DeepFeature: Iterative Context-aware Feature Generation for Wearable Biosignals

TL;DR

DeepFeature tackles the challenge of deriving task-specific, context-aware features from multimodal wearable biosignals by integrating three feature-generation sources (direct LLM prompts, context-guided literature-informed prompts, and operator-based feature composition) with a robust, multi-layer feature-to-code translation and an iterative performance-driven refinement loop. The framework uses filtering and verification to ensure executable, reliable feature extraction, and employs ranking-based feature elimination plus model-performance feedback to progressively improve feature sets. Empirical results across eight healthcare tasks on four datasets show DeepFeature achieves an average AUROC improvement of 4.21-9.67% over baselines and outperforms state-of-the-art methods on five tasks, with ablations confirming each component’s contribution. The approach emphasizes offline feature engineering with acceptable system overhead, enabling robust, context-aware deployment on wearable biosignal platforms and offering a scalable path for diverse clinical applications.

Abstract

Biosignals collected from wearable devices are widely utilized in healthcare applications. Machine learning models used in these applications often rely on features extracted from biosignals due to their effectiveness, lower data dimensionality, and wide compatibility across various model architectures. However, existing feature extraction methods often lack task-specific contextual knowledge, struggle to identify optimal feature extraction settings in high-dimensional feature space, and are prone to code generation and automation errors. In this paper, we propose DeepFeature, the first LLM-empowered, context-aware feature generation framework for wearable biosignals. DeepFeature introduces a multi-source feature generation mechanism that integrates expert knowledge with task settings. It also employs an iterative feature refinement process that uses feature assessment-based feedback for feature re-selection. Additionally, DeepFeature utilizes a robust multi-layer filtering and verification approach for robust feature-to-code translation to ensure that the extraction functions run without crashing. Experimental evaluation results show that DeepFeature achieves an average AUROC improvement of 4.21-9.67% across eight diverse tasks compared to baseline methods. It outperforms state-of-the-art approaches on five tasks while maintaining comparable performance on the remaining tasks.

Figures (16)

• Figure 1: While training ML models using raw wearable biosignals can be unsatisfactory, DeepFeature automates context-aware feature generation from multi-modal biosignals (e.g., GSR, ECG, EEG), enabling diverse downstream healthcare tasks.
• Figure 2: Performance comparison of features directly generated by LLMs (with and without feedback) and expert-validated features on WESAD dataset.
• Figure 3: Error distribution in code generation and execution across three independent trials (Repetition 1--3). The errors are categorized as follows: Error 1 (import failures), Error 2 (runtime crashes), Error 3 (incomplete implementations), and Error 4 (logical flaws).
• Figure 4: Overview of DeepFeature. DeepFeature leverages LLMs to generate feature descriptions by combining task settings with expert knowledge. These features are translated into executable code to transform raw biosignals into tabular data. After being enhanced with mathematical operators, the tabular data is used to trained downstream models. The model's performance then feeds back to the LLMs to refine the feature set in the next iteration.
• Figure 5: Multi-source feature generation.