SAFEVAC is a critical multi-jurisdictional network to monitor adverse events following immunisation (AEFI) and hosts the online spontaneous vaccine safety surveillance reporting systems of Victoria (SAEFVIC) and Western Australia (WAVSS). Launched in 2012 the platform has undergone continuous improvements, however the challenges of the COVID-19 roll-out highlighted limitations in intuitive and consistent data entry; standardisation of AEFI reaction terms and data entry fields; and efficient management of rapid, high volume of reports that required workarounds, such as R programming, to prioritise AEFI of specific interest.

To address these challenges, we embarked on a phased redesign of the SAFEVAC platform, through a commitment to strengthen vaccine safety systems from the Victorian and Western Australian Departments of Health. Following hybrid cloud infrastructure enhancement completion in 2024, the next phase aims to modernise the user interface for intuitive and efficient data entry, enhancing capacity during high-volume periods, reducing reporting errors through standardised terminology, and improving interoperability with health information systems. Machine learning will be integrated for automated alerts of priority AEFIs, ensuring efficient workflows and timely response.

Stakeholder engagement — including interviews and workshops with clinicians that assess AEFI reports, epidemiologists conducting surveillance, and AEFI reporters — informed wireframes, user story maps, prototypes, and visual identity. A development task resource was created to address current challenges, and collaboration with legal and IT teams ensured ethical and security compliance. As the platform redevelopment progressed, we apply a hybrid agile process that integrates regular feedback cycles into the build process.

Ultimately the interior redesign of SAFEVAC’s digital house ensures the platform continues to meet the growing demands of public health, allowing for faster, more accurate reporting of vaccine safety events. With improvements in user experience, scalability, and system integration, SAFEVAC will provide a robust, future-proof solution for vaccine safety monitoring and safer immunisation practices across Australia.

Background:
Sensitive and specific signal detection in vaccine safety surveillance is essential for timely responses to adverse events. Individual statistical methods and data sources typically emphasise sensitivity or specificity at the other’s expense. We propose a Naïve Bayes model (NB) to address this.

Methods:
Four signal detection methods, Proportional Reporting Ratio (PRR), Bayesian Confidence Propagation Neural Network (BCPNN), Farrington and Maximized Sequential Probability Ratio Test (MaxSPRT), were compared with the NB incorporating all methods.

The NB models probabilistic relationships between the signal and detection methods, using expert-estimated parameters. NB leaf nodes (one for each detection method) indicate whether a signal is present, and a target node (parent to all leaves) reflects the overall signal likelihood. When a signal is detected by any method, the NB computes the signal’s probability based on all four detection methods. This is tested against an expert-specified threshold indicating a minimum probability of concern, reporting a signal if the threshold is exceeded.

We compared the NB and the four methods using POLAR-GP general practice data for menstrual changes after the introduction of COVID vaccines from 01/01/2021 to 28/03/2023.

Results:
Of 117 weeks, signals were detected by PRR (weeks 12 and 14), BCPNN (weeks 11-16, 27-30), MaxSPRT (weeks 26-117) and Farrington (weeks 16-18, 72-75).
The NB, with threshold set at 2% (highly signal sensitive), reports signals on weeks 16-18, 26-117 and in 24 out of 31 possible combinations of evidence of the methods. With a 20% threshold (highly conservative), NB reports signals on weeks 16, 27-30, 72-75 and in 20 out of 31 combinations.

Conclusion:
Our model predicts signals through probabilistic inference based on expert prior knowledge, available evidence, and the characteristics of each method. The BN provides a potentially powerful tool for synthesising diverse data sources, supporting more reliable signal detection to help decision-makers assess risks.

Background:
The abstract screening process in systematic reviews is time-consuming and labor-intensive, involving a high volume of literature and complex reporting standards. This research aims to evaluate the performance of large language models (LLMs) in automating abstract screening for a systematic review on prevalence and serotype distribution of nasopharyngeal carriage of Streptococcus pneumoniae in Vietnam. A total of 286 abstracts were screened, of which 27 were selected for full-text review.

Method:
A standardized prompt structure, simulating a typical senior-junior researcher interaction, was developed to assess the performance of eight proprietary LLMs (OpenAI GPT-4o, o1, and o1-mini; Google Gemini 2.0-flash and 1.5-pro; Anthropic AI Claude 3.7-sonnet and 3.5-haiku). The LLMs' screening results were compared against the consensus results of the human expert screening process, which served as the ground truth. Performance metrics included Pearson R correlation coefficient, Cohen's Kappa, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), proportion missed, and workload saving. Total time and cost (USD) for prompt generation and screening were also calculated.

