Fig 1.
(a) High-level view of the system: captured EEG data is first filtered and fed into a neural network. If a neural event is detected with sufficient confidence (>threshold), a decision is made to initiate brain stimulation. The types of delays introduced by each component are denoted in parenthesis: C/V denote Constant/Variable delays, and H/S denote Hardware/Software delays. (b) A detailed implementation scheme, and two possible implementations: (c) an FPGA prototype based on off-the-shelf components, and (d) a Coral-based implementation that uses a custom printed circuit board. Plans are available on our GitHub page.
Fig 2.
Real-time stimulation example of the Portiloop on sleep spindles.
The output of the ANN (likelihood between 0 and 1, magenta) is displayed in the lower part of the Figure. When it crosses an adjustable detection threshold (horizontal grey line, here set to 0.84), the Portiloop sends a stimulus (vertical grey line). The optimal target for this stimulus is the beginning of the sleep spindle (vertical cyan line). Thus, the variable software delay is visible here between the vertical cyan and grey lines. The sections of the signal in red mark are sections wrongly detected as spindles (false positives), and the areas in black those that are not, or not yet, detected as spindles but were identified as spindles by experts (false negatives).
Fig 3.
Portiloop signal processing pipeline for extracting relevant inputs for the ANN.
The selected filters introduce an identical software delay of 40 ms in both branches. (Note that power features are computed offline only for the SpindleNet architecture and are not represented in this diagram).
Table 1.
Quantitative results. Our different models and ablations are compared under “Online Detection” using the nomenclature “mean (std)”, and superscripts for referencing rows in the text.
In rows (4) and (5) we replace an input of our 2-input model by a copy of the other, in row (7) we remove time-dilation, in rows (8) and (9) we train our model only on phase 1 or phase 2 (i.e., young subjects or old subjects), and in row (10) we train a regressor to evaluate it as a classifier.
Fig 4.
Search space of the single-input architecture, found with PMBO.
The hardware cost is the number of trainable parameters in the neural architecture, and the software cost is 1−f1- score of the fully- trained model. Black dots: non-Pareto-optimal models tested by the algorithm. Red dots: Pareto-optimal models found by the algorithm. Red line: Pareto front. The researcher would select a configuration from the Pareto front, which represents optimal trade-offs between both costs.
Fig 5.
Final single-input ANN architecture.
The dimensions of each layer are provided in parenthesis using the PyTorch nomenclature.
Table 2.
Delays measured in the Portiloop (sleep spindle configuration).
Fig 6.
Detection threshold trade-off.
(a) Evolution of the stimulation performance with respect to the chosen detection threshold on the ANN output value. A threshold of 0.84 yields the optimal trade-off; however, researchers may wish to select different parameters according to experimental objectives. (b) Distribution of stimulation delays for a classifier with 0.5 and 0.84 thresholds, respectively. Increasing the threshold yields longer delays. Note that delays are negative when spindles are stimulated in advance of human expert annotation.