Machine learning can make risk-based quality management more sensitive and timelier, but it also produces more signals than most teams can act on. The value of AI in RBQM depends less on the algorithm and more on how well a sponsor defines what matters and who responds.
Risk-based approaches are now firmly embedded in the regulatory framework. ICH E6(R3) emphasizes quality by design, critical-to-quality factors and proportionate oversight. Centralized monitoring, where data is reviewed remotely for anomalies and trends, is a core tool. It is also a natural home for AI.
How is AI used in centralized monitoring?
Statistical methods have supported central monitoring for years: key risk indicators, quality tolerance limits and outlier detection across sites. Machine learning extends this by detecting subtler patterns. Examples include unusual consistency in vital signs that may suggest fabricated data, digit preference in recorded values, or subtle drift in adverse event reporting rates at a site over time.
These are genuine advances. Some data integrity issues that once surfaced only at an on-site visit, or not at all, can now be caught in weeks.
The signal problem
More detection creates a new challenge. An algorithm that flags forty sites a month for review is not helpful if the central monitoring team can only investigate eight. Without calibration, teams start dismissing alerts. Alert fatigue is a quality risk in its own right.
The answer is disciplined design. Tie each signal to a critical-to-quality factor identified during protocol development. Define thresholds deliberately and revisit them as the study matures. Document why a flagged signal was closed without action, because an inspector may ask.
Keep judgment in the loop
AI can tell you a site looks different. It cannot tell you why. A statistical anomaly may reflect fraud, a training gap, a different patient population or a legitimate clinical practice variation. Determining which requires a CRA, a data manager or a medical monitor who understands the context.
Sponsors should make sure their RBQM plan specifies who reviews AI-generated signals, within what timeframe, and with what authority to escalate. This is especially important when the CRO runs central monitoring, since sponsors often receive a summary dashboard rather than the underlying logic.
What sponsors should look for
When evaluating a CRO's RBQM approach, ask to see a real, redacted example of how a machine-generated signal was investigated and resolved. Ask how many signals were generated on a comparable study and what percentage led to action. Ask how thresholds were set. The answers reveal whether the process is operationally mature or simply technically impressive.
RBQM was always about focusing attention where risk lives. AI sharpens the lens. People still decide where to point it.
