Sheng-Feng Sung

Background: Retinal vascular occlusions (RVaO) are signs of systemic vascular disease and share risk factors with stroke and myocardial infarction. The cardiovascular risk of RVaO subtypes and their relationship to transient ischemic attack (TIA) remain unclear. Methods: Using the TriNetX database, we compared the 1‐year incidence of major adverse cardiovascular events (MACE), including ischemic and hemorrhagic stroke, acute myocardial infarction, and death, between patients with RVaO and those with TIA. Cox regression was used to analyze the association between risk factors and MACE risk. Results: We identified 91 936 patients with RVaO and 392 069 patients with TIA. After propensity score matching, each cohort included 90 698 patients. The RVaO cohort had a significantly lower 1‐year MACE risk than the TIA cohort (hazard ratio [HR], 0.77 [95% CI, 0.74–0.80]), with reduced risks of ischemic stroke (HR, 0.64 [95% CI, 0.60–0.67]), hemorrhagic stroke (HR, 0.82 [95% CI, 0.71–0.94]), and acute myocardial infarction (HR, 0.89 [95% CI, 0.83–0.96]). All‐cause mortality was similar. Subtype‐specific analysis showed that patients with central retinal artery occlusion had higher risks of MACE (HR, 1.13 [95% CI, 1.02–1.24]) and all‐cause mortality (HR, 1.23 [95% CI, 1.02–1.49]) than patients with TIA. Within RVaO, age, male sex, retinal artery occlusion, ischemic heart disease, chronic kidney disease, heart failure, and nicotine dependence were important predictors of MACE risk. Conclusions: Patients with RVaO have a lower 1‐year MACE risk than TIA, but this varies by subtype. central retinal artery occlusion has a prognosis similar to TIA, requiring urgent assessment and prevention. These findings highlight the importance of multidisciplinary care in RVaO management.

Objective: This study was undertaken to evaluate the predictive, stratification, and prognostic utility of polygenic risk scores (PRSs) for epilepsy in three Taiwanese cohorts comprising more than 7000 cases among more than 600 000 individuals, thereby addressing the limited evidence on PRS performance in non‐European populations. Methods: Leveraging results from the latest multiancestry genome‐wide association study (GWAS), we calculated PRS for all epilepsy, genetic generalized epilepsy (GGE), and focal epilepsy (FE) in two population cohorts—Taiwan Biobank–National Health Insurance Research Database (TWB‐NHIRD; N ~ 105 K) and Taiwan Precision Medicine Initiative (TPMI; N ~ 500 K), with cases defined by International Classification of Diseases codes and medications—and a clinically ascertained sample (Taiwanese subset of the Epi25 consortium [Epi25‐TWN]; N = 1140). PRS performance was assessed through lifetime risk models, age‐at‐onset stratifications, and phenome‐wide association studies (PheWASs). Results: PRS performance varied substantially by cohort, epilepsy type, and age at onset. GGE PRS had the strongest effect on GGE in Epi25‐TWN (per‐unit odds ratio [OR] = 1.65 [95% confidence interval (CI) = 1.24–2.23], p = 8.3 × 10 ⁻⁴ , R ² = 3.2%) but a weaker effect in TWB‐NHIRD and TPMI (ORs = 1.09–1.13 [95% CI = .98–1.23], p = .002–.057, R ² = .14%–.22%). FE PRS effects on FE were more consistent but modest (ORs = 1.02–1.16 [95% CI = .97–1.33]). No prognostic value of PRS was found in predicting epilepsy from syncope or unspecified seizures. Individuals in the top 5% GGE PRS had >4‐fold increased risk in Epi25‐TWN but <1.5‐fold in the others. Restricting to earlier onset cases strengthened PRS signals in population cohorts, particularly for GGE, with PheWAS analysis revealing epilepsy‐specific associations. PRS effects were generally comparable but attenuated relative to European studies. Significance: Cohort design and phenotyping accuracy strongly influence the assessment of the predictive value of epilepsy PRS, highlighting caution in clinical interpretation. Incorporating richer phenotypic data and more diverse GWASs will be crucial to enhancing its applicability and clinical potential.

Postoperative pneumonia is a significant complication, highlighting a patient's ongoing vulnerability. While traditional tools focus on short-term outcomes, the perioperative period offers a unique "stress test" window to identify high-risk patients. This study developed and validated a machine-learning-based prognostic framework to predict pneumonia risk up to one year after surgery. This retrospective study examined 11,655 surgical encounters at a tertiary hospital. Multiple machine learning algorithms, including random forest (RF), extreme gradient boosting, support vector machine, multilayer perceptron, and penalized logistic regression, were compared using 5-fold cross-validation. Class imbalance was handled using random oversampling (ROS) and undersampling. Models were tested on a separate set, and Shapley additive explanation (SHAP) analysis identified key predictors to improve clinical understanding. Postoperative pneumonia occurred in 238 encounters (2.04%) within 365 days, peaking in the second postoperative month. The RF model with ROS (1:4 ratio) achieved the highest performance with an area under the receiver operating characteristic curve of 0.886, sensitivity of 85.4%, specificity of 77.4%, positive predictive value of 7.4%, and negative predictive value of 99.6%. SHAP analysis identified preoperative hemoglobin, European Society of Cardiology surgical risk, age, American Society of Anesthesiologists Physical Status class, and estimated glomerular filtration rate as key predictors of long-term vulnerability. Machine learning facilitates prognostic stratification of patients at high risk of long-term vulnerability. By functioning as a high-sensitivity secondary screening tool, this model allows clinicians to safely "rule out" low-risk individuals and concentrate intensive surveillance and resources on the high-risk cohort, thereby improving long-term outcomes.