Result:
LLMs demonstrated varying performance in abstract screening. Sensitivity ranged from 0.333 to 0.889, specificity from 0.888 to 0.992, PPV from 0.396 to 0.818, and NPV from 0.935 to 0.986. Cohen's Kappa values ranged from 0.439 to 0.702, indicating moderate to substantial agreement with human experts. The proportion missed ranged from 1.2% to 6.5%, while workload saving ranged from 0.804 to 0.899. The total cost was minimal (less than $1) for 6 out of 8 models, with only one model (OpenAI’s o1) exceeding $31. Total time for screening ranged from 380.1 to 5987.4 seconds.

Conclusion:
LLMs show potential for automating abstract screening for systematic reviews, significantly saving time and resources. However, their performance varies, and careful consideration is needed when selecting and implementing LLMs for this purpose. Future research should focus on optimizing LLMs as well as prompt engineering for specific topics and exploring their integration into existing workflows.

Introduction

Vaccine administration errors, particularly those involving incorrect site and route, pose a significant concern for vaccine efficacy and patient safety. Research indicates that some healthcare providers pinch the skin before injection, which prevents the vaccine from reaching muscle tissue as intended, potentially resulting in subcutaneous administration instead. Another common issue is administering the vaccine at the wrong site, such as the shoulder, biceps, triceps, or regions above or below the deltoid muscle. These errors can lead to clinical consequences like the formation of lumps, nerve damage, shoulder injury related to vaccine administration (SIRVA), and reduced immunity. They can inadvertently shape public understanding of proper immunization practices and confuse untrained immunizers. Furthermore, such errors may erode patient confidence in vaccination.

Since vaccine administration errors can impact patient safety, it is essential for vaccine training materials to be accurate. This includes accurate depictions of vaccine administration, given that images play a key role in how people learn, and in providing context and guidance in professional practice, including in medical education.

Methods

Internet images illustrating the immunization process in the upper limb area were retrieved from vaccination-related websites with no restrictions on age, sex or location of vaccinees. NVivo14 software was used to code images by two independent coders, according to 3 main categories: vaccination site, injection angle, and whether the vaccinator was wearing gloves, with detailed subcategories where needed. Krippendorff's alpha reliability coefficient of >0.8 was achieved.

Results

In this study, we investigated 372 images, retrieved from 34 Vaccine Safety Net (VSN) member websites, 12 international health organizations, 11 international news websites, four medical websites and two government health websites. It was found that images of vaccine administration frequently contained errors. This study discusses potential factors contributing to improper administration and recognises the importance for health institutions and media to incorporate visuals that reflect best practices in healthcare and health communication.

Conclusion

Our findings suggest that incorrect vaccination techniques are widespread across trusted online resources. Identifying common missteps in visual representations of vaccination could facilitate the development of automated tools to detect misleading content, ultimately promoting safe vaccination practices.

Background:
South Australia (SA)’s ongoing statewide Meningococcal B Vaccination Program (the Program) commenced in 2018 and provides 4CMenB (Bexsero) to children aged six weeks to 12 months, and young people aged 15 to 17 years. Our retrospective study aimed to evaluate the safety of 4CMenB, using routinely collected data on adverse events following immunisation (AEFI).

Methods:
We collated and classified AEFI reports relating to 4CMenB received by the Communicable Disease Control Branch (CDCB) of SA’s Department for Health and Wellbeing between the Program’s commencement on 1 October 2018 and 30 June 2022. Reports were included if they related to participants in the Program or those who received 4CMenB through the National Immunisation Program (NIP).

Learning objectives:
• Understand SA’s vaccine safety surveillance system, with particular reference to the 4CMenB vaccine.
• Evaluate the characteristics of the AEFI reports received for the Program/NIP.
• Appreciate the opportunities and challenges presented by using routinely collected surveillance data for research.

Results:
306 AEFI notifications were received during the study period, equating to a reporting rate of 69.9 notifications per 100,000 doses. Most reported AEFI were known, common, non-serious, adverse events (85% of reports), followed by serious AEFI (13%) and suspected unexpected serious adverse reactions (SUSAR) (2%). No new safety signal was ascertained from review of serious AEFI and SUSAR. A high proportion (83%) of children under two years with reported AEFI were administered prophylactic acetaminophen.

