Publication: Prognostic prediction in moderate traumatic brain injury using hybrid deep learning with resting-state electroencephalogram
Loading...
Date
2024-07-01
Authors
Nor Safira Elaina, Mohd Noor
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Outcome prognostication of traumatic brain injury (TBI) is critical to clinical decision making and healthcare policy-making. Visual interpretation of electroencephalogra phy (EEG) is time-consuming, subjective, and limited to human recognizable features. This research presents two studies. In the first research, a prognostic model was con structed by random undersampling boosting decision trees (RUSBoost) to EEG signals to select important features to discriminate between good and poor prognoses. In a resting-state without cognitive demands, 32 EEG dataset were acquired from 18 EEG recordings of patients with moderate TBI during follow-up examinations with eyes closed to reduce α-band activity. Patient outcome at 4-10 weeks to 12 months was di chotomized into poor (GOS score ≤ 4) and good prognoses (GOS score = 5) based on the Glasgow Outcome Scale (GOS). RUSboost outperformed cSVM, DT, and kNN in absolute power spectral density (PSD) with AUCδ = 0.73, and AUCα = 0.69. There fore, this research was extended to develop a novel prognostic model using 60 seg ments of raw and preprocessed resting-state EEG seqeunces as input. An architecture of hybrid bidirectional long short-term memory (BiLSTM) and ensemble-AdaBoost that can accurately distinguish between a good and poor prognoses was proposed. Hy bridization of two or more algorithms may improve forecasting. Following this way, the advantages of BiLSTM and AdaBoost algorithms can be utilized simultaneously. This study demonstrated that the deep integration method of BiLSTM-AdaBoost and BiLSTM-Random Forest (RF) perform better than single LSTM, BiLSTM, AdaBoost, and RF models. The proposed hybrid BiLSTM-AdaBoost model beats all implemented and existing models with the most remarkable mean AUC value of 99.51 ± 1.36% of preprocessing and 98.91 ± 2.47% of raw input. The proposed prognostic model is computationally efficient, processing raw EEG input in 4.11 seconds and EEG prepro cessed input in 5.71 seconds. This hybrid deep learning (DL) framework has the poten tial to revolutionize EEG-based TBI outcome prognostication. The hybrid DL model emerges as an automatic method for predicting TBI outcomes, suggesting promising implications for clinical use.