Instance-based Learning for ICD10 Categorization

In the framework of the CLEF 2018 eHealth campaign, we investigated an instance-based approach for extracting ICD10 codes from death certificates. The 360,000 annotated sentences contained in the training data were indexed with a standard search engine. Then, the k-Nearest Neighbors (k-NN) generated out of an input sentence were exploited in order to infer potential codes, thanks to majority voting. Compared to a standard dictionary-based approach, this simple and robust k-Nearest Neighbors algorithms achieved remarkable good performances (F-Measure 0.79, +13% compared to our dictionary-based approach, +70% compared to the official baseline). This purely statistical approach uses no linguistic knowledge, and could a priori be applied to any language with similar performance levels. The combination of the k-NN with a dictionary-based approach is also a simple way to improve the categorization effectiveness of the system. The reported results are consistent with inter-rater agreements (79-80%) for diagnosis encoding as achieved by trained professional staff. Any significant improvement should therefore be questioned.