Technical Summary: Robust Mitigation of Age-Dependent Confounding Effects via Sample-Difficulty Decorrelation

1. Problem Statement

In medical image classification, deep neural networks often rely on "shortcuts"—readily available cues like patient age—rather than learning true pathological markers. Age acts as a confounder because it correlates with both imaging morphology (input appearance) and disease prevalence (label distribution). This leads to systematic, age-dependent performance disparities:

• Overdiagnosis: Higher false-positive rates (FPR) in age groups with higher disease prevalence.
• Underdiagnosis: Lower true-positive rates (TPR) in age groups with lower disease prevalence.

These disparities are exacerbated by train–test distribution shifts in age. Conventional mitigation strategies often enforce strict age invariance (e.g., removing age from the feature space or discretizing it). However, the authors argue that strict invariance is undesirable because age also encodes clinically meaningful variation. The goal is to suppress spurious age-linked shortcuts without discarding diagnostically relevant age information.

2. Methodology

The proposed framework intervenes in the training dynamics by targeting spurious age–difficulty trends rather than enforcing full statistical independence. The method proceeds through the following stages:

2.1 Empirical Observation and Warm-up

Following a one-epoch warm-up phase, the authors analyze the relationship between sample difficulty (quantified by gradient magnitude $g_i$, equivalent to absolute prediction error $|p_i - y_i|$) and patient age.

• Observation: A strong linear dependence exists between age and difficulty, conditioned on the label ($y$).
  • For positive cases ($y=1$), samples from older patients are "easier" (lower $g_i$) than those from younger patients.
  • For negative cases ($y=0$), samples from older patients are "harder" (higher $g_i$).
• Implication: The model over-relies on age as a predictive shortcut, leading to the observed TPR/FPR disparities.

2.2 Weighted Slope Penalty (Decorrelation)

Instead of enforcing invariance, the method penalizes the slope of the linear relationship between normalized age ($z_i$) and sample difficulty ($g_i$).

• Model: $g_i \approx \alpha_{y_i} + \beta_{y_i} z_i$, where $\beta_{y_i}$ represents the age–difficulty slope for each class.
• Objective: Minimize the squared slope estimate $\hat{\beta}_y^2$ within each mini-batch. This flattens the dominant age-related slope without removing non-linear age information.
• Robustness via Huber Weights: Direct penalization is unstable due to annotation noise and clinical variability. The authors use Weighted Ordinary Least Squares (WOLS) to estimate slopes.
  • Trend Affinity: Weights ($w_i$) are derived from a Huber regression fit performed once after the warm-up phase.
  • Samples aligning with the dominant trend receive full weight; outliers (samples deviating significantly from the trend) are downweighted. This prevents the decorrelation penalty from destabilizing optimization or penalizing highly informative samples.

2.3 Age Coverage Modulation

To ensure stability under limited age diversity within mini-batches, an Age Coverage Score ($C$) is introduced.

• Definition: $C$ is the normalized variance of the normalized age values in a mini-batch.
• Mechanism: The decorrelation penalty is scaled by $C$. If a mini-batch lacks age diversity (low variance), the penalty is attenuated.
• Benefit: This ensures the model only reduces the age–difficulty slope when there is sufficient evidence of a trend in the batch, improving robustness against age imbalances and train–test shifts.

2.4 Total Loss Function

The final objective combines standard Binary Cross-Entropy ($L_{BCE}$) with the coverage-aware slope penalty:
$$L_{total} = L_{BCE} + \lambda \sum_{y \in \{0,1\}} C_y L_{slope,y}$$
where $L_{slope,y} = \hat{\beta}_y^2$.

3. Experimental Setup

• Datasets: Two radiology benchmarks exhibiting strong age-related unfairness:
  1. CheXpert Pleural Effusion (CXP PE).
  2. MIMIC Cardiomegaly (MIMIC Car).
• Distribution Shift: Experiments utilize an age-based exponential sampling scheme to create train–test shifts. The training set is skewed toward younger populations, while the test set shifts toward older cohorts (controlled by intensity parameter $\gamma \in \{0, 4, 8\}$).
• Baselines: Compared against Empirical Risk Minimization (ERM), Resampled Batch, Adversarial Training (Adv), Just-Train-Twice (JTT), and GroupDRO.
• Metrics:
  • Fairness: $\Delta Sep_{10}$, measuring the relative change in TPR/FPR per decade of age (lower is better).
  • Accuracy: AUC (Area Under the Curve).

4. Key Results

• Fairness Improvement: The proposed method achieved the lowest $\Delta Sep_{10}$ across all shift intensities ($\gamma$) on both datasets.
  • On MIMIC (Car), it reduced the average $\Delta Sep_{10}$ by $\sim 35\%$ compared to ERM (20.82% vs. 32.26%).
  • On CXP (PE), it reduced the average $\Delta Sep_{10}$ by $\sim 56\%$ compared to ERM (7.82% vs. 17.90%).
• Robustness to Shift: Unlike adversarial training, which degraded significantly under high shift intensity (worsening from 21.73% to 32.22% on MIMIC), the proposed method remained stable (17.69% to 23.48%).
• Accuracy Trade-off: The method achieved the most favorable fairness–accuracy trade-off. It maintained the lowest $\Delta Sep_{10}$ with minimal AUC degradation (e.g., $-0.15\%$ on CXP compared to ERM), whereas other baselines incurred significant performance losses (e.g., JTT dropped $\sim 3.6\%$ on CXP).
• Ablation Study: Removing either the "Trend Affinity" (Huber weights) or the "Coverage Score" resulted in degraded fairness and slightly reduced AUC, confirming the necessity of both components.

5. Significance and Claims

The paper claims that its framework offers a robust mitigation strategy for age-dependent confounding in medical imaging. Its primary significance lies in:

1. Selective Mitigation: It successfully suppresses spurious linear age–difficulty shortcuts without enforcing strict age invariance, thereby preserving clinically meaningful, non-linear age information.
2. Stability Under Shift: By utilizing a "one-slope" strategy and coverage-aware modulation, the method remains effective even when the test population has a significantly different age distribution and disease prevalence than the training population.
3. Practical Efficacy: It reduces age-dependent TPR/FPR disparities (over- and under-diagnosis) with minimal impact on overall diagnostic accuracy (AUC), addressing a critical source of systematic bias in medical AI.