Background: Stroke remains one of the leading causes of mortality and long-term disability worldwide. Atrial fibrillation (AF) is a major and often underdiagnosed risk factor for ischemic stroke as it is frequently asymptomatic and may remain undetected until a catastrophic cerebrovascular event occurs. The lack of timely identification and preventive treatment for AF substantially increases stroke risk. Although previous studies have proposed various predictive models for AF detection, many rely primarily on structured clinical variables and are developed using data from a single institution, which limits their generalizability and real-world applicability across different health care settings. Objective: The objective of this study was to develop a robust and generalizable AF risk prediction model for patients with stroke using electronic medical records. By integrating structured clinical variables with features derived from unstructured clinical text, this study aimed to construct a more comprehensive representation of patient health status. Furthermore, this study emphasized systematic internal and external validation, along with calibration assessment, to evaluate model stability and generalizability across multiple hospital datasets, thereby supporting its potential use in routine clinical practice. Methods: This study analyzed datasets from 2 hospitals in Taiwan: Landseed International Hospital (LIH), with 3988 patients, and Chia-Yi Christian Hospital (CYCH), with 5821 patients. We applied 5 feature engineering techniques to extract features from unstructured electronic medical record data, addressed data imbalance using 6 distinct resampling methods, and used 9 classification algorithms to compare model performance across both internal and external validation sets. This study identified the top 20 most important features from the best-performing models for both the LIH and CYCH datasets. Results: The optimal predictive model for LIH was based solely on structured variables, whereas the model for CYCH achieved superior results by integrating structured variables with text-derived variables obtained from unstructured clinical notes using term frequency–inverse document frequency. Notably, feature importance analysis consistently identified the ratio of E- to A-wave velocities, left atrial size, and age as the top 3 predictive factors across both datasets, underscoring their critical role in AF risk assessment among patients with stroke. Conclusions: This study demonstrated the development of predictive models for AF in patients with ischemic stroke. Notably, the integration of structured variables with variables derived from unstructured clinical text improved predictive performance in selected model configurations. Rigorous internal and external validation processes confirmed the superior performance of ensemble learning–based machine learning models compared with alternative algorithms, underscoring the potential of this approach for AF risk prediction.

Introduction: Acute dizziness accounts for approximately 4% of emergency department (ED) visits, with stroke often missed. Current methods for stroke detection in dizzy patients have notable limitations, with vestibular strokes missed in a substantial proportion of ED visits. This study aimed to develop a machine learning (ML) tool to assess stroke risk in patients with acute dizziness. Methods: We developed an ensemble model combining four ML algorithms using structured electronic medical record data and unstructured ED physician notes. Model performance was evaluated on a holdout test set and compared with the ABCD ² score using area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI), integrated discrimination improvement (IDI), and decision curve analysis. Results: The ensemble model achieved the highest AUC at 0.880, significantly outperforming the ABCD ² score (AUC 0.673) and individual ML models. The ensemble model demonstrated superior calibration with the lowest Brier score and showed greater clinical utility across different risk thresholds. Features extracted from unstructured clinical text substantially enhanced model performance, with models combining structured and unstructured data consistently outperforming those trained on structured data alone. Conclusions: Our ensemble prediction model effectively stratifies stroke risk in ED patients with acute dizziness. By integrating natural language processing of clinical notes with structured patient data, the model offers a more accurate risk assessment than traditional methods. The implementation of this tool could improve patient outcomes by directing advanced neuroimaging to high-risk patients while avoiding unnecessary testing in low-risk patients, ultimately enhancing patient safety and optimizing resource utilization.

