Automated algorithms for seizure forecast: a systematic review and meta-analysis

Figures

• 

  Fig. 1. Illustration of the concepts of (a) SPH and SOP, where a correct prediction corresponds to an onset that occurs after SPH and within SOP; and (b) forecast horizon, where the forecasts are equally spaced in time in contrast with the task of seizure prediction

• 

  Fig. 2. PRISMA flow diagram depicting the number of records identified, included and excluded, as well as the reasons for exclusion. Four databases and registers were searched, resulting in 15 included reports, with 3 additional reports being included at a later stage. A total of 18 reports were included in the review

• 

  Fig. 3. Summary of reported input data across the 18 papers included in the review, along with how the data were given to the algorithms, when applicable (i.e., raw vs features vs cyclic profile). ACC accelerometry, BVP blood volume pulse, EDA electrodermal activity, EEG electroencephalography, HR heart rate, TEMP temperature, ToD time of day, Other sleep, activity, weather

• 

  Fig. 4. Summary of types of input data reported across the 43 algorithms, along with the category of algorithmic approach used. SMPS Surrogate leasures of the preictal state, CDE cyclic distribution of events, ML machine learning, DL deep learning, GLM Generalized linear lodel, PM phase modeling

• 

  Fig. 5. Forest plot of AUC of seizure forecast algorithms, overall and stratified by the type of input data given to the algorithms. The dashed line illustrates the estimated overall AUC and the diamonds represent either the overall or subgroup summary. Additional information provided includes the horizon of the forecast, the sample size, total number of seizures, and the train–test approach (which was either (pseudo-)prospective, P, or retrospective, R). AUC area under the ROC curve. (*) Identifies the algorithm proposed by [27], since they only reported the median number of seizures (143, IQR of 13-1233)

• 

  Fig. 6. Forest plot of BSS of seizure forecast algorithms, overall and stratified by the type of input data given to the algorithms. The dashed line illustrates the estimated overall BSS, and the diamonds represent either the overall or subgroup summary. Additional information provided includes the horizon of the forecast, the sample size, total number of seizures, and the train–test approach (which was either (pseudo-)prospective, P, or retrospective, R). BSS Brier Skill Score. (*) Identifies the algorithm proposed by [27], since they only reported the median number of seizures (143, IQR of 13–1233)