Introduction: South Korea operates the National Oral Health Screening Program (NOHSP) biennially. The program primarily focuses on screening for dental caries and does not sufficiently include oral mucosal examinations for oral cancer screening. Although a few studies have attempted to evaluate cancer prevention effects of the program, evidence remains limited.

Aims: We aimed to assess the effectiveness of regular NOHSP screening in reducing the incidence of oral cancer. Based on a sequential trial emulation, we sought to adequately control for baseline differences between cohorts.

Methods: From a nationwide screening-based claims database in South Korea, we identified participants aged 40-79 with no oral cancer history from 2010 to 2014. They were followed up to 14 years, until deviation from the plan, censoring, death, or oral cancer occurrence, whichever came first. Propensity scores were matched between the screened (≥3 times within five-year grace period) and non-screened groups. We estimated hazard ratios (HRs) and their 95% confidence intervals (95%CIs), considering individuals as a cluster unit, using two approaches: pooled Cox models with or without stratification by trial; and a weighted Cox model after clone-censor-weighting (CCW).

Results: The matched dataset consisted of 22,398 pairs, with no significant differences observed in baseline characteristics between the two groups (all P>0.1). The median follow-up times were 11.45 years for the screened group and 9.03 years for the non-screened group. The pooled Cox model without trial stratification yielded an HR of 0.867 (95%CI=0.552-1.360). When considering trials as strata, the HR was 0.866 (95%CI=0.552-1.359). The HR estimated using the CCW approach was 0.833 (95%CI=0.468-1.482).

Conclusions: Regular NOHSP screening was associated with a lower estimated incidence of oral cancer, although the association was not statistically significant. Further improvements are needed to enhance early detection across a broader range of oral conditions.

Keywords: Sequential target trial; Screening program; Cancer prevention

Introduction: Non-adherence to lipid-lowering therapy is common and contributes to preventable cardiovascular morbidity and mortality. Standard adherence measures summarise medication coverage but not how patients disengage over time. We developed BRIDGE, a barrier-informed Bayesian model, to predict adherence trajectories and identify plausible intervention targets.

Aims: The overarching aim of BRIDGE is to provide point of care decision support to clinicians to facilitate better adherence to cardioprotective medicines and prevent future CVD events.

Methods: This nationwide, multicentre model-development and prospective external-validation study used de-identified primary-care records from IQVIA-network general practices. Incident lipid-lowering therapy initiators were identified using a 365-day pre-index washout. Adherence trajectories were derived from prescription records using group-based trajectory modelling. BRIDGE used a monotone-spline multinomial-softmax architecture over 108 routinely available features, with World Health Organization-aligned hierarchical priors informed by patient-reported barriers. The model was developed in 2014-2017 initiators and externally validated in the 2018-2020 cohort. Generalisability was assessed using temporal validation in later initiators and geographic-temporal validation excluding Queensland, Victoria, and Western Australia from training. BRIDGE was benchmarked against random forest and XGBoost.

Results: The development cohort included 50,857 adults. Four trajectories were identified: gradual decline (10.4%), early discontinuation (40.6%), rapid decline (28.8%), and persistent adherence (20.2%). In temporal validation, BRIDGE showed stronger discrimination than XGBoost and random forest (AUROC 0.723 [0.720-0.726]) and matched the best accuracy (0.550 [0.544-0.557]). In geographic-temporal validation, BRIDGE led across discrimination, accuracy, and calibration, with AUROC 0.721 (0.716-0.725), accuracy 0.558 (0.549-0.568), and Brier score 0.618 (0.606-0.629). Counterfactual analyses suggested trajectory-specific intervention priorities.

Conclusions: BRIDGE provides interpretable, behaviour-informed prediction of lipid-lowering therapy adherence trajectories from routine primary-care records, identifying different modes of preventable cardiovascular treatment failure early enough to support tailored follow-up, medication review, and shared decision making.

