Table 1 PULSE program recommendations for addressing core barriers surrounding the development and implementation of AI-augmented Clinical Decision Support (AI-CDS)

The secondary use of health data resources for purposes that may influence future patient care decisions requires appropriate justification and informed patient consent

Transparently develop and disclose a justifiable need to patients for the secondary use of their health data to support AI-CDS, inclusive of informed patient consent and research ethics board (REB) oversight.

Inconsistent data access practices can lead to non-compliant data usage and inappropriate data migration

Centralize data access under common legal data sharing agreements (DSA) established with the data custodian. Establish custodian approved environments for investigator and/or partner access to curated data resources with appropriate compute resources.

Lack of clear governance for data protection, usage, and monitoring

Establish objective governance structures for establishing and maintaining policies for data protection, acceptable usage, access security, and use monitoring.

Inconsistent or unreliable de-identification and data migration processes can threaten patient privacy

Develop and validate scalable, privacy-compliant, and auditable data de-identification and migration pathways for structured and unstructured data to destination servers, inclusive of data lineage and provenance logs.

Lack of structured data stewardship and oversight processes for post-access data monitoring

Establish data stewardship processes for the monitoring and reporting of data resource usage across academic, clinical and external partner activities.

Poorly validated or inconsistent data transformations limits real-world implementation of trained AI-CDS

Establish, validate and version-lock data transformations, quality assurance processes, and pre-processing pipelines for the generation of reproducible data products appropriate for model training and clinical inference. Execute longitudinal surveillance for concept drift.

Lack of data interoperability prevents the adoption and external validation of trained AI-CDS

Develop and maintain common data models as versioned data schemas that are compliant or contextualized by validated data ontologies, ensuring interoperability with international and National Institute of Standards and Technology (NIST) data standards.

Poor and bureaucracy-ladened access processes to health data prolongs innovation cycles and prevents appropriate model surveillance

Provide hierarchical, administratively controlled, and auditable access to regularly refreshed and curated de-identified data products for approved investigators or strategic partners within intuitive and cloud-accessible environments.

Lack of systematic documentation and tracking of program outputs limits support for ongoing investment

Implement tracking and reporting of program outputs to assess impact on scientific productivity, intellectual property development, technology transfer, and healthcare outcomes.