International Journal Of Medical Informatics[JOURNAL]
Sun
200 papers
Data-driven decision support in hospital resource planning: an artificial intelligence-based model proposal for emergency department demand.
Demir E, Ayik YZ, Özbilen M
Int J Med Inform · 2026 Aug · PMID 42143518 · Publisher ↗
BACKGROUND: The sustainability of service quality in healthcare systems is directly related to accurate resource planning, especially in emergency departments with high unpredictability. This study aims to analyze the im...
BACKGROUND: The sustainability of service quality in healthcare systems is directly related to accurate resource planning, especially in emergency departments with high unpredictability. This study aims to analyze the impact of meteorological factors on emergency department visits and propose a highly accurate and explainable artificial intelligence-based decision support model for hospital management. Within the scope of the research, a large dataset of approximately 1.5 million records from two different public hospitals in the Eastern Black Sea region of Turkey was used. METHODS: As a method, comprehensive feature engineering was performed on the raw data; calendar variables, meteorological lags, and historical application trends were derived. The meaningfulness of the input variables was initially verified using Correlation and Granger Causality analyses, and the final variable selection was performed using the SHAP (SHapley Additive exPlanations) method, which is an explainable artificial intelligence (XAI) approach. The SHAP-based feature selection step was performed independently as a pre-processing filter, and the resulting feature set was then locked and applied uniformly to all 22 models, ensuring a fair comparison. In the study, a total of 22 models from three different groups, including Machine Learning, Deep Learning, and Time Series methods, were tested comparatively. RESULTS: A scenario-based evaluation strategy was followed to measure the adaptation of models to dynamic data structures. According to the findings, the 7-day "Walk-Forward" (WF-7d) update scenario, which simulates real-life conditions, emerged as the most optimal strategy, reducing the average error by 9.62% compared to static models. The Prophet model, which demonstrated the best performance, achieved the highest success with values of 34.55 ± 5.66 MAE (5.54% ± 1.41% MAPE) in the Ordu State Hospital data and 45.84 ± 7.51 MAE (5.47% ± 1.39% MAPE) in the Education and Research Hospital data. Additionally, it was found that the SVM and CatBoost models, which have low error rates, maintained generalizability in both institutions. CONCLUSION: The proposed system has the potential to increase operational efficiency by providing healthcare administrators with a proactive decision support mechanism for critical processes ranging from staff scheduling to bed capacity management.
Data-driven clustering of chronic pain profiles using Swedish national registry data: Towards individualized decision support in interdisciplinary rehabilitation.
Thomas I, Nyberg R, LoMartire R … +7 more , Bohman T, Tseli E, Ärnlöv J, Grimby-Ekman A, Vixner L, Hagelberg M, Äng B
Int J Med Inform · 2026 Aug · PMID 42139941 · Publisher ↗
BACKGROUND: Chronic pain affects 20-30% of adults and is a leading cause of disability and societal cost. Interdisciplinary, team-based treatment (IDT) is the most comprehensive approach, yet outcomes vary widely, and lo...
BACKGROUND: Chronic pain affects 20-30% of adults and is a leading cause of disability and societal cost. Interdisciplinary, team-based treatment (IDT) is the most comprehensive approach, yet outcomes vary widely, and long-term benefits are, on average, modest. We aimed to develop clinically interpretable patient clusters from routine pre-treatment intake data and to validate them externally using independent national registry indicators, as a foundation for data-driven clinical decision support. METHODS: We analyzed a nationwide cohort of 90,505 patients entering specialist IDT in Sweden. A theory-informed unsupervised approach was used to cluster biopsychosocial intake features from the Swedish Quality Registry for Pain Rehabilitation using k-means clustering. Internal validation assessed stability and separation, while external validation tested concordance between questionnaire-derived cluster structures and pre-intake sick-leave trajectories and medication prescriptions derived from national registers using the Mantel statistic and logistic regression. RESULTS: Eight distinct clusters were identified, characterized by differing constellations of pain severity, psychological distress, functional status, and pain duration. Registry indicators tracked with cluster burden: higher-severity clusters showed greater sick leave and more medication prescriptions. Concordance between questionnaire-based and registry-based distance matrices was moderate to strong (Mantel r = 0.65; p = 0.0016) and cluster membership was significantly associated with the registry-based features. Three pre-intake sick-leave trajectories (high/stable, medium/stable, and low/increasing) were observed and differed across clusters. CONCLUSIONS: Population-scale unsupervised clustering of routine patient-reported data, externally validated with independent national registries and supported by longitudinal sickness-absence patterns, yields clinically interpretable subgroups with strengthened construct validity. This provides a scalable foundation for patient stratification and the development of future clinical decision-support tools to better target and monitor IDT in real-world care.
Machine learning models for predicting readmission after stroke: A systematic review and meta-analysis.
Alajlouni Y, Zayed YS, Nofal Y … +2 more , Musleh A, Ghannam M
Int J Med Inform · 2026 Aug · PMID 42134253 · Publisher ↗
BACKGROUND: Hospital readmission following stroke poses a significant challenge for healthcare systems. Machine learning (ML) offers the potential to improve prediction models for readmission risk, surpassing traditional...
