Google Unveils AI Symptom Assessment Agent More Accurate Than Clinicians

SymptomAI, developed by Google Research and DeepMind, is a conversational AI agent that listens to a user's symptoms and actively asks questions to draw out information. To solve the information incompleteness problem inherent in the existing 'user-guided' approach, it adopts an active elicitation strategy of asking follow-up questions, similar to a doctor.

The results of a large-scale real-world study conducted over 9 months among Fitbit users across the United States are as follows.

• High Accuracy: The differential diagnoses (DDx) presented by SymptomAI were statistically significantly more accurate than the diagnoses of actual clinicians who reviewed the same conversation records (OR = 2.56, p < 0.001).
• Performance Improvement: The active elicitation strategy boosted diagnostic accuracy by an average of 27.57% compared to the default approach of existing chatbots.
• Real-World Validation: The agent demonstrated high performance not on a curated dataset, but in real symptom-reporting situations where general users lacking medical knowledge conveyed their complaints in a disorganized manner.

This study shows that the diagnostic capabilities of LLMs can function as a practical tool that improves actual healthcare accessibility, going beyond simple benchmarks.