Improving Seizure Forecasting with Deep-Learning Models: Integrating Multiday Seizure Patterns with Long-Term Subcutaneous EEG
Abstract number :
2.128
Submission category :
4. Clinical Epilepsy / 4B. Clinical Diagnosis
Year :
2023
Submission ID :
364
Source :
www.aesnet.org
Presentation date :
12/3/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Benjamin Brinkmann, PhD – Mayo Clinic Hospital
Mona Nasseri, PhD – Mayo Clinic Hospital; Tal Pal Attia, MS – Mayo Clinic Hospital; Jordan Clark, MS – Mayo Clinic Hospital; Vladmir Sladky, Bachelor – Mayo Clinic Hospital; Kremen Vaclav, PhD – Mayo Clinic Hospital; Gregory Worrell, MD, PhD – Mayo Clinic Hospital; Jie Cui, PhD – Mayo Clinic Hospital
Rationale:
The unpredictability of seizures is an overwhelming challenge for patients living with epilepsy. The ability to forecast seizures could have a positive impact on their daily lives. While seizure forecasting based on intracranial EEG (iEEG) has been established, it is not suitable for most patients. Our previous work demonstrated the feasibility of forecasting seizures using patient-specific long-short-term-memory (LSTM) deep-learning models. However, reducing false positive alarms (FPA) on non-seizure days is necessary to avoid patient alarm fatigue and to improve specificity of acute pharmacotherapy.
Methods:
Our model was based on subcutaneous EEG (sqEEG) signals, a novel, minimally invasive method for long-term recording of brain signals. The LSTM model architecture consisted of five consecutive LSTM layers. The output was a fully connected layer with a sigmoid activation function for classification. Inputs included two channels of raw sqEEG signals, two channels of FFT amplitude of the signals, one channel of time-of-day of the signals, and, importantly, an additional channel for time-from-last-seizure (TFLS). The use of TFLS signal integrates circadian and multiday cycles of seizure information into the model. This is motivated by recent evidence indicating the existence of periodicity in epileptic brain activity at multiple timescales and is a major innovation in this investigation.
Results:
We investigated the effects of adding TFLS signals on the model performance on two subjects (S1 & S2). S1 recorded 12 seizure days out of 152 days of total recordings and S2 recorded eigh seizure days out of 54 recording days. For S1, before applying TFLS, the model acquired an AUC of 0.80, a sensitivity of 0.67, p-value < 0.001, and ρw (proportion time in warning) of 0.11. After applying TFLS, the AUC decreased to 0.70, the sensitivity to 0.58, p-value to < 0.01 and ρw to 0.21. However, the percentage of non-seizure days without false positives (non-SZ days w/o FP %) significantly increased from 7.91% to 58.33%. For S2, before applying TFLS, the model had an AUC of 0.71, a sensitivity of 0.11, p-value of 0.50 and a non-SZ days w/o FP % of 34.78%. After applying TFLS, the AUC decreased to 0.58, the sensitivity to 0.00, p-value to 1.00 and ρw to 0.03. However, the non-SZ days w/o FP % increased from 34.78% to 71.74%, indicating an improvement in non-seizure day warning performance.
Clinical Epilepsy