BMC Medical Informatics and Decision Making最新文献

Background: Synthetic data generation (SDG) has emerged as a critical enabler for data-driven healthcare research, offering privacy-preserving alternatives to real patient data. Temporal health data - ranging from physiological signals to electronic health records (EHRs) - pose unique challenges for SDG due to their complexity, irregularity, and clinical sensitivity.

Objective: This review systematically examines SDG methods for longitudinal and time-series health data. Its aims are to (1) propose a lightweight taxonomy to support orientation across the SDG landscape along five structural dimensions, (2) characterize the major synthesis techniques and their alignment with temporal structures and data modalities, and (3) synthesize the utility and privacy evaluation strategies used in practice.

Methods: A systematic literature review was conducted following PRISMA guidelines across four major databases (ACM, arXiv, IEEE Xplore, Europe PMC) for publications from 2017 to 2025. Eligible studies proposed or applied SDG techniques to healthcare-relevant temporal data with sufficient methodological transparency. Structured data extraction and thematic analysis were used to identify modeling trends, evaluation metrics, and domain-specific requirements, complemented by a comparative synthesis of SDG methods.

Results: A total of 115 studies were included. Deep generative models - especially Generative Adversarial Networks (GANs), Autoencoders (AEs), and diffusion-based methods - dominate the field, with increasing adoption of autoregressive and hybrid simulation approaches. Event-based EHR data are most commonly targeted, while continuous and irregular time series remain underexplored. Utility evaluations vary widely, with strong emphasis on descriptive statistics and predictive performance, but limited attention to inferential validity and clinical realism. Privacy assessments are sparse and inconsistently reported: only 30% of studies included any metric, and just around 6% implemented differential privacy (DP), often without parameter disclosure. This limited adoption may reflect technical challenges, limited expertise, and the absence of regulatory incentives.

Conclusions: Synthetic temporal data play an increasingly vital role across clinical prediction, public health modeling, and Artificial Intelligence (AI) development. However, SDG research remains fragmented in terminology, evaluation practices, and privacy safeguards. Responsible-AI considerations - such as fairness, transparency, and trust - along with evidence on clinical adoption remain underexplored but are critical for future integration. This review provides a unified conceptual and methodological framework to guide future research, standardization efforts, and interdisciplinary collaboration for responsible, effective use of synthetic health data.

Background: Clinician burnout is a growing global concern, with heavy clinical documentation workload identified as a major contributor. Clinical documentation tasks, though essential for patient care and communication, are time-consuming and cognitively demanding. Recent advances in artificial intelligence (AI), particularly natural language processing and large language models, are being explored as potential tools to alleviate documentation burden, yet their quantitative impact has not been systematically assessed.

Methods: We performed a systematic review and meta-analysis, registered on PROSPERO (CRD420250653291) and guided by PRISMA. Eligible studies included frontline health professionals using AI tools for clinical note creation, with comparators being usual practice or pre-implementation baseline. Primary outcomes were documentation burden, workload, burnout, and time spent on documentation. Searches were conducted in PubMed, Web of Science, Scopus, and key journals. Effect sizes were synthesized using standardized mean difference (SMD) under a random-effects model, with subgroup analyses by study design, AI tool type, task type, editing status, and data origin.

Results: Of the 23 studies included, 12 were non-randomised studies with a concurrent control and 11 employed a before-and-after comparison design. The study participants varied in specialties and were mainly from ambulatory settings, including physicians, surgeons, pediatricians, and ICU specialists. Heterogeneity in results across included studies was considerable, and the methodological quality of the available studies was generally low. Pooling results of the 14 studies yielded an overall standardized mean difference (SMD) of -0.71 (95% confidence interval [CI]: -0.93 to -0.49), indicating a moderate reduction in documentation workload and related burnout. Based on results of studies in which clinicians reviewed and edited AI-generated drafts, AI applications reduced documentation time, similarly representing a moderate effect size (SMD= -0.72, 95% CI -0.99 to -0.45). The quality of notes generated by AI tools was at least comparable to those prepared manually by clinicians.

Conclusions: AI technologies offer promising benefits for reducing clinical documentation burden. However, their implementation must be accompanied by rigorous quality control and ongoing evaluation in practical settings to optimize their effectiveness and safeguard patient care outcomes.

Purpose: Congenital hypothyroidism (CH) is a common cause of severe intellectual disability, affecting approximately 1 in 2,000 newborns globally. Treatable with early intervention, congenital hypothyroidism has long been a target of newborn screening programs. Current thyroid stimulating hormone (TSH) based programs suffer from low positive predictive value, resulting in unnecessary diagnostic investigations. Congenital hypothyroidism screening has proven challenging for machine learning previously due to massive class imbalance and having a single well known predictor, preventing acceptable screening sensitivity. This study represents the most comprehensive evaluation of machine learning for congenital hypothyroidism screening to date.

Methods: Analyzing data from 616,910 infants screened by Newborn Screening Ontario between 2019 and 2024. 12 classification and 12 resampling algorithms were trained using 4 different optimization metrics, for a total of 576 distinct models evaluated using stratified 5-fold cross-validation to ensure robustness. Models were optimized for sensitivity and then positive predictive value using various metrics. Model explainability was assessed using SHAP values and feature importances.

Results: We were able to create a model achieving 16.8% PPV while maintaining 100% sensitivity using a RUSBoost classifier and Gaussian Noise resampling. This represents a 60% improvement in positive predictive value over the current approach. TSH remained the dominant predictor as in current screening, but our model was able to include minor amounts of additional information from other features to improve performance.

Conclusion: These machine learning algorithms show no missed cases of CH and are able to significantly improve performance across robust testing. The findings suggest that machine learning offers a promising avenue for refining TSH-based CH screening processes, reducing false positives, and alleviating unnecessary stress and costs associated with current methods used by the majority of newborn screening programs globally.

Background: Patient-reported symptomatic adverse events (AE) are increasingly collected in oncology clinical trials to characterize treatment tolerability and inform clinical decision making using the Patient-Reported Outcomes (PRO) version of the Common Terminology Criteria for Adverse Events (PRO-CTCAE®). Although there are numerous analysis methods and graphical approaches used for PRO-CTCAE data, the current PRO literature is limited in unified reporting and graphical approaches as well as public-facing analysis tools.

Results: Collaborative efforts from the Standardization Working Group of the National Cancer Institute Cancer Treatment Tolerability Consortium worked to develop the R package, ProAE. Testing and validation of widely used methods were implemented in the R package and deployed to various open-source outlets including the Comprehensive R Archive Network (CRAN).

Conclusion: ProAE is a free and publicly available collection of standardized statistical analysis tools for PRO-CTCAE and other PRO data used in patient care and research. The ProAE package provides oncology researchers with an efficient and modern means to apply the published analysis approaches, including hypothesis testing, descriptive and inferential tables, and longitudinal graphics, without the need for costly software or licensing.