International Journal of Medical Informatics最新文献

Objective: Preventable errors in clinical documentation and decision-making remain a major threat to patient safety, yet the role of open-source large language models (LLMs) as practical "second reviewers" in general Internal Medicine remains unclear.

Methods: We prospectively assembled 102 real-world Emergency Internal Medicine reports (de-identified) and either inserted or confirmed realistic errors across four categories (diagnostics/investigations, medication/therapy, process/communication/follow-up, other). Three LLMs (open-source Deepseek-v3-r1 and GPT-OSS-120b, and closed-source OpenAI-o3) were prompted with a uniform system instruction to (i) localize the predefined error and (ii) recommend corrections. Two blinded Internal Medicine specialists independently graded outputs for error localization (0-1) and recommendation quality (Likert 1-4); disagreements were resolved analytically, and analyses used the more conservative rater. Three human clinicians independently reviewed subsets of the same cases to provide a comparator.

Results: Using the conservative rater, correct error localization was 72.5% (74/102; 95% CI 63.2-80.3) for Deepseek-v3-r1, 79.2% (80/101; 95% CI 70.3-86.0) for o3, and 65.7% (67/102; 95% CI 56.1-74.2) for GPT-OSS-120b (Cochran's Q p = 0.033). Pairwise McNemar tests favored o3 over GPT-OSS-120b (p = 0.020; Holm-adjusted p = 0.060); other contrasts were not significant. Recommendation quality was high for all models (median 4/4), with mean ± SD scores of 3.73 ± 0.49 for Deepseek-v3-r1, 3.65 ± 0.64 for o3, and 3.51 ± 0.73 for GPT-OSS-120b. Inter-rater agreement was excellent for GPT-OSS-120b (κ = 0.94 for detection; κ_w = 0.85 for quality), substantial for Deepseek-v3-r1 (κ = 0.75; κ_w = 0.47), and lower for o3 (κ = 0.31; κ_w = 0.14). All models frequently flagged additional clinically useful issues (≥99% of reports).

Conclusion: In real-world Internal Medicine reports with realistic, expert-defined errors, state-of-the-art open-source LLMs approached the performance of a leading closed model and clearly outperformed clinicians in error detection, while providing predominantly guideline-concordant corrective recommendations. Given their advantages for privacy, customizability, and potential local deployment, open models represent credible candidates for privacy-preserving "second-reviewer" support in Internal Medicine. Prospective, workflow-embedded trials that also quantify specificity on error-free notes, alert burden, and patient outcomes are now warranted.

Background: Advancement in digitalization in the health sector have created numerous opportunities for cardiovascular disease (CVD) self-management but also challenges, especially for older adults with lower digital health literacy. Reviews have examined impacts of digital health technology interventions on health outcomes without examining the role of training provided. The aim of this review is to synthesize evidence about the impacts of digital literacy training (DLT) and its characteristics as a component of digital interventions related to cardiovascular health on patient reported outcome and experience measures among older adults with CVD.

Methods: In accordance with the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, a search of MEDLINE, EMBASE, CINAHL, and PsycINFO databases for articles published between inception to March 31, 2025 was conducted. Empirical studies reporting digital health technology training with adults (M age 60 + years) with CVD were eligible for inclusion. Articles included were quality-rated using the Mixed Methods Appraisal Tool. Data were extracted according to the DLT and health technologies alongside patient-reported outcome (i.e technology- and health-related) and experience measures.

Results: Of the 56 included studies (totaling 7698 participants), DLT varied considerably, with 51 describing in-person training. Two studies (totaling 519 participants) examined the role of training with positive impacts on technology- and health-related outcomes. In many of the remaining studies, positive technology-related outcomes were evident but could not be linked back to DLT separate from the overall intervention. In studies (n = 10) where training was evaluated, feedback from patients largely affirmed the training was needed.

Discussion: The collective evidence suggests DLT overall is useful and needed in digital interventions for older adults with CVD. More work is needed to elucidate the distinct role of DLT characteristics and to determine for whom and under what conditions DLT impacts health and technology-related outcomes.

Registration: The protocol for this review was registered Aug 12, 2024 in Open Science Framework (OSF) (See: https://osf.io/unhd9).

