Increased nocturnal arterial pulsation frequencies of obstructive sleep apnoea patients is associated with an increased number of lapses in a psychomotor vigilance task.

OBJECTIVES: Besides hypoxaemia severity, heart rate variability has been linked to cognitive decline in obstructive sleep apnoea (OSA) patients. Thus, our aim was to examine whether the frequency domain features of a nocturnal photoplethysmogram (PPG) can be linked to poor performance in the psychomotor vigilance task (PVT).
METHODS: PPG signals from 567 suspected OSA patients, extracted from Type 1 diagnostic polysomnography, and corresponding results of PVT were retrospectively examined. The frequency content of complete PPGs was determined, and analyses were conducted separately for men (n=327) and women (n=240). Patients were grouped into PVT performance quartiles based on the number of lapses (reaction times ≥500 ms) and within-test variation in reaction times. The best-performing (Q1) and worst-performing (Q4) quartiles were compared due the lack of clinical thresholds in PVT.
RESULTS: We found that the increase in arterial pulsation frequency (APF) in both men and women was associated with a higher number of lapses. Higher APF was also associated with higher within-test variation in men, but not in women. Median APF (β=0.27, p=0.01), time spent under 90% saturation (β=0.05, p<0.01), female sex (β=1.29, p<0.01), older age (β=0.03, p<0.01) and subjective sleepiness (β=0.07, p<0.01) were significant predictors of belonging to Q4 based on lapses. Only female sex (β=0.75, p<0.01) and depression (β=0.91, p<0.02) were significant predictors of belonging to Q4 based on the within-test variation.
CONCLUSIONS: In conclusion, increased APF in PPG provides a possible polysomnography indicator for deteriorated vigilance especially in male OSA patients. This finding highlights the connection between cardiorespiratory regulation, vigilance and OSA. However, our results indicate substantial sex-dependent differences that warrant further prospective studies.

Introduction

Obstructive sleep apnoea (OSA), characterised by repeated breathing cessations, intermittent hypoxaemia and arousals from sleep, is one of the most prevalent sleep disorders [1]. It has a strong association with various comorbid conditions, such as cerebrovascular, cardiovascular, metabolic and neurological diseases [2, 3]. It has also been linked to deteriorated outcomes in different domains of cognitive functioning [4], for example in the psychomotor vigilance task (PVT) [5]. PVT is a measure of vigilance and the ability to sustain attention. Due to the low learning effect, low-labour test protocol and minimal subjectivity, PVT is recognised as a reliable test to evaluate neurocognitive performance [6].

Previous studies show that severe intermittent hypoxaemia, quantified by the duration, depth and frequency of oxygen desaturations, is associated with daytime sleepiness and cognitive decline in OSA patients [7-10]. However, desaturation metrics and other parameters quantifying the severity of OSA are not fully capable of explaining the impaired performance in PVT in OSA patients having shallow and short desaturations [10]. This suggests that additional factors could explain the associations between OSA and poor PVT performance. Therefore, novel biomarkers could be extremely useful for patients with relatively stable nocturnal oxygenation despite repetitive breathing cessations.

Nocturnal intermittent hypoxaemia can be quantified with a pulse oximeter. It provides noninvasively information on blood oxygen saturation (SpO) based on the absorption of red and infrared light in the blood volume [11]. In sleep medicine, the photoplethysmogram (PPG) derived from the original absorption signal is less used than SpO [12]. However, PPG has great potential in the long-term monitoring of cardiorespiratory function, as it is an unobtrusive measurement technique that contains information on the amplitude and frequency of the arterial pulsations alongside derived SpO [11, 12]. Therefore, frequency domain investigations of the PPG could be beneficial as it combines the information of amplitude and frequency changes without peak detection from the PPG, which has high uncertainties compared to ECG-based R-peak detection [13].

OSA induces abnormal heart rate variability during sleep due to apnoeas and hypopnoeas causing desaturations and arousals from sleep [14]. Moreover, OSA disrupts normal sleep cycles and reduces the time spent in deeper stages of sleep [3]. Deeper stages of sleep are associated with higher parasympathetic drive and lower heart rate; conversely OSA is associated with high nocturnal sympathetic drive and higher heart rates [3, 15]. Furthermore, non-dipping nocturnal heart rate and increased resting heart rate in 24-hour Holter ECG have been linked to worsened cognitive performance [16]. As PPG is shown to be a reliable surrogate for ECG [17], we hypothesise that the information content of the PPG signal can be linked to poor performance in PVT.

