The problem: interdisciplinary teams building AI for healthcare struggle to share knowledge, set expectations, and work inside data-restricted environments. The audience: qualitative researchers, UX/implementation leads, and project managers seeking practical methods for thematic analysis of collaboration data. The payoff: this post refracts the July 22, 2026 PLOS One interview study through the lens of AI-enabled qualitative research to give extractable steps, direct quotes, and software mappings for faster synthesis using the primary keyword "qualitative analysis of AI collaboration in healthcare".
Key Takeaways
The central finding: according to the July 22, 2026 PLOS One paper, early-stage AI-in-healthcare consortia face amplified interdisciplinary friction because AI novelty, computational limits, and data-safe-haven constraints disrupt normal collaboration and knowledge sharing.
- The study interviewed 13 participants across three UK research consortia between January 19, 2023 and March 3, 2023, with 9 face-to-face and 4 remote interviews, according to PLOS One (published July 22, 2026).
- Each consortium employed about 20–30 academic and healthcare researchers, as reported in the PLOS One paper, increasing coordination complexity across work packages.
- Practical bottlenecks included constrained compute (PLOS One quotes a participant noting their Safe Haven supplied only two GPUs) and data access limits that forced extra translation work for non-data scientists.
- Direct voices from the study: a clinician said, "But yeah, we'll have to see how it works" (P10, clinician), and a data scientist noted, "It is good to know, let the clinicians know that all the models don't work every time" (P2, data scientist).
What happened: methods and core observations
Answer: the PLOS One study conducted an inductive thematic analysis of 13 semi-structured interviews to map how stakeholders collaborate on AI development in healthcare.
According to PLOS One (published July 22, 2026), the authors recruited 13 participants across three NIHR-funded consortia between January 19, 2023 and March 3, 2023 to capture early-stage AI development practices.
According to PLOS One, interviews lasted one hour each (9 in-person, 4 remote), transcripts were produced via Otter.ai and manually corrected, and coding followed Braun and Clarke's reflexive thematic analysis using Taguette for initial coding.
According to PLOS One, the analytic output identified three high-level themes: AI novelty and disciplinary differences, AI knowledge sharing across tools and meetings, and inclusion of clinicians and PPIE coordinators to ground real-world healthcare perspectives.
Findings snapshot
| Date / Source | Metric | Value | Implication |
|---|---|---|---|
| Jan 19–Mar 3, 2023 (PLOS One) | Interviews conducted | 13 (9 face-to-face, 4 remote) | Small purposive sample rich in domain knowledge for qualitative theme generation |
| Consortia description (PLOS One) | Consortium size | 20–30 researchers each | Large team sizes increase coordination and communication overhead |
| Published Jul 22, 2026 (PLOS One) | Peer-reviewed output | Thematic analysis with three core themes | Evidence-based basis for recommending workflow and tooling adjustments |
| Study logistics (PLOS One) | Participant reimbursement | £25 voucher each | Standard ethics practice for interview studies |
| Infrastructure constraints (quoted in PLOS One) | Available GPUs in Safe Haven | 2 GPUs procured for all projects versus need for 4+ | Compute scarcity slows iteration and knowledge feedback loops |
Implications for qualitative researchers: qualitative analysis of AI collaboration in healthcare
Answer: qualitative researchers should treat early-stage AI collaborations as socio-technical systems and plan interviews, artifact capture, and synthesis accordingly.
According to PLOS One (published July 22, 2026), AI novelty creates divergent expectations across roles, so qualitative studies must sample clinicians, data scientists, statisticians, and PPIE coordinators to capture meaning-making across groups.
According to PLOS One, data access restrictions and computational constraints create translation tasks where data scientists convert Jupyter notebooks into Excel or slide summaries for clinicians and patients; qualitative coding must therefore include artefact lineage (original script → processed summary).
Practical decision: collect meeting artifacts, code versions, and communication traces (emails, GitHub commits, slide decks) in addition to interviews, because PLOS One found that the substantive content exchanged in meetings was a central site of coordination and loss of detail.
Ethics note: this post focuses on research practice and does not provide clinical or diagnostic advice; follow local governance for patient data and consent as in the PLOS One project.
