International Journal Of Medical Informatics[JOURNAL]
Sun
200 papers
A scoping review on conversational AI in mental health: A human-centered perspective.
Li J, Li H, Zhu H … +5 more , Walter H, Banaschewski T, Li C, Li X, Zhang Y
Int J Med Inform · 2026 Jul · PMID 42048909 · Publisher ↗
BACKGROUND: Conversational AI offers scalable mental health support, with large language models (LLMs) enabling personalized interactions. Human-centered design is critical in this domain, yet a comprehensive synthesis f...
BACKGROUND: Conversational AI offers scalable mental health support, with large language models (LLMs) enabling personalized interactions. Human-centered design is critical in this domain, yet a comprehensive synthesis from this perspective is lacking. This review maps conversational AI research in mental health across the patient journey and develops a human-centered taxonomy to guide future design. METHODS: Following PRISMA guidelines, we conducted a comprehensive search across fifteen multidisciplinary databases. We systematically analyzed the literature across six dimensions: research foci, mental disorder types, target populations, AI technologies, data sources, and evaluation metrics. A consensual taxonomy research method was employed to develop a human-centered design framework. RESULTS: Of 10,293 identified records, 677 studies met the inclusion criteria. Analysis reveals a marked increase in publications since 2020, predominantly from computer science (449 studies), followed by medicine (148) and social sciences (80). Research is skewed toward detection (23%) and intervention (66%) stages, with prevention (8%) and maintenance (3%) receiving less attention. Mood, anxiety, and stress-related disorders are the most investigated conditions. LLMs have emerged as the predominant AI technology, particularly within intervention and maintenance stages. Data sources continue to rely heavily on text-based inputs, with multimodal approaches still limited in adoption. Evaluation metrics vary significantly by discipline, reflecting limited cross-disciplinary integration. Through thematic synthesis, we developed a human-centered taxonomy comprising four primary dimensions: Emotional Sensitivity to Users, User-Centric Interaction Design, Human-AI Collaboration and Capability Enhancement, and Ethics and Accountability, with a total of thirteen sub-dimensions. CONCLUSIONS: This review provides a comprehensive, human-centered mapping of conversational AI research in mental health across the patient journey. Critical gaps remain in stage coverage, disorder diversity, population inclusivity, multimodal data integration, and interdisciplinary evaluation. The proposed taxonomy offers a structured framework to align AI development with human-centered principles, fostering empathetic, ethical, effective, and equitable mental health support.
Generative artificial intelligence use in medical research: A medical student vs. physician survey from Saudi Arabia.
Alhowaish TS, Alshafi NI, Aldekhyyel RN … +4 more , Mesallam TA, Farahat M, Temsah MH, Malki KH
Int J Med Inform · 2026 Jul · PMID 42044622 · Publisher ↗
INTRODUCTION: Generative artificial intelligence (GenAI) such as ChatGPT, is increasingly used to generate research ideas, aid in literature reviews, and support clinical reasoning. While its adoption is rapidly growing...
INTRODUCTION: Generative artificial intelligence (GenAI) such as ChatGPT, is increasingly used to generate research ideas, aid in literature reviews, and support clinical reasoning. While its adoption is rapidly growing worldwide, concerns persist regarding accuracy, bias, data privacy, and academic integrity. Understanding how physicians and medical students perceive and use these tools is critical for developing effective guidelines. This study aimed to assess familiarity, usage patterns, perceived benefits, and ethical concerns surrounding ChatGPT and other GenAI platforms among physicians and medical students in Saudi Arabia. METHODS: A cross-sectional survey was conducted between April 19 and July 7, 2025. The study included a total of 417 respondents: 182 physicians from multiple Saudi major academic hospitals and 235 medical students from six universities across Saudi Arabia. Recruitment was via online and on-site methods using QR codes and survey links. A validated questionnaire addressed demographics, prior research experience, ChatGPT and other GenAI use, perceived benefits, applications, and ethical concerns. Data were analyzed using IBM SPSS version 24, applying Chi-square and Fisher's exact tests to compare groups. We compared intact groups as they exist in practice; findings are interpreted in light of inherent differences in age and clinical experience, without formal adjustment. RESULTS: The use of GenAI in research was higher among medical students (73%) compared to physicians (59%). Most medical students had used ChatGPT (95% vs 81%, p < 0.001), whereas physicians more often reported using other GenAI tools (48% vs 29%, p < 0.001). Physicians most often used GenAI for academic writing (81%), while students preferred summarizing findings (75%). Nearly all students (99%) and physicians (89%) found ChatGPT easy to learn, and a majority in both groups acknowledged improvements in research efficiency (85% vs 80%), time reduction (84% vs 91%), and overall quality (77% vs 74%). Physicians expressed greater concern over data privacy (73%), while students emphasized accuracy and reliability as the key issue (77%). CONCLUSION: Both physicians and medical students are increasingly embracing GenAI for research purposes. Nonetheless, prevalent concerns about reliability and ethics highlight the need for clear guidelines and training programs to ensure responsible use of GenAI in future medical research. Group contrasts are interpreted in light of inherent differences in age and clinical experience.
Mapping digital health initiatives and technological requirements: A proposal from Sao Paulo State for digital maturity assessment.
Carvalho GB, Buto MS, Fontes AS … +8 more , Garcia ML, Viana N, Oliveira AC, Caliman RA, Lima RR, Andrade MC, Cerri GG, Carvalho CR
Int J Med Inform · 2026 Jul · PMID 42035737 · Publisher ↗
INTRODUCTION: Digital transformation in health is driven by Information and Communication Technologies integration and guided by institutional guidelines. Despite multiple models, no consensus exists on assessing facilit...
INTRODUCTION: Digital transformation in health is driven by Information and Communication Technologies integration and guided by institutional guidelines. Despite multiple models, no consensus exists on assessing facility-level digital health maturity, including organizational capacity and workflows. OBJECTIVE: To present the preliminary results of Mapping Digital Health (DH) initiatives and Technological Requirements (TR) conducted in public healthcare facilities in São Paulo State. METHODS: This descriptive observational study was conducted from March to August 2025 in public healthcare facilities (n = 6,319) across 17 Regional Health Departments in Sao Paulo State. A general sample target of 3,719 facilities was calculated, with specific targets by type: primary healthcare units (n = 2,570), specialty outpatient clinics (AME) (n = 62), hospitals (n = 529), emergency care units (n = 381), and prison units (n = 177). Two questionnaires (26 and 28 items across six and seven domains, respectively) were structured based on key technical and policy reference documents and covered essential domains for assessing DH maturity and TR. Responses on a five-point Likert scale were used to calculate maturity indices classified as incipient, developing, collaborative, or advanced. RESULTS: Questionnaires responses included 5,325 for DH and for 5,069 TR questionnaires. Prison units and AME met 100% of their response targets, while PHU, emergency care units, and hospitals achieved around 80%. Overall, Sao Paulo State scored 42.0 for DH and 45.7 for TR, both classified as "developing," considered an intermediate stage and reflects a moderate maturity. Prisons units and AME exhibited the highest scores in DH maturity and TR, respectively. CONCLUSION: Sao Paulo State shows an intermediate stage of digital health and technological requirement maturity, despite inequalities across facility typologies. The facility-level maturity index enabled a detailed analysis and may contribute to future public policies, training initiatives, and infrastructure investments.
Letters to Editor: Refining DL for rectal cancer downstaging.
Liu X, Huang Y
Int J Med Inform · 2026 Jul · PMID 42035736 · Publisher ↗
Abstract loading — click title to view on PubMed.
Enhancing home care safety: a randomized controlled trial of VR-based training for informal caregivers.
Gil-Hernández E, Carrillo I, Pérez-Esteve C … +4 more , Arroyo A, Guilabert M, Ballester P, Mira JJ
Int J Med Inform · 2026 Jul · PMID 42034932 · Publisher ↗
BACKGROUND: Europe's aging population and the move to home-based long-term care place growing demands on informal caregivers, who often lack formal training. This substantially increases both caregiving and medication er...
BACKGROUND: Europe's aging population and the move to home-based long-term care place growing demands on informal caregivers, who often lack formal training. This substantially increases both caregiving and medication errors and caregiver burden. Virtual reality (VR) enables experiential training but is seldom tailored to non-professional caregivers or evaluated in real-world conditions. OBJECTIVE: To analyze the ability of a brief training based on VR to reduce informal caregivers' burden and their caregiving/medication errors at home. METHODS: Two-arm randomized controlled trial in three Spanish regions. Informal caregivers were randomized to structured VR training or usual materials; N = 140 (70/70) caring for people with chronic conditions. Assessments at baseline and 3 months, aligned with the Kirkpatrick model: L1 satisfaction; L2 video-based error detection; L3 self-reported caregiving/medication errors; L4 emotional burden. The intervention delivered 18 immersive scenarios reflecting common home-care tasks. RESULTS: Satisfaction was high in the intervention arm (≥90% positive on usefulness, relevance, and applicability). Level 2: the intervention group improved error recognition in video scenarios (mean identified errors 5.41 to 6.64; mean change + 1.23; P = 0.0001), with 46/70 (65.7%) showing improvement (χ = 33.114; p < 0.0001). Level 3: self-reported errors decreased in the intervention group (62 to 23) but increased in controls (46 to 77); the time-by-group interaction was significant (F = 11.53; P = 0.0009). Level 4: emotional burden shifted toward lower categories at follow-up in the intervention group (χ = 17.73; P = 0.0014). Complementary measures showed an increase in COM-B total score from 6.38 to 7.43 (P = 0.0017), with improvements in Opportunity (P = 0.0325) and positive trends in Capability and Motivation. CONCLUSIONS: A short, structured VR training improved recognition of unsafe practices and reduced self-reported caregiving/medication errors among informal caregivers, with concurrent reductions in emotional burden. Findings support integrating immersive, user-centered training into caregiver support programs to enhance the safety and quality of home care. CLINICALTRIALS: gov NCT05885334.
From prediction to practice: early implementation of a machine learning-based hospitalization prediction tool in the emergency department.
Raven W, de Hond A, Vermeire J … +6 more , Schinkelshoek L, Mulder L, van Someren A, Gaakeer M, de Jonge E, de Groot B
Int J Med Inform · 2026 Jul · PMID 42030768 · Publisher ↗
BACKGROUND: Machine learning (ML) models can accurately predict hospital admissions in emergency departments (EDs), but real-world adoption remains rare. We evaluated the feasibility and early implementation of an ML-bas...
BACKGROUND: Machine learning (ML) models can accurately predict hospital admissions in emergency departments (EDs), but real-world adoption remains rare. We evaluated the feasibility and early implementation of an ML-based hospitalization prediction tool in an ED of a tertiary care centre using the Medical Research Council (MRC) framework. METHODS: A prospective mixed-methods study was conducted at the ED of the Leiden University Medical Centre (∼23,000 annual visits). A validated ML-based hospitalization prediction tool was deployed via a web-based dashboard for four months. The dashboard was designed with stakeholders' input obtained during focus groups. Staff use, perceptions, and implementation barriers were assessed through pre-/post-implementation surveys with Likert-scales and open questions, and were analyzed using Mann-Whitney U-tests. Operational outcomes (ED length of stay [LOS], hospitalization rates) were extracted from the Netherlands Emergency department Evaluation Database and compared across pre- and post-implementation periods via logistic regression adjusted for confounders. RESULTS: Tool utilization was low despite an initial positive attitude towards the tool: 67% of staff reported rarely or never using it. Post-implementation surveys indicated a decline in perceived utility and reduced concern about AI replacing clinical roles. Reported barriers included lack of electronic health record integration, absence of linked actions, and misalignment with existing workflows. Hospitalization rose from 31.8% to 33.3% (p = 0.027), while ED-LOS increased during the implementation period (19.8% to 26.1%, p < 0.001); these changes could not be attributed to the tool given limited adoption. CONCLUSION: This early implementation study demonstrated low adoption of a ML prediction tool for hospitalization at the ED. Context-specific implementation barriers were identified. Guided by the MRC framework, the findings offer concrete strategies to enhance future implementation, including EHR integration, embedded action protocols, and role-specific responsibilities.
Diagnostic accuracy and deployment readiness of AI for burn depth assessment across imaging modalities: a PRISMA-DTA systematic review and meta-analysis.
Wang Y, Jiang Y, Zhang Y … +1 more , Liu S
Int J Med Inform · 2026 Jul · PMID 42025001 · Publisher ↗
BACKGROUND: AI-based burn depth assessment is rapidly emerging, yet evidence for diagnostic accuracy, generalizability, and deployment readiness remains unclear. METHODS: We performed a PRISMA-DTA-aligned systematic revi...
BACKGROUND: AI-based burn depth assessment is rapidly emerging, yet evidence for diagnostic accuracy, generalizability, and deployment readiness remains unclear. METHODS: We performed a PRISMA-DTA-aligned systematic review and meta-analysis. We synthesized accuracy outcomes and extracted deployment-relevant features, including validation design, reference-standard family, analytic unit, and subgroup reporting. Six studies (4,897 observations; mixed image-, wound-, and patient-level units) were included; four studies (N = 2541) formed the prespecified primary DTA dataset. We planned bivariate random-effects models where feasible and conducted exploratory subgroup analyses by imaging modality, skin tone, and age. Two additional studies were included only in exploratory analyses after prespecified approximate 2 × 2 reconstruction. RESULTS: In the primary dataset, the prespecified bivariate model did not converge because of sparse/extreme 2 × 2 patterns and limited study numbers; therefore, no pooled sensitivity or specificity was generated. Study-level sensitivity ranged from 0.50 to 1.00 and specificity from 0.54 to 0.97. In the expanded exploratory dataset (6 studies), pooled sensitivity was 0.924 (95% CI 0.788-0.975) and specificity 0.877 (95% CI 0.701-0.956), but these estimates are hypothesis-generating because they rely partly on approximately reconstructed 2 × 2 data. Exploratory descriptive analyses suggested possible modality-related variation and lower specificity in darker skin, although subgroup evidence was sparse and non-confirmatory; pediatric evidence was limited to a single within-study stratum. CONCLUSIONS: The evidence base is insufficient for deployment-ready use of AI burn depth assessment. The primary dataset did not support hierarchical pooling, and exploratory pooled estimates should be interpreted cautiously because they rely on reconstructed 2 × 2 data and mixed analytic units. More importantly, this review clarifies the evidence gaps separating promising algorithmic performance from deployable clinical decision support. Future studies should prioritize standardized reference standards, patient-level external validation in independent institutions, and reporting that supports safe integration into clinical workflows.
Architectural and translational perspectives on clinical decision support systems for rare disease diagnosis: a scoping review.
Özçetin E, Baş SE, Özpay F
Int J Med Inform · 2026 Jul · PMID 42025000 · Publisher ↗
BACKGROUND: Rare diseases remain difficult to diagnose because of phenotypic heterogeneity, limited clinical familiarity, and fragmented health data infrastructures. Clinical decision support systems (CDSS) have emerged...
BACKGROUND: Rare diseases remain difficult to diagnose because of phenotypic heterogeneity, limited clinical familiarity, and fragmented health data infrastructures. Clinical decision support systems (CDSS) have emerged as promising tools to support earlier recognition and more consistent diagnostic reasoning. However, the literature spans diverse technological paradigms, making it difficult to understand how these systems collectively contribute to clinical decision-making and their translational implementation. OBJECTIVE: This scoping review aimed to map diagnostic CDSS developed for rare disease diagnosis and to examine how data infrastructures, phenotype-driven reasoning frameworks, and artificial intelligence-based approaches contribute to clinical decision support and their translation into practice. METHODS: A PRISMA-ScR-guided scoping review was conducted. Searches were performed primarily in PubMed (MEDLINE), with supplementary screening in Google Scholar; the final search was completed on 30 November 2025. Records were screened in two stages and eligible studies were charted according to CDSS type, data sources, analytical methods, validation strategies, explainability features, interoperability elements, and reported evidence of clinical integration. Findings were synthesized using a taxonomy-based thematic approach rather than through quantitative pooling. RESULTS: The reviewed literature clustered into four main technological paradigms: information-retrieval systems, phenotype- and ontology-driven reasoning tools, data-driven predictive models based on EHRs and AI methods, and interoperable infrastructures such as federated learning and knowledge graphs. In addition, a separate group of studies addressed clinical evaluation and translation readiness across these paradigms. Across these areas, the field showed substantial methodological diversity, but evidence for external validation, workflow-level integration, and real-world clinical implementation remained limited. Interoperability, explainability, and governance were recurring challenges across paradigms. CONCLUSIONS: Rare disease CDSS research is moving from isolated diagnostic tools toward broader, interconnected diagnostic ecosystems. Progress toward clinically actionable implementation will depend on standardized data representations, stronger cross-institutional validation, explainable outputs aligned with clinical workflows, and interoperable infrastructures supported by appropriate governance. This review provides a taxonomy and conceptual framework to support the translational development of rare disease diagnostic CDSS.
