Fig. 5: LSTM RNN Prediction of state transitions based on large-scale BOLD signatures.

a Computational pipeline of the LSTM RNN model for state transition prediction. The model input was the BOLD signals (tPCs) preceding state transitions, and the output was the brain state after state transitions. Details were described in Methods part. b Mean (+/− SEM, n = 500 bootstrap sampling, same in all further analysis) prediction accuracy (ACC) of the LSTM RNN model on the validation dataset with different numbers of layers and hidden units. c Prediction accuracy on the validation dataset was primarily related to the gap times preceding state transitions. Results were based on the LSTM RNN model with 1 hidden layer and 50 hidden units. Gray dot and colored shade represented an individual result and the corresponding best quadratic fitting. d Confusion matrix between empirical and predicted seven categories showed high prediction accuracy on the testing dataset based on 10 s input length, 0 s gap times, 1 hidden layer and 50 hidden units (used in all further analysis, except the gap time). e–h Gap time dependent test accuracy and regional sensitivity on brain state predictions: AW to NREM (e), NREM to AW (f), NREM to REM (g) and REM to AW (h). Upper panel, high prediction accuracy of state transition preceding electrophysiology defined transition time. Note the null dataset was constructed by shuffling categories. Gray shadows: 95% confidence intervals on the null control dataset. Middle panel, distributions of discriminate times of brain state prediction compared to null dataset. Lower panel, sensitive regions of LSTM RNN model on brain state predictions (FDR corrected p < 0.05 and Cohen’s d > 0.3). Results in the axial view were shown in Supplementary Fig. 15. Font sizes of sensitive regions were scaled according to the changes of prediction accuracy. Abbreviations of these regions were listed in Supplementary Data 1. Source data are provided as a Source Data file.