Introduction: The clone-censor-weight (CCW) method is increasingly applied in target trial emulation (TTE) to mitigate immortal time bias. However, inconsistencies in its methodological implementation pose challenges for valid causal inference using real-world data.

Aims: We systematically mapped current CCW practices in TTE studies to identify key methodological gaps throughout the core steps of CCW implementation.

Methods: We searched Medline, Embase, Cochrane Library, and Web of Science for observational studies (2016–2025) utilizing CCW. Two independent reviewers extracted 38 methodological items per study. Descriptive statistics were used to summarize the frequencies of key methodological features.

Results: The final analysis included 121 eligible studies, predominantly within infectious diseases (27.3%), oncology (19.8%), and cardiovascular diseases (12.4%). Drawing on key findings from the 38 extracted methodological items, we describe core CCW implementation practices and identify several critical methodological gaps. Clones per individual ranged from 2 to 81 (median: 2); grace periods ranged from 1 to 365 days (median: 56). Only 39 studies (32.2%) conducted sensitivity analyses on the grace period length, and 72 (59.5%) provided a clinical or literature-based justification. Artificial censoring weights were primarily estimated via pooled logistic regression (43.8%), Cox regression (22.3%), or standard logistic regression (15.7%). Notably, 54 studies (44.6%) failed to account for time-varying covariates during artificial censoring step. To handle extreme weights, 56 studies (46.3%) used truncation, 14 (11.6%) used only stabilization, and 51 (42.1%) reported no method. For confidence interval estimation, 88 studies (73.3%) utilized non-parametric bootstrapping, while 27 (22.5%) failed to report corrections for cloning-induced clustering.

Conclusions: Despite the growing adoption of CCW, significant methodological gaps persist. Infrequent sensitivity analyses on grace periods, the failure to include time-varying covariates in censoring models, and underreporting of clustering corrections threaten real-world causal inference. Standardized CCW reporting guidelines are urgently needed.

Keywords: clone-censor-weight, target trial emulation, immortal time bias, methodological review

Background
Per International Council for Harmonisation (ICH) M14, use of real-world data for decision-making requires clear rationale and reproducible identification of the study population. This is challenging in thyroid eye disease (TED), where diagnosis, disease activity, and severity are variably captured due to both disease complexity and differences in healthcare practice and structured documentation across health care systems.
Objectives
To operationalize TED identification in Taiwan using a patient journey-based electronic health record (EHR) approach.
Methods
Guided by a patient journey-based approach, we identified individuals with Graves’ disease and ophthalmology visits in the Chang Gung Research Database (CGRD) between 2016 and 2024, reflecting expected eye-care visits among TED patients. Structured data extraction from the claims-based module of the EHR and review of free-text physician’s notes in the electronic medical charts were conducted on a random sample of 500 eligible patients to identify physician-confirmed TED cases (the gold standard) and to evaluate the implications of relying on structured data alone. We developed candidate algorithms based on diagnostic codes, examination codes, or their combinations using the structure data. We estimated sensitivity, specificity, positive predictive value (PPV), and F-score, a harmonic mean between sensitivity and PPV, against the physician-confirmed TED diagnosis.
Results
Chart review confirmed 102 (20.4%) TED cases. The performance (i.e. sensitivity, specificity, PPV and F-Score) of selected candidate algorithms were: diagnosis-based algorithm (proptosis or diplopia or eyelid retraction): 56.9%, 95.0%, 74.4% and 64.4%; examination-based algorithm using Hertel examination or Hess chart test: 79.4%, 94.0%, 77.1% and 78.2%; diagnosis/exam-combined algorithm: 85.3%, 91.2%, 71.3% and 77.7%.
Conclusions
This work demonstrates a practical framework for applying ICH M14 principles to operational definitions through integration of structured and unstructured EHR data. These algorithms offer different options for TED identification and can be selected according to the specific research objective and available data sources.

Introduction: Risperidone is widely used in people with dementia and has an approved indication for behaviours and psychological symptoms of dementia (BPSD) in some countries. However, its clinical use is challenging, as treatment effects are modest and heterogeneous, while risks remain substantial.

