Fairness and explainability review of a health underwriting model

The model has reasonable discriminative power, but AUC 0.81 is not enough to justify deployment. The validation findings are the main concern. Predicted claim rates are 22% higher than observed rates in three lower-income pin codes, and the denial rate in those areas is 1.8 times the portfolio average. That pattern could reflect genuine risk differences, but it could also reflect proxy discrimination, missing variables or overfitting to historical underwriting bias.

Pin code and occupation category are common proxies for socioeconomic status. A third-party credit score is also a sensitive feature in insurance underwriting and may not be permitted or may create unfair outcomes. The regulator requires explainability for adverse decisions, and a gradient-boosted model is not inherently explainable. The proposed manual review only above the 95th percentile is too narrow: most adverse decisions will occur below that threshold, and those customers will have no human review.

The model owner's timeline is ambitious. Deploying next month would not allow time to fix the fairness issues, test the explanations with underwriters, or put monitoring in place. My initial view is that deployment should not be approved as proposed. A limited pilot with controls could be considered, but only after the feature set is reviewed and the model is recalibrated.

I would require the following conditions:

• Remove or justify sensitive proxy features. Pin code, occupation category and credit score should be tested for disparate impact. If they cannot be justified by a legitimate actuarial relationship, remove them or use them only in a way that does not proxy for protected characteristics.
• Recalibrate the model by segment. The 22% overprediction in lower-income pin codes must be corrected, and the denial rate should be within a documented tolerance of the portfolio average.
• Provide reason codes for every adverse decision. Use an explainability method that underwriters and customers can understand, and test it for stability.
• Require human review for all declines and for any case near the threshold, not just above the 95th percentile. Give customers an appeal route.
• Set a pilot with a defined population and a kill switch. Compare outcomes against the current process.
• Put monitoring in place for drift, segment-level denial rates, complaints and claim experience. Report monthly to the model risk committee.
• Document the model, the validation, the fairness testing and the decisions. Obtain sign-off from compliance and legal.

If these conditions cannot be met before the proposed launch date, the launch should be delayed. The insurer should not trade speed for fairness and explainability.

To the model owner, I would be direct but constructive. I would acknowledge the model's predictive power and the commercial pressure, then explain that the validation results create regulatory and conduct risk. The 22% overprediction and the 1.8 times denial rate are not minor calibration issues; they could lead to customer harm and regulatory scrutiny. I would propose a joint remediation plan with clear dates: feature review, recalibration, explainability testing and a pilot. I would offer to work with the data science team rather than simply block the project.

To the distribution team, I would explain that the model is not yet approved and that no adverse decisions should be based on it. I would describe the customer-facing risks in practical terms: a customer in a lower-income area may be declined without a clear reason, which could lead to complaints and reputational damage. I would ask them to identify any existing practices that rely on the model and to prepare for a manual review process during the pilot.

Both audiences need the same core message: the model may be useful, but it is not ready. The decision is about protecting customers and the insurer, not about rejecting data science. I would put the conditions and the timeline in writing so there is no ambiguity.