AI Model Aims to Predict Dialysis Risk Before a Crisis

Yet many patients still “crash” onto dialysis. They arrive in hospital very sick and need urgent treatment. Those unplanned starts are linked with worse outcomes. Roughly 40% of patients begin dialysis this way. That pattern is what led Dr. Gregory Hundemer, a nephrologist and clinician-researcher at The Ottawa Hospital and assistant professor at the University of Ottawa, to ask a straightforward question: can we tell who is likely to need dialysis soon, and act before it becomes urgent?

Kidney clinics generate a steady flow of lab results and clinical data. Physicians look at trends, symptoms and experience to judge how things are progressing. But when you are following many patients over long periods of time, subtle patterns can be difficult to spot consistently. Dr. Hundemer and his team set out to build an artificial intelligence (AI) model that estimates a patient’s chance of needing dialysis within the next six to twelve months. Not years from now, but in the near term when there is still time to plan. 

With access to 15 years of patient data, Dr. Hundemer saw an opportunity to use machine learning (ML) to detect patterns that are hard to track reliably in routine care, in a field that has traditionally relied on more conventional prediction tools. He partnered with The Ottawa Hospital and University of Ottawa colleagues Ran Klein, Ayub Akbari, and Chris McCudden, along with James Green from Carleton University, and with PhD trainee Martin Klamrowski who has been leading work on both the model itself and the interface clinicians would use.

For Dr. Hundemer, the collaboration has been as important as the technology. “It’s the most interdisciplinary team I’ve worked with,” he said. “I don’t have the AI or machine learning expertise at all. I have the kidney expertise and they have the AI and ML expertise. Bringing those strengths together has been refreshing. You work with people who have different expertise and perspectives than you do, and you get a better product in the end.”

When the team compared their model to more traditional approaches, the results were encouraging. Dr. Hundemer said the AI-based method outperformed conventional models. One reason is that it can account for real-world complexity, including competing outcomes such as death, which is something many older models did not handle well. In a population with serious illness, that matters. He noted that early data from testing is quite promising, as validation work continues.

In practice, the tool is designed to support routine care. When a patient comes into the clinic, the physician reviews the same lab results and clinical information they always have. The model adds a short-term risk estimate based on changes over time, generating a patient’s odds of needing dialysis within six to twelve months.

The team has been thinking carefully about how different users would interact with the tool so that the user interface meets their different needs. Patients, for example, do not want to see complex trend lines or raw lab values. They want a clear answer. Clinicians want to understand what is driving the risk estimate. Administrators may use the tool to flag outliers, learn from missed cases and improve the model over time.

The team has begun integrating the model within kidney care at The Ottawa Hospital using the Epic platform. The project was also selected by The Ottawa Hospital as one of the initial research initiatives to be studied through Epic Cosmos, which draws on data from healthcare centres worldwide and supports broader validation. Provincial validation is underway in Ontario and Alberta. The team has also engaged stakeholders, including clinicians, patients and the Ontario Renal Network, as part of planning for future rollout.

The project has received support from multiple sources, including the Faculty of Medicine’s Artificial Intelligence (AI) Seed Funding Program. The AI Seed Funding Program is a uOttawa Faculty of Medicine initiative that offers funding for one-year projects that build AI capacity in health and medicine. It supports projects that move AI into real-world use and is backed by the AI Knowledge Translation Fund. For Dr. Hundemer’s team, the seed funding helped advance the model locally, supported testing with provincial datasets and contributed to development of the user interface. Other project funders included The Ottawa Hospital Academic Medical Organization and Canadian Institutes of Health Research.

Next steps depend on additional support. The team plans to conduct a cost-effectiveness analysis to better understand what implementation would mean for the health system compared with the costs of unplanned dialysis starts and hospital care. They are working toward further grants to support an interventional trial and are planning a clinical trial. If those steps are successful, the model could be adapted beyond a single centre. Commercialization is also part of the broader vision for the project, but the priority remains rigorous validation first.

If the evidence holds, the tool could be implemented nationally and even internationally. For Dr. Hundemer, the objective is practical. If clinicians can see which patients are likely to need dialysis in the near future, they can prepare, with fewer rushed starts, more time to plan and better care when it matters most.