexists with specific reported difficulties within each role. Methods: We analyzed a program feedback web survey completed between 11 July and 29 July, 2025 by hub clinicians and spoke providers. Prespecified item profiles covered review of summaries and suggested diagnoses, the perceived relation of suggestions to the clinician's own diagnosis, ease of use and intention to use (hub), and symptom-capture utility, question quantity, question relevance, summary presentation and satisfaction (spoke). Within-role contrasts were reported as absolute differences in acceptance with Newcombe 95% CIs. Short open comments underwent content analysis. Results: The cohort comprised 190 hub and 662 spoke respondents from 13 and 15 states (21.4% and 19.3% of eligible providers). Among hub respondents, 76.8% regularly reviewed the symptom summary and 58.9% the suggested diagnoses; 72.6% judged suggestions matched (32.6%) or appropriate but different (40.0%), 7.9% mostly inappropriate, and 11.1% reported that suggestions were mostly not generated. Overall, 85.3% intended to keep using the tools. Affirmative intention was 98.4% when suggestions matched, 66.7% when they were mostly inappropriate (difference vs matched or appropriate -23.2 percentage points, 95% CI -48.5 to -4.0) and 61.9% when mostly absent (-25.7, 95% CI -47.1 to -7.3); technical issues, reported by 38.4%, made little difference (-2.8, 95% CI -14.1 to 7.3). Among spoke respondents, 63.4% rated question quantity appropriate, with miscalibration in both directions (15.6% too many; 16.6% too few); 89.9% found the questions relevant and 91.8% found the summary clear. Satisfaction was 75.3% and did not differ materially between those reporting too many questions and those reporting an appropriate number (-2.7, 95% CI -12.8 to 6.0). Comments located spoke difficulties in question burden and missing complaints, and hub difficulties in freezing and in the scope and availability of suggestions. Estimates were stable across cohort, ambiguity and state-exclusion sensitivity analyses. Conclusions: Acceptance of eSanjeevani's CDSS was high in both roles. For hub clinicians it tracked the perceived quality and availability of AI suggestions rather than technical friction; spoke workers accepted the question burden, but its two-directional miscalibration points toward complaint-specific, adaptive questioning. These provider reports do not measure diagnostic accuracy. They identify log-based and audit-based indicators that the next evaluation cycle should measure directly.

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.

Objective To evaluate the information quality, reliability, readability, and content deficits of ChatGPT, DeepSeek, and Doubao in answering frequently asked questions about bee stings. Methods Twenty-five high-frequency patient questions were selected based on Google Trends, domestic health platforms, and allergy guidelines, and submitted to the three LLMs, yielding 75 responses. Two emergency physicians independently assessed quality using GQS, mDISCERN, QAMAI, EQIP, and PEMAT-U/A. Readability was evaluated using six indices (ARI, FRES, GFI, FKGL, CLI, SMOG). Content deficits were classified qualitatively. Twenty emergency department patients evaluated the comprehensibility and credibility of AI responses. Results Significant differences were found among the three models on all quality metrics (all P < 0.01). DeepSeek achieved the highest mDISCERN score (median 21.00); ChatGPT led on QAMAI (25.00) and EQIP (75.00%); Doubao ranked lowest on all quality indicators but highest on PEMAT-U (73.00%). No model met the sixth-grade reading level standard. Eleven content deficits were identified: outdated medical advice (45.45%), omissions (36.36%), and hallucinations (18.18%). Doubao received the highest patient preference (45.0%). Conclusion The three LLMs exhibited distinct strengths: DeepSeek in source reliability, ChatGPT in overall quality, and Doubao in understandability. All models had excessively high readability levels, and outdated medical advice was the predominant content deficit. Objective To evaluate the information quality, reliability, readability, and content deficits of ChatGPT, DeepSeek, and Doubao in answering frequently asked questions about bee stings. Methods Twenty-five high-frequency patient questions were selected based on Google Trends, domestic health platforms, and allergy guidelines, and submitted to the three LLMs, yielding 75 responses. Two emergency physicians independently assessed quality using GQS, mDISCERN, QAMAI, EQIP, and PEMAT-U/A. Readability was evaluated using six indices (ARI, FRES, GFI, FKGL, CLI, SMOG). Content deficits were classified qualitatively. Twenty emergency department patients evaluated the comprehensibility and credibility of AI responses. Results Significant differences were found among the three models on all quality metrics (all P < 0.01). DeepSeek achieved the highest mDISCERN score (median 21.00); ChatGPT led on QAMAI (25.00) and EQIP (75.00%); Doubao ranked lowest on all quality indicators but highest on PEMAT-U (73.00%). No model met the sixth-grade reading level standard. Eleven content deficits were identified: outdated medical advice (45.45%), omissions (36.36%), and hallucinations (18.18%). Doubao received the highest patient preference (45.0%). Conclusion The three LLMs exhibited distinct strengths: DeepSeek in source reliability, ChatGPT in overall quality, and Doubao in understandability. All models had excessively high readability levels, and outdated medical advice was the predominant content deficit.

