VerHealth: Vetting Medical Voice Applications through Policy Enforcement
summary
A tool called VerHealth builds a static analyzer plus a dynamic, ML-driven tester (rule-based checks and a CNN for disclaimers) that performs multiple-round dialogues and 855,000 probing questions to exercise voice health apps. It analyzes 813 health skills and finds concrete failures: 244 skills store physical or mental health data (30.23%) and 86.36% missing disclaimers; experts flag that over half of responses need more follow-up questioning and many suggestions lack relevance.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:31Motivation
- 1:35Challenges
- 3:01System Overview
- 3:17Static Module
- 3:35Dynamic Module
- 3:52Key Insights
- 5:20Ensuring safety in future skills
related talks
24mLeveraging Voice in Healthcare After COVID
09mEnhance Healthcare Chatbot Transparency
35mBuilding Trust & Enabling Innovation for Voice Enabled IoT
15mMore empathy and health data protection in AI
07mVOICE 2021 | Chris Butler, IPsoft
31mAI For Good: Fighting Health Insurance with AI
22mAI Powering New Health Solutions
16mGenAI in Healthcare Claims: Processing and Detecting Fraud
