Towards an Ontology-based Medication Conversational Agent for PrEP and PEP

Muhammad Amith, Licong Cui, Kirk Roberts, Cui TaoACL 2020

TL;DR

This paper introduces an ontology-driven conversational agent designed to provide information and counseling about PrEP and PEP, two crucial medications for HIV prevention. By leveraging ontologies to model domain knowledge and dialogue flow, the system aims to overcome communication barriers and improve adherence to these life-saving treatments.

What Problem Does This Paper Address?

HIV continues to be a significant global health concern, and while preventive measures like PrEP (Pre-Exposure Prophylaxis) and PEP (Post-Exposure Prophylaxis) exist, patient adherence to these medications is often low. This is attributed to a lack of clear information, communication challenges between patients and providers, and the time-intensive nature of personalized counseling. Existing solutions like social networks and telehealth have limitations. This paper addresses the need for an accessible, automated system that can provide consistent, reliable information and counseling on PrEP and PEP, thereby improving patient understanding and potentially adherence.

Methods

The core of this work is an ontology-based approach for building a conversational agent. The system utilizes two main ontologies:

  • Ontology of PrEP and PEP (OPP): This ontology serves as a knowledge base, storing atomic facts about PrEP and PEP, including dosing, benefits, harms, and costs, derived from patient-level sources like brochures and websites.
  • Patient Health Information Dialogue Ontology (PHIDO): This ontology models the dialogue flow, defining the sequence of utterances between the system and the user. It links to the OPP to retrieve and present health information.

The dialogue engine is built upon a finite state transition network that encodes the dialogue flow within the ontology. It employs lightweight Natural Language Processing (NLP) techniques to discern user utterances and a question-answering (QA) subsystem that leverages the ontology to provide answers. The system is designed with a deterministic, planned approach to ensure key counseling points are covered.

Datasets

The paper does not explicitly mention specific datasets used for training the NLP components. However, the Ontology of PrEP and PEP (OPP) was populated with information from "patient-level sources (brochures and websites)." The evaluation of the dialogue system's functionality was performed using the Trindi Tick list, a benchmark for assessing dialogue systems across 12 criteria.

Key Insights and Ideas

  • Ontologies as a Foundation for Health Dialogue: The paper strongly advocates for using ontologies to represent complex medical information and dialogue structures, enabling more robust and interpretable conversational agents.
  • Bridging Information Gaps: The system directly tackles the problem of insufficient patient information by providing a structured and accessible knowledge source for PrEP and PEP.
  • Automated Counseling for Adherence: By automating the counseling process, the agent can offer consistent, real-time support, potentially improving adherence rates where human resources are limited.
  • Potential for Personalized Interaction: While the current version has limitations in handling personalized context, the ontology-based approach lays the groundwork for future integration of user context for more tailored interactions.

Key Results

The paper presents a qualitative assessment of the system's functionality using the Trindi Tick list. The system demonstrated capabilities in:

  • Answering questions: It can answer questions related to PrEP and PEP, with examples showing ranked answers for a question about what PrEP protects against.
  • Handling basic dialogue flows: The system can manage introduction, health information delivery, and closing segments.
  • Responding to various utterance types: It can handle confirmations, disconfirmations, and some forms of clarification.

However, the paper also highlights areas where the system performs "Partially yes" or "No," indicating areas for improvement.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Handling Personalized Context: The current system is not sensitive to personal contextual information, such as previously resolved information or specific demographic tailoring. Future work aims to integrate user context ontologies to address this.
  • Negatively Specified Information: The system does not intentionally handle negatively specified information (e.g., "What if I do not have insurance?"). Future work could improve the analysis of such utterances.
  • No Answer Scenarios: When the system cannot answer a question, it currently waits for user input. Implementing timeouts or more sophisticated fallback mechanisms is suggested.
  • Ambiguous Designators: Similar to handling personalized context, the system's ability to deal with ambiguous designators is limited by ontology coverage.
  • Expanding Ontology Content: The OPP is in draft format and needs expansion to include details on where to obtain PrEP/PEP and information on support organizations.
  • Modeling Standard Practices: Future work will focus on modeling typical patient-provider counseling practices and conducting simulation studies with providers.
  • Voice Interface Integration: The current system is text-based; a future goal is to interface with a voice interface for a more natural user experience.

Summary updated

BibTeX
@inproceedings{amith2020towards,
  author = {Muhammad Amith and Licong Cui and Kirk Roberts and Cui Tao},
  title = {{Towards an Ontology-based Medication Conversational Agent for PrEP and PEP}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure