Healthify and OpenAI Collaboration for AI Health Coaching

Healthify, India's largest health platform, has partnered with OpenAI to integrate advanced generative AI into its health tracking and coaching services. This collaboration allows Healthify to scale its personalized coaching and simplify dietary tracking for its 40 million users, aiming to prevent metabolic disease and promote sustainable weight loss.

Overcoming Limitations of Traditional ML Systems

Healthify previously relied on a combination of hierarchical LSTMs, custom NLU systems, and Convolutional Neural Networks (CNNs) for its AI features. While these systems enabled the launch of Ria (a virtual nutritionist) and Snap (a food photo recognition tool), they faced significant technical hurdles:

  • Recognition Accuracy: The Snap feature struggled with photos containing multiple food items and required multiple iterations to achieve 80% accuracy for single Indian foods.
  • Complex Query Handling: Ria was rules-based and could not answer complex, cross-functional queries regarding the relationship between nutrition and other health metrics.
  • Scalability and Localization: Expanding into new markets, such as Southeast Asia, took two years due to the effort required to localize models for local languages, foods, and exercise routines.

Technical Integration of OpenAI Models

Healthify replaced and augmented its existing pipelines with an ensemble of OpenAI models to improve accuracy and global scalability:

GPT-4 Vision for Food Tracking

Healthify integrated GPT-4 Vision to replace the previous CNN-based pipeline. This model allows the system to recognize multiple food items from across the world instantly, matching human-level recognition accuracy. To maintain data privacy and precision, Healthify uses an ensemble of proprietary models alongside GPT-4 Vision, applying custom heuristic models for final food recommendations.

GPT-4 Turbo and GPT-3.5 for Personalized Coaching

Ria, the virtual nutritionist, now utilizes an ensemble of fine-tuned GPT-3.5 and GPT-4 Turbo models. This allows the AI to interpret Healthify's extensive literature and a user's unique history to provide personalized advice. Ria can now correlate data from Continuous Glucose Monitors (CGM), food logs, and wearable sleep logs to answer complex queries such as "How have my glucose levels affected my sleep yesterday?"

Embeddings for Data Mapping

To bridge the gap between the food names generated by GPT-4 and Healthify's internal food database, the team utilized the OpenAI Embeddings model. By using cosine similarity matching between the GPT-identified food name and Healthify's own food embeddings, the platform achieved high accuracy in food matching.

Whisper for Coach Copilot

Healthify utilized Whisper to enhance its Coach Copilot, the assistant used by human coaches to manage their clients more efficiently.

Impact on User Engagement and Health Outcomes

The integration of OpenAI's API has led to measurable increases in user engagement and retention:

  • Increased Food Tracking: Users now track food 50% more often using Snap, which Healthify expects will correlate to a 50% increase in fitness outcomes.
  • Enhanced Coaching Engagement: Conversation lengths with Ria have doubled, with some interactions exceeding 200 messages.
  • Coach Efficiency: Human coaches can respond to clients in half the time, and clients engage with AI-supported coaches 18% more.

According to Stanford research based on Healthify's data, AI-enabled human coaching helps clients lose 70% more weight than AI coaching alone.

Future Goals: Autonomous Health Agents

Healthify aims to launch in 20 countries within a year, significantly reducing the localization time compared to previous traditional ML systems. The company's long-term goal is to develop autonomous health agents that proactively analyze health data to make food, sleep, and exercise recommendations. These agents will be capable of performing actions such as ordering food or booking gym classes with user permission.

Sources