Digitally shared medication lists - A scoping review.
Josendal AV, Bergmo TS, Clausen SS … +5 more , Hammar T, Lind KF, Manskow US, Vidas L, Bülow C
Int J Med Inform · 2026 Jul · PMID 42024999 · Publisher ↗
BACKGROUND: Access to accurate information about prescribed medications is crucial for effective patient care, particularly during care transitions, as discrepancies in medication records can lead to inappropriate prescr...
BACKGROUND: Access to accurate information about prescribed medications is crucial for effective patient care, particularly during care transitions, as discrepancies in medication records can lead to inappropriate prescribing and adverse drug reactions. In response, several countries are developing systems for digitally shared medication lists (SMLs) to improve medication management. This scoping review synthesizes existing evidence on SMLs, defined as systems that electronically transfer and aggregate real-time information about patients' current medications across provider organizations, excluding systems limited to one-to-one transfers. METHODS: A Preferred Reporting Items for Scoping Review (PRISMA) guided approach was followed. We searched CINAHL, EMBASE, PubMed, and Web of Science in January 2025. Pairs of reviewers independently screened and extracted data on characteristics, aims, types of systems, and outcomes related to SMLs. RESULTS: A total of 24 studies were included examining 11 distinct SML systems across eight countries, categorizing them into three main types: 1) National, centralized medication lists/databases; 2) Shared electronic health record (EHR) systems; and 3) Complementary or temporary solutions. The findings indicate that SMLs can enhance patient safety by providing healthcare professionals with more accurate medication information, enabling automated safety checks, and reducing discrepancies. However, challenges related to data quality, usability, and unclear responsibilities for maintaining SML accuracy remain. CONCLUSIONS: SML systems have been implemented in only a few countries and need better integration with clinical workflows to achieve desired effects. Future research should focus on the impact of patient engagement on medication safety, and the long-term effects of SMLs on clinical outcomes.
Decision support systems (DSS) for predicting hypertensive events using real-world telemonitoring data.
Kum HC, Tong CW, Lavu S … +6 more , Beaulieu JL, Hayek MA, Lawley MA, Erraguntla M, Stonebraker E, Mortazavi BJ
Int J Med Inform · 2026 Jul · PMID 42024998 · Publisher ↗
OBJECTIVE: This study contributes to the limited literature on applying machine learning (ML) to telemonitoring data for timely decision support systems (DSS). We propose an end-to-end DSS framework to enable clinicians...
OBJECTIVE: This study contributes to the limited literature on applying machine learning (ML) to telemonitoring data for timely decision support systems (DSS). We propose an end-to-end DSS framework to enable clinicians to intervene proactively by generating actionable information from daily telemonitoring data. The goal was to predict the probability of patients having an adverse event in the next 7 days using real-world data. We focus on the framework for effective development of a human-computer hybrid system that utilizes ML models to predict adverse hypertensive events and evaluate different ML models to determine optimal settings. Our goal is to create a system that enhances collaboration between ML and clinical expertise in hypertension management. MATERIALS AND METHODS: ML models (i.e., Logistic Regression, Random Forest (RF), XGBoost, Fusion Neural Network) were trained and evaluated with tenfold cross-validation, using AUCPR, AUCROC, and F1 score to study optimal configurations. In addition, different methods for calculating daily risk scores were compared. Finally, SHapley Additive exPlanations (SHAP) was used to investigate feature importance. RESULTS AND DISCUSSION: A total of 345,072 sliding windows from 2766 patients were used to study key questions for predicting hypertensive events. First, all ML models had comparable results at AUCPR ≃ 75% and AUCROC ≃ 87% with XGBoost having slightly better results. The optimal period of prediction to use as an input was 10 days, regardless of the ML models. Second, averaging the results across models gave the best performance for a single daily composite risk score. Finally, the key features observed in XGBoost and RF were daily Systolic Blood Pressure (SBP) readings, SBP variance, and Diastolic BP variance. However, key features may differ across different ML models. CONCLUSION: We demonstrate the need to report rigorous transparent details and evaluations to develop successful systems that can provide proactive alerts to manage adverse events.
Integrating actigraphy with demographic data enhances cognitive performance prediction: a multimodal UK biobank analysis using machine learning.
Ghiasi MM, Falck RS, Liu-Ambrose T … +2 more , Galea L, Tam RC
Int J Med Inform · 2026 Jul · PMID 42019230 · Publisher ↗
BACKGROUND: Wrist-worn actigraphy enables continuous, non-invasive measurement of rest-activity patterns and has growing relevance for cognitive health. However, it remains unclear whether actigraphy provides meaningful...