Introduction: Artificial intelligence (AI) technologies are increasingly being integrated into pulmonary rehabilitation (PR) to improve individualization, real-time monitoring, and adherence in individuals with chronic respiratory diseases. However, their clinical impact on exercise capacity remains unclear. This systematic review and meta-analysis aimed to evaluate the effectiveness of AI-supported PR programs compared to usual care in improving exercise capacity and respiratory function in adults with chronic respiratory diseases.

Methods: This systematic review and meta-analysis followed PRISMA guidelines and was registered with PROSPERO (ID: CRD420251075622). A comprehensive search was conducted across five electronic databases (PubMed, Web of Science, Scopus, Cochrane Central Register of Controlled Trials (CENTRAL) and PEDro) from inception to July 2025. Statistical analyses for the meta-analysis were conducted using RevMan 5.4.

Results: Three eligible RCTs with a total of 456 participants were included. Pooled analysis showed a significant improvement in 6-minute walk distance (6MWD) after AI-assisted PR group compared to control (MD: 22.08 m; 95% CI: 4.96-39.20; p = 0.01). Moderate heterogeneity was observed (I² = 40%). No meta-analysis was conducted for respiratory function due to insufficient pre-post data. Risk of bias was generally low, though participant blinding was absent in all studies. Methodological quality was good, with a mean PEDro score of 6.0 ± 0.82.

Conclusion: AI-supported PR can significantly improve exercise capacity in individuals with chronic respiratory diseases. Despite promising results, high-quality studies in different pulmonary patient groups are needed to address existing limitations, particularly regarding standardization, cost-effectiveness, and clinical integration of AI-technology.

Background and objective: Peripheral artery disease (PAD) is an atherosclerotic disorder prevalent in the elderly that leads to peripheral function decline and body composition changes. Current diagnostic approaches lack sensitivity for early PAD detection and staging. This study aimed to develop and validate machine learning (ML) models of clinical and CT-based radiological features to improve PAD diagnosis and severity stratification.

Methods: A retrospective multicenter study was conducted using data from two institutions. Clinical and radiological features (including volumetric body composition and muscle texture parameters extracted from calf and thigh segments) were analyzed. Participants were randomly divided into training (70%) and test (30%) sets, stratified by PAD status. Models with different ML algorithms were developed and compared. Model interpretability was assessed with Shapley additive explanation (SHAP) analysis, and performance was evaluated through receiver operating characteristic analysis, Hosmer-Lemeshow testing, Brier score and calibration curves.

Results: This study comprised 342 participants, divided into training (n = 176), test set (n = 76) from Institute 1, external validation (n = 90) from Institute 2. Three models were developed: clinical model (CM), radiological model (RM), and combined clinical-radiological model (CRM). The calf-based CRM using random forest algorithm achieved area under the curves of 0.871 (training), 0.870 (test), and 0.828 (validation), demonstrating good calibration (p ≥ 0.05) and the low Brier score. SHAP analysis visually interpreted feature contributions toward PAD diagnosis and staging.

Conclusions: The CRM model effectively integrated calf-derived radiological and clinical features into a noninvasive, interpretable tool for PAD diagnosis and severity stratification, demonstrating strong clinical applicability.

Background: Nasal obstruction measurement instruments are widely used in the field of nasal surgery. There are various scales that measure nasal obstruction and they differ regarding the number of items, their wording, and the type of response options. In order to pool the data and analyze it together, it is necessary to harmonize it so that we can compare participants' nasal obstruction scores irrespective of instrument they filled in. Data harmonization is still not used in the field of nasal obstruction assessment.

The aim: The aim of this study was to find the best harmonization method in terms of predicting the scores on a target instrument based on the scores from another instrument as precise as possible in the case of four different nasal complaints instruments. A method was sought to find a transformation of scores on the NOSE, Utrecht-Q and SCHNOS that makes them equivalent to ENFAS scores.

Methods: A total of 1324 unique patients completed all four measurement instruments. We tried linear equating, Item Response Theory (IRT), and the following machine learning methods: linear regression, random forest regression, support vector machine regression, and neural network. We used the root-mean-square error (RMSE) of differences between predicted and observed scores to evaluate the quality of harmonization in 5-fold cross-validation.

