Annual Meeting Abstracts: View

  • (Abst. 2.050), 2018
  • Recurrent Neural Networks for Forecasting Epileptiform Electrographic Activity 24 Hours in Advance
  • Authors: Sharanya Arcot Desai, NeuroPace, Inc.; Thomas K. Tcheng, NeuroPace, Inc.; and Martha J. Morrell, Stanford University / NeuroPace, Inc.
  • Content:

    Rationale: The ability to forecast seizures has the potential to improve quality of life for persons living with epilepsy. Advances in deep learning have enabled computers to learn complex patterns directly from data. In this retrospective study, Recurrent Neural Networks (RNNs), specifically Long Short Term Memory (LSTM) networks, were used for forecasting patient-reported clinical seizures and epileptiform activity from data obtained using the NeuroPace® RNS® System. Methods: The RNS System delivers brain responsive stimulation as adjunctive treatment for adults with medically intractable partial onset seizures in no more than two seizure foci. An implanted neurostimulator senses and records electrocorticographic (ECoG) activity from up to 2 leads (8 contacts) and provides stimulation over those contacts when specific ECoG patterns are detected. Five patients in the RNS System clinical trials were included in this analysis because they had =2 years (2.02 to 2.95 years) with no changes to neurostimulator settings. Epileptiform events (EEs) are defined as hourly counts of detections using patient-specific settings programmed by the treating physician, and long epileptiform events (LEEs) are defined as EEs exceeding a pre-specified duration. A patient-specific LSTM network was trained per patient for 250 epochs to predict EEs in the latter 40% of data, with the initial 60% used for training. Each network consisted of 2 stacked LSTM layers with 150 hidden units per layer. The Adam optimizer algorithm was used with an initial learning rate of 0.005 (decay factor 0.2 every 125 epochs). Correlations between the predicted and actual measures evaluated the LSTM networks forecasting performance on EE counts. In Patient #5, two additional LSTM networks were trained. The first forecasted LEEs and the second forecasted patient-reported clinical seizure rates (CSRs). These LSTM networks consisted of 3 stacked LSTM layers with 150 hidden units per layer, 750 training epochs, and an initial learning rate of 0.005. All LSTM networks were trained to predict hourly event counts 24 hours in advance. Results: LSTM networks predicted circadian patterns of epileptiform events with high accuracy in all 5 patients. The Pearson’s correlation coefficients (CCs) computed between predicted and actual EEs were significant (p<0.00001;Figure 1) in all 5 patients (CCs 0.76, 0.48, 0.69, 0.8, 0.58, respectively). In Patient #5, where clear multiday rhythms were apparent in addition to circadian rhythms, the LSTM network accurately forecasted EE circadian and multiday rhythms (Figure 1: last row). In this patient, LSTM networks predicted LEEs and CSRs, but the predicted LEEs lagged actual LEEs by =6 hours, and the predicted CSRs led the actual CSRs by 1-6 days. Only 1/15 instances of LEE clusters and 2/14 instances of CSR clusters were unpredicted. Conclusions: We have demonstrated that recurrent neural networks are able to learn and forecast patterns of epileptiform activity from ECoG recordings. Future analysis will aim to improve prediction accuracy and validate these findings using prospective data. Funding: None
  • Figures:
  • Figure 1
  • Figure 2