Question

A healthcare startup deploys a Vertex AI model that classifies dermatology images as benign or requiring specialist referral. Regulators require that the automated system minimize harmful false negatives (missed cases needing referral). Product wants to reduce clinician workload but not at the cost of patient safety. Which monitoring and operational strategy best balances responsible AI risk with efficiency?

Accepted Answer

Route predictions where the model's confidence falls below a calibrated threshold to a human reviewer, and continuously monitor the false-negative rate on reviewed cases to recalibrate the threshold.

Answer

Fully automate all decisions using the model's argmax output and only audit a random 1% sample monthly to control cost.

Answer

Set the classification threshold to maximize overall accuracy, since accuracy is the metric regulators care about most.

Answer

Retrain the model weekly on all new data without human review, assuming continual learning will eliminate false negatives over time.