📄 health informatics

Demographic Calibration Gaps in Breast Cancer Risk Prediction: Introducing the Demographic Calibration Gap Score

This paper introduces the Demographic Calibration Gap Score (DCGS) to demonstrate that standard global calibration methods fail to address systematic prediction errors across racial and gender subgroups in breast cancer risk models, particularly under distributional shifts, thereby highlighting the need for subgroup-specific calibration metrics to prevent biased clinical decisions.

Original authors: Eniolade, M.

Published 2026-06-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eniolade, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: The "Average" Lie

Imagine you are a weather forecaster. You tell your town, "On average, there is a 70% chance of rain tomorrow." If you look at the whole town, you might be right. But what if your forecast is perfect for the sunny suburbs, yet completely wrong for the rainy downtown area?

In the world of medical AI (specifically for breast cancer), researchers have been very good at building models that predict who has cancer. They are like weather forecasters who get the "average" right. However, this paper points out a dangerous blind spot: Most studies only report the "average" accuracy. They don't check if the model is lying to specific groups of people, like Black women or men.

If a model is "calibrated," it means when it says "70% chance of cancer," it is actually right 70% of the time. If it's not calibrated, a doctor might think a patient is safe when they aren't, or vice versa. The paper argues that while the average model might look good, it could be systematically wrong for specific racial or gender groups, leading to unfair and dangerous medical decisions.

The New Tool: The "Demographic Calibration Gap Score" (DCGS)

To fix this, the author, Michael O. Eniolade, invented a new measuring stick called the Demographic Calibration Gap Score (DCGS).

Think of it like a fairness ruler.

Old way: You measure the whole class to see if the average student passed.
New way (DCGS): You measure the difference between the best-performing group and the worst-performing group.

If the ruler shows a big gap (specifically, if the error rate differs by more than 5 percentage points between groups), the model is considered "unfair" or "clinically dangerous," even if the average looks fine.

The Experiment: A "Stress Test"

The author didn't just talk about this; he ran a tough test to see how well current tools work.

The Training Ground (Wisconsin Dataset): He taught five different computer models using a dataset of breast cancer cells (like teaching a student with a specific textbook). These models did great on the test they were trained for.
The Real World Test (MIMIC-IV Dataset): Then, he took those same models and tried to use them on a completely different group of patients from a hospital emergency room.
- The Analogy: Imagine teaching a student to drive using a video game, and then immediately handing them the keys to a real truck in a snowstorm. The features are totally different (game controls vs. real pedals).
The Result: As expected, the models got confused. Their overall accuracy dropped significantly because the "textbook" didn't match the "real world."

The Shocking Discovery: "Global" Fixes Don't Work

The researchers tried to "fix" the confused models using standard methods (like Platt Scaling and Isotonic Regression).

The Analogy: Imagine a teacher trying to help a whole class of students who are struggling with a math test. The teacher gives one single hint to the entire class to improve their average score.
The Reality: The paper found that this "one-size-fits-all" hint often made things worse for specific groups.
- It might lower the error for White patients but accidentally make the error higher for Black patients.
- It might fix the error for women but leave men with terrible predictions.
- Key Finding: Even when the "average" error went down, the gap between the groups often stayed huge. The global fix was like putting a bandage on a broken leg; it looked better from a distance, but the underlying injury remained.

The "Subgroup" Attempt

The author also tried a different approach: giving a different hint to each racial group (Subgroup-targeted calibration).

The Result: This helped a little bit, but not enough. In some cases, it actually made the gap wider.
Why? The author explains that for some groups, there just weren't enough patients in the test data to teach the model a new, specific rule. It's like trying to teach a new language to a student when you only have 30 minutes of class time; the lesson is too rushed and ends up confusing them more.

The "Confidence" Trap

The paper also looked at "Conformal Prediction," which is a way for AI to say, "I am 90% sure my answer is in this range."

The Result: When the models were tested on the new hospital data, their confidence collapsed. Instead of being 90% sure, they were only right about 25% of the time.
The Gap: Even worse, this low confidence wasn't shared equally. Some groups got "covered" by the safety net, while others were left completely exposed.

The Bottom Line

The paper concludes with a simple, powerful message:
You cannot fix fairness by just looking at the average.

If you build a medical AI, you must check if it works equally well for everyone. The author provides a free, open-source tool (the DCGS library) that allows researchers to easily measure these gaps. If the gap is too big (over 0.05), the model is not ready for real-world use, no matter how good its "average" score looks.

In short: A model that is "mostly right" can still be dangerously wrong for specific people. We need a new ruler (DCGS) to measure that unfairness before we let these tools into hospitals.