Thus, the main aim of this study is to examine whether the frequency domain features of nocturnal PPG can be linked to deteriorated vigilance in OSA patients. In addition, we aim to examine whether larger within-test variation in reaction times is associated with spectral features of PPG. As earlier literature shows that PVT performance is highly sex-dependent [10, 18], we perform the analyses separately for men and women.

Methods

The initial clinical cohort comprised 912 consecutive suspected OSA patients undergoing polysomnography (PSG) in the Sleep Disorders Center of Princess Alexandra Hospital (Brisbane, Australia). The Institutional Human Research Ethics Committee of the Princess Alexandra Hospital approved the retrospective data collection (HREC/16/QPAH/021 and LNR/2019/QMS/54313). The cohort comprised all patients from the normal clinical inflow of suspected OSA patients between 2015 and 2017. From the initial population, 567 patients were selected into the studied population based on the following inclusion criteria: 1) complete demographic and comorbidity information; 2) successful PSG and completed PVT; 3) apnoea–hypopnoea index (AHI) ≥5; and 4) total sleep time in PSG ≥4 h. PSG recordings were conducted and scored with the Compumedics Grael acquisition system and Compumedics ProFusion PSG 4 software (Compumedics, Abbotsford, Australia). PSG recordings were manually scored by experienced sleep technicians; all scorings were conducted in conformity with the American Academy of Sleep Medicine (AASM) 2012 guidelines [19]. None of the included patients had central sleep apnoea, i.e. the proportion of central apnoeas did not exceed 50% of the total number of apnoeas.

The PPG signal analysis protocol is illustrated in figure 1. First, complete PPG signals, measured with a transmissive Nonin Xpod 3011 finger pulse oximeter, were exported from each patient. Next, PPG signals were decimated from the original 256 Hz sampling frequency to 64 Hz with an anti-aliasing Chebyshev filter to reduce computational load. Signals were further filtered using an 8th order Chebyshev Type 2 lowpass-filter with a 60 dB stop-band ripple threshold and 6 Hz cut-off to eliminate unphysiological noise from decimated PPG. Signals were divided into 512 segments to enable comparison between spectrograms and to retain the temporal changes in frequency content throughout the night (figure 1). For all patients, the median segment length was 52.2 s, varying from a minimum of 38.7 s to a maximum of 68.2 s. For all segments, the power spectrum was computed using Welch's periodogram with eight partitions and 50% overlap utilising the Hamming window to reduce the side-lobe leaking effect. Power spectrums were analysed in a physiologically relevant arterial pulsation band between 0.5 and 4 Hz corresponding to heart rates between 30 and 240 bpm. Power spectrums were further converted into spectrogram images, with the x-axis representing the corresponding segment of the night and y-axis corresponding to PPG signal frequency content (figure 1). The power of each frequency is displayed via a colour map in the spectrogram image.

FIGURE 1

Illustrative description of the decomposition of photoplethysmogram (PPG) signal. The complete preprocessed PPG signal is first divided into 512 segments. All segments are equal in length except the last segment, the length of which is the number of remaining sampling points if the signal is not divisible by 512. The frequency content within the nth segment is determined using Welch's method: the segment is divided into eight parts with a 50% overlap. For every part, a power spectral density estimate is computed and then averaged. After this procedure is completed for every 512 segments, they are ordered into a spectrogram image for temporal interpretation of the frequency content within the signal.

The PVTs were conducted as part of standard clinical protocol utilised in Princess Alexandra Hospital. PVTs were timed between 19:00 and 21:00 h on the evening prior to the PSG study. The Psychology Experiment Building Language (PEBL) PVT programme was used on an ASUS Transformer Pad with an attached keyboard [20]. PVT was performed using a 10-minute protocol. The protocol comprises 121 single-trial visual stimuli occurring randomly at 2 to 10 s intervals. Patients were instructed to use the index finger or thumb of their dominant hand and press the response button as soon as they saw the stimulus on the screen. Full trial series were exported from every patient and data on the PVT performance are presented in table 1. From the series, the number of lapses (reaction times ≥500 ms) was computed. In-test variation in repeated reaction times was quantified using Sample Entropy [21]. Before Sample Entropy analyses, the PVT trial series was normalised to achieve zero-mean with a standard deviation of 1 to reduce the baseline effect. The Sample Entropy estimates were computed using the following parameter settings: a template vector with a length of 5 and a distance-tolerance threshold of 0.4.

