Back

EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings

Rethmeier, N.; Serbetci, N. O.; Möller, S.; Roller, R.

2020-07-26 intensive care and critical care medicine
10.1101/2020.07.21.20157610 medRxiv
Show abstract

Recent medical prognostic models adapted from high data-resource fields like language processing have quickly grown in complexity and size. However, since medical data typically constitute low data-resource settings, performances on tasks like clinical prediction did not improve expectedly. Instead of following this trend of using complex neural models in combination with small, pre-selected feature sets, we propose EffiCare, which focuses on minimizing hospital resource requirements for assistive clinical prediction models. First, by embedding medical events, we eliminate manual domain feature-engineering and increase the amount of learning data. Second, we use small, but data-efficient models, that compute faster and are easier to interpret. We evaluate our approach on four clinical prediction tasks and achieve substantial performance improvements over highly resource-demanding state-of-the-art methods. Finally, to evaluate our model beyond score improvements, we apply explainability and interpretability methods to analyze the decisions of our model and whether it uses data sources and parameters efficiently.1

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.