From Biased Selective Labels to Pseudo-Labels: An Expectation-Maximization Framework for Learning from Biased...

Area of Science:

• Machine Learning
• Causal Inference
• Biostatistics

Background:

• Selective labels arise from decision-making processes, like diagnostic tests.
• Labeling biases can vary across subgroups, leading to unfair imputation of unlabeled data as 'negative'.
• Standard machine learning models may amplify these existing labeling biases.

Purpose of the Study:

• To address the problem of disparate censorship in machine learning.
• To develop and validate an algorithm that mitigates disparate censorship bias.
• To improve the fairness and accuracy of machine learning models trained on biased clinical data.

Main Methods:

• Propose Disparate Censorship Expectation-Maximization (DCEM), an algorithm inspired by causal models of selective labels.
• Theoretically analyze DCEM's mechanism for mitigating disparate censorship effects.
• Validate DCEM using synthetic datasets and a real-world sepsis classification task.

Main Results:

• DCEM demonstrates improved bias mitigation, measured by the area between ROC curves.
• The algorithm maintains high discriminative performance (AUC) compared to baseline methods.
• Similar positive results were observed in the sepsis classification task using clinical data.

Conclusions:

• DCEM is an effective method for learning from data with disparate censorship.
• The proposed algorithm successfully mitigates labeling bias amplification in machine learning models.
• DCEM offers a promising approach for developing fairer and more reliable AI in healthcare.