Aims: We aimed to develop and evaluate clinical prediction models to estimate individual response to risperidone for BPSD.

Methods: Individual participant data from four randomised trials of risperidone use in dementia were harmonised. Three trials (N=1,327) were used for model development and one (N=337) for external validation. Symptom improvement was assessed using the Behavioural Pathology in Alzheimer’s Disease (BEHAVE-AD) scale. Potential risperidone-associated adverse drug events (ADEs) included major cardiovascular events, extrapyramidal symptoms, and somnolence. Outcomes included improvement in BEHAVE-AD score (overall, psychosis, aggression) and composite benefit–risk categories (low ADE risk–high therapeutic benefit and high ADE risk–low therapeutic benefit). Predictors included risperidone use, demographic factors, and clinical characteristics. Conditional inference forest, random forest, and XGBoost models were trained and evaluated in internal–external cross-validation and external validation. Individualised treatment effect (ITE) was assessed using concordance-statistic for benefit.

Results: In the external dataset, our models can predict overall BPSD (c-statistic: 0·76; 95% CI: 0·70 to 0·83) and psychosis (c-statistic: 0·81; 95% CI: 0·75 to 0·87) improvement at week 8. Models distinguished patients benefiting from risperidone for overall BPSD (c-for-benefit: 0·74; 95%CI: 0·66 to 0·82) and psychosis (c-for-benefit: 0·71; 95%CI: 0·60 to 0·80) outcomes. They can identify those with low ADE risk-high therapeutic benefit for psychosis outcomes (c-statistic: 0·76; 95% CI: 0·69 to 0·83). Performance for aggression-related outcomes was limited.

Conclusions: Clinical prediction models can predict the ITE of risperidone for overall BPSD and psychosis. This approach could support personalised treatment decisions in BPSD, although further real-world validation is required before clinical implementation.

Keywords: antipsychotics, machine learning, clinical prediction models

Introduction: Dengue fever is a major public health concern in tropical and subtropical regions, including Taiwan. Severe dengue can lead to life-threatening complications and substantial mortality. Currently, no specific antiviral treatment is available, and management remains largely supportive. Identifying existing medications with potential protective effects against dengue may facilitate drug repurposing and reduce disease burden.

Aims: To systematically identify medications potentially associated with a lower risk of dengue fever using a hypothesis-generating data-mining approach.

Methods: We conducted a population-based case-control study using Taiwan’s National Health Insurance Research Database (NHIRD). Individuals with incident dengue fever between 2013 and 2024 were identified as cases and matched with up to 10 controls on age, sex, calendar time, general health status, and healthcare utilization. The index date was defined as the dengue diagnosis date for cases and the corresponding matched date for controls. Tree-based scan statistics were used to compare prior medication exposures, organized according to the hierarchical Anatomical Therapeutic Chemical (ATC) classification system, within 30-, 91-, 183-, and 365-day windows before the index date. Signals were evaluated at the 4-, 5-, and 7-digit ATC levels, with p-values <0.01 considered statistical alerts.

Results: A total of 88,605 incident dengue cases and 723,671 matched controls were included. 1378 and 1513 drugs were identified before index date from cases and controls. Significant alerts identified medications with anti-inflammatory or immunomodulatory properties, including antimalarials (P01B) and aminoquinolines (P01BA), purine derivatives (C04AD), and pitavastatin (C10AA08). Additional signals included vitamin D and analogues (A11CC), cough suppressants and expectorants other than opium derivatives (R05FB), opium alkaloid-based cough suppressants (R05DA05), and zopiclone (N05CF01).

Conclusions: The study has identified existing medications associated with a lower occurrence of dengue fever, provided potential candidates for drug repurposing and generated hypotheses for future pharmacoepidemiologic and mechanistic validation studies.

Keywords: dengue fever; drug repurposing; tree-based scan statistics