HealthGod- Personalised AI Health & Longevity Coach

Inspiration

People have MANY questions about their day-today HABITS around preventative health. Many people have their bio-markers data logged, and want to take better habits backed by research. But Scientific research around Preventative Health, wellbeing and Longevity (Diet, Fitness, Habits) is VERY complicated, humengous and changing very often and frequently. It is Not possible for common people to keep up with these updates and research. Preventative health is hence currently taking a back seat over medical health AFTER getting health issues. We want to democratice access to information related to preventative health, fitness and longevity.

What it does

Knows your bio-markers (heart-rate, BMI, step count, sleep score, blood glucose, workout activity, etc) by connecting to services like Apple Health, Fitbit, Blood test results, user input, etc. Answers your day-to-day questions about Preventative Health, Well-being and Longevity around Habits, Diet, Fitness, Longevity, etc. We are releasing this as a ChatGPT Plugin, to which people can ask questions about preventative health, well-being and Longevity.

For example "I am drinking coffee 3 times a day most days. I don't have issues with sleeping. But is that okay? What does research tells about this?"

"I am planning to do intermittent fasting to reduce weight. Is it suggestible for me?" the AI knows the age, health condition (glucose levels, diabetic or Not, demographics etc) of the user, and gives suggestion based on the user's personal bio-marker information, science research, podcasts, etc with citation.

How we built it

We trained a model on Scientific Research papers and journals like Arxiv, BioArxiv, etc. and implemented an index and a search layer to fetch the relevant documents and passages for the query. In the final step. we used GPT-4 to collate the results and attribute every line in the result to relevant research paper or podcast episode. This will prevent the possibility of Hallucinations.

Apart from training on Research papers, We also trained the AI model on podcasts of Dr. Eric Berg DC, Andrew D. Huberman, and Peter Attia by fetching the transcripts using Whisper API, etc and then created FAISS Embeddings. Connecting all of this with GPT plugin.

Challenges we ran into

Data gathering (extracting text from transcripts;; AI able to understand who is the speaker) was time consuming.

We have to create our OWN index, our OWN search algorithm, which does not use GPT, so that we are able to attribute relevant sources, and prevent possibility of Hallucinations.

Accomplishments that we're proud of

We are able to answer Personalised Health , Wellbeing and Longevity questions around Preventative Health, with high accuracy, and able to CITE every answer with scientific papers, Journals, articles and Podcasts

What we learned

GPT is prone to hallucinations particularly for science based research and findings, where it gives its own answers and citations which do Not exist. We had to create our own Index, our own Search algorithm, model , and use GPT only for the final summarisation part.