PRIMED-AI Error Mitigation and Technical Management Plan

The Error Mitigation and Technical Management Plan as described below will be required of all applicants to the Validation Center (RFA-RM-27-015), Data-to-Model Academic-Industrial Partnerships (RFA-RM-27-012), and Model-to-Clinic (RFA-RM-27-013) PRIMED-AI NOFOs.

Page limit: 3 pages

The applicant must provide a detailed plan to describe how they will mitigate error and how their project will be monitored and managed both during development and once the CDS tool is established. The following components must be included as an attachment where required. The components listed as optional are included here for awareness of the Program’s goals more generally. The plan must address error mitigation in data collection, data preparation, model development, evaluation and deployment, as well as technical management details for each of the following:

Data Collection – provide a detailed description around how error introduction will be avoided when collecting data to develop multimodal AI tools. When compiling multimodal data, the data collection process should avoid error introduction as much as is feasible for the proposed work while maintaining maximal benefits and minimal harms for patients/participants. When obtaining multimodal data, the following should be considered: 

• Adequate (scientific and statistical) representation of relevant patient populations or data points, including relevance to target patient population the model is being developed for. 
• Incomplete or missing data, which could impact the quality or applicability of the data to the project. Missing data are a common occurrence, particularly when combining several data types, and can have a significant effect on the usefulness, applicability, and/or conclusions that may be drawn from the data. 
• Errors associated with survivorship, or the reliance on existing data at the expense of other data (e.g. historical data, attrition data, etc.). 
• Data augmentation and/or creation of synthetic data. 
• Missing data modalities, and how inclusion or exemption of some of these data modalities could affect the model. 

Data Curation and Harmonization – the PRIMED-AI program expects that all CDS tools developed over the course of the program utilize multimodal data in order to effectively personalize care for the intended context of use. Using data from disparate sources will require data curation and harmonization, or the process of aligning and standardizing data from various sources into a consistent format to create a single, unified view to ensure data quality, consistency, accessibility, and minimize the introduction of errors. Applicants must provide a summary of the methods they plan to employ to harmonize the data. All award recipients are expected to work collectively to ensure data utilized by PRIMED-AI tools is harmonized to the standards of the PRIMED-AI program, to be established by the PRIMED-AI SC in collaboration with the NIH PRIMED-AI program staff. 

Multimodal data Usage – These NOFOs requires leverage of existing data, shared data and/or collected data along with associated metadata to aid in developing multimodal AI CDS tools for clinical use. It is expected that errors will be mitigated whenever possible when utilizing this data – applicants are required to provide a summary detailing how they will identify and correct potential errors in existing multimodal datasets, minimize error introduction, and perform error mitigation. Potential sources for introduction of error include incomplete or erroneous data sets, data provenance issues, system compatibility issues between environments, inaccurate documentation of data, and tolerance levels related to quality control measures. Applicants are responsible for minimizing the introduction of error at all steps of the project. 

Tool/Software Model/Algorithm Development – Development of multimodal AI tools utilizing patient data provides the opportunity to create personalized treatments for all patients. However, as multimodal AI software tools are developed, it is important to understand when and where error might be introduced. Applicants proposing projects under these NOFOs are expected to utilize Good Software Engineering practices when developing multimodal AI tools. These practices include but are not limited to readability (software written to promote understanding by others), resiliency (software designed to prevent or mitigate failures within the system or components of the system), and reusability (software that can be utilized by various users or in other environments). It is also expected that multimodal AI tools will be developed by multi-collaborative teams to ensure that all stakeholders are familiar with the development of the tool and have input into all steps of its development. Applicants must describe how they will address these needs during the development of their tool/software model/algorithm. 

Tool/Software Model/Algorithm Sharing – Resources generated by the PRIMED-AI Consortium may include, but are not limited to, multimodal AI models, protocols, computational algorithms, SOPs, etc. PRIMED-AI tools should be shared with the aim of limiting restrictive permissions, and to ensure all patients and clinicians have access to relevant clinical tools as much as is feasible. In addition, tools should be shared in such a way that minimizes error introduction and provides avenues for error mitigation in future uses of the tools – this must be accomplished through validation and reporting on appropriateness of the tool in coordination with the Validation Center. Applicants must provide a short summary for how the proposed project will address these sharing needs, both with the Validation Center within the PRIMED-AI Consortium and externally. All tools generated by the Consortium should be made accessible as soon as possible, and no later than the time of an associated publication, or the end of the award/support period, whichever comes first. If possible, all tools should be posted and available on the PRIMED-AI web portal managed by the Logistics Center. 

Clinical Implementation – The goal of the PRIMED-AI Program is to develop artificial intelligence (AI) technologies to enable integration of clinical imaging data with a variety of other health data to support the clinical decision-making process. It is expected that multimodal AI tools developed in response to these NOFOs and during the program would ultimately be adopted into clinical workflows. To ensure that these tools are relevant to patient populations, widely available, and will function effectively in environments in which they are implemented, applicants should address error mitigation and technical management of these tools in the clinical environment. It is expected that applicants will perform real-world implementation validation via sandboxing, test runs with expected patient populations, and through validation with the Validation Center. Where applicable, it is required that applicants ensure that tools developed in their application have all appropriate Institutional Review Board (IRB) approvals and an appropriate Data Safety and Monitoring Board (DSMB) in place prior to beginning work with human participants for CDS tools. Approvals obtained from both the IRB and DSMB must take into account any needs of the project that require multi-site tool testing, all relevant patient populations that the tool might be deployed to, and any other multimodal AI specific requirements of the project. Applicants must provide a detailed plan to address these points and provide evidence that their proposed tool will support clinical implementation once fully developed.

Performance Monitoring and Sustainability – The performance of AI systems drifts over time, necessitating ongoing performance surveillance. This is especially important for multimodal AI systems that utilized patient data for clinical decision support. Applicants are expected to consider the lifecycle of multimodal AI tools they propose to develop and ensure that they appropriately address how the tool will be monitored to detect drifts in accuracy beyond the initial deployment. They must also provide a description for sustaining the model after the end of the project.

All award recipients under the PRIMED-AI Program are required to attend Steering Committee meetings and participate in joint activities, including the development of standardized protocols and pipelines. Through these interactions, this error mitigation and technical management plan may be revised throughout the award period to reflect established policies and updated guidance.