Improving Reliability of Clinical Models Using Prediction Calibration

摘要

The wide-spread adoption of representation learning technologies in clinical decision making strongly emphasizes the need for characterizing model reliability and enabling rigorous introspection of model behavior. In supervised and semi-supervised learning, prediction calibration has emerged as a key technique to achieve improved generalization and to promote trust in learned models. In this paper, we investigate the effectiveness of different prediction calibration techniques in improving the reliability of clinical models. First, we introduce reliability plots, which measures the trade-off between model autonomy and generalization, to quantify model reliability. Second, we propose to utilize an interval calibration objective in lieu of the standard cross entropy loss to build classification models. Finally, using a lesion classification problem with dermoscopy images, we evaluate the proposed prediction calibration approach against both uncalibrated models as well as existing prediction calibration techniques such as mixup and single-shot calibration.