BACKGROUND: Wrist-worn actigraphy enables continuous, non-invasive measurement of rest-activity patterns and has growing relevance for cognitive health. However, it remains unclear whether actigraphy provides meaningful predictive value for cognitive function beyond demographic factors. Large-scale evaluations are needed to clarify its utility. OBJECTIVE: This study evaluated the effectiveness of 24-hour actigraphy in improving machine learning prediction of cognitive performance, using the computerized version of Digit Symbol Substitution Test (DSST). Complementary objectives included identifying influential predictors for explainability and assessing performance across sexes to evaluate fairness. METHODS: UK Biobank data included 24-hour actigraphy, demographics, and DSST scores. Random Forest (RF) and Extra Trees (ET) binary classifiers (low vs high DSST) were first trained using demographic features to establish baseline models. Multimodal models then integrated demographics with (1) raw actigraphy features and (2) sine-transformed actigraphy parameters capturing circadian rhythmicity. After preprocessing, the final sample comprised 42,707 participants. Performance was evaluated on a held-out test set (20%) using precision, recall, F1-score, ROC AUC, and PR AUC. RESULTS: The baseline demographic-only ET model demonstrated ROC AUC = 0.66 and PR AUC = 0.75. Adding raw actigraphy features improved ET performance (ROC AUC = 0.76; PR AUC = 0.83), outperforming all other models. In contrast, sine-transformed actigraphy features provided no additional benefit. Age, income, and education were the strongest predictors, while activity levels at 7:00-7:59 and 17:00-21:59 were the most informative actigraphy-derived features. Sex-stratified analyses showed slightly improved identification of lower cognitive performance in males and higher cognitive performance in females. CONCLUSIONS: This novel, large-scale study demonstrated that raw actigraphy features can enhance predictive performance of ET classifier beyond demographics alone. These results support the integration of wearable-derived activity data into population-level and personalized cognitive assessment frameworks. Observed sex-dependent performance differences further highlight the need for sex-aware modeling strategies.
Preoperative frailty for predicting in-hospital mortality in patients after cardiac surgery: an interpretable Machine learning model based on a retrospective multicenter cohort study.
Ge YN, Ouyang YM, Li QQ … +5 more , Huang J, Li D, Wang CL, Zhu YL, Wang JF
Int J Med Inform · 2026 Jul · PMID 42008965 · Publisher ↗
BACKGROUND AND OBJECTIVE: Accurate prediction of mortality risk after cardiac surgery is crucial for perioperative management. The laboratory-based frailty index (FI-lab) has been associated with adverse outcomes after s...
BACKGROUND AND OBJECTIVE: Accurate prediction of mortality risk after cardiac surgery is crucial for perioperative management. The laboratory-based frailty index (FI-lab) has been associated with adverse outcomes after surgery, but its predictive value for mortality after cardiac surgery remains to be validated. This study developed a novel interpretable machine learning model integrating FI-lab and routine clinical variables, with rigorous external validation using two independent critical care databases for reliable, clinically transparent risk stratification. METHODS: Data were obtained from the MIMIC-IV and eICU databases. FI-lab was used for frailty evaluation. Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to select predictive variables, and six machine learning models were developed. Synthetic Minority Oversampling Technique (SMOTE) addressed class imbalance. Model performance was evaluated by receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) values were used for model interpretation. RESULTS: The internal cohort included 7,955 patients, and the external validation cohort included 3,021 patients. LASSO regression identified 37 predictor variables. The Gradient Boosting model performed best, achieving an internal test set AUC of 0.900 (95% CI: 0.868-0.933), sensitivity of 0.844, and specificity of 0.800; external validation set AUC of 0.813 (95% CI: 0.768-0.857), sensitivity of 0.750, specificity of 0.760. SHAP analysis showed that chronic kidney disease (CKD), chronic heart failure (CHF), BUN, FI-lab score, and sex were the most important factors. DCA showed good clinical net benefit in internal validation and had potential application within clinically relevant low-threshold ranges in external validation. CONCLUSION: The FI-lab-based Gradient Boosting model allows for highly accurate prediction of in-hospital mortality risk in patients after cardiac surgery. This externally validated model provides robust evidence for individualized perioperative risk assessment.
Construction and interpretability evaluation of a prediction model for radiographic pneumonia in children at outpatient and emergency departments.
Donglei N, Yanhong S, Zhaoyang L … +2 more , Na L, Baojing W
Int J Med Inform · 2026 Jul · PMID 42000696 · Publisher ↗
BACKGROUND: Radiographic confirmation is crucial for pediatric pneumonia diagnosis, but chest X-ray overuse in outpatient and emergency settings raises concerns about unnecessary radiation and resource utilization. This...
BACKGROUND: Radiographic confirmation is crucial for pediatric pneumonia diagnosis, but chest X-ray overuse in outpatient and emergency settings raises concerns about unnecessary radiation and resource utilization. This study aimed to develop and validate a machine learning-based prediction model to assist in chest X-ray decision-making for children with suspected respiratory tract infection. METHODS: A retrospective cohort of 4,822 children with respiratory infections from Xuchang Central Hospital (Oct 2023-Jun 2024) was analyzed. Four feature selection methods identified key predictors, which were used to train an XGBoost model. External validation used an independent cohort. Model performance, interpretability, and clinical utility were assessed via ROC curves, SHAP analysis, and decision curve analysis. RESULTS: Four predictors were selected: hs-CRP, age, neutrophil count, and lymphocyte count. The model achieved an AUC of 0.773 in the test set (sensitivity 64.8%, specificity 74.6%, NPV 90.2%) and 0.751 in external validation. SHAP analysis highlighted hs-CRP and age 8-12 years as key contributors. Decision curve analysis showed superior net benefit at thresholds 0%-40%. A risk cutoff of 0.20 identified 56.5% of patients as low-risk. CONCLUSION: his XGBoost model, incorporating four objective variables, demonstrates good predictive performance for radiographic pneumonia in pediatric outpatient and emergency settings. Its high negative predictive value supports its use as a screening tool to reliably exclude pneumonia and reduce unnecessary chest X‑rays. The model has been deployed as a web-based tool for clinical use.
