795,000 people die or are permanently disabled due to medical misdiagnosis in the US per year. Modern medicine can often focus on aggregates and averages, neglecting rarer diseases. This talk tries to push cutting-edge LLMs into precision-medicine, to turn simple equations into excellence at work.

From Equations to Excellence - Optimizing Generative AI for Precision Medicine

Problem Statement and Importance

According to BMJ Quality & Safety (from British Medical Journal, also stated on reports from Johns Hopkins Medicine) approximately 795,000 people die or are permanently disabled due to misdiagnosis a year in America alone. Current medicine focuses on aggregates and averages, and often neglects rarer diseases.

This program uses applied Generative AI and prompt engineering techniques to help fight against the norm, and push for better care on all fronts. The rare disease mentioned and tested here is Immune Thrombocytopenic Purpura (ITP), and it is a rare disease according to Johns Hopkins Medicine. This may also be referred to as Thrombocytopenia with Immune Thrombocytopenic Purpura, as thrombocytopenia is simply having a low platelet count. Another one of the most common names used is Idiopathic Thrombocytopenic Purpura. A quick description of ITP is that it is a disease where the body kills its own platelets.

ITP specifically was picked because of my own personal experience with misdiagnosis being able to cause serious damage. A member in my own family has ITP, and when he broke his arm and had to undergo surgery his condition caused surprise complications.

However, Generative AI models (such as Google’s Gemma 3n), as can be demonstrated can diagnose a rare disease, and one that has made nearly every injury amplified to the extremes.

This project tries to make sure that other families like mine won’t deal with the same situation and that more families can plan and be more informed regarding their diseases. This could help family members keep themselves and their loved ones safe, and not have to be sprung with problems in a life or death scenario.

Additionally, human doctors can have human problems, meaning that they can be fatigued, be overloaded from work, and be under stress. All of these factors can make humans perform worse.

Generative AI doesn’t have these problems. Furthermore, using GenAI is cheaper, easier, and faster than manually going to a doctor, scheduling an appointment, and paying hefty prices or waiting in long lines.

Framework

This project first parses the patient file into a JSON, before using internal reasoning to make diagnoses.

This is done because this project does not only hold the potential to save lives through prediction, but it also solves a problem that plagues many applications of technology in healthcare, which is unstructured data. By structuring the data so that it is very easy to understand and in a standardized format, it is more easy to use and implement.

The Use of Structured Data

According to the National Institutes of Health, 80 percent of medical data is unstructured. If we need to use data to create more products used in medicine, especially diagnosis prediction, it needs to be in a standard, recognized, and understandable format. Currently, some data can be written in barely spoken languages, may never be translated, and be very hard for professionals across the globe to use in products that can change lives.

This project helps to solve that. It parses the file into a JSON, outputting structured data, that AI can then diagnose.

This pushes the frontiers of precision medicine, structured data, and LLM/GenAI Power, made by a Richmond-based developer who wants to help the community.

Technical Requirements

Agreement to the Gemma 3n T&C, Google Chrome, and Kaggle. ## Why I am the Best Person to Speak on This Subject I am a Richmond-based developer that loves working with applied AI, especially for the greater good in fields of medicine. Some notable projects of mine include “DoctorBot”, which was a Gemini-powered system to help doctors and used Google Search powered RAG systems. My second project for Google DeepMind’s Gemma 3n Impact Challenge is what inspired this talk, where I made an LLM powered diagnosis support framework that used the on-device Gemma 3n with this “Parse-Analyze” framework that has yielded such great results. I have also chaired at PyData Virginia, where I oversaw sessions on the use of AI to predict diseases and Python to study health disparities. These projects show my experience in precision medicine, how I have worked with cutting-edge tools to push the frontier, and why I am one of the best people to speak on the subject.