A Hybrid Artificial Intelligence System for Automated EEG Background Analysis and Report Generation

TL;DR

An innovative hybrid artificial intelligence system for automatic interpretation of EEG background activity and report generation that combines deep learning models for posterior dominant rhythm (PDR) prediction, unsupervised artifact removal, and expert-designed algorithms for abnormality detection is proposed.

Abstract

Electroencephalography (EEG) plays a crucial role in the diagnosis of various neurological disorders. However, small hospitals and clinics often lack advanced EEG signal analysis systems and are prone to misinterpretation in manual EEG reading. This study proposes an innovative hybrid artificial intelligence (AI) system for automatic interpretation of EEG background activity and report generation. The system combines deep learning models for posterior dominant rhythm (PDR) prediction, unsupervised artifact removal, and expert-designed algorithms for abnormality detection. For PDR prediction, 1530 labeled EEGs were used, and the best ensemble model achieved a mean absolute error (MAE) of 0.237, a root mean square error (RMSE) of 0.359, an accuracy of 91.8% within a 0.6Hz error, and an accuracy of 99% within a 1.2Hz error. The AI system significantly outperformed neurologists in detecting generalized background slowing (p = 0.02; F1: AI 0.93, neurologists 0.82) and demonstrated improved focal abnormality detection, although not statistically significant (p = 0.79; F1: AI 0.71, neurologists 0.55). Validation on both an internal dataset and the Temple University Abnormal EEG Corpus showed consistent performance (F1: 0.884 and 0.835, respectively; p = 0.66), demonstrating generalizability. The use of large language models (LLMs) for report generation demonstrated 100% accuracy, verified by three other independent LLMs. This hybrid AI system provides an easily scalable and accurate solution for EEG interpretation in resource-limited settings, assisting neurologists in improving diagnostic accuracy and reducing misdiagnosis rates.

Figures (8)

• Figure 1: Workflow of the proposed hybrid AI system for automated EEG background interpretation and report generation.
• Figure 2: Neighbor electrode example and artifact repair using the neighbor-HBOS method. (A) F3 and its neighboring electrodes: Fp1, F7, and C3. (B) EEG signal before artifact repair. (C) EEG signal after artifact repair using the neighbor-HBOS method.
• Figure 3: Example of features used for PDR prediction. (A) Original EEG with a labeled PDR value of 9.5 Hz for both the right and left hemispheres. The ensemble model's predicted values are 9.4 Hz for the right hemisphere and 9.5 Hz for the left hemisphere. (B) Feature maps for the right and left PDR, displayed as images. The right PDR map includes electrodes [T6, O2, P4, T5, O1, P3], while the left PDR map includes electrodes [T5, O1, P3, T6, O2, P4]. (C) Power spectral density (PSD) features for each electrode, showing the power distribution across frequencies from 3 to 15 Hz.
• Figure 4: Overview of the deep learning pipeline for PDR prediction model. The input dataset X contains power spectral density (PSD) values with a frequency band range of 3 to 15Hz, window size of 0.25Hz, PDR range of 4 to 12Hz. The dataset is normalized with X values scaled by the maximum X value and Y values scaled by (y-4)/8. The normalized dataset is split into a training set (70%) and test set (30%). The training data is used to train a CNN model, which could be GoogleNet, ResNet, or an ensemble approach. The trained model outputs predictions between 0 and 1.
• Figure 5: The workflow for generating and verifying EEG reports using large language models (LLMs). The process begins with an input prompt that combines the EEG features. This prompt is then passed to the Google Gemini 1.5 Pro API for text generation. The API generates a report text, which subsequently undergoes an accuracy verification step using an ensemble of three LLMs: Gemini 1.5 Flash, Claude 3 Sonnet, and GPT-4o. These models independently assess the report's accuracy, and the final verification result is determined based on the majority agreement among the three LLMs.