Scholarlink

Reliable and personalised support for chronic illness self-management remains a significant challenge, particularly in the unmonitored intervals between clinical consultations. Existing digital solutions often involve a trade-off: either compromising privacy through remote data storage or lacking the adaptability required for meaningful interaction. This paper introduces TrackWise, a proof-of-concept conversational agent (CA) designed to provide tailored support without persistent data storage. Architecturally, TrackWise is a lightweight, client-side single-page application (SPA) built with React and TypeScript. It operates entirely within the user’s browser and leverages Google’s Gemini-2.5-Flash large language model (LLM) via the official @google/genai SDK. This serverless design ensures all user data remains ephemeral—stored solely in volatile React component state and erased upon session termination—thus adhering to a strict privacy-by-design model. The system’s adaptive behaviour is governed not by model fine-tuning but by a layered prompt-as-knowledge framework, which encodes safety constraints, context-awareness, and functional roles within structured instructions. Early implementation demonstrates that this approach effectively regulates safe interaction boundaries while enabling features such as streaming dialogue and on-demand summarisation. These findings demonstrate the feasibility of prompt engineering as a practical and auditable method for deploying privacy-preserving, client-side CAs in healthcare. The study offers a design pattern for secure, real-time AI support tools to augment chronic care between clinical visits.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TrackWise: A Privacy-Preserving Client-Side Conversational Agent for Ambient Chronic Illness Support

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Seif Allah El Mesloul Nasri,
  • Manolya Kavakli-Throne

摘要

Reliable and personalised support for chronic illness self-management remains a significant challenge, particularly in the unmonitored intervals between clinical consultations. Existing digital solutions often involve a trade-off: either compromising privacy through remote data storage or lacking the adaptability required for meaningful interaction. This paper introduces TrackWise, a proof-of-concept conversational agent (CA) designed to provide tailored support without persistent data storage. Architecturally, TrackWise is a lightweight, client-side single-page application (SPA) built with React and TypeScript. It operates entirely within the user’s browser and leverages Google’s Gemini-2.5-Flash large language model (LLM) via the official @google/genai SDK. This serverless design ensures all user data remains ephemeral—stored solely in volatile React component state and erased upon session termination—thus adhering to a strict privacy-by-design model. The system’s adaptive behaviour is governed not by model fine-tuning but by a layered prompt-as-knowledge framework, which encodes safety constraints, context-awareness, and functional roles within structured instructions. Early implementation demonstrates that this approach effectively regulates safe interaction boundaries while enabling features such as streaming dialogue and on-demand summarisation. These findings demonstrate the feasibility of prompt engineering as a practical and auditable method for deploying privacy-preserving, client-side CAs in healthcare. The study offers a design pattern for secure, real-time AI support tools to augment chronic care between clinical visits.