↓Skip to main content

Semantic Enrichment of Explanations of AI Models for Healthcare

·1 min

TL;DR #

Semantic Enrichment of Explanations of AI Models for Healthcare connects Machine Learning explanations with the clinical language used by clinicians.

Problem #

Feature-importance explanations can identify which variables influenced a prediction, but they do not always provide enough context for a clinician. In healthcare, the relevant evidence may be expressed in clinical notes using concepts and terminology that are not directly represented by the model features.

What did we do? #

We enriched explanations using medical ontologies, including SNOMED CT, and natural-language processing of clinical notes. The method links features that contribute to a prediction with semantically related information in the patient’s notes.

Main findings #

Adding clinical semantics can make model explanations more informative than a list of feature names or scores. It helps connect a model’s internal evidence with concepts that healthcare experts already use.

Resources #

Cite this paper #

@inproceedings{corbucci2023semantic,
  title = {Semantic Enrichment of Explanations of AI Models for Healthcare},
  author = {Corbucci, Luca and Monreale, Anna and Panigutti, Cecilia and Natilli, Michela and Smiraglio, Simona and Pedreschi, Dino},
  booktitle = {Discovery Science},
  year = {2023},
  doi = {10.1007/978-3-031-45275-8_15}
}