# Directive Explanations for Monitoring the Risk of Diabetes Onset: Introducing Directive Data-Centric Explanations and Combinations to Support What-If Explorations

**Authors:** Aditya Bhattacharya, Jeroen Ooge, Gregor Stiglic, Katrien Verbert  
**arXiv:** [2302.10671](https://arxiv.org/abs/2302.10671)  
**Date:** 2023-02-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.10671v1/

## Key points

- Developed an explanation dashboard to predict diabetes onset and provide explanations to healthcare experts
- Included data-centric, feature-importance, and example-based explanations for monitoring patient health records
- Compared different explanation methods in terms of understandability, usefulness, actionability, and trust
- Participants preferred data-centric explanations providing local insights with a global overview
- Visualizations were crucial for quick understanding, with color-coded representations appreciated for easy interpretation
- Some participants found difficulty in interpreting data-distribution charts without textual annotations
- Directive explanations were important in guiding healthcare experts towards actionable insights from patient health records
- Identified limitations such as offline predictions and the need for evaluation with specialists dedicated to type-2 diabetes care
- Future research will focus on personalizing directive counterfactual explanations and analyzing the utility of different explanation methods used in isolation
- Emphasized the significance of visually directive data-centric explanations in assisting healthcare experts with decision-making based on machine learning predictions

## Summary

Our study developed an explanation dashboard to predict the risk of diabetes onset and provide explanations to healthcare experts. The dashboard included data-centric, feature-importance, and example-based explanations to assist nurses and physicians in monitoring patient health records and recommending measures to minimize risk. Through qualitative and mixed-methods studies involving 11 healthcare experts and 51 diabetic patients, we compared different explanation methods in terms of understandability, usefulness, actionability, and trust. Our findings revealed that participants preferred data-centric explanations that provided local insights with a global overview. Visualizations played a crucial role in facilitating quick understanding, with participants appreciating color-coded representations for easy interpretation. However, some participants expressed difficulty in interpreting data-distribution charts without textual annotations. The study also highlighted the importance of directive explanations in guiding healthcare experts towards actionable insights from patient health records. We identified limitations such as offline predictions and the need for evaluation with specialists dedicated to type-2 diabetes care. Future research will focus on personalizing directive counterfactual explanations and analyzing the utility of different explanation methods used in isolation. Overall, our research underscores the significance of visually directive data-centric explanations in assisting healthcare experts with decision-making based on machine learning predictions for patient care. By tailoring visual representations of explanation methods, we aim to enhance the usability and effectiveness of AI-driven decision support systems in healthcare settings.