National scale-up of decentralised hepatitis C point-of-care testing in Australia: a multilevel qualitative process evaluation with exploratory sentiment analysis

Background The Australian Hepatitis C Point-of-Care Testing Program is a national effort to scale up hepatitis C virus (HCV) point-of-care testing across 125 community and prison settings. We examined how multilevel implementation determinants shaped adoption, routine delivery, local adaptation, and anticipated sustainment of the program, and explored the feasibility of using natural language processing (NLP)-driven sentiment analysis as a complementary method for examining implementation determinants. Methods We conducted a multilevel qualitative process evaluation consistent with Medical Research Council guidance and informed by the updated Consolidated Framework for Implementation Research (CFIR). Data sources included program meeting records, protocol documents, and semi-structured interviews. We used deductive-inductive thematic analysis, coding to CFIR 2.0 constructs, followed by inductive generation of determinants within constructs. Two researchers independently rated each determinant from − 2 (major barrier) to + 2 (major enabler), with disagreements resolved by consensus. We then used an offline transformer model to generate complementary sentiment scores and compared the two metrics using Pearson’s r and Cohen’s κ. Evidence strength scores summarised the breadth and triangulation of supporting data across roles, jurisdictions, and source types. Results We analysed 28 program documents, records from 124 meetings, and 28 interviews with 30 stakeholders. We identified 571 determinants, most commonly in the Outer Setting, Innovation, and Individuals domains. Adoption was enabled by the perceived relative advantage of point-of-care HCV testing, including its speed, flexibility, portability, and fit with outreach-oriented care, and by partnerships, leadership, and peer-supported engagement. During routine delivery, local adaptation, and anticipated sustainment, constraints were concentrated in organisational and system contexts, including limited connectivity and interoperability, staffing instability, workflow pressures, funding uncertainty, and stigma. Human ratings and NLP-derived sentiment aligned for 80.9% of determinant records (κ = 0.64, 95% CI 0.59–0.69), against 91.5% between human raters pre-consensus (κ = 0.83, 95% CI 0.79–0.87). Discordant cases surfaced divergence between a determinant’s implementation influence and the tone in which it was discussed. Conclusions Stakeholder accounts indicate that national scale-up of HCV point-of-care testing needs support for workflow integration, data connectivity, workforce capacity, financing, and responses to stigma. After human coding, the offline pipeline generated sentiment scores automatically, reaching moderate concordance with human ratings.