Article Figures & Data
Figures
- Figure 1.
Exemplary LFP recordings in the deep S1 of a GAERS (right) as well as simultaneously recorded LFPs in the deep S1 and VPM of a WAG/Rij rat (upper left panel and lower left panel, respectively). Arrows indicates the onset of the SWD, determined according to the criteria outlined by van Luijtelaar and Coenen (1986), taking the peak of the first spike of twice the background as reference for SWD onset.
- Figure 2.
Wavelet analysis for SWD prediction. Relative sensitivity (A) and average false alarm rate (C) of SWD prediction for different combinations of recording sites in the cortico-thalamic system, obtained by the Maksimenko et al. (2017) algorithm. LFPs, simultaneously recorded in the cortico-thalamic system of WAG-Rij rats, were analyzed in combinations of either two or three recording sites. Results from all 85 combinations are presented in Table 1. To avoid Type II errors, all combinations of recording sites were grouped as either CC (two intracortical recording sites in S1), CT (one cortical recording site in S1 and one thalamic recording site), TT (two intrathalamic recording sites), CCC (three intracortical recording sites in S1), CCT (two cortical recording sites in S1 and one thalamic recording site), CTT (one cortical recording site in S1 and two thalamic recording sites), TTT (three intrathalamic recording sites), or MCCC (three intracortical recording sites in the secondary motor cortex), respectively. B, D, Results of post hoc comparison verified by ANOVA, with *** indicating significance at a p < 0.001 level and * indicating significance at a p < 0.05 level, for sensitivity of prediction (B) and false alarm rate (D), respectively. E, Relationship of false alarm rates and average sensitivity of SWD prediction for different combinations of recording sites in the cortico-thalamic system of WAG/Rij rats, analyzed by the Maksimenko et al. (2017) algorithm. Note highest sensitivity with a low false alarm rate for prediction based on three intracortical recordings in S1 (blue triangle) that outperforms all other combinations of recording sites. Further note the negative correlation between both indicators of SWD prediction performance (r = −0.716; p < 0.001), indicating that higher SWD prediction sensitivity at any given combination of recording sites does not occur at the trade-off of a high false alarm rate.
- Figure 3.
SWD prediction in two genetic rat models of absence epilepsy. A, B, Average sensitivity of SWD prediction (A) and false alarm rate expressed in number of false positives per hour (nFP/h; B) achieved by the Maksimenko et al. (2017) algorithm assessed in 4-h lasting LFP recordings, obtained in Layers IV, V, and VI of S1 in GAERS and WAG/Rij rats. C–E, Comparison of wavelet spectra of true and false positive predictions. An exemplary LFP trace depicting a pre-SWD -> SWD transition is presented in C. Onset of SWD is marked by red vertical line termed 2. The corresponding spectrogram of a true positive detection identified in intracortical LFP recordings in S1 of a GAERS is shown in D. Time period −0.5–0 (red rectangle termed 1) features the analysis window (window size 500 ms) in which the true positive precursor is detected. An exemplary spectrogram of a false positive detection is shown in E. Again, Time period −0.5–0 features the analysis window (window size 500 ms) in which the false positive precursor is detected. F, Statistical comparison of the product of wavelet energy, assessed in the frequency bands W(5–10 Hz), W(3–5 Hz), and W(7–20 Hz) (Maksimenko et al., 2017), between true and false positives in WAG/Rij rats. E, Statistical comparison of the product of wavelet energy, assessed in the frequency bands W(5–10 Hz), W(3–5 Hz), and W(7–20 Hz) (Maksimenko et al., 2017), between true and false positives in GAERS; * indicates a significant difference verified by ANOVA at level of * p < 0.05.
- Figure 4.
Differentiation between true and false positives by a random forest machine learning algorithm. A, Schematic representation of the random forest machine learning algorithm for differentiation between true positive and false positive predictions. After wavelet analysis of either two or three simultaneously recorded LFP traces, the wavelet energies (W(5–10 Hz), W(3–5 Hz), and W(7–20 Hz)) extracted in each trace are fed to a random forest composed of 1000 decision trees. Final classification of the random forest is yielded from a majority voting of the different trees. B, Out-of-sample performance (expressed as balanced accuracy) of random forests. Training in an undersampling approach on wavelet spectra derived from recordings in Layers V and VI of S1 (CC); recordings in Layers IV, V, and VI of S1 (CCC); recordings in Layers IV and VI of S1 and VPM (CCT); recordings in Layer VI of S1, VPM, and RTN (CTT); recordings in VPM, cRTN, and Po (TTT) of WAG/Rij rats at a sensitivity of 60%; and recordings in Layers IV, V, and VI of S1 of GAERS at a sensitivity of 60% (GCCC) or 90% (GCCC90%). Numbers in GAERS groups (1844, 161, 145) refer to the different amount of true/false positive fragments, with which the random forest was trained. Stars in B indicate a significant classification above chance as validated by surrogate statistics with * indicating significance at a p < 0.05 level. C, Table of achieved average balanced accuracies achieved by analysis of the different combinations of recording sites. D, Statistics between group comparison of balanced accuracies performed with ANOVA with * indicating significance at a p < 0.05, **p < 0.01, and ***p < 0.001 level. E, Relation between classification accuracy and the number of incorporated trees in the random forest.