Patient narratives can reveal aspects of health that standardized measures miss, yet collecting open-ended responses at population scale remains difficult. We evaluated whether speech-to-text (STT) could enable richer responses in online health surveys without the resource demands of interviews. Across four surveys involving people with post-COVID-19 condition, healthy adults, people who engage in sex work, and older adults, participants could answer open-ended questions by typing or using STT. We compared response length, lexical diversity, stopword proportion, and total unique content words, alongside STT uptake and participant feedback. Quantitative comparisons were made using a mixed-effects model including modality (STT vs typed) as fixed effect and a random intercept per participant. STT uptake varied markedly across populations (0.7-80.0%). Spoken responses were estimated to contain more characters than typed responses in two studies, by 245.0 and 891.8 characters on average (both p<0.001), and contained more unique content words, despite lower lexical diversity and higher stopword proportions. Participants generally valued STT for its convenience and spontaneity, but uptake was constrained by context, privacy, interface design and individual preferences. STT may enable population-scale collection of richer health narratives, provided implementation is tailored to the study population and research context.

BackgroundPatient education is critical to safe and effective postoperative bariatric surgery care. Large language models (LLMs) may improve access to personalized patient education but evidence regarding direct patient use in clinical settings remains limited. We evaluated BariBot, an LLM-powered educational application for patients undergoing bariatric surgery. MethodsWe conducted a prospective, single-arm mixed-methods feasibility study at a tertiary academic medical center from May 2024 to May 2025. Adult patients undergoing bariatric surgery used BariBot independently on postoperative day 1. The application provided conversational postoperative education and adapted language complexity to participants educational attainment. Outcomes included feasibility, usability, acceptability, perceived trustworthiness, and anticipated future use. Usability was assessed using the System Usability Scale (SUS). Immediately after BariBot use, participants completed a semi-structured qualitative interview assessing acceptability, perceived trustworthiness, and anticipated future use. Interview transcripts underwent thematic analysis. ResultsSixteen participants completed the BariBot use session, SUS assessment, and post-intervention interview. No technical failures prevented application use, and no participant requested early discontinuation. The median SUS score was 91.3 (IQR, 81.3-98.1), consistent with excellent perceived usability. Qualitative analysis showed that participants valued the applications immediacy, structured responses, conversational context, ability to address follow-up questions, judgement-free nature, and future availability between clinical encounters. Trust was strengthened by alignment with prior clinical guidance and perceived institutional association, although participants emphasized the need for transparency, clinician oversight, and use as an adjunct rather than a replacement of clinician guidance. ConclusionsBariBot was feasible to implement and was associated with high perceived usability and acceptability postoperatively after bariatric surgery. Larger, longitudinal, and comparative studies are needed to evaluate safety during unsupervised use and determine whether LLM-based educational applications improve patient understanding, engagement, and clinical outcomes.

Objective: To estimate Long COVID undercoding in electronic health records (EHRs) against federal survey prevalence, and to evaluate a predictor of later coding that anyone can rebuild from counts a research enclave releases. Materials and Methods: Among 23,790 All of Us adults with COVID-19, we compared coded shares with Household Pulse Survey prevalence. Among 23,300 adults uncoded at day 90, we predicted a first U09.9 code by day 455. We fitted naive Bayes and linear discriminant models outside the enclave from released counts and compared them with an elastic net fitted inside to the same patients and features. Tests were a random 20% of patients (4,680, 50 coded) and 18 held-out EHR sites (5,760, 60 coded). Results: Under survey-comparability assumptions, about 1 in 9 adults who ever had Long COVID carried its code, and the women-to-men coding ratio was 0.79 (95% sensitivity interval, 0.65-0.95). The released models' areas under the receiver operating characteristic curve (AUROCs) were 0.68 (95% CI, 0.60-0.75) on the patient test and 0.64 (0.55-0.72) at held-out sites, 0.049 (0.001-0.104) above the elastic net. On the patient test, the elastic net had a higher area under the precision-recall curve (difference 0.062, 0.003-0.122). Released probabilities overestimated coding risk at held-out sites. Discussion: Fitting from released counts did not consistently lower discrimination. Coded outcomes inform disease discrimination only under stated coding assumptions. Conclusion: Long COVID is substantially and unevenly undercoded, and a model rebuilt from public counts predicts its later coding with moderate discrimination.

