Introduction: In self-controlled case series (SCCS) studies, properly specifying the focal window is crucial to avoid biased estimates. This is challenging when limited clinical information is available to inform design decisions.

Aims: Using COVID-19 vaccine as a case study, we compare three data-driven methods for identifying optimal SCCS focal windows across two databases.

Methods: We studied individuals aged ≥12 years hospitalised for myocarditis, pulmonary embolism (PE), or thrombocytopenia and received 1-2 doses of BNT162b2 or ChAdOx1 vaccine using the Malaysian surveillance and administrative data and the Taiwan’s National Health Insurance Research Database (NHIRD). Using the SCCS design, we scanned focal window lengths (day 1-70 post-vaccination) using three methods: (i) maximum and graphical approach (Xu 2011) using the maximum incidence rate ratio (IRR) or the decreasing "change point" within a linear relationship between IRR and the inverse of average person-time at risk, (ii) linear-quadratic spline-based method (Campos 2017) identifying knot with the minimum error on IRR-inverse window length plots, and (iii) scan-based test (Hunsberger 2015) detecting temporal event clusters by likelihood ratio statistics.

Results: For myocarditis, Xu and Campos identified a 6–8-day focal window for BNT162b2 first dose, while Hunsberger suggested a 11-20-day window across databases. All methods estimated a 4-day window for BNT162b2 dose 2 in both databases. For thrombocytopenia, signals appeared only in the NIHRD database: Hunsberger identified a 34-day window for BNT162b2, while Xu and Campos suggested an 8-9-week window for ChAdOx1. For PE, no signal was identified across all scanned windows in both databases.

Conclusion: The optimal focal windows across methods and databases are consistent when signals and clustering of events are strong post-vaccination, while weaker signals increases variability between methods. The variations between databases further suggest that results are influenced by the specific risk profiles of underlying populations.

Keywords: self-controlled case series; vaccine safety; covid-19 vaccine

Background: Vaccination is one of public health’s greatest achievements, second only to clean water in preventing death and disease globally. While intensified scrutiny obviously accompanies new vaccines, variants, or use in new populations, sustained monitoring of established vaccines is equally critical to ensure detection of rare adverse events, identify risks in sub-populations and rapidly recognise anomalies by brand, batch, provider-group, recipient characteristics, or even spurious findings. Because safety signals may emerge from multiple pathways, effective vigilance demands diverse surveillance systems and analytic methods to ensure that no meaningful signal goes unnoticed.

Aim: Describe Victoria’s routine weekly vaccine safety vigilance supporting safe immunisation program delivery.

Methods: Victoria operates an integrated, multi-modal vaccine safety surveillance system. Weekly spontaneous reports to SAEFVIC are assessed using disproportionality analyses of vaccine–adverse event pairs and reporting rates per 100,000 doses administered, stratified by sex, age group, vaccine target disease, and, where relevant, vaccine brand. Serious AEFI reports undergo line-list review for epidemiological clustering. Telephone helpline contacts and emergency department presentations are monitored using modified CUSUM methods to detect deviations from historical patterns in post-vaccination contacts and presentations within 28 days of vaccination. Monthly VaxPulse social listening monitors online personal adverse-event discussions and misinformation trends for prioritised vaccines. Signal hypotheses are evaluated for cross-method consistency, biological plausibility, known associations, and clinical significance to determine escalation, investigation, or return to routine surveillance.

Results: The framework integrates passive reporting, syndromic surveillance, healthcare utilisation data, and social listening to support timely signal detection and evaluation. Multiple complementary methods enable assessment of potential signals across diverse data sources and facilitate rapid prioritisation of issues requiring further investigation.

Conclusion: Integrated, multi-source, sex- and age-disaggregated routine vaccine safety surveillance strengthens pharmacovigilance capacity, supports timely signal detection and evaluation, and provides ongoing assurance for immunisation program safety.

