Heet D's Senior Project Blog

Project Title:  Zero-Shot and Few-Shot Learning for Rare Disease Diagnosis BASIS Advisor: Ms. Briggs Internship Location: ASU/Virtual Onsite Mentor: Professor Jianming Liang.

Project Abstract

In a small clinic, a mother clutches her child’s hand as another round of inconclusive test results comes in. For years, she’s searched for answers to a rare disease no one seems to recognize. Doctors don’t know how to treat it and there is little to no data on past cases. I believe emerging techniques like zero-shot and few-shot learning are poised to change stories like hers, offering faster and more accurate diagnoses where traditional methods fail. A rare disease is defined as a condition that affects less than 200,000 people. But despite each disease being uncommon individually, in total they affect about 300 million people worldwide. This low individual prevalence combined with the high collective impact poses multiple challenges for both doctors and patients. My research evaluates zero-shot and few-shot learning as a potential solution. These machine learning techniques allow AI models to identify diseases with little to no training data, which is necessary for rare diseases. My research will evaluate the performance of models like GPT-4, CLIP, and BioBERT. I will use publicly available image datasets, like Chest X-ray images and Skin Cancer MNIST which are available on Kaggle, along with textual datasets. The models will be tested in both zero-shot and few-shot learning settings, and the performance will be assessed based on metrics like accuracy, bias, and generalizability. The results will provide insight into the effectiveness of each model, technique, and data type, highlighting the potential of zero-shot and few-shot learning to be used as real-world tools for supporting healthcare professionals in diagnosing rare diseases more accurately—and ultimately, saving lives.