Prediction to Practice: What an Autism Prediction Project Highlights for Trustworthy Medical AI

With funding from the Precision Child and Youth Mental Health Collaboratory, researchers from the Children’s Hospital of Eastern Ontario (CHEO) Research Institute, the Ottawa Hospital Research Institute, and the University of Ottawa investigated whether an AI-enabled prediction model could identify young children more likely to have autism, allowing them to be prioritized for comprehensive clinical assessment.

The team used modern AI methods designed to find subtle patterns across large, linked datasets, together with explainability tools that help show which factors may be driving a prediction. They leveraged data from the Better Outcomes Registry and Network (BORN) Ontario, the province’s perinatal, newborn and child registry, linking it with other provincial administrative health data sets, conducting the analysis at ICES.

What they found (see: Transformer-based deep learning ensemble framework predicts autism spectrum disorder using health administrative and birth registry data) was encouraging: the strongest AI model correctly identified about seven in ten children who had autism and correctly ruled out a little under six in ten who did not. While not perfect, the results suggest the model may help detect patterns that could one day support triage decisions and identify children for closer clinical attention. The model continues to be refined and validated to further improve performance before potential clinical use.

Beyond the Algorithm: Applying Systems Thinking

Study team members are also researchers at BORN Ontario, a provincial program with deep experience leveraging data and insights to drive improvements across the system of care. BORN has an AI research lab, is building an interoperable digital health record, and brings a mature systems perspective that extends well beyond model development to questions of workflow integration, governance, transparency, and scale.

“Building the algorithm is only the first step,” says Dr. Christine Armour, Investigator at CHEO Research Institute and co-Medical Director of Prenatal Screening Ontario, Clinical Genetics, at BORN Ontario and co-lead of the study. “The bigger challenge is turning it into a usable health system tool that clinicians can trust and that can work in real time.”

On its own, the initiative is a strong medical AI project tied to a real need in pediatric care. But it also creates an opportunity to reflect on a wider set of questions. Moving from research to real-world use means thinking through issues such as model explainability, data governance, workflow, scale, data quality, stigma, equity, and transparency. With stewardship of standardized, population-level data and close integration with clinical and policy partners, BORN is poised to support the equitable deployment of AI across Ontario, stewarding personal health information and addressing risks related to bias, stigma, and access while designing guardrails that promote fairness and trust.

Fit for Purpose: Triage vs. Diagnosis

A model can perform well on historical data, but the output must appear at a point in care when it can change a child’s course of care. Clinicians need enough confidence in the results to use them. That includes understanding a crucial distinction: the goal is triage, not diagnosis. Like other screening tools, from mammography to colorectal cancer tests, the model doesn't need to be perfect to be useful. It just needs to offer an acceptable level of precision to identify and navigate those who may need care to the right place at the right time.

The Risk of Labels and the Need for Governance

The human implications are just as important as the modelling. A predictive signal can easily be heard as a label. If a child is flagged as "high likelihood" before a clinician has ever seen them, who sees that information? How is it worded? Does it stay in the record forever?

Responsible data management in pediatric settings goes beyond standard privacy compliance. It requires careful consideration of the impact of a positive screen so as to minimize unnecessary medicalization, parental anxiety, and stigma. A triage signal should remain just that: a tool to identify and expedite care needs rather than becoming a durable label that impacts a child’s educational or social trajectory before a formal diagnosis is made. This is especially relevant for autism spectrum disorder, which some families may see as a difference rather than disorder.

Equity: Risks and Opportunities

Fairness cannot be assumed; it must be engineered. Models trained on health system utilization data can inadvertently reinforce inequities in access that exist in the system. Families who face barriers to care due to geography, language or socioeconomic status may appear differently in the data, not because their children need less support, but because services have been harder for them to reach.

However, there is also a significant opportunity for equity. Currently, recognition of signs and symptoms or referral practices can vary widely depending on where a family lives or which provider they see. A province-wide prediction model, built on standardized data, could help democratize access to expertise. It could offer a consistent, objective baseline that helps spread standards of care more fairly, ensuring that a child in a rural community has the same chance of being flagged for early assessment as a child in a major urban centre.

“We have the data and a systems-level view,” says Heather Howley, BORN’s Director of Research. “The early question is not only whether an algorithm works, but how to build a tool that can be used across a system of care.”

Explainability and Transparency for Trust

Clinician trust grows when they can understand which variables are contributing most to a prediction. Explainable AI methods can show how a model works so clinicians can determine whether the results make sense in the context of a child’s care. 

Public trust also depends on clarity. In the context of system-wide screening or triage, the conversation shifts from individual consent for every data point to a broader framework of transparency and notice. Families need plain-language explanations of what is happening: what data is being used, what the output means, what it does notmean and what it will not be used for.

“People’s health information has actually been at work in the system for a long time,” says Alicia St. Hill, Executive Director of BORN Ontario. “To enhance trusted and ongoing use, we must ensure health information is used to help people achieve better health outcomes. We have to be clear about how it is being used, and where the guardrails are and how tools like this could support care across the system.”

Research teams need more structure and guidance on ‘what’s next’ after a promising study. Who needs to be involved? What evidence is enough for a pilot? What governance is needed? These are some of the questions that are being tackled in a series of articles by the Ottawa Medical AI Research Institute (OMARI) focused on the pathways from AI research to Software as a Medical Device (SaMD).

The autism prediction project stands as an exciting example of practical medical AI research grounded in a real problem, built on Canadian data, that can help children and their families. It also highlights the kind of role a program like BORN can play in supporting future perinatal and pediatric AI work.

“BORN holds a level of depth, granularity and accuracy in maternal-newborn data that is unparalleled,” says Dr. Mark Walker, Scientific Director of BORN Ontario and Vice-Dean, Internationalization and Global Health at the University of Ottawa. “For AI, we have only scratched the surface of what that could make possible.”

More broadly, the project helps clarify what trustworthy medical AI projects need to move into care: clarity on purpose, strong data, explainability and transparency, subgroup monitoring to ensure fairness and thoughtful, ethical data governance. By bringing together people, resources and practical guidance, prediction models can move from promising research to useful tools that help health systems act sooner, more fairly and with greater confidence.