In a recent preprint (available through Cornell Universityâs open access website arXiv), a team led by a Lawrence Livermore National Laboratory (LLNL) computer scientist proposes a novel deep learning approach aimed at improving the reliability of classifier models designed for predicting disease types from diagnostic images, with an additional goal of enabling interpretability by a medical expert without sacrificing accuracy. The approach uses a concept called confidence calibration, which systematically adjusts the modelâs predictions to match the human expertâs expectations in the real world.
In practice, quantifying the reliability of machine-learned models is challenging, so the LLNL researchers introduced the “reliability plot,” which includes experts in the inference loop to reveal the trade-off between model autonomy and accuracy. By allowing a model to defer from making predictions when its confidence is low, it enables a holistic evaluation of how reliable the model is.
However, more important than increased accuracy, prediction calibration provides a completely new way to build interpretability tools in scientific problems, Thiagarajan said. The team developed an introspection approach, where the user inputs a hypothesis about the patient (such as the onset of a certain disease) and the model returns counterfactual evidence that maximally agrees with the hypothesis. Using this “what-if” analysis, they were able to identify complex relationships between disparate classes of data and shed light on strengths and weaknesses of the model that would not otherwise be apparent.