BACKGROUND: Hospital readmission following stroke poses a significant challenge for healthcare systems. Machine learning (ML) offers the potential to improve prediction models for readmission risk, surpassing traditional statistical methods. However, the performance of ML models in such context has not been systematically evaluated. We aim to evaluate the performance of ML models in predicting post-stroke hospital readmission, as well as identifying the most important readmission predictors. METHODS: A comprehensive systematic literature review and meta-analysis was conducted.Studies included were those utilizing ML models for stroke readmission prediction. The primary outcome was the predictive performance of ML models for all time stroke readmission as reported by the Area Under the Receiver Operating Characteristic Curve (AUROC). Pooled AUROC values were calculated using a random-effects model. RESULTS: A total of eleven studies involving 380,254 patients were analyzed. Mean age of the patients ranged between 64.5 and 79.7 years. In total, 49 ML models were reported with Logistic regression (LR) and Random Forest being the most used. The number of input predictive variables used for model training ranged from 6 to 10,047. The overall pooled AUROC for predicting hospital readmission was 0.74 (95% CI: 0.69 to 0.78). For 30-day readmission, the pooled AUROC was 0.73 (95% CI: 0.67 to 0.79) while for 90-day readmission, it was 0.75 (95% CI: 0.69 to 0.80). The most frequently reported variables as being top predictors of readmission were Length of stay (LOS) (3 studies), Age (2 studies), National Institutes of Health Stroke Score (NIHSS) (2 studies), HbA1c (2 studies), Homocysteine blood level (2 studies). CONCLUSION: Machine learning models demonstrate moderate predictive performance in stroke readmission risk prediction. Future research should focus on validating and refining existing models and adopting unified methodological approaches to aid in drawing more accurate conclusions.
Safety and diagnostic accuracy of large-language model application of PECARN head injury algorithm.
Simper MA, Ahmad FA, Abraham J … +3 more , Maddox TM, Flotken J, Rudloff JR
Int J Med Inform · 2026 Aug · PMID 42127532 · Publisher ↗
BACKGROUND: Clinical decision algorithms used by clinicians guide evidence-based decisions and actions. Automated tools can help with the adoption and sustainability of these algorithms, and automation is especially need...
BACKGROUND: Clinical decision algorithms used by clinicians guide evidence-based decisions and actions. Automated tools can help with the adoption and sustainability of these algorithms, and automation is especially needed in the emergency setting which requires quick decision making in a chaotic environment. This preimplementation study evaluated whether a large language model (LLM) can apply the Pediatric Emergency Care Applied Research Network (PECARN) Head Injury Algorithm safely and accurately on emergency department notes, a preliminary step in the development of a clinical decision support tool for treating pediatric head injuries. METHODS: We studied the safety and capability of an LLM, Generative Pretrained Transformer (GPT), to apply the PECARN Head Injury Algorithm using notes of patient encounters aged 3 months - 18 years who presented to a pediatric emergency department (ED) with chief complaint of "head injury". A dataset of 24 patients was curated to include a diverse range of symptoms for developing models and a dataset of 122 was randomly selected for testing. We developed and compared four LLM models to extract clinical features from clinical notes. Primary outcomes were safety, measured as negative predictive value (NPV), and accuracy of the LLM models compared to gold standard pediatric emergency medicine (PEM) physicians. Secondary outcomes were the accuracies for the nine features used in the PECARN algorithm. RESULTS: All models demonstrated high NPV comparable to PEM physicians. The GPT model with the highest combination of NPV and accuracy was the prompt-engineered "Optimized Features Model" (NPV = 0.98, accuracy = 0.89), which performed similarly to that of the ED clinicians in both NPV (0.99) and accuracy (0.92). CONCLUSIONS: Our LLM-based tool for a clinical decision algorithm demonstrated high accuracy and NPV. While promising, further studies on scalability and feasibility are needed to ensure LLM-based digital health tools encourage safe, effective care for pediatric patients in the ED.
Predicting length of stay in the pediatric intensive care unit at a tertiary center in Saudi Arabia using machine learning.
Alowa M, Alqassab A, Al-Rumaih D … +4 more , AlJubab H, Alsoqati A, Alzamanan M, Al Khalifah A
Int J Med Inform · 2026 Aug · PMID 42127531 · Publisher ↗
BACKGROUND: Prolonged stay in pediatric intensive care units (PICUs) is associated with increased mortality risk, elevated healthcare costs, and diminished critical care capacity. Accurate early prediction of length of s...
BACKGROUND: Prolonged stay in pediatric intensive care units (PICUs) is associated with increased mortality risk, elevated healthcare costs, and diminished critical care capacity. Accurate early prediction of length of stay (LOS) may facilitate resource allocation, discharge planning, and family counseling. Traditional regression-based models have demonstrated limited performance because of the complex, non-linear nature of pediatric critical illnesses. OBJECTIVES: To develop and internally validate machine-learning models that predict PICU LOS using admission-time clinical data and to identify key predictors using explainable artificial intelligence techniques. METHODS: This retrospective cohort study included all eligible PICU admissions at a tertiary center in Saudi Arabia (2013-2022). LOS was categorized into short, intermediate, and prolonged stay using percentile-based binning. Multiple supervised machine learning algorithms were trained on a stratified set with cross-validation hyperparameter tuning and evaluated on an independent held-out test set. Performance was assessed using accuracy and micro-averaged multiclass area under the curve and interpretability via SHapley Additive exPlanations. RESULTS: Data from 6,090 admissions were analyzed. The Light Gradient Boosting Machine and Categorical Boosting models demonstrated the best performance, achieving micro-averaged multiclass areas under the curve of 0.826 and 0.832, respectively. Discrimination was strongest for short- and prolonged-stay categories, with lower performance for intermediate stays. Key predictors of prolonged stay included early mechanical ventilation, admission source, post-operative status, physiological instability, and comorbidity burden. CONCLUSIONS: Machine-learning models using admission-time data can reliably classify PICU LOS, particularly at the extremes of stay duration. This explainable, data-driven approach may support early risk stratification and inform operational decision-making in pediatric critical care.
Development and usability testing of 'Eating Smart' - A mobile application for promoting healthy eating in Chinese colorectal cancer survivors and high-risk populations.
Wai Chung KC, Takemura N, Ho MM … +7 more , Tak Lam WW, Lee AM, Yee Chan WY, Lam S, Foo CC, Shum NF, Tak Fong DY
Int J Med Inform · 2026 Aug · PMID 42119218 · Publisher ↗
OBJECTIVE: Dietary modifications are critical for reducing colorectal cancer (CRC) incidence and mortality. While our prior dietary intervention demonstrated efficacy in improving dietary patterns among Chinese CRC survi...
OBJECTIVE: Dietary modifications are critical for reducing colorectal cancer (CRC) incidence and mortality. While our prior dietary intervention demonstrated efficacy in improving dietary patterns among Chinese CRC survivors, scalability limitations necessitated adaptation into a mobile application (app). This study aimed to develop the Eating Smart app to reduce consumption of red/processed meats and refined grains, and evaluate its usability among Chinese CRC survivors and high-risk individuals. MATERIALS AND METHODS: A mixed-methods study was conducted in three phases: (1) prototype development (adapting materials from our previous dietary intervention, reviewed by an expert advisory group), (2) app development, and (3) usability testing. Participants, comprising 100 colon polyp clinic attendees and 11 CRC survivors, engaged with the app for approximately two weeks. Usability was assessed using the validated improved Chinese version of the Mobile Health App Usability Questionnaire (I-C-MAUQ) and semi-structured interviews. The I-C-MAUQ score ranges from 1 to 7, with a lower score indicating better usability. RESULTS: In phase 1, a culturally aligned, theory-based prototype was adapted from materials used in a prior randomised controlled trial. Building on the theoretical and cultural considerations from phase 1, phase 2 iteratively developed a functional application comprising 4 major modules (Healthy Cooking, Progress Tracking, Educational Pages, and Discussion Forum). Participants reported good overall app usability, with a mean I-C-MAUQ score of 2.49 ± 0.24 (range 1-7), reflecting a tendency to agree with usability statements. Qualitative interviews encapsulated strengths, including intuitive navigation, culturally tailored dietary recommendations, and peer support via the discussion forum. Key optimization areas included streamlining the food logging process (e.g., utensil-based portion estimation and pre-specified meal options for composite dishes) and integrating stoma-friendly recipes. CONCLUSIONS: The Eating Smart app is the first standalone app offering culturally tailored dietary support for Chinese CRC survivors and high-risk individuals. While further evaluation on intervention effectiveness is needed, initial feedback in this study supports its usability. Iterative optimizations - particularly simplified food logging features, predefined meal options, stoma-friendly recipes, and nutritionally labeled recipes - are warranted. Future work should optimize the app and explore its efficacy.
Development and validation of a clinlabomics-based machine-learning model for noninvasive risk stratification of moderate-to-severe OSA.
Liu P, Guo Z, Sheng Y … +5 more , Tan L, Li B, Zhang W, Min H, Lyu X
Int J Med Inform · 2026 Aug · PMID 42119217 · Publisher ↗
PURPOSE: Obstructive sleep apnea (OSA) is a highly prevalent sleep disorder strongly associated with adverse cardiometabolic and neurocognitive outcomes. Polysomnography (PSG), the diagnostic gold standard, is not a read...
PURPOSE: Obstructive sleep apnea (OSA) is a highly prevalent sleep disorder strongly associated with adverse cardiometabolic and neurocognitive outcomes. Polysomnography (PSG), the diagnostic gold standard, is not a readily accessible test. Therefore, accurate risk stratification tools independent of PSG are critically needed to optimize clinical triage and timely intervention. METHODS: We retrospectively analyzed 408 patients who underwent PSG and routine clinical evaluation at the Second Xiangya Hospital (2018-2025). Seventy-one demographic, anthropometric, hematologic, and biochemical variables were screened. After imputation and collinearity adjustment, univariate logistic regression and LASSO regression identified eight key predictors. Nine machine learning (ML) algorithms were trained and internally validated (7:3 split). Model performance was assessed using discrimination, calibration, and decision curve analyses. Interpretability was evaluated with SHAP values. RESULTS: Among the 408 participants, 276 (67.6%) had moderate-to-severe OSA (AHI ≥ 15). Age, BMI, neck circumference, Mallampati classification, glucose, fibrinogen, AST/ALT ratio, and anion gap were independent predictors. Among ML models, XGBoost achieved the best performance (AUC: 0.852 training, 0.861 validation). Calibration and decision analyses confirmed clinical utility. SHAP identified anion gap, BMI, and FIB as dominant contributors. CONCLUSIONS: We developed and validated an interpretable XGBoost model using routinely available anthropometric and laboratory data for risk stratification of moderate-to-severe OSA. This model, implemented as an online tool, may enable clinical risk stratification and triage for high-risk individuals (particularly patients undergoing PSG evaluation), optimize PSG resource allocation, and support timely targeted intervention.
Knowledge, attitudes, and practices toward artificial intelligence in medicine among Chinese physicians: A cross-sectional study from January to March 2024 with analysis of influencing factors.
