TLDR: CLARITY is an AI-driven platform that uses a hybrid Finite State Machine and Large Language Model architecture to improve patient-to-specialist routing, clinical consultations, and severity assessment. It has demonstrated superior first-attempt routing precision and significantly shorter consultation times compared to human interactions in real-world deployments, enhancing healthcare efficiency and accessibility.
In the rapidly evolving landscape of healthcare, the integration of artificial intelligence holds immense promise for transforming patient care. A new AI-driven platform, CLARITY (Clinical Assistant for Routing, Inference, and Triage), is designed to significantly improve how patients are routed to specialists, how clinical consultations are conducted, and how the severity of a patient’s condition is assessed.
Developed by a team including Vladimir Shaposhnikov, Aleksandr Nesterov, and Ilia Kopanichuk, CLARITY addresses critical bottlenecks in traditional healthcare systems, such as long wait times and inefficient referrals. The platform aims to streamline processes, making healthcare more accessible and efficient for both patients and providers.
Understanding CLARITY’s Approach
CLARITY stands out due to its unique hybrid architecture. It combines a Finite State Machine (FSM) for managing structured dialogue flows with collaborative agents that utilize Large Language Models (LLMs). This combination allows the system to analyze symptoms effectively and prioritize referrals to the most appropriate medical specialists. The modular design, built on a microservices framework, ensures that CLARITY is safe, efficient, robust, and easily scalable to fit various healthcare workflows and IT solutions.
The system’s core components work together to manage patient interactions and provide accurate assessments. The Chat Manager, implemented as an FSM, guides the conversation through different stages: Initialization (greeting and initial complaint collection), Information Collection (gathering symptoms and medical history), Diagnosis (generating preliminary diagnoses and specialist recommendations), Moderation (preventing unsafe dialogue), Emergency (handling urgent cases), and Free Dialogue (for non-standard requests).
Key Services Powering CLARITY
Several specialized services underpin CLARITY’s functionality:
- Moderator: This service checks for prohibited topics in user requests or model responses, ensuring conversations remain appropriate and safe, especially given the sensitive nature of medical data.
- Emergency Detector: It identifies whether a user requires emergency medical care, processing chat history to determine criticality and triggering predefined protocols for urgent cases.
- Readiness Estimator: This module determines the optimal time to transition from gathering information to making a referral, preventing premature or delayed actions.
- Question Detector: It classifies user messages to identify clarifying questions, allowing for flexible information gathering.
- Information Collector: This module gathers comprehensive patient complaints and medical history, using an LLM-based question generator and a dialogue search system to ensure relevant and non-redundant questioning.
- Medical Specialty Selector: A crucial component that generates possible diagnoses, identifies suitable medical specialists for each diagnosis, and provides clear explanations for both the diagnosis and the referral.
- Answer Generator: This module handles arbitrary user queries and open dialogues, adapting LLMs to the medical domain with specialized prompts to provide coherent and contextually relevant responses.
Real-World Impact and Performance
CLARITY has already demonstrated impressive results in real-world deployments. During a two-month pilot, over 55,000 user dialogues were completed on a large-scale national inter-hospital IT platform. Validation results indicate that CLARITY surpasses human-level performance in first-attempt routing precision, and consultations are up to three times shorter than those with a human.
Specifically, the system achieved 77% precision for first recommendations and 96% recall for top-3 recommendations in specialist routing. The Emergency Detector, while conservatively set to minimize false alarms, identified 7.4% of cases as requiring urgent intervention, aligning with emergency department triage statistics. User engagement metrics were strong, with a 32.4% conversion rate to actual appointments and 80% positive interaction ratings, highlighting user trust and satisfaction.
The mean consultation time of 2 minutes and 13 seconds represents a significant improvement in operational efficiency without compromising diagnostic accuracy. The system’s ability to manage dialogue, with a near-zero repetition rate in information collection and only 3.4% non-target dialogues, showcases its sophisticated conversation management.
Also Read:
- Enhancing Access to UK Clinical Guidelines with AI: A RAG System for Healthcare
- BayesianRouter: A Smart Approach to Aligning Language Models with Human Preferences
Future Considerations
While CLARITY marks a significant advancement, the researchers acknowledge several limitations. These include the current limited interpretability of LLM outputs, the undisclosed identity of the specific LLM used (due to anonymity constraints), and potential limitations in generalizing to rare or atypical conditions. Future work will focus on improving interpretability, expanding datasets for broader generalization, and refining dialogue depth based on user feedback. The full research paper can be accessed here: CLARITY: Clinical Assistant for Routing, Inference, and Triage.
CLARITY represents a promising step forward in leveraging AI to create more accessible, efficient, and personalized healthcare services, setting a new standard for reliable and user-centric medical assistance.