Conclusion:
These results are consistent with existing safety data and provide further support for the safety of the 4CMenB vaccine. Challenges encountered in using data collected for routine surveillance for research could be mitigated by designing data collection instruments with future research applications in mind.

Understanding the short-term safety of scheduled National Immunisation Program (NIP) and additional vaccines in young children is crucial for guiding immunisation policies and improving coverage. We assessed the safety of childhood vaccines administered at 2, 4, 6, and 12 months, comparing children receiving standard NIP vaccines alone, meningococcal B (MenB) alone, or both in one visit. Data were collected from children aged 2 – 12 months vaccinated between 1 April 2022 and 30 June 2024 at AusVaxSafety sentinel sites via parent surveys three days post-vaccination.
At the 2-month schedule, 15,088 responses were received (85% standard NIP [DTPa-HepB-IPV-Hib; Pneumococcal; Rotavirus], 6.5% MenB alone, 8.5% both), with adverse events following immunisation (AEFI) reported in 23%, 31%, and 26%, respectively. At 4 months (14,109 responses: 75.6% standard NIP [DTPa-HepB-IPV-Hib; Pneumococcal; Rotavirus], 10.1% MenB alone, 14.3% both), AEFI were reported in 30%, 36%, and 48%. At 6 months (9,726 responses: 69.5% standard NIP [DTPa-HepB-IPV-Hib], 18% MenB alone, 12.4% both), AEFI were reported in 20%, 36%, and 43%. At 12 months (12,447 responses: 72.4% standard NIP [Meningococcal ACWY; Measles-mumps-rubella; Pneumococcal], 13.4% MenB alone, 14.2% both), AEFI were reported in 28%, 47%, and 48%.
Fever was highest in the NIP+MenB group at all time points (16% at 2 months, 31-34% at 4-12 months), compared to 16-29% for MenB alone, and 9-16% for standard NIP. Medical attendance ranged from 0.9-1% for standard NIP, 1-3% for MenB alone, and 1-2% for NIP+MenB across all schedules.
AEFI rates varied by age and vaccine type, with higher proportions in children receiving MenB alone or with standard NIP vaccines. However, severe AEFI were rare. These findings support the co-administration of scheduled childhood vaccines and MenB to minimise GP visits and enhance vaccine coverage. Ongoing surveillance remains essential to maintain confidence in childhood immunisation.

Introduction:
Serious adverse events following immunisation (AEFI) are rare but require investigation to identify causes to enable appropriate responses and maintain public confidence in vaccine safety. Timor-Leste has observed an increase in reported AEFI, from less than 5 annually before 2020, to ≥30 cases annually from 2020 reaching a peak of 59 cases in 2024. We examine 2024 AEFI trends, potential causes, and recommendations.

Methods:
Serious AEFI are reported to the Ministry of Health and National Immunization Technical Advisory Group for investigation. AEFI surveillance is paper-based, supplemented by WhatsApp notifications. Causality assessments are conducted for select cases.

Results:
In 2024, 59 AEFI were reported, all in infants (aged <1 year) following routine immunisations. Of these, 56 (95%) were classified as serious, including 52 classified as abscesses, resulting in one fatality. Cases were distributed nationwide and occurred throughout the year. Two clusters of two events were identified in two different municipalities. Most of the cases (47 cases; 80%) occurred following hepatitis B birth dose vaccination. Nine cases underwent thorough investigation, with seven causality assessments completed. Findings indicated vaccination technique errors, with three involving recent health worker graduates volunteering to deliver vaccination. Two cases were linked to traditional medicine use, though this occurred post-symptom onset. Three events were considered coincidental. Limited testing was conducted with one sample testing positive for Staphylococcus aureus.

Discussion:
The increase in reported AEFI may partially reflect improved surveillance, but further investigation is needed. Potential causes include vaccine administration errors and improper handling of multi-dose vials resulting in contamination. Corrective actions include onboarding and refresher training, restricting vaccine administration to trained staff, and reinforcing sterile practices. Health education and risk communication are required to improve the timeliness of case detection and reporting. Sustainable funding is needed to intensify AEFI surveillance efforts and facilitate prompt investigations and response.