Keywords: vaccine safety, pharmacovigilance, signal detection

Introduction
Evidence on influenza vaccination and septic shock in elderly heart failure patients remains limited. Standard self-controlled case series (SCCS) analyses are vulnerable to informative censoring, event-dependent exposure, and depletion of pre-vaccination events reflecting healthy vaccinee bias in this high-mortality population.
Aims
To assess the association between influenza vaccination and septic shock, and how sequentially correcting these assumption violations affects the estimates.
Methods
We identified septic shock cases aged ≥65 with heart failure vaccinated against influenza between October 2022 and September 2023 using nationwide claims data. Septic shock required a first inpatient diagnosis with concurrent vasopressor use, and heart failure by any prior diagnosis and medication use in the preceding year. Five analyses were applied sequentially, beginning with a standard SCCS without a pre-exposure period, then adding a 21-day one. Modified SCCS, which requires no pre-exposure period, then sequentially addressed event-dependent exposure using Horvitz-Thompson weighting, informative censoring by extending follow-up to study end regardless of death, and depletion of pre-vaccination events by adding unvaccinated cases. Sensitivity analyses varied the pre-exposure period from 28 to 98 days in 7-day intervals.
Results
Among 4,185 septic shock cases, 1,656 were vaccinated. IRRs for days 0–14, 15–56, 57–120, and 121–180 were 0.86, 1.52, 1.55, and 1.60 (standard SCCS), 0.30, 0.61, 0.86, and 1.15 (pre-exposure period added), 0.01, 0.01, 0.04, and 0.11 (Horvitz-Thompson weighting), 1.04, 1.33, 1.34, and 1.24 (informative censoring corrected), and 0.57, 0.75, 0.89, and 0.92 (fully adjusted model). Across sensitivity analyses, IRRs declined steadily without convergence.
Conclusions
Influenza vaccination was associated with a waning protective effect against septic shock. Modified SCCS outperformed standard approaches, which failed to adequately address event-dependent exposure even with varying-length pre-exposure periods. Correction for informative censoring and healthy vaccinee bias was additionally required for unbiased estimates in this high-mortality setting.
Keywords: Influenza vaccination, septic shock, self-controlled case series

Introduction: Timely responses to potential vaccine adverse event (AE) signals are essential for translating real-world data into public health decisions. Signal detection methods should be sensitive to identify emerging safety concerns while remaining specific to minimise false alerts. We present a novel augmented Naïve Bayes model (AugNB) that combines established signal detection approaches, improving performance over individual methods alone.

Aims: Developing AugNB for AE signal detection in general practice (GP) data by integrating complementary statistical methods, improving predictive performance compared with individual approaches, and validating against expert-led signal investigation outcomes.

Methods: AugNB incorporated four signal detection methods: Proportional Reporting Ratio (PRR), Bayesian Confidence Propagation Neural-Network (BCPNN), Farrington, and Maximised Sequential Probability Ratio Test (MaxSPRT). When any method identifies a potential signal, AugNB estimates the probability of a true safety problem. Performance was evaluated against each individual method and unaugmented NB using POLAR primary care data for 16 Adverse Events of Special Interest (AESIs) and four priority vaccines (influenza, COVID-19, shingles, and RSV), linked to prior investigation outcomes from January 2017 to March 2024.

Results: AugNB outperformed MaxSPRT, BCPNN, PRR, Farrington, and NB using leave-one-out cross-validation achieving 81% accuracy, 87% F1 score (harmonic mean of precision and sensitivity), and 75% AUC (measure of the model’s ability to distinguish signals from non-signals) —exceeding the strongest baseline (NB: 75%, 82%, 71%). This included detection of menstrual changes (rejected by PRR) and rejection of appendicitis (detected by PRR, BCPNN and Farrington) following COVID-19 vaccination, matching expert-led investigations.

Conclusions: AugNB improved signal classification performance and may support prioritisation of vaccine safety signals for further investigation. It offers a transparent, data-driven framework synthesising multiple surveillance methods, strengthening timely vaccine safety surveillance, and supporting evidence-based public health decisions in routine pharmacoepidemiologic surveillance.

