Few-Shot Optimization for Sensor Data Using Large Language Models: A Case Study on Fatigue Detection.

Area of Science:

• Machine Learning
• Signal Processing
• Wearable Technology

Background:

• Few-shot prompting is efficient for limited labeled data but sensitive to example selection quality.
• Sensor-based classification tasks, like fatigue detection, require nuanced example selection due to complex data patterns and variability.

Purpose of the Study:

• To introduce a novel few-shot optimization method, Hybrid Euclidean Distance with Large Language Models (HED-LM), for improved example selection.
• To enhance the performance of sensor-based classification tasks, particularly fatigue detection, by optimizing the selection of training examples.

Main Methods:

• HED-LM employs a hybrid pipeline: filtering candidate examples using Euclidean distance and re-ranking them with contextual relevance scores from Large Language Models (LLMs).
• The method was validated on a fatigue detection task using accelerometer data, known for overlapping patterns and high inter-subject variability.

Main Results:

• HED-LM achieved a mean macro F1-score of 69.13 ± 10.71% in fatigue detection.
• This significantly outperformed random selection (59.30 ± 10.13%) and distance-only filtering (67.61 ± 11.39%), showing relative improvements of 16.6% and 2.3%, respectively.

Conclusions:

• Combining numerical similarity (Euclidean distance) with contextual relevance (LLMs) enhances the robustness of few-shot prompting.
• HED-LM provides a practical approach for real-world sensor-based learning and has potential in healthcare monitoring, activity recognition, and industrial safety.