How Sensitive Data Affects Fairness and Accuracy in Medical AI Models

Table of Links

1. Abstract and Introduction

2. Related work

3. Methods

   3.1 Positive-sum fairness

   3.2 Application

4. Experiments

   4.1 Initial results

   4.2 Positive-sum fairness

5. Conclusion and References

3.2 Application

To put this fairness notion into practice and show the difference with traditional group fairness, we compare three models which use sensitive attributes to a baseline model. The way sensitive attributes are used by the model is known to have an impact on the fairness and performance of the model [3,39,41,11]. Therefore, we make use of models that explicitly include sensitive attributes, or conversely, remove any demographic encoding from the input data.

The four models are trained on a multi-label classification problem of findings in chest radiography (CXR). In all settings, a Densenet-121 [13] backbone is used, which was empirically determined to give the best performance for this problem. The exact model architectures are shown in figure 2 and described below:

– M1: a baseline classifier using the images as input and trained to predict the targeted CXR findings associated to our dataset. The model comprises a backbone to extract the image features and a finding branch consisting of a fully connected layer and a binary cross entropy loss for each finding.

– M2: a classifier using both the images and race features as input. The race information comes in the form of a categorical variable, which we convert to a one-hot vector and feed to a fully-connected layer. We concatenate the features from the fully connected layer and the image features before forwarding to finding branch. The model is trained end-to-end.

– M3: a classifier using the images as input only, but trained to predict image findings as well as the race group (i.e. this model aims to exploit the race encodings present in the images). For this model, we modify the final layer of the baseline classifier by adapting the loss function to optimize the two tasks: CXR findings and race group. We also transform race information to one-hot encoded vector to apply multi-class loss. The race classification branch is made of a fully-connected layer and a cross entropy loss function. The final loss is calculated by adding finding loss and the race loss with a loss weight λ.

– M4: a classifier using the images as input, trained to predict image findings, while minimizing the use of race information encoded in the image. For this model, we implement the gradient reversal technique described in [28]. We apply the gradient reversal layer before the race branch.