Results: The ML methods gave overall the best results (the lowest RMSEs) and outperformed IRT (which is considered as a common choice for data harmonization in psychometrics).

Conclusion: The ML methods led to the best quality of the results, confirming their strong potential for data harmonization. This study shows that next to linear equating and IRT that are commonly used for data harmonization, we can also use ML methods for the same purpose and, by doing so, to even increase the quality of the harmonization in certain use cases.

Background: This systematic review compiles evidence and examines how various artificial intelligence (AI) approaches, including machine learning (ML), natural language processing (NLP), meta-heuristic optimization, and explainable AI (XAI), are utilized to predict and diagnose coronary artery disease (CAD). We aim to identify the most commonly used models, evaluate their performance, and explore how interpretability and optimization enhance their usefulness in clinical practice.

Method: A thorough search was conducted across five major databases (PubMed, Scopus, IEEE Xplore, ACM Digital Library, and SpringerLink) to identify relevant studies published between January 2022 and August 2025, in accordance with the PRISMA guidelines. Dual independent reviewers performed study selection and data extraction. The quality of the included studies was evaluated using a checklist based on QUADAS-2. Data were collected on study characteristics, model types, validation methods, and performance metrics, which will be the cornerstone of the analysis.

Results: Sixty-one studies met the inclusion criteria. ML and deep learning models demonstrated strong performance and achieved high accuracy in benchmark datasets, but showed limited clinical validation. Transformer-based models (e.g., BioBERT, ClinicalBERT) showed high efficacy for medical text analysis, but require substantial data and computational resources. Meta-heuristic algorithms (e.g., Genetic Algorithms, Particle Swarm Optimization) effectively improved model efficiency but were rarely applied to unstructured clinical narratives. XAI tools (e.g., SHAP, LIME) improved model transparency, though most studies highlight a need for more rigorous evaluation.

Conclusion: Integrated ML, NLP, meta-heuristic optimization, and XAI hold significant promise in advancing the diagnosis of CAD by improving both accuracy and interpretability. However, challenges such as data scarcity, limited external validation, and a lack of standardized, clinician-centric explainability impede clinical adoption. Future research should focus on hybrid frameworks validated for large, diverse, and real-world datasets.

Objective: To explore the potential of causal rule mining (CRM) as a complementary method to outcome monitoring in identifying plausible causal patterns that may explain undesired clinical outcomes or elevated care consumption in cardiac surgery.

Methods: In this proof-of-concept study, CRM was applied to data from 1,068 patients who underwent elective isolated coronary artery bypass grafting between January 2016 and March 2021 at a single heart center and its referral network in the Netherlands. Outcomes of interest included: 1-year and 120-day mortality, in-hospital stroke, 30-day deep sternal wound infection (DSWI), 30-day re-explorations, 1-year coronary reinterventions, event-free survival, 30-day emergency department (ED) visits, postoperative length of stay, and preoperative fractional flow reserve (FFR) testing. Causal rules were considered relevant if both the odds ratio (OR) and its 95 % confidence interval (CI) were > 1. Identified rules were independently reviewed by clinical experts.

Results: CRM identified 114 significant rules. Five rules were rated as 'new and interesting' and two additional rules were included based on special interest. In follow-up discussions, clinical experts agreed that three rules warrant further clinical investigation: (1) the absence of fractional flow reserve (FFR) testing reducing the likelihood of coronary reintervention, (2) absence of red blood cell (RBC) transfusion during admission reducing the likelihood of 30-day re-explorations, and (3) RBC transfusion increasing the likelihood of 30-day re-explorations.

Conclusion: CRM helped identify potential explanations for certain outcomes and care consumption, providing structured input for hypothesis-driven quality improvement and supporting efforts to enhance patient value.

Introduction: Age and hypertension are key drivers of renal impairment, predisposing older hypertensive adults to faster kidney function decline and higher mortality. We aim to develop an interpretable machinelearning model to predict 4-year chronic kidney disease (CKD) risk in this population.

