Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multimodal deep learning approach to predicting neurological recovery from coma after cardiac arrest

Felix H. Krones, Ben Walker, Guy Parsons, Terry Lyons, Adam Mahdi

TL;DR

This study shows the efficacy and limitations of employing transfer learning in medical classification and reveals that the performance of the model is strongly linked to the selection of a decision threshold and exhibits strong variability across data splits.

Abstract

This work showcases our team's (The BEEGees) contributions to the 2023 George B. Moody PhysioNet Challenge. The aim was to predict neurological recovery from coma following cardiac arrest using clinical data and time-series such as multi-channel EEG and ECG signals. Our modelling approach is multimodal, based on two-dimensional spectrogram representations derived from numerous EEG channels, alongside the integration of clinical data and features extracted directly from EEG recordings. Our submitted model achieved a Challenge score of $0.53$ on the hidden test set for predictions made $72$ hours after return of spontaneous circulation. Our study shows the efficacy and limitations of employing transfer learning in medical classification. With regard to prospective implementation, our analysis reveals that the performance of the model is strongly linked to the selection of a decision threshold and exhibits strong variability across data splits.

Multimodal deep learning approach to predicting neurological recovery from coma after cardiac arrest

TL;DR

This study shows the efficacy and limitations of employing transfer learning in medical classification and reveals that the performance of the model is strongly linked to the selection of a decision threshold and exhibits strong variability across data splits.

Abstract

This work showcases our team's (The BEEGees) contributions to the 2023 George B. Moody PhysioNet Challenge. The aim was to predict neurological recovery from coma following cardiac arrest using clinical data and time-series such as multi-channel EEG and ECG signals. Our modelling approach is multimodal, based on two-dimensional spectrogram representations derived from numerous EEG channels, alongside the integration of clinical data and features extracted directly from EEG recordings. Our submitted model achieved a Challenge score of on the hidden test set for predictions made hours after return of spontaneous circulation. Our study shows the efficacy and limitations of employing transfer learning in medical classification. With regard to prospective implementation, our analysis reveals that the performance of the model is strongly linked to the selection of a decision threshold and exhibits strong variability across data splits.
Paper Structure (10 sections, 1 equation, 5 figures, 2 tables)

This paper contains 10 sections, 1 equation, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Data
  4. Models and architecture
  5. Scoring
  6. Results
  7. Discussion
  8. Main findings and limitations
  9. Previous work
  10. Future research

Figures (5)

  • Figure 1: An example of EEG recordings (bottom) for patient 994 at the 8th hour of channel F8 with good outcome. The top displays the corresponding spectrogram (squared amplitude in decibel units relative to peak power).
  • Figure 2: A schematic diagram of our model architecture. Blue: input data; red: filters and aggregation; yellow: pre-defined features; green: trainable models.
  • Figure 3: Relative feature importance plot for the locally best performing model M6. Here 'agg.' means aggregated over channels and time using the mean prediction and a majority voting for the DenseNet features.
  • Figure 4: Accuracy (blue), FPR (red dashed) and FNR (red dashed-dotted) of the 'Poor' outcome label for different decision thresholds of model M6.
  • Figure 5: ROC curve (blue solid) of our final model (AUC = 0.854) with the indicated (black line) 5% FPR threshold.