TABLE 1

Psychomotor vigilance task performance data in all patients (n=567), in men (n=327) and in women (n=240)

	All patients	Male OSA patients	Female OSA patients
Median RT ms	378.0 (341.0–440.8)	362.4 (332.6–410.8)	405.1 (363.0–477.0)
RRT ms−1	2.6 (2.2–2.9)	2.7 (2.4–3.0)	2.4 (2.1–2.7)
Slowest 10% RT ms	676.0 (544.3–1025.0)	622.0 (518.3–910.5)	753.5 (578.5–1221.0)
Fastest 10% RT ms	297.0 (277.4–334.2)	289.0 (272.5–317.2)	313.0 (285.5–352.0)
Lapses	13 (5–35)	10 (4–21)	19 (9–51)
Sample Entropy	0.31 (0.12–0.71)	0.22 (0.11–0.66)	0.44 (0.16–0.80)

All values are presented as the median (interquartile range). All performance metrics differed significantly between men and women based on Wilcoxon rank-sum test. OSA: obstructive sleep apnoea; RT: reaction time; RRT: mean reciprocal reaction time.

The studied patient population was first divided into men and women. After separation, men and women were grouped to quartiles based on PVT performance as PVT does not have standard clinical thresholds for outcomes. Based on preliminary analyses, differences between quartiles 1, 2 and 3 were relatively small in men. In women, the difference between quartiles 1 and 2 and between quartiles 3 and 4 was small. For clarity and to achieve more coherent and informative results, we decided only to present the differences in PPG frequency content in patients with normal and notably impaired vigilance. Thus, only the best (Q1) and the worst (Q4) quartiles based on the number of lapses and Sample Entropy are compared in detail. All reported comorbidities are based on medical records and interviews in a sleep clinic. Group-median spectrograms were defined for Q1 and Q4. Stationary group-median power spectrums were extracted from the spectrogram. In addition, peak-power frequencies indicating the dominant arterial pulsation frequency (APF) were computed from individual spectrograms and from the median spectrograms of Q1 and Q4. For plotting purposes, the APF curves were smoothed via a moving average filter having a 25-point window. The cumulative distribution function (CDF) of APF in the peak-frequency curve was computed from median spectrograms together with 95% confidence intervals via Kaplan–Meier estimates. Statistical difference of cumulative distribution functions between Q1 and Q4 was computed using a two-sided Kolmogorov–Smirnov test. To evaluate the possible confounding effect of reported comorbidities and to analyse multivariate models' sensitivity and specificity, stepwise logistic regression with Bayesian Information Criterion (BIC) was used within the whole population to assess the probability of belonging to Q4. The model was adjusted for sex, age, body mass index, chronic obstructive pulmonary disease (COPD), hypertension, depression, smoking status and subjective sleepiness assessed with the Epworth Sleepiness Scale (ESS). Sleep stage distributions and parameters describing OSA severity (table 2) were investigated by inputting them to regression models separately. After estimating the most significant predictor variables, their performance was evaluated by computing receiver operating characteristic (ROC) curves and corresponding area under the curve (AUC). All data analyses and statistical testing were performed using MatLab (ver. 2018b, MathWorks Inc., Natick, MA, USA) with custom-made functions and functions in Statistics and Machine Learning Toolbox and Signal Processing Toolbox.

TABLE 2

Demographic, polysomnographic (PSG) and comorbidity data and statistical analyses between the best-performing (Q1) and worst-performing (Q4) quartiles based on lapses and sample entropy of reaction times in a psychomotor vigilance task

