TY - GEN
T1 - Deep Learning for Classifying Mild Traumatic Brain Injury for the Need of CT Scans Using EEG Signals
AU - Nelavagal Sridhara, Deepika
AU - Hareesha, K. S.
AU - Hegde, Ajay
AU - Menon, Girish
AU - Srinivasan, Siddharth
AU - Murthy, Arjun Anand
AU - Swamy, P. T.
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.
PY - 2025
Y1 - 2025
N2 - Traumatic brain injury (TBI) can result from either a minor bump on the head or significant damage to the brain. MRIs and CT scans are the standard medical procedures, which is unnecessary in the event of a negative result. This paper proposes deep learning approaches for classifying patients into CT positive and CT negative categories and the effects of EEG segmentation for patients with mild TBI (mTBI). The proposed algorithms are convolutional neural networks (CNNs), long short-term memory (LSTM), and gated recurrent units (GRUs) to perform classification. In the proposed architectures, preprocessing of EEG is not performed. The proposed architectures of CNN, LSTM, and GRU resulted in high-performing accuracy of 96±0.3%, 96.74±0.30%, and 97±0.20%, respectively. The segments with a length of 120 fared better than other segments, according to the results of the various segments. Furthermore, our research shows that the models performed better when the EEG data segments were shorter. This suggests that the models could be applied in real-time scenarios to produce faster outcomes, especially in emergency rooms.
AB - Traumatic brain injury (TBI) can result from either a minor bump on the head or significant damage to the brain. MRIs and CT scans are the standard medical procedures, which is unnecessary in the event of a negative result. This paper proposes deep learning approaches for classifying patients into CT positive and CT negative categories and the effects of EEG segmentation for patients with mild TBI (mTBI). The proposed algorithms are convolutional neural networks (CNNs), long short-term memory (LSTM), and gated recurrent units (GRUs) to perform classification. In the proposed architectures, preprocessing of EEG is not performed. The proposed architectures of CNN, LSTM, and GRU resulted in high-performing accuracy of 96±0.3%, 96.74±0.30%, and 97±0.20%, respectively. The segments with a length of 120 fared better than other segments, according to the results of the various segments. Furthermore, our research shows that the models performed better when the EEG data segments were shorter. This suggests that the models could be applied in real-time scenarios to produce faster outcomes, especially in emergency rooms.
UR - https://www.scopus.com/pages/publications/105018795886
UR - https://www.scopus.com/pages/publications/105018795886#tab=citedBy
U2 - 10.1007/978-981-96-4008-9_8
DO - 10.1007/978-981-96-4008-9_8
M3 - Conference contribution
AN - SCOPUS:105018795886
SN - 9789819640072
T3 - Lecture Notes in Networks and Systems
SP - 95
EP - 108
BT - Advances in Health Informatics, Intelligent Systems, and Networking Technologies - Proceedings of HINT 2024
A2 - Jeyabose, Andrew
A2 - Jeyabose, Andrew
A2 - Balas, Valentina Emilia
A2 - Balas, Valentina Emilia
A2 - Fernandes, Steven L.
PB - Springer Science and Business Media Deutschland GmbH
T2 - International Conference on Health Informatics, Intelligent Systems, and Networking Technologies, HINT 2024
Y2 - 13 March 2024 through 14 March 2024
ER -