Contextualized Test-Time Adaptation for Clinical Data Utilizing In-Context Learning and Neighbour Graph
Keywords: test-time adaption, neighbour graph, graph learning, in-context learning, clinical data
TL;DR: This paper enhances test-time adaptation in clinical data by utilizing a graph-based methodology and intelligent in-context learning, focusing on intelligent prompt selection and effective exploitation of query-prompt relations.
Abstract: In the vast realm of clinical data, challenges arise in accommodating domain shifts and achieving efficient test-time adaptation. Unlike conventional Machine Learning (ML) models that focus only on source data, this paper proposes an innovative approach to effectively adapt the model at inference time, powered by in-context learning. Our technique utilizes personalized patient data, knowledge graphs, and external knowledge to construct prompts for patient features, symptoms, and diseases. Leveraging a graph-based methodology, we introduce a novel test-time adaptation technique that selects and retrains the model based on the nearest data points in a pre-established graph. We further explore the applicability of diffusion models for superior source distribution estimation. Experimental results demonstrate significant improvements in the efficiency of test-time adaptation for clinical data, marking a stride towards personalized healthcare.
Submission Number: 10
Loading