[2024-07] ALS Care Assistant

Note: To protect commercial confidentiality, public information is limited.

Project Introduction

During the development of CareNurse N1, we collaborated with medical experts (including Prof. 冯国栋 and Prof. 闫炳苍), healthcare institutions, and 300+ families. Our research identified specialized caregiving as the critical factor in extending ALS patients’ lifespans – particularly regarding swallowing difficulties, pressure ulcer management, and related complications.

A fundamental challenge emerged: nearly all families lacked essential medical knowledge and caregiving skills for managing severe disability caused by this rare disease.

To address this gap, we engineered an AI-powered Q&A assistant using Retrieval-Augmented Generation (RAG) technology. This solution empowers patients and caregivers to understand medical conditions and access evidence-based care protocols.

For users less proficient with browsers, we integrated a WeChat chatbot enabling natural language queries via mobile messaging:

Key Achievements & Technical Insights

• Evaluated multiple RAG frameworks (LangChain, FastGPT), with FastGPT selected for its superior UI/UX

• Optimized RAG performance through dual focus:

  1. Retrieval Effectiveness

  • Employed segmented text and custom Q&A pairs with rigorous quality control:
    • Ensured semantic coherence in segmented content
    • Maintained precise question-answer alignment
  • Combined semantic or hybrid retrieval methods to enhance precision/recall

  2. Response Quality

  • Tested LLM prompting strategies for:
    • Contextual citation of retrieved content
    • Personality calibration
  • Implemented security protocols:
    • Safety model screening
    • Workflow-based negative response filtering
• Pioneered multimodal database integration for visual queries:
  • Curated image data by removing extraneous background elements
  • Developed VLM-powered workflow:
    1. Identify user’s object of interest
    2. Perform targeted image segmentation
    3. Retrieve context-specific information