Methods: Our study incorporated 4,142 hypertensive patients from the Health and Retirement Study (HRS) 2010 and 2012 cohorts for model development and internal validation, with additional temporal validation performed within the HRS 2006 and 2008 cohorts. External validation was conducted using three distinct subcohorts derived from the China Health and Retirement Longitudinal Study (CHARLS) database. Feature selection was implemented through an integrated LASSO-Boruta algorithm, followed by model construction using eight machine learning approaches. Discriminative performance was rigorously evaluated through multiple metrics, including receiver operating characteristic (ROC) curve analysis, accuracy, sensitivity, specificity, and Brier score. The optimal model underwent interpretability analysis via SHapley Additive exPlanations (SHAP) to elucidate decision-making mechanisms and was subsequently deployed as a web-based clinical prediction tool.

Results: Using a combined LASSO-Boruta strategy, we identified nine routinely available predictors for model development. In the training set, SVM achieved the highest AUC (0.735), closely followed by XGBoost (0.734); notably, in the temporal validation cohort, XGBoost was the only model with an AUC > 0.700 (0.702). Overall performance metrics derived from confusion matrices, together with Brier scores, suggested that XGBoost provided a favorable balance between sensitivity and specificity while maintaining acceptable probabilistic calibration. Calibration curves further suggested that XGBoost showed relatively stable agreement between predicted and observed risks across datasets, supporting its selection for subsequent SHAP-based interpretation and web deployment; SHAP identified age as the leading contributor to CKD risk.

Conclusions: We developed an interpretable model using routine clinical indicators to predict 4-year CKD risk in elderly hypertensive adults, with applicability across Asian and Caucasian populations.

Background: Perinatal Mental Disorders (PMDs) are common during pregnancy and the first postpartum year, with negative consequences for women, their partners, and infants, as well as broader societal costs. While numerous interventions have been developed to prevent PMDs, there remains a need for a universal, personalized, and cost-effective solution integrated into routine maternal care. The e-Perinatal study aimed to address this gap. This paper describes the design of the e-Perinatal intervention, delivered via a dedicated mobile health app.

Methods: Guided by the Medical Research Council framework, the e-Perinatal app integrates Self-Determination Theory, Normalization Process Theory, and Patient and Public Involvement perspectives. Existing evidence was reviewed, and stakeholders participated in the co-development of digital micro-interventions (DMs). A clinical rule-based algorithm was implemented to generate personalized recommendations across four pathways (1) weekly content delivery, (2) user preferences, (3) individual risk profile, and (4) PMD monitoring.

Results: The e-Perinatal app includes: 1) DMs focused on psychological, physical activity, and healthy lifestyle domains; 2) a personalized recommendation engine; 3) a social support section; 4) mental health monitoring; 5) an 'SOS' button for assistance; and 6) an appointment reminder tool. In total, 332 evidence-based DMs were developed for women and their partners and delivered in text, audio, and video formats. A clinical rule-based algorithm tailors recommendations according to user characteristics and perinatal stage, employing adaptive content filtering to optimize personalization.

Conclusion: the e-Perinatal app is a personalized mHealth intervention toprevent PMDs within routine maternal care. The intervention combines evidence-based strategies, personalized recommendations, and adaptive digital content to prevent PMDs. Future research will assess effectiveness, implementation, and real-world impact of e-Perinatal intervention for PMD prevention.

Immersive technologies offer promising capabilities for chronic disease management, but their implementation and specific applications across the chronic care continuum remain limited. This study examines how immersive technologies are being utilized across various chronic disease contexts through a scoping review. Using a comprehensive mapping of literature published between 1995 and 2024, we identified 2,012 relevant articles from major databases using WHO and CDC-defined chronic disease keywords and finally focused on 127 studies for detailed manual review. Our approach combined text analytics (BERTopic modelling) with manual synthesis. This methodology revealed eight key themes where immersive technologies are being applied: medical procedures, training and education for healthcare professionals, substance use disorder therapy, cognitive rehabilitation, physical rehabilitation, exergaming and biofeedback, navigation and spatial therapy, and pain, stress, and anxiety management. These themes reflect the growing use of immersive technologies to support diverse activities in chronic care settings. The findings highlight the breadth of immersive technology applications across multiple points in chronic care. Our study introduces a thematic framework for understanding immersive applications in healthcare and identifies research directions and opportunities for future investigation. Future research should explore long-term integration into clinical workflows, as well as inclusivity and adoption across diverse populations.