Why Building Healthcare AI Is Also a Collaboration Problem
By Duncan Reynolds
When we talk about artificial intelligence in healthcare, attention tends to fall on the technology: the algorithms, the data, and whether a model can make an accurate prediction. But developing healthcare AI is not simply a technical exercise. It requires data scientists, clinicians, statisticians, researchers, patients and public contributors to work together. These groups bring different expertise, use different tools, and often have different ideas about what counts as useful evidence.
A recent PLOS One study by Rafael Henkin and colleagues asks what happens when these people actually try to collaborate. The researchers interviewed 13 members of three UK research consortia developing AI for people with multiple long-term health conditions. Their central finding is straightforward: adding AI can make already difficult interdisciplinary collaboration even harder.
AI means different things to different people
One problem was that AI was relatively new to many participants. People entered projects with different levels of technical knowledge and, importantly, different expectations about what AI could achieve.
For some researchers, experimenting with many alternative models was a normal part of the work. For clinicians or epidemiologists, the more pressing question might instead be: which approach should we actually use, and does it improve on a simpler method?
These are not merely differences in technical knowledge. They reflect different professional ways of judging worthwhile research. The study found that AI’s novelty heightened existing differences between disciplines and sometimes created expectations that were technically unrealistic.
This matters because a team does not automatically have a shared understanding simply because everyone is working on the same project. Developing that shared understanding becomes part of the research itself.
Collaboration requires translation
The second major issue concerned how AI work was communicated.
A data scientist might naturally work in Python, R, GitHub or a Jupyter Notebook. Those tools make sense when collaborating with another programmer. They may be considerably less useful when discussing findings with a clinician or a patient group.
Participants therefore regularly changed both what they communicated and how they communicated it. Technical outputs might be converted into spreadsheets or presentation slides. Mathematical detail might be reduced for a large interdisciplinary meeting. More explanation might be added when presenting findings to patient and public involvement groups.
The authors argue that this translation is not a minor presentation task. It is part of the collaborative work required to make AI research possible.
There is also a potential cost. Every time technical work is condensed or reformatted for another audience, somebody decides which details matter and which can be left out. The researchers raise the possibility that information may be lost during this process.
Better collaboration, then, is unlikely to come simply from finding the right software. People on all sides have to learn enough about each other’s methods, language and priorities to communicate meaningfully.
Healthcare data are not just numbers
Perhaps the most important contribution of the study concerns the role of clinicians and patient perspectives.
Electronic health records can easily appear to researchers as enormous datasets containing diagnoses, prescriptions, test results and other variables. But those data were produced through encounters between patients and healthcare systems.
Clinicians helped data scientists interpret whether patterns found in the data made sense in practice. Something statistically interesting might turn out to be clinically obvious, implausible or irrelevant. Conversely, clinical knowledge might reveal why an apparently strange pattern was entirely reasonable. The study therefore describes clinicians as providing an important form of “sense checking”.
Patient and public perspectives posed a related but different question: what does this analysis mean for patients?
That question could redirect researchers’ attention. Patients might care about outcomes such as quality of life that were poorly represented, or entirely absent, in the available datasets. Their involvement could therefore expose a gap between what researchers were able to measure and what mattered to people living with illness.
In this sense, patient involvement did more than make research more accessible. It helped reconnect abstract datasets with the people and experiences those data represent.
Some important work is easy to overlook
The paper also draws attention to patient and public involvement and engagement, or PPIE, coordinators.
These coordinators frequently acted as translators between technical researchers and patient contributors, helping researchers decide what information to present and how to present it. Yet participants appeared to have limited awareness of how much work this required.
That observation has wider implications. Interdisciplinary projects often depend on forms of coordination and translation that are less visible than modelling, data analysis or publishing papers. If this work is essential to making collaboration function, projects need to allocate time and resources to it rather than treating it as an informal extra.
The infrastructure shapes the science too
There was another constraint that had little to do with misunderstandings between professions: the practical environment in which healthcare AI is developed.
Sensitive health data are commonly held within secure research environments. These safeguards are important, but the environments can limit available programming languages, computing power and access to data.
That can create an unusual problem for AI research. A team may have permission to analyse valuable healthcare data but lack sufficient computing resources to run the models it wants to test. Restrictions can also mean that only some collaborators can access the data directly, requiring additional summaries, slides or simulated data so that others can participate in discussions.
The result is that technical infrastructure, information governance and collaboration become tightly connected. The paper therefore treats AI development as a sociotechnical activity: technical choices cannot be separated neatly from institutions, professional practices and social relationships.
What should healthcare AI projects do differently?
The authors stop short of proposing a definitive recipe for successful interdisciplinary AI. Their evidence comes from only 13 participants in three UK consortia, all studied relatively early in their projects. Patient and public contributors themselves were not interviewed, so claims about their perspectives come through researchers and PPIE coordinators rather than directly from patients.
Those limitations matter. The findings should not automatically be assumed to apply to every healthcare AI project.
Nevertheless, the study offers several practical starting points. Teams should establish realistic expectations about what AI can and cannot do, build AI literacy across disciplines, devote proper resources to translating technical work into accessible forms, and recognise the additional coordination required to bridge professional groups.
Perhaps the broader lesson is that successful healthcare AI depends on much more than producing a good model.
A technically sophisticated algorithm can still sit within a poorly functioning collaboration. And a research team can have access to excellent data while lacking the shared language, infrastructure or relationships needed to make good use of them.
Building healthcare AI is therefore also about building the conditions under which people with very different forms of expertise can understand one another, challenge one another and decide together what useful AI should look like.