From Few to None: Exploring Few-Shot, One-Shot, and Zero-Shot Deep Learning in Clinical Settings

Organizers: Ahmed P. Tafti, Yu-Chiao Chiu, Yanshan Wang, Yufei Huang, Dana Tudorascu

Affiliation: University of Pittsburgh

Abstract:

Deep learning algorithms have made significant advances in a wide range of clinical applications, including medical imaging informatics, clinical natural language processing, and health data sciences. However, traditional deep learning methods often require large columns of manually annotated and gold-standard data, which can be expensive and very time-consuming in clinical settings. In recent years, there has been growing interest in developing deep learning methods that can learn from a few samples or even no samples at all, known as few-shot and zero-shot learning, respectively. The current tutorial at IEEE BHI 2023 aims to mainly provide a professional forum to share state-of-the-art in few-shot, one-shot, and zero-shot learning in two different settings: (1) Medical Imaging Informatics and Transcriptomics, and (2) Clinical Natural Language Processing. In this tutorial, which is also equipped with a hackathon, we will be exploring the latest advances in few-shot deep learning. We will start with an overview of the basic concepts and techniques involved in these areas, and we will then dive into specific applications of few-shot, one-shot, and zero-shot learning in medical imaging, image embeddings of molecular (RNA-Seq) data, and clinical natural language processing. Participants will have the opportunity to work on hands-on coding exercises and challenges, using popular deep learning frameworks such as PyTorch. They will learn how to build, train, validate, and test few-shot models, and how to evaluate their performance on different datasets. By the end of the tutorial, participants will have a solid understanding of the state-of-the-art in few-shot deep learning and will achieve practical experience in implementing and validating these techniques.

How to tame your ChatGPT: Complex Reasoning with LLMs using Few-shot prompting

Organizers: Aman Madaan

Affiliation: Carnegie Melon University

Abstract:

Few-shot prompting, a technique where a model is presented with a few examples (or ‘shots’) of a task, and then ‘prompted’ with a similar task, has been transformative in the field of natural language processing. It has enabled us to leverage the power of Large Language Models (LLMs) such as ChatGPT and PaLM2 to effectively respond to a wide array of tasks without requiring extensive task-specific training. However, employing few-shot prompting for complex reasoning tasks introduces unique challenges. The performance of these models can be highly sensitive to the prompt’s specifics, and they may display peculiar failure modes.

This tutorial aims to present techniques that can help navigate these challenges. We will start with an overview of LLMs and the principles of few-shot prompting. Following that, we will explore various strategies that have been developed to enhance the robustness and performance of few-shot prompting. We will discuss ‘prompt-design’ techniques, like chain-of-thought prompting and least-to-most prompting, that guide an LM to include reasoning steps in its output. We’ll also introduce dynamic prompting techniques that adapt the prompt for each specific input, like selecting the best fitting training examples for the prompt and modifying the prompt based on feedback from similar tasks.

Towards the end, we will delve into the use of code-generation for generating structured outputs (like graphs and plans) from LLMs. The tutorial will be interspersed with hands-on exercises that offer a practical understanding of these techniques. These exercises will include tasks using biomedical and health informatics data. The tutorial is designed to be broadly accessible, and only some familiarity with Python is necessary.

The Single-Cell Spatial Transcriptomics Analysis (ScSTA) Cookbook

Organizer: Arun Das

Affiliation: University of Pittsburgh

Abstract:

This tutorial aims to provide attendees with comprehensive knowledge and hands-on experience in the data analysis of sub-cellular resolution single-cell spatial transcriptomics (ScST) technologies such as Nanostring CosMx and 10X Genomics Xenium. ScST is a cutting-edge technology that profiles single-cell transcriptomics with spatial information of cells/transcripts in intact tissues. While it enables researchers to understand gene expression patterns in the context of tissue structure and disease pathology, it poses unique challenges in data analysis. This tutorial will cover each step of the analysis workflow and showcase typical use cases. It intends to accelerate the adoption of spatial single-cell transcriptomics analysis and foster breakthroughs in our understanding of complex biological systems.

Objectives:

• Introduce participants to the principles and significance of ScST.
• Familiarize attendees with the components of the Single-Cell Spatial Transcriptomics Analysis (ScSTA) pipeline, including image registration, cell segmentation, cell type identification, spatial gene expression analysis, spatial correlation analysis, and more.
• Provide practical experience in using relevant computational tools and software for ScST.
• Illustrate the application of ScST in disease pathology research, biomarker discovery, and patient outcome prediction.

Proposed Agenda:

Module 1: Introduction to Single-Cell Spatial Transcriptomics technology

• Overview of single-cell spatial transcriptomics and its limitations.
• Importance of incorporating spatial information for deeper insights.
• Introduction to the ScSTA pipeline and its key components.

Module 2: Image Registration and Cell Segmentation

• Familiarizing the popular libraries for image registration and cell segmentation.
• Importance of maintaining good image registration and cell segmentation in whole-slide-image (WSI) tissues.

Module 3: Data Preprocessing, Quality Control, and Cell Typing

• Cleaning and normalization of the segmented single-cell spatial data.
• Addressing technical variability and batch effects.
• Quality control measures for reliable downstream analysis.
• Using a scRNA-seq reference genomic profiles to carry out cell type transfer to spatial single-cell transcriptomics datasets.
• Using marker-based cell typing for ScST datasets.

Module 4: Spatial Mapping, Neighborhood Analysis, and Visualization

• Extracting spatial features such as cell density, average gene expression, and neighborhood cell type composition.
• Identifying spatial hotspots using spatial correlation methods.
• Developing visualization techniques for spatial mapping of the extracted spatial features.

Module 6: Hands-on Practical

• Guided tutorial using real-world single cell spatial transcriptomics dataset.
• Step-by-step walkthrough of the ScST analysis pipeline.
• Interactive assistance and troubleshooting.

Module 7: Future Directions and Challenges

• Emerging trends in single-cell spatial omics technologies.
• Ongoing challenges and opportunities in the field.
• Collaboration possibilities and resources for continued learning.