Objective: Retrospective evaluation of clinical pathway adherence using electronic medical record (eMR) data is limited by heterogeneous data structures and statistical measures that do not account for patient-specific context, care sequence, or timing. We designed a modular event-based architecture to dynamically evaluate the New South Wales Adult Sepsis Pathway over time. Method: Data from eMRs were collected from adult emergency department encounters across four hospitals between July 2022 and June 2024 and processed through three sequential layers: event-stream transformation, pathway-relevance filtering, and state-based pathway evaluation. Heterogeneous records were standardized and organized into a chronological timeline for each encounter. The framework then updated the pathway state as new clinical information became available and assessed whether recommended care was delivered within the relevant timeframe and sequence. Result: The architecture processed 451,124 encounters. This included 57,631,084 clinical events extracted and transformed into encounter-level event streams. After filtering for pathway relevance, 43,951,146 events remained (76.3%). A total of 129,907 (34.6%) encounters entered the sepsis pathway, 50,118 (38.6%) entered YELLOW state, indicating progression to suspected sepsis, and 3,349 (2.6%) encounters progressed to RED state from GREEN, indicating the initial shock. Of 50,118 encounters entering YELLOW state, 2,426 (4.8%) progressed to shock, 8,700 (17.4%) encounters were recognized with sepsis through a blood culture order, and 4,287 (8.6%) encounters were recognized with sepsis through an antibiotic order. This demonstrates the ability of the architecture to prioritize higher-acuity states during evaluation rather than going through an intermediate state. Conclusion: The architecture developed shows the feasibility of transforming fragmented routine eMR data into a reusable longitudinal representation that supports context-dependent pathway evaluation. Separating data preparation from pathway rules improves transparency, maintainability, and adaptability and provides a foundation for future interoperable, near-real-time pathway monitoring

Importance: Conditions that are hard to diagnose, several more common in women, may be missing from electronic health records (EHRs), hiding patients from research cohorts and case finding. Objective: To measure how often the EHR contains a diagnosis code for chronic conditions that participants report, and to compare this with national survey benchmarks. Design, Setting, and Participants: This cross-sectional study used the All of Us Research Program Registered Tier Dataset v9. Participants were adults with EHR visits in both 12-month periods from July 2021 to June 2023 who completed a health history survey. We analyzed the data in October 2026. Exposures: Self-reported clinician diagnosis of 10 conditions with ongoing visits or treatment, as well as sex at birth, race and ethnicity, age, and years of EHR records. Main Outcomes and Measures: Share of participants reporting current care who had a qualifying diagnosis code, alone and relative to hypertension and type 2 diabetes, and the ratio of age-standardized coded shares to national survey prevalence. Results: Of 161,480 respondents, 114,610 (73,870 female [64.5%], median [IQR] age 63 [50-72] years) had at least 5 years of EHR records before the survey. Among those reporting current care, 84.6% (95% CI, 84.2%-85.0%) with hypertension and 87.0% (95% CI, 86.5%-87.6%) with type 2 diabetes had a code, compared with 41.3% (95% CI, 38.9%-43.8%) for endometriosis, 39.4% (95% CI, 34.3%-44.8%) for Lyme disease, and 24.4% (95% CI, 23.2%-25.6%) for chronic fatigue. Counting ICD-9-CM and other mapped codes, 55.3%, 51.5%, and 28.3%, respectively, had any qualifying record. Women had lower coded shares than men for hypertension, type 2 diabetes, obesity, depression, and anxiety (crude ratios 0.93-0.96, within-health-system ratios 0.98-1.02) and higher shares for migraine and fibromyalgia (within health systems, 1.11 [95% CI, 1.07-1.15] and 1.41 [95% CI, 1.29-1.57]). Age-standardized coded shares among active patients were 1.1 to 2.4 times national prevalence of the same self-reported diagnoses. Conclusions and Relevance: In this cross-sectional study, fewer than half of participants reporting ongoing care for chronic fatigue, Lyme disease, or endometriosis had a qualifying ICD-10-CM code. Linking each participant's report to their own record measured this gap directly.

