Navigating AI Research at PHSA

This page provides a framework for responsible AI research and a toolkit with resources and infrastructure available to anyone doing research within PHSA who aims to design, develop, or evaluate AI-based interventions for healthcare.

The AI Research Lifecycle

The AI research lifecycle separates the research process into three phases, each with its own objectives and focus (adapted from McCradden et al. (1)).

Before initiating AI research at PHSA that aims to impact clinical populations or services, it is critical to define the research hypothesis in partnership with clinical or administrative leaders to ensure that the research meets a real need within our healthcare settings. In addition, post-deployment monitoring aimed at auditing model performance and identifying when a model update is needed is a critical step to close the loop of development and should be planned in consultation with the PHSA PDHIS Digital Innovations team and clinical or administrative program leaders. Post-deployment monitoring needs to be embedded within the program adopting AI, is necessary for the success of AI in our health system and can lead to further research. These initial and final processes are represented as a feedback loop in the lifecycle shown below.

AI research lifecycle_Goals for Each Phase V2.jpg

A toolkit for AI Research at PHSA

The downloadable AI research Toolkit aims to support responsible AI research across the AI research lifecycle. In this toolkit you will find links to infrastructure available through PHSA that can support AI research and useful guidelines and frameworks, including researcher guidance on bias and transparency in AI research. The guidance table on bias and fairness provides considerations and actions for researchers at each phase of the AI research lifecycle to identify and, where possible, mitigate bias so that research output is equitable.

The toolkit is meant to be a living document. An interactive and expanded version is available to anyone with a PHSA account here. We encourage anyone conducting AI research at PHSA to review, edit and add to this knowledge base.

Phase 1: AI Model Development and Validation

‎

The focus of this phase is to define the hypothesis you will test with AI methods and then develop and validate a model(s) to test this hypothesis. Your hypothesis should be grounded in a clinical need and an understanding of the context in which you plan to implement the AI tool in the future.

Data used in this phase of research is retrospective. Depending on the modelling approach, large datasets may be needed to develop an initial model/algorithm. Engagement with operational program leaders and the PHSA Digital Health Innovations team to determine if your project aligns with PHSA priorities and guidelines is recommended at this stage.

Regulatory approvals such as ethics approval and data access requests must be completed. Researchers can access the AI/ML ethics checklists for the PHSA Research Ethics Boards that are required in addition to the standard RiSE application when submitting a proposal that includes AI/ML model development in the AI research Toolkit.

At the end of this phase, you will have tested your hypothesis, developed and internally validated your model on your test dataset plus other data unseen by the model, if possible. This is often called a proof of concept.

You will find information on accessing retrospective PHSA data for research and suggestions for how to access needed compute infrastructure in the PHSA AI research Toolkit. Other useful resources for this Phase include the TRIPOD-AI statement and FUTURE-AI, which provides guidance on best practices for model development and reporting.

Phase 2: Silent Prospective Validation

In this phase of AI research, the focus shifts to evaluating impact of the proposed AI on clinical workflow and measuring performance of the AI model when applied to local data and populations. Model validation activities in Phase 2 of the AI research lifecycle are sometimes referred to as external validation. This Phase also includes development and testing of needed data pipelines and solution architecture to enable the AI tool to work in the healthcare setting. Aspects of this Phase can be integrated with Phase 1. For example when model development includes external validation with datasets from additional centres similar to the target implementation site. Workflow studies are also not needed if the AI tool being developed is not intended for use in clinical decision making or care.

Data used for this phase are prospectively collected, and the output of the AI tool is blinded to the user so that it does not impact clinical outcomes or decision making. Model performance and output is audited to confirm that it is working as expected.

At the end of this phase, you will have recalibrated the model as needed, estimated model performance in context to establish a baseline for future research and evaluation phases and defined how the algorithm can be best integrated into clinical workflows.

Information on accessing needed compute infrastructure at PHSA, connecting with Data Analytics support, frameworks for model output auditing and other resources needed during this phase can be found in the PHSA AI Research Toolkit.

‎

Phase 3: Clinical Implementation and Evaluation

The focus of the third phase of AI research is to generate evidence of model effectiveness and evaluate implementation success in a real clinical or operational setting.

In this phase, the AI's outputs are visible to its users and may be used to directly or indirectly impact clinical care and decision-making. Ensuring proper approvals, which may include ethics and/or privacy and security review, is critical prior to starting this phase of AI research.

Data is collected prospectively to measure the tool's impact on pre-defined clinical effectiveness and implementation outcomes. Performance auditing remains important. If the research outcome is positive, at the end of this phase, an evidence-based and valid AI tool will be ready for scale-up and sustainment. If outcomes are negative or not as expected, researchers may return to Phases 1 and 2 to iteratively update the model until desired outcomes are achieved.

During Phase 3, researchers should engage with operational leaders and PHSA Provincial Digital Health and Information Services to begin planning for sustainment of the AI intervention after the research has succeeded.

Resources in the PHSA AI research Toolkit relevant to this phase of research include recommendations and guidelines for evaluation, monitoring and reporting of AI implementations in healthcare.

Conducting Research at PHSA