A Scalable Framework for Closed-Loop Neuromodulation with Deep Learning

Figures

• 

  Figure 1.. Novel framework for scalable closed-loop neuromodulation with deep learning. (a) Open-loop training data collection consists of multiple trials (Trial) across a cohort (Participants) where single trials record performance changes by stimulation (yellow and pink periods) as compared with no stimulation (gray). Note each stimulation intervention (e.g., Type 1: yellow, Type 2: pink periods) is preceded with a period of no stimulation (gray) where biomarkers (brain state) are collected. Training data can be arranged into biomarkers pre intervention (training Input X) and stimulation response (Labels y). (b) Multiple varying model architectures (spanning model space M) ingest multi-channel and multi-sample input data. Note, each set of model ensembles are trained on different intervention types, which allows for interchanging intervention types as needed. Next, the independent selection of stimulation decision (S) rules allows for the incorporation of extraneous factors such as stimulation cost. (c) In a closed-loop deployment, biomarker measurements (X(t)) from a novel participant are passed to each ensemble (e.g. No Stimulation, Stimulation Type 1, Stimulation Type 2) of the Model…

• 

  Figure 2.. Dataset summary and model architecture. (a) The GX dataset was collected in two phases or experiments. Experiment 1 consisted of a parameter space exploration where 9 different stimulation conditions were tested. This was followed with Experiment 2 which tested two parameters (F30 and M30) from Experiment 1 on a different participant cohort with 20 trials of each parameter. (b) The behavioral task for the GX dataset consisted of a 70-min compensatory tracking task (CTT) where participants used a trackball to keep a moving circle at the center of the screen. The radial distance or deviation from the center of the screen at each frame determined participant’s behavioral performance. (c) EEG (light blue) and stimulation sites for frontal (orange), and motor (purple) stimulation. (d) MRI-derived head model used to visualize stimulation electrode placement. Trial-wise percent improvement derived from change in deviation for all participants in Experiment 2 for (e) F30 and (f) M30. Trial-wise implementation of stimulation trials for (g) Experiment 1 repeated three times for each participant to cover all 9 parameters and (h) Experiment 2 repeated two times for each participant to cover…

• 

  Figure 3.. Data label distribution and model prediction metrics. Distribution of training and testing data used for (a) No Stimulation, (b) F30, and (c) M30 stimulation conditions. Confusion matrices for Model 1 and 2 for (d) No Stimulation, (e) F30, and (f) M30. (g) Scores for Models 1 and 2 in terms of accuracy; and weighted scores for precision, recall, and F1 with SEM. (h) Correct and incorrect classifications for Models 1 and 2 with their respective percent change in deviation. Boxplots are indicated for correct classifications only. (i) Precision-recall curves across test data input type and models with the area under the curve (AUC) computed for the average precision recall across participants with SEM.

• 

  Figure 4.. Simulation of closed-loop responsiveness predictions. (a) A total of 1056 input test data trials (15 channels X 3000 samples) were individually passed to all 6 models. (b) The predictions from all 6 models were aggregated and compared post prediction. (c) Model predictions (with SEM) for 10 exemplary trials where the probability of improvement (probability of trial classed as Better) is averaged over predictions for each respective model (i.e., predictions from Models 1 and 2 for NoStim, F30, and M30). Using simple decision rules, model predictions are compared. (d) Stimulation decisions for 10 example trials in c. (e) To assess simulation outcomes, stimulation decision predictions of test trials where no stimulation, F30, and M30 stimulation were applied were divided into four classes: correct assessment, potential improvement, wrong assessment, and missed opportunities. (f) These four classes were then aggregated across all test data.