[Multi-task learning with sleep features for interictal epileptiform discharge detection: a model development and validation study].

Abstract

Objective: To establish and validate an automated detection model for interictal epileptiform discharges (IED) through a multi-task learning algorithm that integrates sleep features, providing more precise electroencephalogram (EEG) interpretation support for clinical practice. Methods: Based on convolutional neural networks, a multi-task learning model Siamese-ES that integrates sleep feature was developed. The dataset comprised EEG recordings from 150 patients at Peking Union Medical College Hospital Epilepsy Center from March 2019 to April 2023, of which 140 cases were diagnosed with epilepsy, and the other 10 cases were non-epileptic patients without IED. There were 79 male and 71 female patients, with an age of 27 (3-87) years. After EEG data preprocessing and time-frequency conversion, EEG features were put into two networks, Twins-Electron and Twins-Sleep, to extract the IED features and deep sleep features respectively. Then the features were fused for IED detection. Siamese-ES and two classic single-task IED detection models were trained on the same dataset EpiSet-260K for model comparisons. Additionally, ablation experiments of sleep features and multi-task learning modes were set up to verify the effectiveness. Results: The EpiSet-260K dataset contained 265, 551 samples. The multi-task learning Siamese-ES model integrated with sleep features showed better precision (71.18%), specificity(98.46%), F1 value(76.26%) and area under the curve [0.978 (95%CI:0.977-0.980)]. Moreover, the ablation experiments confirmed that integration of sleep features through a multi-task learning algorithm achieved better evaluation markers. At 80.00% sensitivity, sleep features can improve precision by 1.19%, and multi-task learning mode can improve precision by 2.18%. Conclusions: Our study demonstrates that the Siamese-ES model effectively improves the performance of IED detection model through sleep features and multi-task learning algorithm. The results provide future research directions for IED detection models in different real clinical scenarios.