Achieving model fairness through distilling knowledge from deep networks: A retrospective cohort study (Preprint)

Abstract

Sepsis has become one of the leading causes of hospital readmission in the U.S. and puts immense pressure on patients’ costs of hospitalization. Differences in risk score predictions in patients who differ by race/ethnicity or by sex introduce unfairness in decision making and create disparities.

To ensure fairness among different race/gender subgroups while maintaining high prediction accuracy, we propose a prediction model with fairness enhancement to estimate the probability of sepsis-caused readmission for emergency room patients at least 72 hours after admission and mitigate the prediction performances among different gender/race groups.

The model is based on a multi-modal encoder-decoder attention mechanism structure and attends to making full use of the temporal characteristics of electronic health records (EHR). Tables of different features and formats (lab test results, vital signals, medications, and demographics) are window-extracted from the EHR database and compose the input. We develop a knowledge-distillation-based approach that utilizes a smaller Machine Learning model to mimic the pre-trained Deep Learning model, to keep the high prediction power while capable of enhancing fairness, and we name this framework KAFENET (Knowledge-distillation Assisted Fairness Enhancing neural NETwork).

The model is evaluated over an emergency sepsis database, and the results show that it manages to keep satisfying prediction performance while successfully mitigating the disparity among minority groups. Compared to both benchmark DL models and ML models, KAFENET reached high accuracy (overall Area Under the Receiver Operating Characteristic curve 0.93) while eliminated all disparities, showing its ability to maintain a great balance between prediction and fairness performance.

This evaluation demonstrates KAFENET’s ability to make hiegh-accuracy predictions with fairness to disadvantaged subgroups in clinical prediction tasks, and has great potential to be generalized to other healthcare prediction tasks.