Knowledge Graph-based Explanation for Biomedical Artificial Inteligence

Motivation

Artificial intelligence (AI) has proven to be incredibly successful in a variety of fields, including the biomedical one. However, one issue is that most advanced AI algorithms are incomprehensible black-box solutions, which means they create useful information without providing any knowledge about their internal workings. This poses significant obstacles, such as explaining a predictive outcome.

One of the historical challenges of AI is explainability. Especially in medicine, explainability can help enhance the trust of physicians in AI systems. To enable explanations, there is a necessity for shared context in much the same way when we understand each other as humans.

So, we need to integrate the data fed to the AI with background knowledge to infer the causality of the AI prediction and generate explanations.

Biomedical Semantic Web technologies represent an unparalleled solution to the problem of human-centric knowledge enabled explanations providing the needed semantic context. They provide large amounts of freely available data in the form of Knowledge Graphs which link data to ontologies.

Objective

Ontologies establish a conceptual model that represents the concepts of a domain and their relationships with one another, in a way that humans and machines can understand.

When both input data and AI outcomes are encoded within the Knowledge Graph, it can act as background knowledge for building explanations closer to human conceptualization.

Our main goal is to explain black-box AI methods, performed in personalized oncology, using Knowledge Graphs.

Data

This project is integrated in the european project- Knowledge At the Tip of Your fingers: Clinical Knowledge for Humanity (KATY).

Clinical and biological data from renal cancer patients is gathered by investigators to produce biomedical datasets. This datasets wil be used by AI models to perform predictions.

A knowledge graph containing twenty eight biomedical ontologies integrates the datasets and will be used to generate several types of semantic explanations.

Knowledge Graphs can support Explainable AI

XAI techniques can be categorized according to how they support human reasoning: inductive reasoning, querying, and similarity modelling. A knowledge graph can support all three of them. We will be focusing on the comparative performance of two complementary types of KG-based explanations.

Inductive Reasoning

Similarity Modeling

Research Questions

What is the best complementary KG-based explanations: line of reasoning or similarity-based?

What is the impact of different KG characteristics including coverage, granularity and semantic richness?

What is the best explanation KG-based or data-based?

Evaluation

Quantitative Evaluation

Evaluate the scientific correctness of the explanations.

Rank explanations

Compare explanations based on semantic similarity.

Qualitative Evaluation

Access the usefulness and usability of the explanations.

Perform user studies.

Compare a visualization tool and a written report.

Authors

Susana Nunes

LASIGE, Faculdade de Ciências

Catia Pesquita