Machine learning reveals interhemispheric somatosensory coherence as indicator of anesthetic depth

Abstract

The goal of this study was to identify features in mouse electrocorticogram recordings that indicate the depth of anesthesia as approximated by the administered anesthetic dosage. Anesthetic depth in laboratory animals must be precisely monitored and controlled. However, for the most common lab species (mice) few indicators useful for monitoring anesthetic depth have been established. We used electrocorticogram recordings in mice, coupled with peripheral stimulation, in order to identify features of brain activity modulated by isoflurane anesthesia and explored their usefulness in monitoring anesthetic depth through machine learning techniques. Using a gradient boosting regressor framework we identified interhemispheric somatosensory coherence as the most informative and reliable electrocorticogram feature for determining anesthetic depth, yielding good generalization and performance over many subjects. Knowing that interhemispheric somatosensory coherence indicates the effectively administered isoflurane concentration is an important step for establishing better anesthetic monitoring protocols and closed-loop systems for animal surgeries.

Keywords: cortico-cortical coherence; depth of anesthesia; gradient boosting; mouse; somatosensory cortex (S1).

Figure 1

Electrocorticogram signals are extracted for estimation of anesthetic depth. (A) Schematic overview of the recording setup. Two electrodes over the somatosensory cortices (CH1 and CH2) were measured against a common reference over cerebellum (Ref), while stimulating the right whiskers and varying the isoflurane concentration. The gray shaded bars in the “Vaporizer” inset mark the first 5 min of each given isoflurane concentration segment which were excluded from statistical analysis. (B) (Top) Craniotomy over barrel cortices. The Ag/AgCl electrodes were placed directly on the dura, then covered in phosphate-buffered saline based agar and two component silicone. (Bottom) Partial craniotomy over cerebellum for the reference electrode, drilled to 20% thickness, and covered as above. (C) Example ECoG traces recorded using the OpenBCI during different administered isoflurane concentrations. Red shaded areas denote the [−0.2, 0.5 s] interval around a stimulus. (D) The evoked responses averaged over all trials in each isoflurane concentration block, zero-aligned at t = 0. The shaded area indicates the 2σ-range of the standard error of the mean.

Figure 2

ECoG signal features display modulation to administered isoflurane concentrations. The leftmost column depicts representative examples of the features found to be most modulated by changes in administered isoflurane concentrations. The middle column depicts violin plots of the respective feature averages over animals across isoflurane concentrations. The rightmost column depicts violin plots of the respective standard deviation within the different isoflurane segments, relative to the mean value in the segment. All violin plots display the minimum, maximum, and median values of the distributions. The regions highlighted in red in the left panels depicts the data used in further analysis shown in the right panels. (A) Interhemispheric somatosensory coherence based on the unfiltered raw-data. The traces illustrate median and 95 % confidence interval calculated over all isoflurane blocks (right panels: 5–40 Hz). (B) Bursting Patterns. Shaded regions in representative example depict signal portions classified as burst. (C) Lempel-Ziv Complexity. Representative examples of signal segments with maximum and minimum LZC. (D) Sample Entropy. Representative examples of signal segments with maximum and minimum sample entropy. Significance testing computed with two-sided Mann–Whitney U-test across isoflurane concentrations with three Benjamini-Hochberg False Detection Rate controls (Benjamini and Hochberg, 1995). Significant results are denoted with ^*p < 0.05, ^**p < 0.01. Precise p-values are listed in Supplementary Table 1.

Figure 3

ECoG features serve as input to a gradient boosting regressor to successfully estimate DoA. Feature importances indicate interhemispheric somatosensory coherence as a critical DoA readout. (A) Overview of feature extraction and estimator workflow. Notch and high-pass filtered signal traces were extracted in 10 s blocks, signal features were calculated (FE), and then the three most recent consecutive blocks provided as input to estimate a target variable ŷ via a gradient boosting ensemble. (B) Example of the estimation of isoflurane concentration. GT denotes the measured ground-truth. The filled area under the predicted curve shows the standard error of the past minute. (C) Distribution of the Mean Absolute estimation Error (MAE) over all animals in each isoflurane regime, corresponding to estimation of the isoflurane concentration. (D) Feature importance over all folds. The interhemispheric somatosensory coherences are the most important features, over almost all folds, corresponding to estimation of the isoflurane concentration. (E) Example estimation of the evoked response attenuation (ERA). GT denotes the measured ground-truth. The filled area under the predicted curve shows the standard error of the past minute. (F) MAE over all folds or animals (n = 11), corresponding to estimation of the ERA. (G) Distribution of the feature importances over all animals, corresponding to estimation of the ERA. All the violin plots indicate the minimum, maximum and median values.