	Male OSA patients#				Female OSA patients¶
	Lapses		Sample Entropy		Lapses		Sample Entropy
	Q1 (<4)	Q4 (>21)	Q1 (<0.105)	Q4 (>0.656)	Q1 (<9)	Q4 (>51)	Q1 (<0.163)	Q4 (>0.800)
Demographic parameters
Patients n	88	80	82	82	66	59	60	60
Age years	51.1 (40.7–61.7)	56.9 (48.4–67.1)	52.8 (40.5–65.3)	55.4 (44.9–62.2)	54.3 (44.7–60.3)	58.8 (50.3–64.8)	54.1 (45.1–62.3)	55.9 (45.5–62.2)
BMI kg·m−2	32.9 (28.7–37.8)	34.7 (29.8–39.5)	34.5 (30.9–40.4)	33.7 (29.5–37.8)	37.0 (32.3–44.4)	36.0 (31.8–47.1)	38.7 (32.5–45.8)	36.0 (29.0–42.7)
Comorbidities
COPD	6 (6.8)	8 (10.0)	5 (6.1)	5 (6.1)	5 (7.6)	8 (13.6)	5 (8.3)	6 (10.0)
Hypertension	26 (29.6)	35 (43.8)	32 (39.0)	30 (36.6)	22 (33.3)	29 (49.2)	22 (36.7)	22 (36.7)
Depression	8 (9.1)	13 (16.3)	6 (7.3)	13 (15.9)	14 (21.2)	15 (25.4)	10 (16.7)	14 (23.3)
Smokers	13 (14.8)	13 (16.3)	14 (17.1)	12 (14.6)	6 (9.1)	10 (17.0)	8 (13.3)	9 (15.0)
ESS score	9 (6–15)	10 (6–14)	8 (5–12)	10 (5–13)	9 (6–12)	13 (6–18)	11 (6–15)	9 (5–13)
PSG parameters
Total sleep time h	5.4 (4.8–6.1)	4.9 (4.5–5.8)	5.3 (4.7–6.1)	5.2 (4.6–5.9)	5.4 (6.0–4.7)	5.6 (4.9–6.4)	5.6 (5.0–6.2)	5.7 (4.9–6.3)
N1 %	15.5 (10.6–23.8)	14.1 (8.7–21.5)	15.4 (9.3–22.1)	14.7 (9.6–25.7)	7.9 (5.7–10.8)	8.1 (5.8–12.8)	8.0 (5.6–13.3)	8.9 (6.0–14.4)
N2 %	47.3 (42.3–52.9)	49.5 (42.6–57.9)	48.2 (42.2–56.1)	45.8 (39.6–54.1)	47.1 (41.4–53.4)	49.0 (40.3–61.2)	47.3 (41.5–53.4)	49.2 (40.4–53.0)
N3 %	16.2 (7.4–23.1)	14.9 (5.0–24.2)	14.6 (7.7–23.8)	15.1 (9.4–23.4)	24.2 (17.3–29.9)	21.6 (12.0–29.3)	23.0 (17.4–29.3)	20.6 (15.2–29.4)
REM %	18.0 (13.7–22.4)	17.0 (12.9–20.6)	18.8 (14.7–22.4)	15.8 (11.6–22.8)	18.5 (12.9–23.3)	19.1 (12.9–22.5)	19.0 (13.0–21.8)	19.2 (12.1–24.8)
AI events·h−1	33.5 (22.4–46.8)	31.6 (22.1–45.6)	31.5 (19.5–42.9)	35.1 (23.1–47.2)	21.3 (30.4–16.1)	24.3 (15.5–33.8)	23.1 (16.0–32.0)	22.7 (18.3–32.4)
AHI events·h−1	29.8 (13.4–49.2)	28.6 (19.9–44.3)	25.0 (15.3–42.3)	28.8 (18.9–53.8)	16.0 (9.3–29.7)	15.8 (10.1–37.7)	17.6 (10.1–27.2)	15.0 (8.9–31.0)
ODI events·h−1	20.4 (6.7–32.9)	21.0 (9.9–37.4)	16.8 (7.5–30.1)	20.0 (8.3–36.6)	11.5 (5.2–26.8)	13.3 (5.6–25.6)	12.5 (5.1–24.6)	9.3 (4.0–21.6)
t90% min	6.8 (1.3–46.8)	15.6 (1.7–98.2)	12.4 (1.4–44.7)	8.9 (0.8–50.4)	4.6 (0.8–32.8)	11.5 (1.0–63.5)	4.3 (0.5–42.8)	3.4 (0.3–45.6)
mDD %	4.9 (4.1–7.4)	5.3 (4.4–7.5)	5.6 (4.1–8.0)	5.1 (4.1–6.6)	5.0 (4.1–6.4)	5.1 (4.3–6.0)	4.7 (4.0–6.2)	4.9 (4.1–6.0)

Data are presented as median (interquartile range) or n (%), unless otherwise stated. OSA: obstructive sleep apnoea; BMI: body mass index; COPD: chronic obstructive pulmonary disease; ESS: Epworth Sleepiness Scale; N1–N3: non-rapid eye movement sleep stages 1–3; REM: rapid eye movement sleep; AI: arousal index; AHI: apnoea–hypopnoea index; ODI: oxygen desaturation index; t90%: time spent under 90% oxygenation; mDD: median desaturation depth. #: n=327; ¶: n=240. Bolded values indicate statistical difference (p<0.05) between Q1 and Q4 in the corresponding sex. Wilcoxon rank sum and Chi-squared tests were used where appropriate.

