Respiratory Mandibular Movement Signals Reliably Identify Obstructive Hypopnea Events During Sleep

Abstract

Context: Accurate discrimination between obstructive and central hypopneas requires quantitative assessments of respiratory effort by esophageal pressure (OeP) measurements, which preclude widespread implementation in sleep medicine practice. Mandibular Movement (MM) signals are closely associated with diaphragmatic effort during sleep. Objective: We aimed at reliably detecting obstructive off central hypopneas events using MM statistical characteristics. Methods: A bio-signal learning approach was implemented whereby raw MM fragments corresponding to normal breathing (NPB; n = 501), central (n = 263), and obstructive hypopneas (n = 1861) were collected from 28 consecutive patients (mean age = 54 years, mean AHI = 34.7 n/h) undergoing in-lab polysomnography (PSG) coupled with a MM magnetometer, and OeP recordings. Twenty three input features were extracted from raw data fragments to explore distinctive changes in MM signals. A Random Forest model was built upon those input features to classify the central and obstructive hypopnea events. External validation and interpretive analysis were performed to evaluate the model's performance and the contribution of each feature to the model's output. Results: Obstructive hypopneas were characterized by a longer duration (21.9 vs. 17.8 s, p < 10^-6), more extreme low values (p < 10^-6), a more negative trend reflecting mouth opening amplitude, wider variation, and the asymmetrical distribution of MM amplitude. External validation showed a reliable performance of the MM features-based classification rule (Kappa coefficient = 0.879 and a balanced accuracy of 0.872). The interpretive analysis revealed that event duration, lower percentiles, central tendency, and the trend of MM amplitude were the most important determinants of events. Conclusions: MM signals can be used as surrogate markers of OeP to differentiate obstructive from central hypopneas during sleep.

Keywords: central hypopnea; hypopnea; mandibular movements; obstructive hypopnea; respiratory effort; sleep apnea syndrome.

Figure 1

This schematic diagram summarizes a 6 steps data analysis plan. (1) Acquisition and preprocessing of individual data in 28 patients. OeP, Esophageal pressure signal; MM, Mandibular movement signal; (2) Manual label scoring based on OeP signal: N, Normal breathing (n = 501), O= Obstructive hypopneas (n = 21861) and C, Central hypopneas (n = 263); (3) Feature extraction and data compilation; (4) Exploratory data analysis; (5) Developing a classification rule based on Random Forest algorithm; (6) Model explanation, to determine the role of each contributor and their interactions to identify the 3 target events.

Figure 2

Surrogate features to describe the characteristics of MM signal during central and obstructive hypopnea events. Mandibular movement signal was recorded as a time series (sampling rate = 10 Hz). The figure presents two distinct series: a 12.5 s central hypopnea event (125 points, in red) and an obstructive hypopnea event (45 s or 450 points, in blue). Each point (a) indicates a single value in the series. The slope of a linear model (c) allows to describe the linear trend of MM signal. A generalized additive model (b,d) with polynomial smoothing spline function was fitted for estimating the time dependent variability of MM signal in 2 series, then the intercept (b) and 4 coefficients of spline function (denoted as S1, S2, S3, and S4) were extracted from the model as surrogate features to characterize the complex trajectories of MM in time. The distribution shape parameters (e) (skewness, Kurtosis, Variance) were estimated to describe the shape of MM signal distribution in each series. Other parameters aimed to describe the centrality (h), including mean, mode and median), lower extremities (g, including the minimum, 5th and 25th centiles) and upper extremities (f, including maximum, 95th and 75th centiles) of MM amplitude. Finally, the event duration (i, measured in second) was also included as characteristic feature.

Figure 3

Characteristics of mandibular movement signal during obstructive and central hypopnea events. Each panel visualizes the distribution of a surrogate feature of MM signal in 3 event types: normal breathing (501 series, green color), central hypopneas (n = 263 series, red color) and obstructive hypopnea (n = 1,861 series, blue color). The surrogate features are presented in 3 groups: (A) The extreme levels of signal amplitude, with lower extremities on left side and upper extremities on the right side; (B) The centrality or location parameter of the signal (left) and distribution shape parameters (right); (C) The event duration, linear trend and coefficients of the smoothing spline time series model (S1–S3). The letter-value boxplot was used to ensure a better description for large data (10). Multiple boxes were drawn, each one represents a pair of lower and upper letter values. The procedure starts with the median, followed with quartiles, and so on. The innermost box is equivalent the conventional boxplot. As moving toward the tails, the boxes became incrementally narrower until we reached the extremes values (outliers, minimum, and maximum). The p_values correspond to a pairwise comparison using t-test with Bonferroni correction. A difference is considered significant if p_value is lower than 10⁻⁶.

Figure 4

Average marginal contribution of the 18 most relevant features to the model's prediction. Each line in this graph corresponds to a single feature and consists of 3 violin plots showing the contribution of that feature to the prediction of 3 target groups. Each column corresponds to a target group (normal, central and obstructive hypopneas). The violin plots and X scale indicate the distribution of the SHAP value, a score assigned to each feature to measure the average marginal contribution of that feature across all possible coalitions with other features to make a certain prediction. The shape and location of the violin plots indicate the impact of each feature on the model's output, or how much a feature may contribute to a certain prediction. A negative SHAP value indicates that the feature participates to rule out that group (by decreasing the predicted probability), whilst a feature with positive SHAP value would increase the probability of that group, thus supporting its identification. Larger absolute SHAP scores (the density plot extends further to the left or right side) indicate more important role of that feature. A zero SHAP value (concentrated density plot) indicates that the feature does not contribute at all to the prediction of that group. The feature values were normalized and mapped to a blue/red color scale (blue = lower values, red = higher values), allowing to estimate the tendency of prediction as the feature value increases or decreases.

Figure 5

Cooperation network among MM signal features in Random Forest classifier. The figure presents a network of all possible combinations among the 23 potential surrogate features of MM signals that occurred in 500 classification rules in the Random Forest model. Pi, ith percentile; IQR, interquartile range; Amp, amplitude of MM signal; Int, intercept of the linear model estimating MM signal amplitude in function of time; S1–S4, spline functions determining the curvilinear trend of MM signal amplitude in function of Time; Skew, skewness; Kurto, Kurtosis; Trend, linear trend of the MM signal. Each node in the network indicates a feature and the link connecting between two nodes indicates that these two features did co-exist in at least 250 decision trees. The color intensity of connection links is proportionate with the frequency of that combination. The nodes are positioned in function of their ability to inter-connect with other nodes, thus the more centralized nodes (colored in yellow; more links), such as event duration (length), 5th centile, linear trend, intercept, variance, and Min were the most important features, because those features did participate in almost every classification rule.