Chen Z, Shu J, Wang D … +7 more , Sun E, Wang B, Qi L, Zhang H, Han S, Wu Y, Wang G
Int J Med Inform · 2026 Aug · PMID 42114372 · Publisher ↗
BACKGROUND: As artificial intelligence (AI) transforms medicine, understanding physicians' knowledge, attitudes, and practices (KAP) toward AI is crucial. However, large-scale nationwide studies among Chinese physicians...
BACKGROUND: As artificial intelligence (AI) transforms medicine, understanding physicians' knowledge, attitudes, and practices (KAP) toward AI is crucial. However, large-scale nationwide studies among Chinese physicians are lacking. This study investigates KAP toward AI among Chinese physicians and analyzes its influencing factors. METHODS: A nationwide cross-sectional online survey was conducted from January 15 to March 14, 2024. Multistage sampling was used to recruit practicing physicians across China. A validated, self-administered questionnaire was used to assess demographic characteristics and KAP. The statistical analyses included descriptive statistics, non-parametric tests, multivariate logistic regression, and Spearman correlation. RESULTS: This study included 1,137 participants, with 346 (30.4%) demonstrating good AI knowledge. While 1,034 (90.9%) held positive attitudes toward AI, only 51.2% agreed that medical AI is safe. The rate of clinical AI implementation remained low, with only 328 physicians (28.8%) reporting good AI practices. While "AI in medicine" is the broad field of study, Chinese physicians strongly prefer the term "AI-assisted medicine" (75.1%) to describe AI's functional role, emphasizing its assistive nature. Multivariate analysis identified male gender (OR 1.85; 95% CI 1.42-2.41) and age 50-59 years (OR 2.20; 95% CI 1.31-3.71) as independent predictors of good AI knowledge. Chief physicians showed more positive attitudes than residents (OR 2.06; 95% CI 1.15-3.69). Factors significantly associated with good AI practices included male gender (OR 1.35; 95% CI 1.02-1.80), doctoral degree (OR 2.50; 95% CI 1.71-3.67), and specialization in obstetrics/gynecology (OR 4.95; 95% CI 1.80-13.60), internal medicine (OR 3.94; 95% CI 1.51-10.26), or surgery (OR 4.62; 95% CI 1.77-12.04), compared to pediatrics. Significant positive correlations were found between knowledge and attitude (r = 0.172), knowledge and practice (r = 0.441), and attitude and practice (r = 0.242) (all P < 0.001). CONCLUSIONS: This study found that among Chinese physicians, senior physicians (aged 50-59 years) demonstrated higher AI knowledge than younger colleagues. Additionally, AI adoption rates varied significantly by specialty, with pediatrics showing lower adoption compared to surgery, internal medicine, and obstetrics/gynecology. These findings support targeted strategies, including specialized education for younger and female physicians and the prioritization of AI tool development for underserved specialties such as pediatrics, to foster responsible AI integration into Chinese clinical practice.
A privacy-centric microservice framework for secure FHIR-based integration of heterogeneous medical data.
Ungurean I, Gherman OI, Lavric A … +1 more , Dimian M
Int J Med Inform · 2026 Aug · PMID 42114371 · Publisher ↗
BACKGROUND: Fragmentation across consumer wearables, hospital messaging systems, telehealth platforms, and cloud services impedes timely data reuse and increases privacy risks, for example through proprietary telemetry,...
BACKGROUND: Fragmentation across consumer wearables, hospital messaging systems, telehealth platforms, and cloud services impedes timely data reuse and increases privacy risks, for example through proprietary telemetry, legacy HL7v2 messages, intermittent home connectivity, and inconsistent consent enforcement. OBJECTIVE: To design and evaluate a privacy-centric microservice framework that integrates heterogeneous medical data into Fast Healthcare Interoperability Resources (FHIR) while sustaining sub-second end-to-dashboard latency under realistic network conditions. METHODS: We followed a security- and privacy-by-design process inspired by secure software development lifecycles, consolidating requirements from the GDPR, EHDS Regulation, and Zero Trust guidance into an explicit threat model. We built container-native services orchestrated with Kubernetes and enforced a Zero Trust posture using device-bound X.509 certificates and mutual TLS. A hot-swappable adapter layer normalized HL7v2/IEEE 11073/proprietary telemetry to FHIR resources and persisted them in PostgreSQL with a row-level security. Performance tests replayed wearable sessions through 50 concurrent Android emulators over Wi-Fi and 4G and a 5% random loss model. The outcomes were processing, ingestion, end-to-dashboard latencies, throughput, and resource utilization. Security probes assessed credential replay/clone, injection attempts, horizontal privilege escalation, and volumetric DoS. RESULTS: The processing interval from NIC reception to storage remained below 100 ms (95th percentile). On 4G (RTT ≈ 75 ms), the ingestion latency was 132 ms; with 5% random loss, the end-to-dashboard The security controls blocked credential replay/clone attempts, neutralized injected SQL, enforced row-level isolation, and reduced the observed request rate in an illustrative 50,000 SYN/s availability probe. These results represent an initial validation rather than a full penetration test; large-scale clinical deployments will require independent security assessments. The outcome is a Zero Trust FHIR reference implementation with an explicit threat model, device-bound enrollment, adapter-based normalization, row-level isolation, observability, and targeted validation. CONCLUSIONS: A zero-trust microservice architecture can deliver EHDS-aligned privacy controls and real-time analytics without compromising performance. Beyond the reference implementation, the work contributes a structured security- and privacy-by-design methodology and explicit threat model that can guide similar digital-health platforms.
Empowering open medium-sized generative language models for effective structured search in biomedical systematic reviews.
Budau L, Finney R, Ensan F
Int J Med Inform · 2026 Aug · PMID 42107249 · Publisher ↗
BACKGROUND: Systematic Literature Reviews (SLRs) are essential in biomedical research, particularly for informing public health policy and clinical decision-making. However, the manual generation of Boolean queries for l...
BACKGROUND: Systematic Literature Reviews (SLRs) are essential in biomedical research, particularly for informing public health policy and clinical decision-making. However, the manual generation of Boolean queries for literature searches is resource-intensive, prone to errors, and difficult to scale. Recent advances in large language models (LLMs) have demonstrated potential, yet most existing approaches rely on zero-shot prompting of commercial models, overlooking the cost-efficiency and domain adaptability of fine-tuned open-source alternatives. METHODS: This study proposes a novel, three-stage framework that employs medium-sized, open-source generative models, specifically BioGPT and BioT5, for automated Boolean query generation over PubMed. We develop and release datasets comprising PubMed article titles, MeSH terms, and keywords, and fine-tune the models using both title-only and title-plus-metadata prompts. We evaluate performance on two benchmark datasets: CLEF TAR and FASS-BSLR. Our experiments include comparisons with state-of-the-art baselines, prompt-based large language models, and ablation studies exploring the effects of training data size, metadata inclusion, and post-processing with PubMed's Automatic Term Mapping. RESULTS: Fine-tuned BioGPT outperforms both traditional TAR models and commercial LLMs across key retrieval metrics. On the CLEF TAR dataset, it achieves a Precision of 0.2544, F1 of 0.2392, MAP@1000 of 0.1424, and NDCG@1000 of 0.2490, which surpasses all baselines. On the FASS dataset, it reaches a Recall of 0.1801 and NDCG@1000 of 0.0900, again outperforming all competing models. While slightly behind BioGPT, BioT5 still outperforms most baselines. Notably, BioGPT's Recall of 0.1801 on FASS is more than twice that of PubMed-Title and PubMed-Keyword, and exceeds GPT-3.5 Turbo, GPT-4, Gemini-2, and Llama-3. CONCLUSION: This work demonstrates that fine-tuned, open-source, medium-sized generative models can match or exceed the performance of much larger commercial LLMs in Boolean query generation for biomedical SLRs. These models offer a cost-effective, privacy-preserving, and scalable alternative for structured retrieval of biomedical scholarly texts.
From decision support to clinical integration: A scoping review of artificial intelligence in prehospital airway management.
Fangfang B, Wenjuan Q, Xiaoting Z … +1 more , Yanghui F
Int J Med Inform · 2026 Aug · PMID 42107248 · Publisher ↗
BACKGROUND: Airway management is a critical component of prehospital emergency care, where rapid decision-making and procedural accuracy are essential for patient survival. In recent years, artificial intelligence has em...
BACKGROUND: Airway management is a critical component of prehospital emergency care, where rapid decision-making and procedural accuracy are essential for patient survival. In recent years, artificial intelligence has emerged as a promising tool. However, the current landscape and translational readiness of artificial intelligence applications in prehospital airway management remain unclear. OBJECTIVE: This scoping review aimed to synthesize existing evidence on the applications of artificial intelligence in prehospital airway management and to identify current research gaps and challenges for clinical integration. METHODS: This review followed the PRISMA-ScR guidelines. A systematic search of PubMed, Web of Science, and EBSCOhost was conducted through February 2026. Two reviewers independently screened studies and extracted data on study characteristics, artificial intelligence methods, clinical applications, and performance metrics. RESULTS: Nine studies published between 2020 and 2026 were included. artificial intelligence applications were categorized into four functional domains: predictive modeling for airway intervention and triage, physiological signal monitoring, natural language processing for clinical documentation analysis, and computer vision for anatomical recognition during intubation. Most studies focused on prediction before airway intervention, while relatively few addressed procedural assistance or post-intubation monitoring. Model performance was generally strong, with reported AUC values ranging from 0.867 to 0.960. However, all studies relied on retrospective data and internal validation; only one study conducted external validation, and none reported prospective trials or fairness assessments. CONCLUSIONS: Artificial intelligence shows considerable potential to support prehospital airway management across multiple stages of care. Nevertheless, current evidence remains exploratory and is limited by methodological constraints, lack of prospective validation, and insufficient integration with clinical workflows. Future research should prioritize multimodal data integration, external validation, and prospective implementation studies to facilitate the safe and effective translation of artificial intelligence into real-world prehospital practice. Protocol registered number: PROSPERO (CRD42026132386).
Factors influencing older adults' acceptance and usability of assistive technology services: A longitudinal multilevel analysis.
Fiorini L, Pani J, Rovini E … +8 more , D'Onofrio G, Iannacone G, Russo S, Giuliani F, Lorusso L, Toccafondi L, Calamida N, Cavallo F
Int J Med Inform · 2026 Aug · PMID 42102445 · Publisher ↗
OBJECTIVE: Assistive technologies have the potential to enhance the quality of life of older adults and their caregivers. However, long-term adoption in real-life settings remains limited, and the non-technological facto...
