Developing Robust Clinical Text Deep Learning Models - A "Painless" Approach.

The success of deep learning in natural language processing relies on ample labelled training data. However, models in the health domain often face data inadequacy due to the high cost and difficulty of acquiring training data. Developing such models thus requires robustness and performance on new data. A generalised incremental multiphase framework is proposed for developing robust and performant clinical text deep learning classifiers. It incorporates incremental multiphases for training data size assessments, cross-validation setup to avoid test data bias, and robustness testing through inter/intra-model significance analysis. The framework's effectiveness and generalisation were confirmed by the task of identifying patients presenting in 'pain' to the emergency department.