Queensland University of Technology has opened a senior research fellow post in health AI implementation, and the successful candidate will be judged on a question most model developers would rather postpone: what happens after the algorithm leaves the training notebook and enters an emergency department, a ward round, a community clinic or a telehealth queue. The role, listed on QUT's recruitment portal and syndicated across academic job boards, asks for experience in four interlocking fields — machine learning evaluation, health services research, implementation science and clinical data governance.

Implementation science is the study of how to get evidence-based tools into routine care. It is the discipline that asks why a model with impressive retrospective results still fails when clinicians have to use it at 2am. Australia has no shortage of health AI prototypes. The bottleneck is the smaller pool of researchers who understand clinical workflows, data quality, governance and regulatory constraints well enough to move a pilot into routine practice.

Implementation is where health AI usually stalls

Start with a well-documented base rate. When researchers externally validated the Epic Sepsis Model at a large Michigan health system, the area under the curve fell to 0.63, down from development-era estimates of roughly 0.76 to 0.83. That is not a purely technical failure. Clinicians were already treating sepsis; the model's reference standard shifted as the hospital changed how it recorded pneumonia and acute illness. The algorithm stayed the same, but the target it was trained to predict no longer matched the live electronic record.

Health AI projects in Australia and elsewhere repeat that pattern. A model passes a pilot evaluation, then stumbles when ward clerks code a comorbidity differently, when a referral pathway changes hands, when the tool adds an extra click to a nurse's already crowded shift, or when the version of the electronic record differs between sites. Implementation researchers call this the deployment gap: the distance between a model's published performance and its performance inside a working clinical service.

That gap is not limited to commercial models. Early in the pandemic, a systematic review of 232 prediction models for COVID-19 found that all of them were at high or unclear risk of bias, largely because of unrepresentative training data and weak validation. The findings did not mean the idea was wrong; they meant the distance between development and deployment was being ignored.

Why Queensland is a sensible place to run that experiment

Brisbane-based QUT already hosts a substantial cluster of health services research, including its Centre for Healthcare Transformation and long-running partnerships with Queensland Health. The senior research fellow position sits inside that applied environment, where the distance between a model's validation report and a hospital's digital record is shorter than in a pure computer science department.

Queensland Health has invested in electronic medical records across its largest sites. Those records generate the kind of live data implementation research needs, but they also create the compliance burden that makes deployment difficult. QUT's task is often the unglamorous middle layer — building the prospective evaluation, governance arrangements and clinician training that determine whether a prediction model reduces harm or simply adds noise. The Michigan validation, published in JAMA Internal Medicine, illustrates the gap precisely.

The listing itself is available on QUT's recruitment portal; salary packaging, closing date and key selection criteria are set out there.

What a competitive application will need to show

Senior research fellow roles at Australian universities rarely turn on a single first-author paper. Selection panels weigh a mix of qualifications, funding, supervision and external relationships. For this position, evidence of implementation will separate genuine candidates from applicants with a long publication list and no deployment experience.

• A doctoral qualification in a relevant discipline, usually health data science, clinical informatics, epidemiology or implementation science.
• A record of external research income, with grants tied to health services or digital health rather than only methodological novelty.
• Experience supervising doctoral students and research staff, and evidence that those collaborations delivered practical outputs.
• Demonstrated ability to work with hospital partners, health services, ethics committees and clinical data custodians.

But the decisive line in the application is implementation. Candidates who have taken a clinical prediction model through human research ethics, data linkage, clinician co-design and prospective evaluation will stand out. The supporting statement should name the workflow changes, governance documents and training materials the applicant built, not just the model's precision and recall.

If the CV still reads like a chronological memoir, the statement is the place to fix it. A tightened CV helps, but the panel needs to see the implementation evidence directly. Practical guidance on framing that evidence is covered in our academic CV advice.

How this hiring signal fits the broader health AI market

Australian health AI is moving from roadmaps into operational budgets. The Australian Digital Health Agency's National Digital Health Strategy frames artificial intelligence as part of safe, seamless care that stays centred on the person. At the same time, the Therapeutic Goods Administration has been regulating software as a medical device when it performs clinical functions rather than only storing records; current guidance is outlined in the TGA's software as a medical device pages.

Those two forces create demand for a particular kind of researcher: someone who can speak implementation science to clinicians, explain model limitations to hospital executives, write a validation protocol that a human research ethics committee will approve, and document the data lineage for auditors. A senior research fellow in health AI implementation is one of the few roles explicitly designed for that translation layer.

Each university hire in this area pulls a little more capacity into prospective evaluation, data sharing agreements and model documentation. Without those roles, health systems end up with two bad defaults: freezing AI out entirely because the evidence is too thin, or adopting tools too quickly and spending years rebuilding clinician trust.

What this means for your lab

For research groups working on clinical AI, the practical lesson from this role is straightforward: document the implementation artefacts alongside the model code. Most labs archive models and papers. Fewer archive the ethics amendments, data dictionaries, clinician feedback logs and decision rules that made the model usable in a live service. When a postdoc leaves, that unwritten knowledge leaves with them.

The exception is groups that treat implementation as a first-class research output. They write an implementation protocol before the model is built, log every data transformation, and keep the governance paperwork in the same repository as the code. Those labs produce candidates who can describe exactly how a tool changed a clinical workflow, not just how it performed in a held-out test set.

For an applicant, that means the QUT panel is offering a different kind of academic currency. A paper in a top machine learning venue still helps, but it does not substitute for evidence that you have managed the messier parts of deployment: the data custodian who says the fields are not collected consistently, the nurse unit manager who asks why the alert fires after discharge, the ethics committee that wants a plan for model drift, and the security team that wants a data flow diagram.

The concrete next step

Before applying, read the full position description and selection criteria on the QUT recruitment portal, then map one implementation project from your own work to each requirement. If you cannot name the workflow change, the data provenance issue, the governance approval or the model monitoring plan that followed, that is the evidence to build next. The AcademicJobs listing carries the application link and current close date.