Predicting disintegration time in fast-disintegrating tablets using machine learning: a data-driven framework based on functional excipient representation.
Momeni M, Tabesh H
Int J Med Inform · 2026 Jul · PMID 41996748 · Publisher ↗
BACKGROUND: Fast-disintegrating tablets (FDTs) are widely used oral dosage forms in which disintegration time is a critical quality attribute influencing drug release and patient compliance. However, formulation developm...
BACKGROUND: Fast-disintegrating tablets (FDTs) are widely used oral dosage forms in which disintegration time is a critical quality attribute influencing drug release and patient compliance. However, formulation development is challenging due to complex and often non-linear interactions between excipient composition, physicochemical properties, and tablet characteristics. Conventional trial-and-error approaches are therefore time-consuming and inefficient. OBJECTIVE: This study aimed to develop a data-driven framework for predicting disintegration time in FDT formulations and to investigate how different representations of excipient information affect predictive performance. METHODS: A dataset of 1982 FDT formulations was analyzed using three alternative excipient representations: (i) identity-based encoding of excipients by chemical name and quantity, (ii) excipient-specific functional representation preserving both identity and functional role, and (iii) functionally aggregated representation summarizing quantities by excipient class. Regression models were used as the primary predictive approach to estimate continuous disintegration time. Classification models were additionally explored by discretizing disintegration time into formulation-relevant intervals. Model performance was evaluated using regression metrics and classification measures including weighted F1-score and the Matthews correlation coefficient (MCC). RESULTS: Deep neural networks achieved the highest predictive performance across all representations, with the best results obtained using the excipient-specific functional dataset (R = 0.86, MAE = 8.53 s). Random forest models also demonstrated stable performance. Functional excipient representation improved prediction compared with identity-based encoding by reducing dataset sparsity while preserving formulation-relevant information. CONCLUSION: Functional excipient representation provides an effective data abstraction strategy that enhances predictive modeling in pharmaceutical formulation datasets and supports more efficient data-driven decision making in early-stage FDT development.
Temporal Knowledge Discovery in Drug-Resistant Tuberculosis: A Decade-Long Machine Learning Analysis From Egyptian Clinical Data.
Elsayed H, Elkhwsky F, Amin W … +1 more , Abdelbaky I
Int J Med Inform · 2026 Jul · PMID 41996747 · Publisher ↗
BACKGROUND: Advanced predictive tools are required due to the worldwide persistence of tuberculosis (TB) and the growing threat of multidrug-resistant tuberculosis (MDR-TB). The rate and precision of traditional diagnost...
BACKGROUND: Advanced predictive tools are required due to the worldwide persistence of tuberculosis (TB) and the growing threat of multidrug-resistant tuberculosis (MDR-TB). The rate and precision of traditional diagnostic techniques frequently experience delays. This study uses machine learning (ML) to identify clinical and longitudinal treatment-history risk factors for drug-resistant tuberculosis (DR-TB) in Egypt using a ten-year dataset (2012-2022). METHODS: Three significant respiratory hospitals in Egypt participated in a retrospective case-control study. Demographics, clinical history, and lifestyle factors were examined in 1,462 patients (677 DR-TB cases and 785 DS-TB controls). Drug resistance was ascertained through phenotypic drug susceptibility testing (pDST) and, where applicable, rapid molecular assays including Xpert MTB/RIF. SHapley Additive exPlanations (SHAP) was used for feature selection. Stratified five-fold cross-validation on the training set, combined with evaluation on an independent held-out test set (20% of the total dataset), was used to build and validate four machine learning models: Random Forest, XGBoost, KNN, and Neural Networks. RESULTS: With 94.54% accuracy, 95.54% recall, and a ROC-AUC of 94.46%, the Random Forest model performed most effectively (cross-validation mean accuracy: 92.64% ± 0.50%) . With an accuracy of 93.17%, XGBoost was closely behind (cross-validation mean accuracy: 92.90% ± 1.98%). Longitudinal treatment-history variables (prior first-line drug use, patient category, and number of previous treatment episodes), geographic region (Governorate), and radiological infiltrations were the most influential predictors. CONCLUSION: Using longitudinal clinical data, ML models showed high efficacy in differentiating DR-TB from DS-TB. A strong framework for early DR-TB prediction is provided by the integration of treatment-history features and geographic risk factors with AI, which may optimize treatment initiation and resource allocation in high-burden settings.
Bias, fairness, and equity in artificial intelligence systems used in dental imaging: A systematic review.
Suganya P, Dupada P, Sruthi KG … +2 more , Panda P, Mohanty JR
Int J Med Inform · 2026 Jul · PMID 41962403 · Publisher ↗
BACKGROUND: Artificial intelligence (AI) is increasingly used in dental imaging for automated interpretation of dental images and as a clinical decision-support system. Although reported diagnostic accuracies are high, l...
