Fig. 4: Intrusions and instability predict sleep quality.
From: A continuous approach to explain insomnia and subjective-objective sleep discrepancy
A Root mean squared error (RMSE) for a ML regression as a function of the number of features included as predictors for each of the 4 sleep parameters, using only SOSD+ and SOSD−. Lines and shaded areas are the average and standard deviation, respectively, out of 120 iterations of a repeated 5-fold cross-validated ML regression algorithm. B Average predicted vs true sleep parameter values corresponding to using all the 16 features. Subject averages were obtained from the predictions in the test samples. Colors denote sleep groups and the diagonal dashed red line is the identity line. All metrics significantly correlated with their prediction (p < 0.001), and the correlation followed WASO > SE > TST > SOL. Predictions were better for SOSD− than for SOSD+ . C Spearman correlation matrix between the features and the sleep quality parameters. Only colored squares were significant (Bonferroni corrected p-value < 0.001). Features are sorted in descending order following the average R² (first row of the matrix). Features are encoded as ‘ave’ for average, ‘H’ and ‘dKL’ for entropy and DKL, respectively, and the last word corresponds to the sleep stage. For example, ave_H_W is the average entropy of wakefulness. Note that average intrusions and instability in wakefulness show the largest average R² and the largest correlations with WASO and SE.
