In recent years, many institutions have introduced machine learning-based medical diagnostic systems. Though these systems have achieved high accuracy in predicting disease conditions, there are biases in predicting results for different population groups in skin disease datasets, as shown in Groh et al., 2021, Kinyanjui et al., 2020, Tschandl et al., 2018. For instance, in Fig. 1, we present a toy example related to this issue. In this example, we show the prediction accuracy for each demographic group in the Fitzpatrick-17k dataset (Groh et al., 2021), while the sensitive attribute is skin tone. We can observe that there exist differences in accuracy among different groups. These kinds of prediction biases can occur when there is an imbalance in the number of disease images of different demographic groups, leading to inaccurate predictions and misdiagnoses. The discriminatory nature of these models can harm society, causing distrust in computer-assisted diagnostic methods among other sensitive groups, such as race or gender (Adamson and Smith, 2018, Ueda et al., 2024). For example, in real-world healthcare scenarios, consider an approach that demonstrates high accuracy in predictions but exhibits diagnostic unfairness, such as biased predictions between males and females. Once this method is applied in clinical situations, all clinical guides from this method can present biased diagnoses, further reducing patient trust in an AI-based healthcare system. This scenario underscores the critical need for fair clinical diagnosis methods.

Recently, various methods have been proposed to mitigate bias in machine learning models. Many methods (Wang et al., 2022, Zhang et al., 2018, Kim et al., 2019, Ngxande et al., 2020) use adversarial training to train networks to learn classifiers while removing the adversary’s ability to classify sensitive attributes to eliminate bias. Another mainstream approach is regularization-based methods, such as Quadrianto et al., 2019, Jung et al., 2021, which use specific loss functions to constrain the model to learn the features unrelated to sensitive attributes. Contrastive learning (Park et al., 2022) and pruning (Wu et al., 2022, Lin et al., 2022) have also received extensive research in recent years. These methods mitigate prediction biases in classification by using demographic information (sensitive attributes, e.g., gender, race, age) in the training data.

Nevertheless, collecting sensitive attribute information is only sometimes feasible due to privacy or legal issues. If the sensitive attributes are missing, methods that rely on them to mitigate prediction bias can fail. For instance, during the pruning phase, if sensitive attributes are not utilized, the pruning-based bias mitigation method “FairPrune” (Wu et al., 2022) will degenerate into an ordinary pruning method like “OBD” (LeCun et al., 1989). The experimental results in Table 1 of “FairPrune”(Wu et al., 2022) indicate that if sensitive attributes cannot be utilized during pruning, the difference in the F1-score for this pruning-based method increases from 0.048 for “FairPrune” to 0.061 for “OBD”. Additionally, through other experimental results and fairness metric scores, we can observe that this pruning-based method could potentially collapse in the absence of sensitive attributes. As a result, the methods that can achieve fair prediction without sensitive attributes are proposed to overcome this issue.

To mitigate prediction bias without using sensitive attributes, mainstream methods inspired by the Rawlsian Max-Min fairness objectives include distributionally robust optimization (DRO) (Hashimoto et al., 2018), adversarial learning-based methods (Lahoti et al., 2020), and fair self-supervised learning (Chai and Wang, 2022). However, these methods often significantly reduce the model’s accuracy as they aim to balance predicted performance among various groups by minimizing the loss for the worst-performing group utility. Other methods, such as knowledge distillation (Chai et al., 2022) or multi-exit training (Chiu et al., 2023b, Chiu et al., 2023a), also address the fairness problem without using sensitive attributes. However, further research is needed for the knowledge distillation method due to a limited understanding of the generality and quality of the knowledge learned, particularly in measuring knowledge quality related to fairness concepts (Gou et al., 2021). As for multi-exit training, the impact on predictive performance may vary due to changes in network architecture, and improvement is only sometimes guaranteed (Chiu et al., 2023a) . Extending this framework to other methods is needed to enhance fair classification outcomes.

Compared to the fairness issue in other fields, such as natural images or chest X-ray images, achieving fairness in dermatological images without sensitive attributes poses an additional challenge for models. In most dermatological images, besides the diseased areas, the majority of the remaining regions consist of skin texture alone. Therefore, compared to other image modalities containing more complex background, models may easily learn features related to sensitive attributes through skin texture or color. To tackle these challenges, in this paper, we are inspired by the observation in Chiu et al. (2023b), which found that models can achieve fairer predictions by using features highly entangled with different sensitive attribute classes for classification. We propose an Attention-based feature ENtanglement regularization method (AttEN), by using the soft nearest neighbor loss (SNNL) as a measure, we can assess the degree of entanglement between different class features in the embedding space, as introduced in Frosst et al. (2019). If the SNNL values across different sensitive attribute classes are low, features become more distinguishable among sensitive attributes, and if they are high, they become indistinguishable. Since we aim to improve predictive fairness across different demographic groups without using sensitive attribute information, we designed a training framework. This framework ensures that the model classifies disease types only based on features related to the diseased part (target attribute) without distinguishing features related to the skin (sensitive attribute) within the same disease category. This design aims to prevent the model from learning to classify diseases based on skin features, which can lead to predictive performance biases among different demographic groups in the same disease category. Specifically, by using attention modules to obtain features that focus on the diseased part and the skin part and then using SNNL as a regularization constraint, the model’s ability to differentiate diseases is enhanced, while its ability to determine skin differences within the same disease category is reduced. Our concept is shown in Fig. 2, where the red dashed arrow represents decreased feature entanglement between different features, and the blue solid line represents increased feature entanglement. For images belonging to the same class, such as those on the left belonging to the “Melanoma” class, the feature entanglement related to the diseased part and skin part should be increased between different images; for images from other classes, such as “Benign keratosis” shown in the figure, the feature entanglement relates to the diseased part should be decreased. To improve the quality of attention maps during training, we also incorporate disease masks generated by the Segment Anything Model (SAM) (Kirillov et al., 2023) into the obtained features, further enhancing the model’s accuracy and fairness.

Our extensive experiments demonstrate that our “AttEN” can achieve fairness without using sensitive attributes. This method is more suitable for skin disease diagnosis than previous bias mitigation methods because sensitive attribute information may involve privacy issues and may not always be available. This method offers better predictive accuracy and fairness than other methods that do not require collecting sensitive attribute information and can maintain robust results across different model architectures. We compare our method with the current state-of-the-art methods that require using sensitive attributes during training and methods that do not require sensitive attributes, showing the best overall performance in these comparisons.

The main contributions of the proposed method are as follows:

We propose a method to improve predictive fairness in dermatological disease classification without using sensitive attribute information.

With theoretical justification, we have confirmed that our approach of capturing features and performing entanglement regularization through SAM and the attention module can enhance the fairness of the learned features.

Through extensive experiments, we show that our approach can improve fairness while maintaining competitive accuracy on both the dermatological disease datasets, ISIC 2019 (Combalia et al., 2019), and Fitzpatrick-17k datasets (Groh et al., 2021).