Health Insights Platform
Can AI Improve Personal Health? / Case
Our client sought to innovate their offerings by integrating a solution for personalized health insights.
Advanced health insights. Using large language models (LLMs) trained on WHO guidelines and medical research, the platform aimed to estimate life expectancy, calculate health indices, and provide personalized recommendations for healthier living.
Data consistency and accuracy. Integrating tools and methods to enhance data consistency and reliability across user health profiles, had to allow the platform to provide tailored and accurate health assessments.
User-focused AI model development. AI models based on user input and scientific data were to deliver meaningful, individualized health insights.
Phase 1: Mock-up development and model testing
In the initial phase, several AI models, including GPT-4o Mini, LLaMA-3, and Falcon 7B, were tested for their effectiveness in providing health insights. Key aspects evaluated were life expectancy and health index accuracy. Consistency metrics were fine-tuned, with significant improvements in the HEALTHY profile category after adjustments, ensuring that the model's life expectancy predictions aligned with user expectations and scientific standards.
This phase also included establishing a data flow (user input → data processing → health metrics generation) that streamlined communication between the user interface and the AI model. The web app allowed users to input key health metrics, which the AI analyzed to provide real-time insights, including a Healthy Lifestyle Index (HLI) and life expectancy estimation.
Phase 2: Prototype development and consistency enhancement
With a model selected, the prototype was developed to implement enhanced consistency measures and parameters.
The LLM used weighted scores across health parameters, including physical activity, dietary habits, sleep quality, and environmental factors. This approach ensured a comprehensive health index calculation aligned with each user’s lifestyle choices.
A base life expectancy was adjusted based on various health factors, with adjustments reflecting habits like smoking, physical activity, and diet quality. This model provided users with a practical and accurate life expectancy projection.
The prototype was integrated with a Python-based API, enhancing real-time processing of user data and enabling the app to return calculated insights almost instantly. This allowed for scalable integration with existing client platforms.
Challenges and solutions
The initial implementation faced issues with inconsistency in model predictions, particularly in life expectancy estimates across various user health profiles. Outputs were especially unstable for the UNHEALTHY profile, impacting the model’s reliability for users with complex health needs. To resolve this, we fine-tuned key parameters, such as seed and temperature, which significantly improved stability and predictability in the output. These refinements allowed the model to provide more accurate and consistent life expectancy predictions, enhancing its usability for a broader range of health profiles.
Data reliability and the need for parameter adjustments posed additional challenges in achieving a stable health index across diverse user profiles. The project required rigorous testing to ensure that the AI model could deliver accurate and repeatable health metrics for each user, regardless of profile variations. By testing a variety of LLMs, we identified models with the best accuracy for each profile. This targeted model selection, combined with tailored API adjustments, allowed the platform to enhance the reliability of health index predictions, addressing the variability across different health profiles.
Providing user-specific health recommendations proved challenging, as crafting guidance that applies universally required careful AI fine-tuning. Without proper customization, the AI might overlook critical health nuances unique to each user. To address this, we based the training on WHO guidelines and a range of validated medical literature, enabling the AI to generate tailored recommendations. This approach ensured that each user received actionable lifestyle suggestions, such as guidance on addressing high BMI or increasing physical activity, supporting users with personalized and relevant health insights.
Outcomes and impact
First, the model provided an accuracy improvement of over 20% in health index consistency, aligning with the goal of delivering reliable, data-driven health recommendations. Next, the real-time insights and science-backed recommendations increased user trust, as reflected in feedback received during testing sessions. Lastly, operational efficiency improved with automated health recommendations, minimizing manual data entry.