OBJECTIVE: Assistive technologies have the potential to enhance the quality of life of older adults and their caregivers. However, long-term adoption in real-life settings remains limited, and the non-technological factors influencing sustained use are not fully explored. This study investigates how socio-demographic, psychological, and caregiving-related variables affect the acceptance and usability of assistive technologies over time. METHODS: Seventy-eight community-dwelling older adults participated in a one-year pilot evaluating a socialization platform and a health/environmental monitoring system. Usability (SUS) and acceptance (Almere Model constructs) were assessed at baseline, 6 months, and 12 months. Baseline socio-demographic and psychosocial variables were used to derive user profiles through k-means clustering. Linear mixed-effects models examined profile-dependent longitudinal trajectories, complemented by regression-based models testing independent predictors. RESULTS: Two distinct baseline profiles were identified, differing in age, digital skills, technostress, loneliness, and perceived support. Profile membership predicted divergent longitudinal trajectories across usability and multiple acceptance domains. The digitally resilient profile demonstrated higher usability and more favourable changes over time. Across regression models, technostress and loneliness consistently emerged as robust negative predictors, independent of time and scenario. Attrition analyses did not reveal systematic baseline differences between completers and dropouts. CONCLUSION: Technology acceptance among older adults is structured by baseline psychosocial and digital configurations rather than exposure alone. Integrating baseline profiling with longitudinal modelling provides a framework to understand heterogeneity in gerontechnology adoption and highlights modifiable factors that may support sustained engagement.
Mobile applications for time and event management in older adults and their careagivers: A scoping review.
Sawyer K, Wang R, Dewalt S … +3 more , Hufnagel A, Comeau AK, Cruz AM
Int J Med Inform · 2026 Aug · PMID 42097113 · Publisher ↗
BACKGROUND: Time and event management is critical for independence in daily living, and mobile health applications are an emerging compensatory strategy that may support older adults as they age. OBJECTIVE: This scoping...
BACKGROUND: Time and event management is critical for independence in daily living, and mobile health applications are an emerging compensatory strategy that may support older adults as they age. OBJECTIVE: This scoping review aims to examine what applications exist for time and event management, the ability of older adults to use these applications and their effects, outcome variables used, and the strength of evidence. MATERIALS: A scoping review was conducted using the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews. A search strategy and study selection process were developed. Six electronic databases, MEDLINE, EMBASE, APA PsycInfo, CINAHL, Scopus, and Web of Science, were used to identify studies exploring the use of mobile applications for time and event management in older adults. RESULTS: The search returned 2,854 results, and 24 studies were identified for inclusion. Most studies examined applications that are not available on the market; those that are available primarily consist of manufacturer-based applications. Study participants often benefited from education, training, and support, and impacts were observed on cognition, functional independence, and self-management. Common outcome variables were related to usability and acceptance, behavioural and log data, function and independence, and quality of life and well-being. CONCLUSION: While a small number of studies report benefits to cognition and daily functioning, there is a low level of evidence for these studies. There is limited evidence for the use of mobile applications for time and event management in older adults and their caregivers.
How improper dataset split hinders model generalizability: a systematic comparison in Human activity recognition and exercise evaluation tasks.
Pe S, Dagliati A, Parimbelli E … +4 more , Vittoria Guerra BM, Sozzi S, Bellazzi R, Nicora G
Int J Med Inform · 2026 Aug · PMID 42097112 · Publisher ↗
BACKGROUND: Human Activity Recognition (HAR) and exercise assessment models are increasingly used in healthcare to support clinical evaluation, rehabilitation, and remote monitoring. However, their real-world applicabili...
BACKGROUND: Human Activity Recognition (HAR) and exercise assessment models are increasingly used in healthcare to support clinical evaluation, rehabilitation, and remote monitoring. However, their real-world applicability critically depends on the ability to generalize across unseen subjects, whose movement patterns may differ substantially due to inter-individual variability. Despite this, many studies adopt random noncross-subject (NCS) data splits, where samples from the same individual appear in both training and test sets, potentially leading to overly optimistic and clinically misleading performance estimates. OBJECTIVE: We investigate (i) how NCS and cross-subject (CS) splits affect performance estimation across machine learning and deep learning models under tasks of increasing complexity, (ii) how data splitting and differences between training and test sets contribute to predictive variance and stability. METHODS: Experiments were performed using a large-scale HAR benchmark dataset (NTU RGB+D 120) and a rehabilitation-specific dataset (IntelliRehabDS). A total of 12 machine learning and deep learning models were trained across both tasks, and their performance was estimated and compared using a simulation-based approach. Predictive variance decomposition, via Generalized Linear Mixed-Effects models, was applied to link the split strategy and differences in training and test instances to model output stability. RESULTS: NCS splits consistently overestimated model performances, with discrepancies increasing alongside task and model complexity. DL architectures, in particular, showed markedly higher NCS performance compared to CS splits, generally with statistical significance. Variance decomposition revealed that greater subject difference between training and test sets often enhances predictive instability, while CS splitting reduces variance by promoting more generalizable representations. CONCLUSIONS: Improper dataset splits can mislead model evaluation, exaggerate generalization capabilities, and undermine clinical trust. Our study provides empirical evidence for computer vision-based rehabilitation models and offers methodological guidance for robust evaluation practices, supporting reproducible and trustworthy AI deployment in rehabilitation and broader healthcare applications.
Large Language Models and the return of knowledge engineering.
Bellazzi R
Int J Med Inform · 2026 Aug · PMID 42091400 · Publisher ↗
Abstract loading — click title to view on PubMed.
Performance of large language models and clinical decision support in perioperative management of oral anticoagulants.
