A natural language processing approach to determine whether streamlined health technology assessment for multi-indication medicines is feasible

Conducting a health technology assessment (HTA) for each indication of a medicine contributes to increasing demand on HTA systems and results in delays in access. Streamlining assessments for subsequent indications has been proposed, but there is limited evidence on how this could be achieved. We aimed to quantify semantic similarity between HTA documents for different indications of the same medicine to explore whether concepts could be carried forward to subsequent assessments. Public Summary Documents (PSDs) from July 2019 to November 2024 were included if they related to new major submissions to the Australian Pharmaceutical Benefits Advisory Committee. Each paragraph was embedded using the Sentence-Bidirectional Encoder Representations from Transformers (SBERT) model. Pairwise Euclidean distances (measuring semantic similarity) were quantified between embeddings across indications for multi-indication medicines and were compared with those for single-indication medicines. Cluster analysis, followed by manual thematic review, identified recurrent concepts across indications of the same medicine. A total of 326 PSDs were included (154 single-indication medicines, 59 multi-indication medicines). Multi-indication medicines demonstrated significantly greater semantic similarity across their indications compared with single-indication medicines. However, recurrence of specific concepts across either all or similar indications of a medicine was infrequent. Despite the greater semantic similarity observed across indications of multi-indication medicines, this did not translate into substantial opportunities to carry forward assessment concepts. Therefore, it is unlikely that any concepts or information from assessments undertaken for previous indications could be carried forward to subsequent indications. The methods described provide a foundation for future research on semantic overlap in HTA and other policy areas.Author summary: When a new medicine is evaluated for public funding, it is usually assessed separately for each medical condition it treats. As medicines are increasingly approved for multiple conditions, this approach places growing pressure on health technology assessment systems and can delay patient access. Some have suggested that parts of earlier evaluations could be reused when the same medicine is assessed for a new condition. However, there has been little evidence to show whether this is feasible. This study uses machine learning-based analysis to examine whether assessment documents for the same medicine, but for different conditions, contain recurring concepts that could realistically be carried forward. We analysed more than 300 reimbursement decisions in Australia. Although assessments of the same medicine were more similar to each other than they were to assessments of unrelated medicines, we found that evaluation concepts rarely repeated consistently across conditions. This indicates that simply reusing content from earlier assessments may not be appropriate in most cases. Beyond this specific application, our approach demonstrates how machine learning can be used to measure similarity across other types of health policy documents. Such methods could help policymakers determine other processes that can be streamlined, such as assessments across different jurisdictions.