Toward a Personalized Real-Time Diagnosis in Neonatal Seizure Detection

Figures

• 

  FIGURE 1.. From generic to personalized neonatal seizure detector. The adaptation of the generic model can be performed on the fly. The optional clinical feedback in the testing stage can be used to purify the hypothesised decisions.

• 

  FIGURE 2.. The ensemble of patient-adaptive GMM and patient-independent SVM SDAs.

• 

  FIGURE 3.. (a) Four different weight functions; here x = PSVM for providing weights for updates to seizure class GMM, x = (1-PSVM) for updating non-seizure class GMM. (b) An example of a sampled sigmoid weighting using 5 clusters.

• 

  FIGURE 4.. Performance of a SDA – measured as the area under the ROC curve.

• 

  FIGURE 5.. Performance of the baseline SVM and adaptive fusion system measured as AUC (top) and AUC90 (bottom) as a function of time.

• 

  FIGURE 6.. Relative improvement in AUC90 between the PA-FUSION and PI-FUSION systems. ‘*’ indicates statistical significance at á set to 1%.

• 

  FIGURE 7.. Real-time functioning of the developed personalised SDA algorithm. A) Probabilistic output of patient-adaptive and patient-independent SDAs with superimposed ground truth. B) 30s of seizure activity detected with higher probability with adaptive fusion system. C) 30s of non-seizure activity detected with lower probability.

• 

  FIGURE 8.. Neonatal seizure detection system diagram.