Agentic Feature Selection via LLM for Epileptic Seizure Detection

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Health & Medical Research · Depth: Expert, quick

Summary

An investigation into LLM-guided feature selection for automated epileptic seizure detection from electroencephalography (EEG) signals used a small instruction-tuned LLM, Qwen2.5-1.5B-Instruct. This LLM agent, receiving five statistical summaries, iteratively selected a feature subset from the Epileptic Seizure Recognition dataset (11,500 samples, 178 features). The agent achieved 96.5% accuracy and 0.911 F1 using 40 features, compared to the best full-feature baseline (SVM-RBF on 178 features) which reached 97.9% accuracy and 0.946 F1. A critical finding revealed that 39 of the LLM's 40 selected features matched the top-39 mutual-information features. Consequently, a deterministic Top-39 MI filter, evaluated with the same Random Forest classifier, yielded identical performance: 96.5% accuracy and 0.911 F1. This work establishes an empirical baseline, suggesting that a 1.5B-parameter LLM behaves similarly to a univariate MI ranker for this task.

Key takeaway

For Machine Learning Engineers developing automated seizure detection systems, this research suggests that current small LLM-based feature selection approaches may not offer significant advantages over established statistical methods like mutual information ranking. Before investing in LLM-driven feature selection, you should benchmark against simpler, deterministic statistical filters, especially for tasks with well-understood feature spaces. Consider exploring larger or domain-specialized LLMs only after validating the limitations of traditional techniques.

Key insights

Small LLMs for feature selection currently mimic univariate statistical ranking methods.

Principles

• LLM-guided feature selection can achieve high accuracy.
• Smaller LLMs may replicate statistical methods.
• Mutual Information is a strong feature selection criterion.

Method

An LLM agent (Qwen2.5-1.5B-Instruct) receives five statistical summaries and performs multi-round reasoning to select a feature subset for binary classification.

In practice

• Evaluate LLM feature selection against strong statistical baselines.
• Consider mutual information for initial feature ranking.
• Explore larger LLMs for advanced feature interactions.

Topics

• Epileptic Seizure Detection
• EEG Signal Processing
• Feature Selection
• Large Language Models
• Mutual Information
• Qwen2.5-1.5B-Instruct

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Predicting Post-Traumatic Epilepsy from Clinical Records using Large Language Model Embeddings — LLM embeddings from clinical records can predict post-traumatic epilepsy with high accuracy.
• LLM-Driven Feature Discovery — LLMs can effectively discover and label qualitative behavioral features from model transcripts without internal access.
• Joint MVMD-based optimal feature selection and FW-LS-TWSVM for motor imagery recognition — Combining MVMD feature selection with FW-LS-TWSVM significantly boosts MI-BCI decoding accuracy and efficiency.
• Lazy or Efficient? Towards Accessible Eye-Tracking Event Detection Using LLMs — An LLM-driven pipeline automates eye-tracking event detection, making complex analysis accessible without specialized programming.
• A Few-Shot LLM Framework for Extreme Day Classification in Electricity Markets — LLMs can effectively classify electricity price spikes with limited data by interpreting natural language prompts of system conditions.
• Seizure-Semiology-Suite (S3): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding — MLLMs struggle with fine-grained medical video interpretation, but domain-specific methods significantly boost diagnostic accuracy for seizure semiology.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.