Keywords: Vaccine safety; Signal detection; Probabilistic Inference

Introduction: Vaccine safety surveillance relies on multiple sources of information, including spontaneous reports, healthcare datasets, and consumer concerns. During Australia's COVID-19 vaccination program, a consumer who developed giant cell arteritis (GCA) following COVID-19 vaccination sought further investigation of a potential association, prompting a signal assessment using multiple Australian healthcare datasets.

Aim: To investigate the potential association between COVID-19 vaccination and incident GCA.

Methods: A staged signal assessment was undertaken. Stage 1 involved review of spontaneous reports of giant cell arteritis (GCA) following COVID-19 vaccination received in Victoria, including reporting rate and disproportionality analyses. Stage 2 involved investigation using Australian general practice (POLAR; ~12 million electronic health records) and linked hospital (VSHL) datasets covering January 2021 to September 2024. Incident GCA cases were identified using coded diagnoses and analysed using a self-controlled case series design. Relative incidence (RI) was estimated for a 42-day post-vaccination risk period compared with baseline periods, with sensitivity analyses using 1–7, 1–14 and 1–28-day risk windows. Analyses were stratified by vaccine platform, age group and sex.

Results: Six spontaneous reports of GCA following COVID-19 vaccination were identified. No disproportionality signal was detected (PRR 0.05, chi-squared 2.75). No increased risk of GCA was observed within 42 days of vaccination in general practice (RI 0.96, 95% CI 0.76–1.21) or linked hospital datasets (RI 0.79, 95% CI 0.68–0.91). Results were consistent across vaccine brands, age groups, and sex.

Conclusion: No evidence of an increased risk of GCA following COVID-19 vaccination was identified. Multi-datasource robust pharmacoepidemiological methods can provide evidence-based answers to support vaccine safety surveillance, address consumer concerns and maintain public trust.

Keywords: Vaccine safety; Pharmacovigilance; Consumer engagement

Introduction
Evidence on mRNA COVID-19 vaccination and incident autoimmune diseases remains limited, with the potential for residual event-dependent exposure in standard self-controlled case series (SCCS) analyses, even with pre-exposure periods incorporated.
Aims
To assess the association between mRNA COVID-19 vaccination and incident autoimmune diseases and to quantify the impact of event-dependent exposure by comparing standard and modified SCCS estimates.
Methods
Two SCCS analyses were conducted using nationwide health insurance claims data linked with a COVID-19 registry, among incident autoimmune cases aged ≥20 who received at least one mRNA COVID-19 vaccination dose between March 2021 and September 2022, without prior autoimmune diagnosis. Autoimmune cases were defined as the first diagnosis across 10 organ system categories, identified from inpatient, outpatient, or emergency settings. The standard SCCS incorporated a 14-day pre-exposure window, while the modified SCCS treated post-event exposure as missing and applied Horvitz-Thompson weighting. Both analyses estimated IRRs for day 1-21 post-dose risk windows, adjusting for SARS-CoV-2 infection and variant period.
Results
Among 295,749 incident autoimmune cases who received at least one mRNA COVID-19 vaccination, the modified SCCS estimated an overall day 1-21 IRR of 1.69 (95% CI, 1.66-1.71). Dose-stratified IRRs were 1.75 (95% CI, 1.72-1.79), 1.49 (95% CI, 1.45-1.52), and 1.80 (95% CI, 1.77-1.84) for the first, second, and third doses, respectively. In contrast, the standard SCCS yielded a lower overall IRR of 1.20 (95% CI, 1.19-1.21), with the third-dose estimate suggesting a protective effect (IRR, 0.95, 95% CI, 0.93-0.97).
Conclusions
mRNA COVID-19 vaccination was associated with increased risk of incident autoimmune disease within 21 days after each dose. The reversed third-dose estimate in standard SCCS underscores the necessity of modified SCCS that address event-dependent exposure rather than relying solely on fixed pre-exposure periods in multi-dose vaccination safety research.
Keywords: mRNA COVID-19 vaccination, incident autoimmune disease, self-controlled case series