Question 290
A hospital classification model has a high sensitivity score but very low precision, causing many unnecessary interventions. What should reviewers examine?
Explanation
Correct Answer
A single score cannot capture competing error costs; deployment thresholds must reflect the harm of both error types.
Incorrect Answers
Run representative missing-feature stress tests and evaluate resulting performance disparities. That operation would be relevant if: A data scientist wants to compare error patterns when one critical input feature is missing during inference. Its corresponding reason or verification is: Robustness checks reveal whether incomplete production records can trigger unsafe or systematically biased predictions. However the problem in this question is: A hospital classification model has a high sensitivity score but very low precision, causing many unnecessary interventions. The requirement instead calls for Assess false-positive harms and the decision threshold under clinical constraints; the question-specific check is: A single score cannot capture competing error costs; deployment thresholds must reflect the harm of both error types.
Evaluate error rates and fairness metrics separately for relevant demographic cohorts. A separate source case supports that approach when: An automated lending classifier has high overall accuracy but significantly higher false-negative rates for one demographic group. The supporting rationale or observable result is: Aggregate performance can hide unequal outcomes, so group-specific analysis is needed before claiming responsible deployment. This question describes a different issue: A hospital classification model has a high sensitivity score but very low precision, causing many unnecessary interventions. Address it through Assess false-positive harms and the decision threshold under clinical constraints, with the required justification or evidence: A single score cannot capture competing error costs; deployment thresholds must reflect the harm of both error types.
Test proxy-feature impact and potentially reduce or constrain unfair reliance on that feature. This is useful for the distinct issue: A responsible AI review finds the model relies heavily on a postcode proxy for protected characteristics. The relevant technical rationale or evidence for that issue is: High predictive utility does not excuse discriminatory proxy behavior; the feature's effects require explicit fairness analysis. Here the reported issue is: A hospital classification model has a high sensitivity score but very low precision, causing many unnecessary interventions. The relevant response is: Assess false-positive harms and the decision threshold under clinical constraints. The first result cannot verify this different correction: A single score cannot capture competing error costs; deployment thresholds must reflect the harm of both error types.