Çalışkan MU, Sarıbaş H, Keskin G … +3 more , Kertmen Ö, Çakmak A, Özbay Y
Int J Med Inform · 2026 Aug · PMID 42070385 · Publisher ↗
BACKGROUND: Perioperative anticoagulant management is critical because of the competing risks of ischemia and bleeding. Large language models (LLMs) and clinical decision support (CDS) have shown substantial advances and...
BACKGROUND: Perioperative anticoagulant management is critical because of the competing risks of ischemia and bleeding. Large language models (LLMs) and clinical decision support (CDS) have shown substantial advances and are increasingly being applied across various domains of cardiology. However, to date, no studies have specifically evaluated the performance of LLMs combined with CDS in perioperative anticoagulant management. METHODS: Two cardiologists developed 40 guideline-based clinical scenarios involving patients receiving oral anticoagulants scheduled for non-cardiac surgery, including 20 direct oral anticoagulant (DOAC) and 20 warfarin scenarios. Three commonly used LLMs (ChatGPT 5.2, Gemini 3.0 Pro, and DeepSeek V3.2) were evaluated in two phases: baseline performance (Phase 1) and performance after integration of structured CDS tables (Phase 2). Responses were assessed for guideline concordance on a per-scenario basis, requiring correct recommendations across all predefined domains. Model performances were compared using Cochran's Q test, with post hoc Dunn-Bonferroni correction, and within-model comparisons were performed using McNemar's test. RESULTS: In Phase 1, guideline-concordant management was achieved in 60% of scenarios by ChatGPT-5.2, 55% by Gemini 3.0 Pro, and 50% by DeepSeek V3.2, with no significant difference among models (p = 0.180). Following CDS augmentation in Phase 2, scenario-level accuracy improved significantly for all models (all p < 0.05), increasing to 80% for ChatGPT-5.2, 100% for Gemini 3.0 Pro, and 85% for DeepSeek V3.2. Comparative analysis in Phase 2 demonstrated a significant difference among models (p = 0.009), driven primarily by superior performance of Gemini 3.0 Pro compared with ChatGPT-5.2 (adjusted p = 0.009). CONCLUSIONS: Baseline performance of LLMs in perioperative anticoagulant management was modest and insufficient to replace clinical judgment. However, integration of structured CDS tools resulted in marked and clinically meaningful improvements in guideline-concordant performance across all evaluated models.
Maximising the patient voice in health care: A scoping review of secondary use of routinely collected qualitative open-ended patient reported experience measure comments.
Engstrom T, Kirwa T, Olsen Q … +4 more , Woods L, Quigley DD, Sullivan C, Pole JD
Int J Med Inform · 2026 Jul · PMID 42066517 · Publisher ↗
INTRODUCTION: Routinely collected patient reported experience measure (PREM) surveys capture patients' experience across many countries and care settings. Secondary use of open-ended PREM comments is not common practice,...
INTRODUCTION: Routinely collected patient reported experience measure (PREM) surveys capture patients' experience across many countries and care settings. Secondary use of open-ended PREM comments is not common practice, however with increasing volume and scope being captured, this should be investigated. We examine how routinely collected open-ended PREM comments are used in research for secondary uses and synthesise considerations. METHODS: Using the JBI approach and PRISMA-ScR guideline, we searched four academic databases for peer-reviewed English-language studies (2010-2024) that used routinely collected open-ended PREM comments to answer questions focused on a specific person, place, time or perspective beyond the broad intent of the original survey. We summarised study characteristics, categories and methods of secondary use and performed a descriptive analysis of study authors' considerations for secondary use of this data. RESULTS: We identified 2,200 unique articles and conducted full-text review of 206 articles, yielding 25 included studies. Studies most frequently used open-ended PREM comments to investigate elements of care (32%, n = 8), time periods (24%, n = 6), examine types of care (20%, n = 5), or treatments (20%, n = 5). Researchers used manual analysis approaches (56%, n = 14), applied sentiment analysis and thematic analysis (each 48%, n = 12). A key strength in the secondary use of open-ended PREM comments is that it reflects what is important to patients' care experiences; while a limitation is the potential bias inherent in survey data (e.g. non-response bias). CONCLUSION: Secondary use of open-ended PREM comments for research is growing, but it is not yet widely used. Using these already collected data for research eliminates the time and cost of additional data collection and incorporates patient voice into research. A formal framework for secondary use of open-ended PREM comments will support incorporating these data into research. Secondary use of PREMs comments could enable more insightful, efficient research and maximise the patient voice in healthcare.
Examining user-AI interaction patterns in health-Information queries.
Wang JT, Chung HW, Do GN … +1 more , Yang AT
Int J Med Inform · 2026 Jul · PMID 42061119 · Publisher ↗
In this study, we examine how individuals utilize generative artificial intelligence (GAI) when seeking health-related information. Using a dataset of user-GAI chat logs available on Hugging Face, we analyzed real-world...
In this study, we examine how individuals utilize generative artificial intelligence (GAI) when seeking health-related information. Using a dataset of user-GAI chat logs available on Hugging Face, we analyzed real-world interactions in which users posed health-related questions to a generative model. We applied a combination of data and text-analytic methods to categorize these interactions, including supervised machine learning techniques such as Support Vector Machines (SVMs). SVMs were selected for their efficiency and strong performance in high-dimensional text classification tasks, and used to identify recurrent themes in user queries and interactions. We found that users frequently consult AI chatbots for symptom exploration, medical education, mental health support, and general health advice. The findings suggest that GAI tools may not only function as informational resources, but also as preliminary support tools that can shape users' health knowledge and encourage them to seek consultation with medical professionals.