Results

Differences in demographics, comorbidities and PSG parameters

The PVT performance data are presented in table 1. Demographic, comorbidity and PSG data of Q1 and Q4 with respect to both PVT outcome variables are presented in table 2. The comparison between men and women shows that women have a significantly higher number of lapses and consistently longer reaction times than men (table 1). Conversely, women suffer from substantially less severe OSA based on AHI, oxygen desaturation index and t90% (table 2). In addition, a substantially lower degree of sleep disruption is seen women, having a higher amount of N3 sleep, a lower amount of N1 sleep and a lower arousal index (table 2).

Subgroup analysis between men in Q1 and Q4 based on lapses revealed that patients belonging to Q4 were significantly older (p=0.02) and had shorter total sleep time in PSG (p=0.02) (table 2). Q4 tended to comprise a higher number of hypertensive (43.8% versus 29.6%, p=0.05) and depressed (16.3% versus 9.1%, p=0.16) patients compared to Q1. In addition, patients in lapses Q4 had a trend towards more severe nocturnal hypoxaemia seen as increased t90% (15.6 versus 6.8 min, p=0.12). In Sample Entropy comparison, Q1 and Q4 were highly similar based on demographics, PSG parameters and comorbidities (table 2).

Female patients in lapses Q4 were subjectively sleepier (ESS: 9 versus 13, p=0.01) compared to Q1 (table 2). Q4 tended to be older (p=0.06), slept longer (p=0.06), and were more likely to have hypertension (p=0.07) compared to Q1. No difference between Sample Entropy Q1 and Q4 were observed (table 2).

Frequency domain analysis of the PPG signal

A comparison of male OSA patients in Q1 and Q4 based on lapse count showed significantly higher APF in Q4 (figure 2a and 2b). The median power spectrum exhibited a clear pulse peak shift towards higher APF and reduced power in Q4 (figure 2c). Furthermore, the CDF of temporal APFs contained significantly (p<0.001) higher values in male patients belonging to Q4 (figure 2d). Median spectrograms for Sample Entropy Q1 and Q4 indicated higher APF and larger frequency variation in males belonging to Q4 with significantly higher temporal APF values (p<0.001) (figure 3a, 3b and 3d). The median power spectrum exhibited only a moderate pulse peak shift towards higher APF but higher power in Q4 (figure 3c).

FIGURE 2

Comparison of male obstructive sleep apnoea patients in a) the best-performing (Q1, n=88) and b) the worst-performing quartiles (Q4, n=80) based on lapse count in psychomotor vigilance task. Median spectrograms (colour map represents power of each pulsation frequency) of males in Q4 exhibit higher arterial pulsation frequency (APF) and variance compared to males in Q1. c) The median power spectrum reveals a clear pulse peak shift towards higher APF and reduced power in Q4. d) The cumulative distribution function (CDF) of APF with 95% confidence intervals contains significantly (p<0.001) higher values in males belonging to the worst-performing quartile. The peak-frequency curve indicates the specific frequency in each segment having the highest power.

FIGURE 3

Comparison of male obstructive sleep apnoea patients in a) the best-performing (Q1, n=82) and b) the worst-performing quartiles (Q4, n=82) based on Sample Entropy (SE) in the psychomotor vigilance task. Median spectrograms (colour map represents power of each pulsation frequency) indicate higher arterial pulsation frequency (APF) and larger variation in males belonging to Q4, together with significantly higher APFs (p<0.001) in cumulative distribution function (CDF). c) However, the median power spectrum exhibits only a moderate pulse peak shift towards higher APF and higher power in Q4. The peak-frequency curve indicates the specific frequency in each segment having the highest power. d) CDF of nocturnal APF.

