Tasty on Fhir – Enterprise Data Management

TastyOnFhir is a project that aims to investigate and analyse various medical cases, using deep learning techniques. It offers physicians the possibility to make appropriate diagnoses based on the symptoms of a patient in the shortest possible time. TastyOnFhir offers an interactive text editor, which uses the Natural Language Processing concepts Named Entity Recognition and Named Entity Linking. The Unified Medical Language System (UMLS) is used as knowledge base. After the physician enters the symptoms, medical entities are recognized and linked to the uniquely matching data of the knowledge base. The data from the database matching the entity is displayed to the physician as a possible diagnosis.

An overview of the results can be found here: presentation.pdf



The goal of our project was to identify 10 error classes of a german-language dataset in the hate speech domain provided by the NoHate project. This analysis was based on manual analysis, as described by Andrew Ng in “Machine Learning Yearning”, and error classes described in papers such as “Challenges for Toxic Comment Classification: An In-Depth Error Analysis, van Aken et al. 2018”. 

We also examined the applicability of “Errudite”, an interactive tool for scalable, reproducible and counterfactual error analysis, for the hate speech domain. Our analysis of the software and later interactions with the developer showed, that Errudite was not applicable for our project due to time constrains.

Based on this result we created our own Jupiter Notebook based software for analyzing and visualizing error classes. With this software we selected based on the previously identified error classes, 3 error classes, that when fully corrected for, could increase the F1 score of the provided BERT-model from 59% to 67%, 70% and 72% respectively.

An overview of the results can be found here: presentation.pdf