Background: Clinical trial eligibility screening using electronic medical records (EMRs) is challenging due to the complexity of patient data and the varied clinical terminologies. Manual screening is time-consuming, requires specialized knowledge, and can lead to inconsistent participant selection, potentially compromising patient safety and research outcomes. This is critical in time-sensitive conditions like acute ischemic stroke. While computerized clinical decision support tools offer solutions, most require software engineering expertise to update, limiting their practical utility when eligibility criteria change. Objective: We developed and evaluated the intelligent trial eligibility screening tool (iTEST), which combines natural language processing with a block-based visual programming interface designed to enable clinicians to create and modify eligibility screening rules independently. In this study, we assessed iTEST’s rule evaluation module using pre-configured rules and compared its effectiveness with that of standard EMR interfaces. Methods: We conducted an experiment at a tertiary teaching hospital in Taiwan with 12 clinicians using a 2-period crossover design. The clinicians assessed the eligibility of 4 patients with stroke for 2 clinical trials using both standard EMR and iTEST in a counterbalanced order, resulting in 48 evaluation scenarios. The iTEST comprised a rule authoring module using Google Blockly and a rule evaluation module utilizing MetaMap Lite for extracting medical concepts from unstructured EMR documents and structured laboratory data. Primary outcomes included accuracy in determining eligibility. Secondary outcomes measured task completion time, cognitive workload using the National Aeronautics and Space Administration Task Load Index scale (range 0‐100, with lower scores indicating a lower cognitive workload), and system usability through the system usability scale (range: 0‐100, with higher scores indicating higher system usability). Results: The iTEST significantly improved accuracy scores (from 0.91 to 1.00, P <.001) and reduced completion time (from 3.18 to 2.44 min, P =.004) compared to the standard EMR interface. Users reported lower cognitive workload (National Aeronautics and Space Administration Task Load Index scale, 39.7 vs 62.8, P =.02) and higher system usability scale scores (71.3 vs 46.3, P =.01) with the iTEST. Particularly notable improvements in perceived cognitive workload were observed in temporal demand, effort, and frustration levels. Conclusions: The iTEST demonstrated superior performance in clinical trial eligibility screening, delivering improved accuracy, reduced completion time, lower cognitive workload, and better usability when evaluating preconfigured eligibility rules. The improved accuracy is critical for patient safety, as the misidentification of eligibility criteria could expose patients to inappropriate treatments or exclude them from beneficial trials. The adaptability and ability of the iTEST to process both structured and unstructured data make it valuable for time-sensitive scenarios and evolving research protocols. Future research should evaluate clinicians’ ability to create and modify eligibility rules using the block-based authoring interface, as well as assess the iTEST across diverse types of clinical trials and health care settings.

Background: Taiwan’s National Health Insurance Research Database (NHIRD) has evolved into a cornerstone of real-world evidence generation. As Taiwan’s National Health Insurance program reaches its 30th anniversary, a comprehensive reassessment of the NHIRD’s development, challenges and future directions is warranted. Objective: To provide an updated review of the NHIRD. Methods: We conducted a narrative review of Taiwan’s NHIRD, synthesizing published studies, government reports, and policy documents from 2019 through 2024. We summarized developments related to database infrastructure, data collection, validation studies, linkage strategies, governance reforms and interoperability initiatives, with a particular focus on their implications for real-world evidence generation and AI-driven research. Results: The NHIRD has additionally incorporated structured laboratory results and medical imaging data, significantly broadening its research capabilities. Validation studies have demonstrated the reliability of International Classification of Diseases, 10th Revision, Clinical Modification codes across various conditions, reinforcing the database’s applicability for epidemiological research. Integration efforts with national registries, surveys and electronic medical records have further enhanced the depth and accuracy of clinical outcome measurements. Nonetheless, critical challenges persist, including data standardization inconsistencies, cybersecurity vulnerabilities, and heightened scrutiny following constitutional court rulings on data governance and privacy rights. In response, Taiwan’s Ministry of Health and Welfare has launched initiatives to address these concerns, notably through the development of a Fast Healthcare Interoperability Resources (FHIR)-based data infrastructure aimed at improving interoperability, data security, and artificial intelligence (AI)-readiness to balance ethical governance with scientific innovation. Conclusion: The NHIRD’s transformation over three decades underscores the importance of continuous investment in data quality, privacy protection, and interoperability. With sustained reforms, the NHIRD will be poised to remain a leading resource for real-world evidence generation and to contribute meaningfully to global health and digital medicine.

Background and Aims: Anticholinergic burden has been linked to cardiovascular events and mortality, but its association with out-of-hospital cardiac arrest (OHCA) remains unclear. This study investigated whether transient increases in anticholinergic burden elevate OHCA risk in middle-aged and older populations. Methods: A nationwide case-crossover, case-time-control, and case-case-time-control study using Taiwan’s National Health Insurance Research Database involved 173 974 adults aged ≥40 years who experienced OHCA between 2011 and 2021. The participants were middle-aged (40–64 years) or older (≥65 years). Anticholinergic burden was assessed using established scales and categorized as 0, 1–2, and 3 or more points. Each participant’s burden during the hazard period (days −30 to −1 before OHCA) was compared to a randomly selected 30-day reference period (from days −180 to −61). Conditional logistic regression was used to estimate odds ratios (ORs) with 95% confidence intervals (CIs). Results: Case-crossover analysis showed more patients having higher anticholinergic burden during the hazard period than during the reference period in both age groups. The ORs for OHCA were 1.48 (95% CI: 1.42–1.55) and 1.56 (95% CI: 1.52–1.61) for scores of 1–2, compared with 0, and 2.03 (95% CI: 1.95–2.11) and 2.21 (95% CI: 2.15–2.27) for scores of 3 or more, compared to 0, in the middle-aged and older groups, respectively. Case-time-control and case-case-time-control analyses consistently showed a dose–response relationship, with results confirmed by sensitivity analyses. Conclusions: Transient increases in anticholinergic burden significantly raise OHCA risk, particularly among individuals with higher burden levels.