QuestionQ118

VerifiedAI Safety, Security, and Governance

A hospital is developing an AI application to assist medical clinicians with treatment decisions. The application uses Amazon Bedrock to analyze patient case histories and recommend diagnoses. The application must sustain sub-500 ms response times to integrate with the hospital’s existing real-time clinical workflow. To comply with privacy regulations, the application must log every personally identifiable information (PII) handling decision for audits. The application must identify and remove PII from responses with at least 99% accuracy.

Following the initial deployment, clinicians report that diagnostic summaries from the application sometimes contain patient names and medical record numbers that were not in the original case-history inputs. An investigation shows that Amazon Comprehend Medical successfully identifies and removes PII from inputs with 95% accuracy, and the application substitutes tokens for every detected entity before sending inputs to Amazon Bedrock. However, the application still generates patient-identifying information in approximately 3–5% of outputs.

The company needs a solution that prevents the application from displaying PII in outputs while satisfying all other operational requirements.

Which solution meets these requirements?

Explanation

Amazon Bedrock Guardrails sensitive information filters can evaluate model responses for PII and be configured to block the response or mask detected entities, including predefined PII types and custom regular-expression patterns. Applying the filter to outputs supplies the required control at the point where the leakage occurs; guardrail intervention results can be recorded as the application’s PII-handling audit decisions.

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!