How machine-learning models can amplify inequities in medical diagnosis and treatment

Abstraction: Four-type taxonomy of ML subpopulation shifts causing medical inequity

Key points:

• MIT researchers (Marzyeh Ghassemi, Yuzhe Yang, Haoran Zhang, Dina Katabi) identify four types of subpopulation shifts: spurious correlations, attribute imbalance, class imbalance, and attribute generalization
• A single unified equation captures all four shift types — the first such coherent framework for understanding the sources of ML bias across subgroups
• Improvements to the neural network classifier layer reduced spurious correlations and class imbalance; improvements to the encoder reduced attribute imbalance; but neither layer adjustment fixed attribute generalization
• Boosting worst-group accuracy (WGA) — the standard gold-standard metric — trades off against "worst-case precision," which is equally important in clinical diagnosis
• Tested 20 advanced algorithms on 12 datasets; disparities persist across age, gender, ethnicity, and intersectional groups even on large chest X-ray datasets
• Presented at ICML 2023 (40th conference, Honolulu); paper title "Change is Hard" signals the difficulty of systemic reform

Connections: Mit · Mit Ibm Watson AI Lab · Algorithmic Fairness · Machine Learning · AI Bias In Healthcare

Source: https://news.mit.edu/2023/how-machine-learning-models-can-amplify-inequities-medical-diagnosis-treatment-0817