How Evidano helps: map from collaborative pain points to AI-enabled qualitative features
Problem: scattered transcripts, meeting notes, and artifact versions slow synthesis
Evidano is an AI-powered qualitative data analysis platform that helps researchers analyze interviews, open-ended surveys, and documents.
Solution: Evidano ingests transcripts, slide decks, and code snippets, then auto-generates thematic, content, and frequency analyses so teams can identify recurring coordination problems faster; this addresses the PLOS One finding that meetings and artifact conversion are central sites of work.
Relevant feature: automated ingest and thematic coding reduces manual Taguette-style steps and preserves artifact provenance. See Evidano features for specifics.
Problem: noisy automated transcripts and domain terms (Otter.ai used in PLOS One)
Solution: Evidano offers transcription with custom dictionaries and PII redaction, which helps reproduce the PLOS One approach but with higher domain accuracy and privacy controls.
Relevant feature link: Evidano speech-to-text for enterprise transcription with custom vocabularies.
Problem: translating technical notebooks into clinician- and patient-facing summaries
Solution: Evidano's AI chat over documents plus visualization tools (co-occurrence networks and hierarchical codes) helps research teams generate concise, audience-tailored summaries and slide-ready outputs, matching the PLOS One observation that data scientists often convert notebooks into Excel or slide summaries for meetings.
Benefit: faster turnarounds on iteration cycles that PLOS One notes are slowed by long model runs and restricted compute.
Problem: tracking differences across stakeholder segments
Solution: Evidano's cross-segment analyses let teams compare themes across clinicians, data scientists, and PPIE coordinators so teams can quantify where expectations diverge, directly addressing the PLOS One theme of disciplinary differences.
Problem: governance and safe-haven constraints for sensitive data
Solution: Evidano provides encrypted storage and clear export controls that help researchers comply with TRE or Safe Haven rules while enabling collaborative qualitative synthesis; see Evidano data security.
FAQ: qualitative analysis of AI collaboration in healthcare
How should I design interviews to study interdisciplinary AI collaboration?
Answer: use purposive sampling that includes data scientists, clinicians, statisticians, and PPIE coordinators and collect artifacts as well as interviews.
Supporting detail: the PLOS One study (published July 22, 2026) recruited 13 participants across three consortia to capture multiple disciplinary perspectives and used meeting artifacts and summaries to triangulate findings.
What artifacts matter most for thematic analysis of AI projects?
Answer: Jupyter notebooks, slide decks, Excel summaries, GitHub repositories, and Safe Haven access logs are primary artifacts to collect.
Supporting detail: PLOS One reports that data scientists used Jupyter Notebooks internally and converted outputs to Excel or slides for clinicians, making the notebook-to-summary pathway a key locus for information loss.
How can researchers capture compute and infrastructure constraints in qualitative data?
Answer: ask participants about run-times, GPU availability, Safe Haven limitations, and decision delays, and code those as structural constraints.
Supporting detail: PLOS One quotes a data scientist describing that a National Safe Haven had only two GPUs for dozens of projects, which materially slowed model iteration.
Can AI tools speed qualitative synthesis without harming rigor?
Answer: yes, when AI tools are used to surface candidate codes and patterns that researchers then validate manually.
Supporting detail: PLOS One used Braun and Clarke's reflexive thematic analysis; Evidano's AI-assisted coding can bootstrap that workflow but should be paired with human reflexivity and validation, mirroring the study's interpretative approach.
Conclusion & Next Steps
Recap: according to PLOS One (published July 22, 2026), early-stage AI-in-healthcare consortia face amplified interdisciplinary friction because AI novelty, compute scarcity, and data governance create new translation work for teams.
Actionable next step: capture interviews plus the notebook-to-summary artifacts, tag infrastructure constraints in your codebook, and run a cross-segment thematic analysis to surface where expectations diverge.
If you want to accelerate that pipeline, Evidano automates transcript ingestion, thematic and cross-segment analyses, and produces presentation-ready summaries that address the practical bottlenecks highlighted in the PLOS One study; learn more at Evidano features.
Try a hands-on next step: Try Evidano for free to upload transcripts, artifacts, and meeting notes and get a first automated thematic synthesis.