Lmst
Login

#XAIl

Tiago F. R. Ribeirotiago_ribeiro
2025-07-18

“Visual-Conversational Interface for Evidence-Based Explanation of Diabetes Risk Prediction”

Este trabalho apresenta uma ‘interface’ visual‑conversacional para explicar predições de risco de diabetes, integrando visualizações interativas e agente conversacional ancorado em evidências científicas. Num estudo com 30 profissionais de saúde, comprovou‑se melhoria da compreensão das avaliações de risco e calibração da confiança no sistema.

📎arxiv.org/pdf/2507.02920

Figure 1: Hybrid query processing architecture for conversational Al integration. Incoming user queries are first processed by a semantic matcher that determines if they match supported analytical operations. Matching queries are parsed by a fine-tuned T5 model to trigger specific backend functions, while non-matching queries are handled by a general-purpose LLM (Claude) that has access to contextual information. The scientific evidence are retrieved by Gemini. (a) The overall interface of our DSS: the patient record data (VC1) is visualized in the left column. In the middle, The user can see the risk assessment of the chosen patient and select different visualizations from the drop-down menu of Analysis Visualization (in this case Recommendations (VC4)). The chat component (CC) in the right column provides responses to the user interactions (such as clicking on "Get Recommendation” button) as well as follow-up questions. Figure 2: The DSS interface components. (a) Shows the main dashboard with patient records, analysis visualizations, and chat component. (b) Displays feature importance and range analysis with scientific evidence integration.

Client Info

Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst