A Systematic Mapping Study on Explainable Artificial Intelligence for Biomedical Data : Emerging Trends in Digital Health, Summer Term 2022

Background: The advent of artificial intelligence in healthcare is seen as a great opportunity for new approaches to applications in medicine. In the future, this digitalization should enable programs to use algorithms to evaluate correlations far more accurately and quickly than with conventional methods. Since the complexity of most programs denies insight into the internal processes and causes distrust among experts and the general public, there is a need to use a number of techniques that bring explainability to artificial intelligences. For this, the term Explainable Artificial Intelligence (XAI) was established. Currently, the distribution of insights from XAI applications for biomedical omics data shows a wide dispersion in various scientific fields such as biology and computer science, leading to a lack of overview of the current state of the research field.
Objective: This research paper provides an overview of the existing XAI applications in literature between 2021 and 2022.
Methods: The methodology of the research paper is based on conducting a systematic mapping study. For this, three databases (Scopus, PubMed, and Web of Science) were traversed using a search string and provided a list of potential literature from 2021 to 2022. ... mehrRelevant literature was then identified through abstract screening and full text analyses. Using a classification scheme, 20 relevant sources were finally classified by text analysis and pie charts and systematic maps were created based on this classified literature.
Results: The results are presented in form of five pie charts and two systematic maps. The models applied were model-specific and model-independent in equal proportions. Likewise, the distribution of the XAI models with respect to the parameters explainability and interpretability was 50% each. Furthermore, a clear focus of the use cases in the omics domain of genomics was visible. In this study, 7 medical use cases emerged in the analyzed literature, such as cancer detection or cancer type classification. In the interplay between AI methods and explainability approaches, a versatile combination of the two concepts was found. Here, explainability focused on the following three methods: feature relevance, transparent model, and visual explanation combined with widely used AI methods. When considering the combination of omics data and AI methods, versatile combinations were also found. In 25% of the cases, the literature dealt with DNA-sequencing combined with a variety of AI methods.
Conclusion: This paper lays the groundwork for further exploration in the aforementioned area. It is recommended to conduct further research due to the limited amount of literature analyzed in this work.