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

Background: Unstructured free-text narratives in electronic health records (EHRs) contain critical clinical information that is not captured by structured standard codes. In Catalonia, primary care notes follow a semi-structured format known as MEAP (in Catalan). However, extracting structured variables using cloud-hosted commercial large language models (LLMs) raises substantial data privacy concerns and incurs high operational costs. Objective: To develop and validate a privacy-preserving local hybrid pipeline combining a lightweight small language model (SLM) with deterministic regular expressions for extracting six uncoded urinary tract infection (UTI) clinical features (fever, nitrites, leukocytes, lumbar pain, abdominal pain, and haematuria) from primary care EHR narratives. Methods: We deployed the open-weight model (Phi4-mini, 3.8B parameters) within the institutional firewall using Ollama. Prompts were iteratively refined in collaboration with clinicians, and regular expressions were applied post-inference. Validation was conducted across two independent arms: (1) a double-blind clinician gold standard review of 60 real patient records (consensus-resolved cases), and (2) an adversarial synthetic dataset of 720 notes enriched with linguistic noise, generated using GPT-4.1 and Grok-4.1. Point estimates and 95% confidence intervals (CIs) were computed for key performance metrics. Results: The pipeline processed 15,498 MEAP narratives from a matched cohort of 2,962 primary care patients and identified 4,663 clinical feature occurrences. Patients progressing to acute pyelonephritis (cases) presented a higher features burden than non-progressing controls (71.3% vs. 55.8%; standardized mean difference [SMD] = 0.326), particularly for fever (33.0% vs. 9.0%; SMD = 0.616) and lumbar pain (29.0% vs. 9.8%; SMD = 0.502). In real-world clinician validation, overall performance yielded 93.8% accuracy (95% CI 90.6%-96.1%), 83.6% sensitivity (95% CI 73.0%-91.2%), 96.6% specificity (95% CI 93.6%-98.4%), 87.1% positive predictive value (PPV; 95% CI 77.0%-93.9%), and 95.5% negative predictive value (NPV; 95% CI 92.3%-97.7%). In synthetic stress-testing, the pipeline demonstrated 87.2% accuracy (95% CI 84.6%-89.6%), 73.9% sensitivity (95% CI 68.9%-78.5%), near-perfect specificity (99.5%, 95% CI 98.1%-99.9%), and 99.2% PPV (95% CI 97.2%-99.9%). Conclusions: A privacy-preserving local hybrid framework combining a compact open-weight SLM with regular expression rules achieves high specificity and precision for extracting six uncoded clinical features from multilingual primary care narratives. Running entirely within institutional servers, this approach keeps patient data secure, meets privacy standards, and avoids Application Programming Interface (API) costs, making it a practical tool for EHR research, though with lower sensitivity for narratively complex descriptions.

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

Background: Machine Learning (ML) has shown promise in histopathological diagnosis, but rare tumours, like duodenal epithelial neoplasms, are challenging because large, labelled datasets are difficult to assemble. We propose a data-efficient ML pipeline for detecting duodenal epithelial neoplasia in whole-slide images. Methods: We assembled a five-hospital dataset of haematoxylin and eosin-stained duodenal biopsy whole slide images (WSIs) at 10x magnification, comprising 8,027 normal and 401 neoplastic biopsies, including adenomas, carcinomas and neuroendocrine tumours. WSIs were divided into patches and encoded using the Hibou-B foundation model. We compared global standardisation with unsupervised site-specific standardisation of patch embeddings. Our pipeline combined principal component analysis, a Gaussian mixture model and a Random Forest classifier. We assessed generalisability using leave-one-hospital-out cross-validation across four hospitals and on an independent test set from a fifth hospital. Results: In leave-one-hospital-out cross-validation, the pipeline achieved a mean accuracy of 90%, sensitivity of 96%, specificity of 83% and an AUC of 0.96. On the external test hospital, it achieved 94% accuracy, 93% sensitivity, 95% specificity, and AUC of 0.97. Site-specific standardisation improved mean cross-validation accuracy by 20 percentage points compared with global standardisation, while the Gaussian mixture pipeline improved accuracy by 12 percentage points compared with mean pooling of patch embeddings. Conclusions: We present a data-efficient ML approach for detecting duodenal epithelial neoplasia from WSIs. Strong performance was maintained across hospitals and on an independent external test cohort despite the limited number of neoplastic training cases, supporting the feasibility of foundation-model-based approaches for rare histopathological diagnoses.