Tables
- Table 1
Combinations of recording sites analyzed by the Maksimenko et al. algorithm and achieved average sensitivities of prediction and false alarm rates
Number of simultaneous
recording sitesCombination
numberArea 1 Area 1 Area 3 Abbreviation in
text and figuresAverage
sensitivityAverage
nFP/h3 1 ctx 4 ctx 5 ctx 6 CCC 61,755 85,962 2 ctx 4 ctx 5 Po CCT 48,392 65,199 3 ctx 4 ctx 5 ATN CCT 45,974 85,363 4 ctx 4 ctx 5 rRTN CCT 44,230 69,947 5 ctx 4 ctx 5 cRTN CCT 50,600 58,431 6 ctx 4 ctx 5 VPM CCT 46,007 61,826 7 ctx 4 ctx 6 Po CCT 50,718 68,470 8 ctx 4 ctx 6 ATN CCT 48,932 82,776 9 ctx 4 ctx 6 rRTN CCT 45,690 71,436 10 ctx 4 ctx 6 cRTN CCT 51,823 60,125 11 ctx 4 ctx 6 VPM CCT 50,269 58,889 12 ctx 5 ctx 6 Po CCT 48,880 79,424 13 ctx 5 ctx 6 ATN CCT 48,345 95,587 14 ctx 5 ctx 6 rRTN CCT 51,354 65,995 15 ctx 5 ctx 6 cRTN CCT 48,963 72,081 16 ctx 5 ctx 6 VPM CCT 48,708 62,217 17 ctx 4 Po ATN CTT 36,121 98,180 18 ctx 4 Po rRTN CTT 35,171 97,276 19 ctx 4 Po cRTN CTT 35,430 95,410 20 ctx 4 Po VPM CTT 34,470 99,526 21 ctx 4 ATN rRTN CTT 38,536 82,294 22 ctx 4 ATN cRTN CTT 34,133 101,225 23 ctx 4 ATN VPM CTT 32,981 99,376 24 ctx 4 rRTN cRTN CTT 35,892 93,150 25 ctx 4 rRTN VPM CTT 33,018 101,039 26 ctx 4 cRTN VPM CTT 37,588 83,424 27 ctx 5 Po ATN CTT 38,046 96,665 28 ctx 5 Po rRTN CTT 36,549 93,522 29 ctx 5 Po cRTN CTT 36,114 97,002 30 ctx 5 Po VPM CTT 34,702 99,814 31 ctx 5 ATN rRTN CTT 40,655 77,191 32 ctx 5 ATN cRTN CTT 36,485 98,925 33 ctx 5 ATN VPM CTT 33,716 98,891 34 ctx 5 rRTN cRTN CTT 37,172 90,429 35 ctx 5 rRTN VPM CTT 33,526 100,798 36 ctx 5 cRTN VPM CTT 38,023 82,687 37 ctx 6 Po ATN CTT 40,751 95,255 38 ctx 6 Po rRTN CTT 38,563 93,038 39 ctx 6 Po cRTN CTT 38,292 95,827 40 ctx 6 Po VPM CTT 36,516 101,606 41 ctx 6 ATN rRTN CTT 43,434 72,403 42 ctx 6 ATN cRTN CTT 37,946 100,546 43 ctx 6 ATN VPM CTT 35,950 98,257 44 ctx 6 rRTN cRTN CTT 40,527 84,918 45 ctx 6 rRTN VPM CTT 35,363 100,356 46 ctx 6 cRTN VPM CTT 38,784 87,363 47 Po ATN rRTN TTT 35,880 103,088 48 Po ATN cRTN TTT 31,342 115,496 49 Po ATN VPM TTT 30,849 115,094 50 Po rRTN cRTN TTT 33,263 109,348 51 Po rRTN VPM TTT 31,252 116,632 52 Po cRTN VPM TTT 30,485 116,111 53 ATN rRTN cRTN TTT 36,646 89,893 54 ATN rRTN VPM TTT 34,497 98,061 55 ATN cRTN VPM TTT 30,137 110,576 56 rRTN cRTN VPM TTT 30,907 115,390 57 Mctx 5a Mctx 5b Mctx 6 MCCC 33,330 129,803 2 1 ctx 4 ctx 5 CC 31,173 211,365 2 ctx 4 ctx 6 CC 34,619 209,386 3 ctx 5 ctx 6 CC 33,612 242,989 4 ctx 4 VPM CT 21,408 123,705 5 ctx 4 ATN CT 20,799 148,887 6 ctx 4 Po CT 21,987 151,854 7 ctx 4 cRTN CT 23,729 122,750 8 ctx 4 rRTN CT 25,276 130,967 9 ctx 5 VPM CT 23,357 120,332 10 ctx 5 ATN CT 22,728 158,520 11 ctx 5 Po CT 24,474 151,471 12 ctx 5 cRTN CT 24,874 130,418 13 ctx 5 rRTN CT 29,267 121,645 14 ctx 6 VPM CT 23,514 146,704 15 ctx 6 ATN CT 24,906 174,084 16 ctx 6 Po CT 25,886 171,314 17 ctx 6 cRTN CT 25,948 145,599 18 ctx 6 rRTN CT 31,349 137,519 19 VPM ATN TT 10,411 157,414 20 VPM Po TT 10,741 186,945 21 VPM cRTN TT 14,999 151,043 22 VPM rRTN TT 15,703 155,311 23 ATN Po TT 12,648 179,928 24 ATN cRTN TT 10,670 165,252 25 ATN rRTN TT 20,339 142,317 26 Po cRTN TT 11,267 171,176 27 Po rRTN TT 17,575 166,227 28 cRTN rRTN TT 21,339 142,157 ctx4, Layer IV of S1; ctx5, Layer V of S1; ctx6, Layer VI of S1; ATN, anterior thalamic nucleus; VPM, vertral-postero-medial thalamic nucleus; Po, posterior thalamic nucleus; rRTN, rostral reticular thalamic nucleus; cRTN, caudal reticular thalamic nucleus; Mctx5a, Layer Va of secondary motor cortex; Mctx5b, Layer Vb of secondary motor cortex; Mctx6, Layer VI of secondary motor cortex.