Artificial intelligence-based reclassification of gastric adenocarcinoma enables prognostic stratification via diffuse-type patch proportion.
Andraș C, Minciună CE, Stan D … +10 more , Tudor S, Manuc T, Dragomir MP, Bitere O, Micu A, Almarii F, Droc G, Calin G, Herlea V, Vasilescu C
Int J Med Inform · 2026 Jul · PMID 42054933 · Publisher ↗
BACKGROUND: Gastric adenocarcinoma (GAC) remains a major global health burden with marked heterogeneity, complicating diagnosis and prognostic assessment. The Laurén classification, though widely used, suffers from inter...
BACKGROUND: Gastric adenocarcinoma (GAC) remains a major global health burden with marked heterogeneity, complicating diagnosis and prognostic assessment. The Laurén classification, though widely used, suffers from interobserver variability, particularly in defining the mixed subtype. Artificial intelligence (AI)-driven image analysis may improve standardization and prognostic assessment in GAC. METHODS: We retrospectively analyzed 404 patients with resected GAC (2015-2022) from Fundeni Clinical Institute. Whole-slide images (WSIs) were annotated by pathologists with expertise and processed into patches for training a two-stage deep learning pipeline based on YOLO26m-cls. The first model (GAC-I) distinguished malignant from non-malignant tissue, while the second (GAC-ST) classified malignant patches as intestinal or diffuse. We developed the diffuse prognostic score (DPS), defined as the proportion of diffuse patches relative to total patches, and correlated it with overall survival (OS). RESULTS: GAC-I and GAC-ST achieved high diagnostic performances, with accuracies of 0.9437 ± 0.0317 (F1 score: 0.9456 ± 0.0243) and 0.8080 ± 0.0833 (F1 score: 0.7528 ± 0.1094). DPS ≥ 0.5 was significantly associated with lower median OS (16.1 months) compared to DPS < 0.5 (42.067 months), association confirmed by multivariate Cox-regression analysis (HR 3.88, p < 0.001) and matched case-control analysis. Groups were balanced across all variables except tumor differentiation, which was more frequently high-grade in DPS ≥ 0.5. After adjustment, DPS ≥ 0.5 remained an independent predictor of mortality (HR 2.684, p = 0.027). CONCLUSION: We developed and validated a robust AI-based framework for automated GAC classification and prognostic stratification using H&E WSIs. DPS is an independent, reproducible marker of OS, supporting its potential integration into clinical pathology workflows to guide personalized treatment.
Ai-enhanced clinical decision support reduces medication errors and adverse drug events in a multicenter teaching hospital network: A prospective randomized controlled trial.
Shakarbaev N
Int J Med Inform · 2026 Jul · PMID 42054932 · Publisher ↗
BACKGROUND: Medication errors remain a leading source of preventable harm in hospitalized patients, contributing to adverse drug events (ADEs), prolonged hospital stay, and avoidable healthcare costs. Although clinical d...
BACKGROUND: Medication errors remain a leading source of preventable harm in hospitalized patients, contributing to adverse drug events (ADEs), prolonged hospital stay, and avoidable healthcare costs. Although clinical decision support systems (CDSS) integrated with electronic health records (EHRs) have demonstrated potential to reduce prescribing errors, rigorous multicenter randomized evidence from low- and middle-income countries (LMICs), as classified by the World Bank income criteria, remains scarce. OBJECTIVES: To evaluate the effect of an AI-enhanced, EHR-integrated CDSS on medication error rates and ADE incidence in hospitalized adults, with additional assessment of alert performance, clinician adoption, and cost-effectiveness. METHODS: We conducted a prospective, parallel-arm, randomized controlled trial (RCT) across four tertiary-care teaching hospitals in Tashkent, Uzbekistan (January 2022 - August 2023). Adult inpatients were randomized 1:1 to CDSS-assisted care (MedGuard-UZ v1.3) or standard care. Primary outcomes were medication error rate per 1,000 patient-days and ADE incidence per 100 admissions. Analyses followed the intention-to-treat (ITT) principle. RESULTS: Among 2,384 randomized patients, the CDSS group demonstrated a 49.2% reduction in medication error rates (3.47 vs. 6.83 per 1,000 patient-days; p < 0.001) and a 47.2% reduction in ADE incidence (4.7 vs. 8.9 per 100 admissions; p < 0.001). Overall alert acceptance was 73.6%, with allergy/contraindication alerts achieving 95.5%. Clinician adoption rose from 42.1% to 88.7% daily active users over 12 months. Length of stay was significantly shorter in the CDSS group (7.1 vs. 7.8 days; p = 0.002), as were 30-day readmissions (11.2% vs. 13.7%; p = 0.041). Estimated return on investment was 489% over 12 months. CONCLUSIONS: AI-enhanced CDSS integration was associated with substantially improved medication safety and selected hospital outcomes in a multicenter LMIC tertiary-care setting. The MedGuard-UZ AI project materials are publicly available at https://github.com/Shakarbayev/MedGuard-UZ. For peer-review reproducibility, the repository state corresponding to this revision has been archived under the tagged release v1.0.0-ijmedi-rct (tag commit: f796c3d). MedGuard-UZ v1.3 denotes the internal trial system version, whereas v1.0.0-ijmedi-rct (tag commit: f796c3d) identifies the public revision archive supporting evaluation of the AI project component in resource-constrained health systems.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe