Toolkit to Create a Personalized Medication Adherence Chatbot

Step 1: Define the Purpose and Workflow

Objective: To assess and support patients’ medication adherence using a structured survey (MMAS-4 & MMAS-8), determine adherence issues, and develop a Medication Adherence Action Plan (MAAP).

Workflow Outline:

• Collect initial adherence data (MMAS-4).
• For/medium adherence, extend to MMAS-8 questions.
• Classify adherence type: Intentional vs. Unintentional.
• Map to the World Health Organization (WHO) dimensions of adherence (e.g., socioeconomic, healthcare team-related).
• Develop a MAAP based on identified needs.

Step 2: Design the Chatbot’s Conversational Flow

MMAS-4 & MMAS-8 Questions:

• Integrate the validated Morisky Medication Adherence Scale (MMAS-4) questions to start, and based on responses, proceed to MMAS-8 if needed.
• Score and code the MMAS8 to define intentional vs. unintentional non-adherence.
• Develop questions that categorize adherence barriers according to WHO dimensions: Socioeconomic, Healthcare Team, Condition-related, Therapy-related, and Patient-related.

Step 3: Train a Generative AI Model

Select a Pre-trained Model:

• Use a language model like OpenAI’s GPT, Gemini or any similar model fine-tuned for healthcare applications.
• Implement the Adherence Survey Logic

Build MMAS-4 and MMAS-8 Question Logic:

• Create a system to administer the MMAS-4 and automatically score responses.
• If the MMAS-4 score indicates low/medium adherence, trigger the MMAS-8 questionnaire.

​Develop Classification Algorithms:

• Use decision logic (e.g., rule-based or probabilistic classification) to determine if non-adherence is intentional or unintentional.

WHO Dimension Mapping:

• Map the responses to one or more of the WHO dimensions using pre-defined rules.

Fine-tune the Model:

• Fine-tune using the prepared dataset, focusing on question interpretation, empathy, and accuracy.

Step 4: Generate the Medication Adherence Action Plan (MAAP)

Create a table for Action Plan Components:

• key adherence steps, actions, responsibility, and completion dates that align with identified issues.
• generative AI will populate the table with creations - unintentional (suggest reminders or pill organizers) or intentional (provide motivational information or consult options).
• Train the AI model to pull from a library of tailored responses.