BACKGROUND: Artificial intelligence (AI) is increasingly used in dental imaging for automated interpretation of dental images and as a clinical decision-support system. Although reported diagnostic accuracies are high, limited attention has been paid to bias, fairness, and equity within such AI-enabled dental systems. In this review, bias refers to systematic errors arising from non-representative or imbalanced training or test datasets; fairness refers to consistent and equitable AI model performance across diverse population groups; and equity refers to the provision of comparable diagnostic value across groups that differ in age, sex, ethnicity, dentition stage, or socioeconomic background. AIM AND OBJECTIVES: The aim of this systematic review was to assess original research on the use of AI in dental imaging, particularly with regard to diagnostic accuracy, methodological quality, and reporting on bias, fairness, and equity. The specific objectives were to: (1) assess diagnostic accuracy; (2) examine demographic reporting and subgroup analyses; (3) determine how bias, fairness, and equity are addressed in model development and validation; and (4) identify methodological priorities for more equitable dental imaging AI research. METHODS: A comprehensive literature search was conducted in accordance with PRISMA 2020 across PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar for studies published from January 2010 to March 2025. The review protocol has been registered in PROSPERO (CRD420261336733; registered 10 March 2026). Original research articles applying AI to dental imaging were included. Risk of bias was assessed using an adapted QUADAS-2 tool, and equity-related reporting was evaluated according to demographic description, dataset representativeness, and subgroup performance analysis. Narrative synthesis was undertaken because of heterogeneity in datasets, models, and outcome measures. RESULTS: Ten original studies met the inclusion criteria. All included studies used deep learning models applied to 2D dental imaging modalities such as panoramic radiographs, periapical radiographs, bitewing radiographs, and intraoral photographs; no eligible CBCT-based studies were identified. AI models demonstrated high diagnostic performance for tooth detection and tooth numbering, and encouraging results for caries detection, gingival assessment, plaque detection, and impacted tooth detection. Most studies demonstrated low methodological risk of bias. However, none of the included studies performed demographic subgroup analysis, only two reported limited demographic summaries, and most lacked sufficient reporting to support any meaningful equity assessment. Specifically, 0/10 studies conducted subgroup analysis, 2/10 provided partial demographic reporting, and 8/10 provided no meaningful demographic reporting. CONCLUSION: AI systems used in dental imaging show strong technical capability but lack adequate evaluation of bias, fairness, and equity. Future research should use representative, multi-centre datasets, report demographic characteristics transparently, and incorporate subgroup performance analysis to support fair and equitable clinical deployment. A fair validation procedure should include an independent test set from demographically diverse groups with subgroup-specific performance reporting, and a representative dataset should reflect the characteristics of the intended real-world clinical population.
Measuring digital health competence among healthcare professionals: A rapid review of assessment tools.
Robertson ST, Hoffman J, Sullivan C … +2 more , Donovan R, Woods L
Int J Med Inform · 2026 Jul · PMID 41962402 · Publisher ↗
INTRODUCTION: Healthcare professionals are increasingly expected to use digital health technologies in their clinical practice, despite limited prior education and training. Measuring digital health competence can assist...
INTRODUCTION: Healthcare professionals are increasingly expected to use digital health technologies in their clinical practice, despite limited prior education and training. Measuring digital health competence can assist organisations to understand workforce learning needs and tailor digital transformation efforts for maximum impact. This study aims to review the extant literature to answer the questions: what validated assessment tools are available to assess healthcare professionals' digital competence and what is the quality of evidence for these tools? METHODS: We used a criteria-led evaluation approach to 1) develop criteria for evaluation, 2) conduct a rapid literature review to identify tools and 3) evaluate tools against key criteria and determine their quality of evidence. We searched PubMed, CINAHL, Google scholar and grey literature up to April 22, 2025. The search strategy included three concepts: 'questionnaire OR survey' AND 'digital competence' AND 'healthcare staff'. Reporting followed PRISMA guidelines adapted for the rapid review methodology. RESULTS: Twenty-eight publications and grey literature met the inclusion criteria, with 61% published in the last 5 years. Most assessments designed for the healthcare workforce were nursing-specific (n = 9/20, 45%). Psychometric properties were reported for 71% of included instruments with varying quality of evidence. Only two tools met the criteria of being valid and scoped to the interprofessional healthcare workforce. CONCLUSION: Understanding which tools are validated and fit-for-purpose is essential for researchers, educators and health services seeking to measure and improve digital health competence among healthcare professionals. There is a need to expand research into interprofessional measures of healthcare workforce digital competence to support effective workforce transformation.
Machine learning for early detection and prediction of sepsis: explainability and key sepsis biomarkers representation-A systematic review.
Papapanagiotou I, Karalis A, Kokkoris S … +8 more , Vrettou CS, Kampouropoulou O, Giannopoulou V, Golemati S, Kotanidou A, Papapanagiotou S, Dimopoulou I, Vassiliou AG
Int J Med Inform · 2026 Jul · PMID 41962401 · Publisher ↗
OBJECTIVE: To systematically review machine learning-based sepsis prediction studies, examining model explainability and the extent to which explanations reflect key sepsis biomarkers. DATA SOURCES: Following the PRISMA...
OBJECTIVE: To systematically review machine learning-based sepsis prediction studies, examining model explainability and the extent to which explanations reflect key sepsis biomarkers. DATA SOURCES: Following the PRISMA guidelines, we reviewed the titles, abstracts, and full texts. The search was conducted in four major bibliographic databases with publication dates from January 1, 2019 to July 16, 2025. STUDY SELECTION: The included studies provided a clear definition of sepsis based on the Sepsis-3 criteria and involved critically ill adult human subjects. DATA EXTRACTION AND SYNTHESIS: Two authors (IP and AKa) independently reviewed and assessed each study. Using statistical methods, we assessed study quality and explainability trends. RESULTS: A total of 37 studies were included. Our analysis revealed a notable temporal increase (≈67% greater odds per year) in the use of explainability methods in sepsis prediction models. However, key sepsis biomarkers (procalcitonin or C-reactive protein) were not among the top predictive features, highlighting a gap between the model output and known sepsis pathophysiology. DISCUSSION: Model attributions often mirror what electronic health records measure most consistently (vital signs) rather than what is most biologically specific, partly due to the high missingness and irregular sampling of CRP/PCT in public datasets. Heterogeneity in feature selection and reliance on local datasets limit generalizability, while sparse code/data sharing constrains reproducibility. CONCLUSION: This review newly quantifies the rise of explainability use in sepsis prediction and identifies a consistent gap between model explanations and key sepsis biomarkers, providing a foundation for future work to bridge data-driven insights with sepsis pathophysiology. SYSTEMATIC REVIEW REGISTRATION NUMBER: CRD420251101470.
Challenges in translating AI-driven ASD/ADHD diagnosis: A methodological systematic review.
Rasool A, Ahmad F, Bunterngchit C … +1 more , Aslam S
Int J Med Inform · 2026 Jul · PMID 41955914 · Publisher ↗
BACKGROUND: Early and accurate diagnosis of neurodevelopmental disorders (NDDs), including autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD), remains a critical challenge in pediatric car...
BACKGROUND: Early and accurate diagnosis of neurodevelopmental disorders (NDDs), including autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD), remains a critical challenge in pediatric care. Traditional methods rely on subjective behavioral assessments that are time-intensive and prone to bias. OBJECTIVE: This systematic review synthesizes biomedical informatics methodologies using deep learning-driven computer vision to enable objective, data-driven diagnostic decision support for pediatric NDDs. METHODS: Following PRISMA guidelines, we searched Web of Science and Scopus (2020-2024), identifying 43 Q1/Q2 studies. Four informatics-focused research questions were addressed: multimodal feature extraction, deep learning architectures, high-performing strategies, and robust data integration challenges. Methodological quality and bias were assessed using the APPRAISE-AI framework. RESULTS: Multimodal fusion and hybrid informatics pipelines dominated (38% of studies), outperforming unimodal approaches by integrating complementary streams-facial imaging (high specificity), EEG/fMRI (superior sensitivity). Transfer learning and fusion techniques were prevalent, but federated learning and explainable AI were underutilized. APPRAISE-AI revealed strong clinical relevance (72.8%) and reporting quality (66.1%), yet substantial gaps in reproducibility (41.0%) and result robustness (45.1%). CONCLUSIONS: AI-driven biomedical informatics holds significant potential to reduce diagnostic delays and costs in NDDs. However, reproducibility, interpretability, and ethical data integration must be improved through standardized, privacy-preserving, and auditable frameworks to enable scalable clinical deployment.
Integration of deep learning and radiomic features from multiplex immunohistochemistry images for reproducible Multi-Outcome prediction in a Multi-Center study of colorectal cancer.
Yin Y, Sun Z, Deng X … +1 more , Fan Q
Int J Med Inform · 2026 Jul · PMID 41955913 · Publisher ↗
OBJECTIVE: To develop and validate a robust, multimodal machine learning framework integrating radiomic and deep learning features from multiplex immunohistochemistry (mIHC) images for comprehensive outcome prediction in...
OBJECTIVE: To develop and validate a robust, multimodal machine learning framework integrating radiomic and deep learning features from multiplex immunohistochemistry (mIHC) images for comprehensive outcome prediction in colorectal cancer (CRC). MATERIALS AND METHODS: This multi-institutional retrospective study included 2,117 CRC patients from seven centers, with 1,548 cases used for model training and internal testing, and 569 for external validation. mIHC-stained whole-slide images targeting six immune markers (CD3, CD8, CD45RO, PD-1, LAG-3, Tim-3) were analyzed from two spatial compartments: tumor center and invasive margin. Radiomic features (n = 71/region/marker) were extracted using HistomicsTK, while 768-dimensional deep features were derived using a pre-trained Vision Transformer (ViT-B/16). Feature robustness across biomarkers was quantified via intraclass correlation coefficients (ICC ≥ 0.75). Selected features underwent multi-step selection (LASSO, MI, RFE) and were fused into a single feature space, followed by PCA-based dimensionality reduction. Five clinical tasks were modeled: tumor recurrence, survival status, overall survival duration, TNM staging, and immune profile classification. Classification models (TabTransformer, XGBoost, TabNet) and survival models (DeepSurv, CoxPH, RSF) were trained using 5-fold cross-validation and tested on independent cohorts. RESULTS: Fused features significantly outperformed individual modalities across all tasks. TabTransformer with LASSO-selected fused features achieved top performance: recurrence (AUC = 95.9%), survival status (AUC = 94.5%), TNM staging (macro-AUC = 91.0%), and immune profile (macro-AUC = 91.0%). For survival regression, DeepSurv achieved a C-index of 0.82 and time-dependent AUC of 0.82. Models exhibited strong generalizability, with negligible performance drop on external datasets. SHAP analysis confirmed feature interpretability, with fused features contributing the most across tasks. CONCLUSIONS: This study demonstrates that fused mIHC-derived radiomic and deep features yield accurate, interpretable, and generalizable predictions for multiple CRC outcomes, supporting their integration into precision oncology workflows.
About
- Frequency
- Sun
- Papers found
- 200
- RSS feed
- Subscribe