Implementing machine learning models within pharmaceutical R&D and clinical data can be a complex challenge. Whilst modern generative and predictive models may need to process large volumes of data quickly and be easily adapted to new use cases, regulators require systems to have full determinism and auditability. This means you need to ensure that all your processes are fully compliant with the ALCOA+ standards (Attributable, Legible, Contemporaneous, Original and Accurate and Complete, Consistent, Enduring and Available).
AI systems must be architected to ensure that models are trained within a compliant framework.
AI Architectural Challenges
Machine learning models are typically designed to be fast, scalable and accurate. When deploying models to GxP environments you run into some compliance issues when using non-compliant models:
Data lineage isn’t clear. Data pipelines will often lose metadata during transformations.
Contemporaneous logs aren’t present. With complex agents you will often store temporary results and checkpoint decisions in memory.
No control over autonomous decision making. If you have a fully autonomous pipeline there could be decisions made without being observed by a human.
As an architect you want to include governance.
ALCOA+ Compliant AI Orchestration
As an engineer when creating AI pipelines, you want to include data governance throughout.
Tamper Proof Logging
All data coming into the pipeline, any model inferences and metadata should be logged. This is done by using append only logs. During data ingestion you will include cryptographic hashes to ensure datasets can’t be altered.
Human-in-the-Loop (HITL)
Any critical decisions such as changes to trial protocols, marking patients for recruitment and writing submissions should have a human verification step enforced. The engine should stop running until an authorised user has signed off the change.
Real-Time Predictive Monitoring
Whenever possible the orchestration layer should monitor your telemetry data. If any anomalies are discovered such as data distributions, model drift or unexpected inferences then you want to include circuit breakers within your pipeline.
Engineers often use specialist middleware to help with the implementation of these architectures. You can find more information on how to integrate them within this guide on ALCOA+ data integrity within pharma AI.
Implementation Strategy
As an engineer you should include these steps when creating your pipeline.
Separate Compute from Compliance. Keep your heavy compute separate from your compliance. Ensure that any state changes are done so using validated orchestration.
Version Data, Code and Artifacts. Ensure that any changes to datasets, feature stores, model weights and pipeline config are versioned. Never run any inferences on unversioned data.
Automate Verification. Create automated regression tests and data validation tools to ensure your system is deterministic and your logs are accurate.
Conclusion
Creating a compliant system within pharmaceutical AI doesn’t mean you have to limit yourself. Creating resilient and transparent software will help you create high performing AI systems. Using the principles above you will ensure that your workflow pipelines are based on tamper proof logs, real time monitoring and human oversight to meet the high standards set by global health authorities.
Top comments (0)