- Table 2
Out-of-sample performance of the random forest (trained in an undersampling approach on spectra derived from three intracortical recordings in S1 of GAERS at a sensitivity of 90%) confronted to spectra derived from 24-h recordings in a separate group of GAERS (n = 9)
Average confusion matrix Predicted as false positive Predicted as true positive False positive 52.46 ± 9.38% 47.54 ± 9.38% True positive 50.66 ± 8.95% 49.34 ± 8.95% Balanced accuracy F1 score Rat 1 47.37% 14.53% Rat 2 53.68% 6.89% Rat 3 47.44% 11.74% Rat 4 49.07% 4.82% Rat 5 59.62% * 9.14% Rat 6 51.06% 5.44% Rat 7 51.93% 7.18% Rat 8 50.13% 4.25% Rat 9 47.82% 9.68% Depicted in the upper panel is the average confusion matrix (±SEM), specifying the percentage of true positives correctly classified as true positives (lower right), true positives incorrectly classified as false positives (lower left), false positives correctly classified as false positives (upper left), and false positives incorrectly classified as true positives (upper right). Lower panel depicts the balanced accuracies and F1 scores for each individual rat. Note that the F1 score reflects the trade-off between false alarm rate/sensitivity. Low F1 scores are reflecting the drop of sensitivity associated to the drop of false alarm rate. As our goal in this work is the latter, the low scores are justified by the high balanced accuracies; * denotes an above chance balanced accuracy of classification as verified by surrogate statistics.
- Table 3
Out-of-sample performance of the random forest (trained in an oversampling approach on spectra derived from three intracortical recordings in S1 of GAERS at a sensitivity of 90%) confronted to spectra derived from 24-h recordings in a separate group of GAERS (n = 9)
Average confusion matrix Predicted as false positive Predicted as true positive False positive 71.38 ± 2.56% 28.62% ± 2.56% True positive 46.00 ± 4.00% 54.00 ± 4.00% Balanced accuracy F1 score Rat 1 70.28%* 46.88% Rat 2 55.14% 7.59% Rat 3 60.13%* 16.60% Rat 4 63.98%* 12.21% Rat 5 63.15%* 12.02% Rat 6 59.70%* 8.64% Rat 7 68.47%* 13.14% Rat 8 59.00%* 6.51% Rat 9 64.38%* 19.71% Depicted in the upper panel is the average confusion matrix (±SEM), specifying the percentage of true positives correctly classified as true positives (lower right), true positives incorrectly classified as false positives (lower left), false positives correctly classified as false positives (upper left), and false positives incorrectly classified as true positives (upper right). Lower panel depicts the balanced accuracies and F1 scores for each individual rat. Note that the F1 score reflects the trade-off between false alarm rate/sensitivity. Low F1 scores are reflecting the drop of sensitivity associated to the drop of false alarm rate. As our goal in this work is the latter, the low scores are justified by the high balanced accuracies; * denotes an above chance balanced accuracy of classification as verified by surrogate statistics.
Extended Data 1
RF oversampling seizure prediction. Download Extended Data 1, ZIP file.
In this issue
