AI-Enabled Conversational Journaling for Advancing Parkinson's Disease Symptom Tracking

TL;DR

Patrika introduces AI-enabled conversational journaling to advance Parkinson's disease symptom tracking by embedding cooperative conversation, clinical interview-style probing, and personalized dialogue within a voice-driven journaling tool. The system combines NLU ( Spacy DIETClassifier), rule-based and LLM-driven response generation, and a retrieval-augmented personalization pipeline to generate relevant, context-aware follow-ups and enrich patient-generated health data. Across two user studies with PwPD, Patrika achieved high intent identification accuracy (~99%), substantial personalization (~81%), and strong clinical relevance of collected data, while participants reported engagement, empathy, and usefulness in symptom and medication tracking. The work demonstrates that AI-driven, voice-based journaling can improve data quality, patient engagement, and potential clinical insights, with implications for broader chronic-care applications and future integration with analytics and EHRs.

Abstract

Journaling plays a crucial role in managing chronic conditions by allowing patients to document symptoms and medication intake, providing essential data for long-term care. While valuable, traditional journaling methods often rely on static, self-directed entries, lacking interactive feedback and real-time guidance. This gap can result in incomplete or imprecise information, limiting its usefulness for effective treatment. To address this gap, we introduce PATRIKA, an AI-enabled prototype designed specifically for people with Parkinson's disease (PwPD). The system incorporates cooperative conversation principles, clinical interview simulations, and personalization to create a more effective and user-friendly journaling experience. Through two user studies with PwPD and iterative refinement of PATRIKA, we demonstrate conversational journaling's significant potential in patient engagement and collecting clinically valuable information. Our results showed that generating probing questions PATRIKA turned journaling into a bi-directional interaction. Additionally, we offer insights for designing journaling systems for healthcare and future directions for promoting sustained journaling.

Figures (7)

• Figure 1: This figure shows the system overview of Patrika. The system comprises three modules: Natural Language Understanding (NLU), response generation, and journal. The NLU module tokenizes the user input and predicts the user's intent. This predicted intent is then passed on to the response generation module, which uses both the intent and the conversation context to predict a relevant journal prompt or response. To personalize the journal prompt, the response generation module queries the journal module, which returns the user's conversation history. The response generation module uses the conversation history to generate a personalized journal prompt, which is then delivered to the user through Alexa.
• Figure 2: This figure shows the workflow of the Natural Language Understanding (NLU) Module of Patrika. The NLU module consists of two components: a parser and an intent interpreter. The parser converts the user's input into a string of words, which are then passed to the intent interpreter. The intent interpreter transforms this word sequence into a matrix representation. This matrix is fed into a transformer-based model called the "DIETClassifier," which compares it to a set of predefined intents. The "DIETClassifier" then predicts the intent that best matches the user's input.
• Figure 3: This figure illustrates the study workflow. After meeting participants in person and obtaining their signed consent, the study proceeded in three stages: (1) in-person onboarding, during which the participants completed a pre-study questionnaire to gather demographic information and their experience with journaling and smart devices. Then they watched a 12-minute tutorial on Patrika's features and journaling examples, followed by a study brief with instructions on how to interact with Patrika, received a tutorial; (2) a two-week period where they used Patrika in their primary place of residence to journal their symptoms; and (3) a post-study interview where we collected their feedback on the conversation flow, session length, satisfaction, enjoyment, and views on personalization using conversation history.
• Figure 4: The system overview of Patrika following improvements made after Study I. First, the NLU module parses the user's input and forwards the information to the LLM intent interpreter as a prompt. The LLM intent interpreter predicts the user's intent from the prompt and sends it to the response generation module. The response generation module uses the predicted intent to trigger an appropriate journaling rule and selects the next relevant journal probe. To personalize this probe, the response generation module queries the journal module using the journal probe and conversation context. The journal module uses a BM25 retriever to retrieve relevant history from the user's conversation history that correspond to the journal probe and the ongoing conversation. This conversation history is then used to personalize the journal probe, which is sent to the user through Alexa.
• Figure 5: This diagram shows information captured via journaling in Study II. To the left of Participant(s) are Symptom(s) expressed and the Follow-up(s) generated by Patrika to further probe and collect information about each symptom. To the right of Participant(s) are Anecdote(s) and Topic(s). Anecdotes are expressions that are not directly related to one of the 12 symptoms included in our study but provide valuable insights about participants, such as information about their medication and positive or negative experiences. Participants are ordered in descending order by the total number of symptoms and anecdotes.