Comparison of female OSA patients in Q1 and Q4 based on lapse count showed that median spectrograms of females in Q4 exhibit only slightly higher APF compared to female patients in Q1 (figure 4a and b). The median power spectrum showed a moderate pulse peak shift towards higher APF and reduced power in Q4 (figure 4c). The cumulative distribution of temporal APFs contained significantly higher values in Q4 than in Q1 (figure 4d). In the comparison between Sample Entropy Q1 and Q4, median spectrograms and median power spectrums were highly similar (figure 5a–c). However, females in Q1 had a trend towards slightly higher (p=0.09) temporal APF values in cumulative distribution (figure 5d).

FIGURE 4

Comparison of female obstructive sleep apnoea patients in a) the best-performing (Q1, n=66) and b) the worst-performing quartiles (Q4, n=59) based on lapse count in the psychomotor vigilance task. Median spectrograms (colour map represents power of each pulsation frequency) of females in Q4 exhibit slightly higher arterial pulsation frequency (APF) compared to females in Q1. c) The median power spectrum shows a small shift towards higher APF in pulse peak location between Q1 and Q4, and reduced power in Q4. d) The cumulative distribution function of temporal APF with 95% confidence intervals contains slightly (p<0.05) higher values in the worst-performing quartile. The peak-frequency curve indicates the specific frequency in each segment having the highest power.

FIGURE 5

Comparison of female obstructive sleep apnoea patients in a) the best-performing (Q1, n=60) and b) the worst-performing quartiles (Q4, n=60) based on Sample Entropy (SE) in the psychomotor vigilance task. Median spectrograms (colour map represents power of each pulsation frequency) indicate similar arterial pulsation frequency (APF) and overall frequency content in both quartiles. c) Furthermore, median power spectrums exhibit similar pulse peak locations and power. d) In the cumulative distribution function of APFs, the difference was not statistically significant (p=0.09). The peak-frequency curve indicates the specific frequency in each segment having the highest power.

Regression analysis

Based on stepwise regression analyses, median APF was found to be a significant predictor of belonging to lapses Q4 (β=0.270, SE=0.109, p=0.013) alongside female sex (β=1.286, SE=0.267, p<0.001), t90% (β=0.050, SE=0.02, p<0.001), age (β=0.031, SE=0.010, p=0.002) and subjective sleepiness (β=0.073, SE=0.024, p=0.002) (table 3). In contrast, only female sex (β=0.748, SE=0.249, p=0.002) and depression (β=0.906, SE=0.362, p=0.012) were significant predictors in belonging to Q4 based on Sample Entropy (table 3). Sleep stage distribution, AHI or arousal index were not associated with increased probability of belonging to Q4 of lapses or Sample Entropy (table 3).

TABLE 3

Stepwise regression models for estimated β-coefficients and corresponding odds for belonging to Q4 instead of Q1

	Lapses				Sample Entropy
	OR	β	Standard error	p-value	OR	β	Standard error	p-value
Female sex	3.62	1.29	0.27	<0.01	2.11	0.75	0.25	<0.01
Age years	1.03	0.03	0.01	<0.01
BMI kg·m−2
Hypertension
Smoking
COPD
Depression					2.47	0.91	0.36	<0.02
ESS score	1.08	0.07	0.02	<0.02
AHI events·h−1
t90% per 10 min	1.05	0.05	0.02	<0.01
AI events·h−1
N1 %
N2 %
N3 %
REM %
Median APF min−1	1.31	0.27	0.11	<0.02

This analysis comprises the whole population (n=567), where quartiles are defined for lapses (Q1: 0–4 lapses; Q4: over 35 lapses) and Sample Entropy (Q1: 0–0.12; Q4: over 0.71) separately. Both models are computed using the logit-link function. Models were constructed using constant starting model, Bayesian Information Criterion (BIC) as a model entering criterion and treating the response variable as a categorical variable. All models were constructed by separately inputting investigated polysomnography-based parameters. Parameters that were excluded from the final regression model based on BIC are left blank. BMI: body mass index; COPD: chronic obstructive pulmonary disease; ESS: Epworth Sleepiness Scale; AHI: apnoea–hypopnoea index; t90%: time spent under 90% saturation; AI: arousal index; N1–N3: non-rapid eye movement sleep stages 1–3; REM: rapid eye movement sleep; APF: arterial pulsation frequency.

