ECHO: A Locally-Deployable Agentic Health Assistant...

ECHO: A Locally-Deployable Agentic Health Assistant... | AI Research

Key Takeaways

ECHO (Enhanced Care & Health Observer) is a locally-deployable conversational assistant designed to support long-term chronic care management.
This paper presents ECHO (Enhanced Care \& Health Observer), a locally-deployable conversational health assistant for long-term chronic care management.
ECHO integrates three complementary software modules developed under shared supervision as a unified system.
A multimodal speech assessment module combining Whisper acoustic encoding and BERT text encoding with cross-attention fusion estimates emotion, depression, and pain, reaching a mean macro F1 of 0.652.
The full system is implemented as a web application that can run entirely on consumer hardware, with no patient data transmitted to external services, supporting compliance with GDPR and KVKK.

Paper AbstractExpand

This paper presents ECHO (Enhanced Care \& Health Observer), a locally-deployable conversational health assistant for long-term chronic care management. ECHO integrates three complementary software modules developed under shared supervision as a unified system. The core module is an agentic chatbot built on a ReAct loop orchestrated via LangGraph, equipped with 17 clinical tools and a temporal knowledge graph for persistent cross-session memory; it achieves a 94.9\% tool-execution pass rate across a 59-scenario benchmark with GPT-5 Mini. A two-stage hybrid safety layer intercepts all incoming queries: a rule-based layer handles explicit crisis signals and jailbreak attempts in under 1ms, while a signed graph neural network (GNN) with APPNP-style propagation classifies boundary cases by clinical intent, achieving 88.8\% accuracy and 90.6\% unsafe recall on a 2,537-query annotated Turkish health dataset while outperforming zero-shot LLM baselines including Llama 3.3 70B. A multimodal speech assessment module combining Whisper acoustic encoding and BERT text encoding with cross-attention fusion estimates emotion, depression, and pain, reaching a mean macro F1 of 0.652. The full system is implemented as a web application that can run entirely on consumer hardware, with no patient data transmitted to external services, supporting compliance with GDPR and KVKK.

ECHO (Enhanced Care & Health Observer) is a locally-deployable conversational assistant designed to support long-term chronic care management. It addresses the limitations of current digital health tools—specifically their inability to maintain medical history across sessions, their passive nature, and their lack of specialized safety guardrails—by running entirely on consumer hardware to ensure patient data privacy and regulatory compliance.

System Architecture

The system functions through three integrated modules:

Agentic Orchestration: An agentic chatbot built on a ReAct loop using LangGraph. It utilizes 17 clinical tools for tasks like medication management, calendar scheduling, and symptom logging.
Temporal Memory: A "Hindsight" knowledge graph that stores patient history. Unlike standard systems that append data, this module uses a structured SQLite layer and a temporal graph to supersede outdated clinical facts with new, timestamped information, ensuring a consistent patient profile.
Hybrid Safety Guardrails: A two-stage filter that intercepts queries before they reach the LLM. A rule-based layer handles explicit crises in under 1ms, while a signed Graph Neural Network (GNN) classifies boundary-case queries based on clinical intent.

Speech Assessment

ECHO includes a multimodal module that analyzes voice input to provide passive health signals. By combining Whisper for acoustic encoding and BERT for text encoding, the system uses cross-attention fusion to estimate a user’s emotion, depression, and pain levels. These estimates are injected into the agent’s context, allowing the system to proactively adjust its tone or suggest symptom logging when it detects distress.

Performance and Results

In a 59-scenario benchmark, the system achieved a 94.9% tool-execution pass rate using GPT-5 Mini. The safety classifier reached 88.8% accuracy and 90.6% unsafe recall on a 2,537-query Turkish health dataset, outperforming zero-shot LLM baselines such as Llama 3.3 70B. The speech assessment module achieved a mean macro F1 score of 0.652, with the most significant performance gains observed in pain detection.

Limitations and Future Directions

The researchers identified several areas for improvement:

Safety: The rule-based crisis lexicon requires more robust adversarial testing, and the GNN requires expanded datasets to better identify rare symptom presentations.
Depression Screening: The current speech module shows limited improvement in depression detection, suggesting that short, utterance-level analysis is insufficient for this task.
Extensibility: The authors propose using the Model Context Protocol (MCP) to allow third-party applications, such as wearable sensors or electronic health records, to integrate with ECHO’s clinical tools.