Technical Summary: Demographic Calibration Gaps in Breast Cancer Risk Prediction

Problem Statement
Current machine learning (ML) research in breast cancer prediction frequently neglects calibration reporting, with systematic reviews indicating that only a small fraction of studies assess calibration, and even fewer examine it across demographic subgroups. While models often achieve high discrimination (AUROC), predicted probabilities that are systematically miscalibrated for specific racial or gender groups lead to biased clinical decisions. Aggregate metrics, such as global Expected Calibration Error (ECE), mask these disparities; a model can be well-calibrated on average while assigning systematically biased probabilities to specific subgroups. Existing fairness metrics like equalized odds address ranking parity but do not capture calibration equity, where a model may assign probabilities that are consistently too high for one group and too low for another.

Methodology
The study introduces the Demographic Calibration Gap Score (DCGS), a metric designed to quantify the variation in calibration error across demographic subgroups. The research employs a two-dataset stress-test design:

Training and Internal Validation: Five classifiers (Logistic Regression, Random Forest, SVM, Gaussian Naive Bayes, XGBoost) were trained on the Wisconsin Diagnostic Breast Cancer dataset (n=569), utilizing FNA nucleus measurements.
External Validation: Models were evaluated on a breast cancer cohort from MIMIC-IV (n=1,316), derived from emergency department triage vitals. This deliberate feature mismatch (cytology vs. vitals) simulates a severe distributional shift typical of real-world deployment.

Calibration Conditions:
The study evaluated models under four conditions:

Base: No post-hoc calibration.
Platt Scaling: Global sigmoid calibration fitted on the Wisconsin validation set.
Isotonic Regression: Global monotone piecewise-constant calibration fitted on the Wisconsin validation set.
Subgroup-Targeted Platt Scaling: Applied only to the MIMIC cohort, fitting separate Platt scalers per racial group (with a global fallback for small subgroups).

Metrics:

DCGS: Defined as $DCGS_{range} = \max_k(ECE_k) - \min_k(ECE_k)$ across $K$ demographic subgroups (race and gender). A threshold of 0.05 is proposed as clinically meaningful, representing a 5-percentage-point difference in accuracy between the best and worst-calibrated subgroups.
Demographic Coverage Gap (DCG): Applied to conformal prediction to measure inequality in prediction set coverage across subgroups.
Standard metrics (AUROC, Brier Score, ECE) were also computed.

Key Results

Internal Performance: On the Wisconsin dataset, all models achieved high AUROC (>0.98) and low ECE (<0.12). Calibration methods showed inconsistent effects on ECE depending on the base model.
External Performance (Distributional Shift): On the MIMIC cohort, global calibration failed to maintain utility. Base and globally calibrated models suffered severe performance drops, with AUROC falling to 0.45–0.57 and Brier Skill Scores turning negative, indicating performance worse than a naive prevalence baseline.
Demographic Calibration Gaps:
- DCGS Threshold Exceeded: DCGS exceeded the 0.05 clinical significance threshold in 28 of 40 model-calibration combinations on the race axis.
- Ineffectiveness of Global Calibration: Neither global Platt scaling nor isotonic regression reliably reduced DCGS below 0.05. In some cases (e.g., XGBoost), global isotonic calibration reduced aggregate ECE but left racial DCGS unchanged.
- Subgroup-Targeted Calibration: While subgroup-targeted Platt scaling improved discrimination (AUROC rose to 0.59–0.75 for some models), it did not consistently close calibration gaps. Notably, applying subgroup-targeted calibration to Naive Bayes amplified the racial gap (DCGS increased to 0.602), demonstrating that subgroup calibration on small, sparse samples with poorly behaved base models can worsen disparities.
Conformal Prediction: Marginal coverage guarantees (target 90%) collapsed to approximately 25% on MIMIC due to distributional shift. Furthermore, the Demographic Coverage Gap (DCG) exceeded 0.15 for all model-variant combinations, indicating that even low aggregate coverage was distributed unequally across groups.

Significance and Claims
The paper claims that reducing population-level ECE through global recalibration does not reliably close demographic calibration gaps. The authors argue that:

Visibility of Disparities: DCGS provides a direct, standardized method for researchers to detect and report calibration disparities that aggregate statistics miss.
Clinical Relevance: A DCGS above 0.05 implies that predicted probabilities have different meanings for different demographic groups, potentially shifting patients across clinical decision thresholds (e.g., biopsy referrals) based on race or gender rather than risk.
Limitations of Current Methods: No single global calibration method tested was sufficient to ensure equity across diverse external cohorts. The study suggests that subgroup-targeted approaches are necessary but require larger per-group sample sizes to be effective.
Resource Availability: The authors have released the DCGS computation library and full evaluation pipeline as open-source Python code to facilitate the adoption of subgroup-stratified calibration reporting.

The work concludes that population-level metrics are insufficient for equitable clinical AI and that researchers must compute and report DCGS on external cohorts to ensure models do not propagate health disparities.