Based on models constructed by logistic stepwise regression, ROC curves and AUC values were determined to evaluate models' sensitivity and specificity in classifying patients to Q1 and Q4 based on lapses (figure 6). Univariate models show that median APF and ESS had the best differentiation capability, however with AUCs of only 0.5960 and 0.6067, respectively. By utilising a multivariate approach including sex, age, ESS score together with median APF or t90%, AUC reached 0.7314 and 0.7349, respectively (figure 6). For comparison, AHI was able to differentiate Q1 and Q4 patients with a sensitivity of 65.5% and specificity of 32.9% with 15 events·h−1 cut-off and was not a significant predictor of belonging to lapses Q4. In addition, ESS was able to differentiate Q1 and Q4 patients with a sensitivity of 25.6% and specificity of 81.7% when using a 16-point cut-off.

FIGURE 6

Receiver operating characteristic curves and corresponding area under the curve (AUC) for the assessment of belonging to the worst-performing quartile based on lapses. Only the significant (p<0.05) predictor variables, determined in stepwise logistic regression, are used in multivariate and univariate estimations of models' sensitivity and specificity. ESS: Epworth Sleepiness Scale; mAPF: median arterial pulsation frequency; t90%: sleep time with oxygen saturation <90%.

Discussion

In this study, we compared PPG-derived APF and frequency domain features between OSA patients within the best and worst PVT performance quartiles. We found increased APF and higher nocturnal variance of the frequencies in the worst-performing male OSA patients compared to the best-performing patients. However, differences in PPG features diminished in female patients and were not as clearly distinguishable. Stepwise regression analysis revealed that higher APF and t90% are associated with a higher number of lapses in PVT. These results imply that increased APF together with more severe nocturnal hypoxaemia may provide a PSG marker for impaired vigilance in male OSA patients. In addition, findings are in line with previous studies, indicating that female sex and older age are independent risk factors for poor PVT performance.

The frequency content of nocturnal PPG exhibits a clear difference between males in the worst-performing (Q4) and the best-performing (Q1) quartiles based on both lapses and Sample Entropy. APF was found to vary more in both Q4s and have significantly higher values compared to Q1s. Furthermore, in lapses Q4s the increased APF was also observed in the group-median power spectrum together with reduced power of the peak frequencies (figures 2c and 4c). Higher APF is due to increased heart rate; OSA and intermittent hypoxaemia (tables 2 and 3) elevate the nocturnal sympathetic activity, heart rate and blood pressure [22-24]. As seen in table 2, males in lapses Q4 had a minor trend for more severe hypoxaemia together with shorter total sleep time. However, increase in t90% significantly elevated the odds of belonging to Q4 based on lapses (table 3). Conversely, this was not found in Sample Entropy Q4 despite the higher APF. Moreover, the association between APF and Sample Entropy diminished in stepwise regression (table 3), and power reduction of peak frequencies in Q4 was not observed (figures 3c and 5c). This implies that the spectrograms, which retain the temporal changes in the PPG signal, associate better to the deteriorated vigilance than median APF. It could also be speculated that there is a substantial inter-individual variation in cardiac response to hypoxaemia and breathing cessations. Older male patients especially seem to be vulnerable to nocturnal sympathetic overdrive (table 1), making APF a useful tool for assessment of the risk for deteriorated vigilance in male OSA patients.

Analysis comprising only female patients showed smaller differences between best and worst performers in the frequency content of PPG. Based on previous studies, women tend to perform worse than males in PVT [10, 18, 25], which was observed in this cohort also (table 1). Women also exhibit a higher resting heart rate and different cardiac regulation sensitivity [26]. Furthermore, in both Q1 and Q4 women had substantially milder OSA, longer total sleep time, a larger amount of N3 sleep and less N1 sleep than men (table 2), which can partly explain the obtained results. Moreover, females had a higher Q4 threshold than men both in lapses and Sample Entropy as well as significantly higher ESS scores in lapses Q4 (tables 1 and 2). These results indicate a clear difference between men and women in vigilance deterioration and baseline sleepiness in best- and worst-performing quartiles. This was however expected as we have previously shown that OSA and OSA-related sleepiness also are sex-dependent [9, 27]. Based on the current body of evidence regarding the association between OSA and impaired vigilance [7, 8, 10], it could be speculated that more severe intermittent hypoxaemia, self-reported sleepiness and fatigue, comorbid depression and older age are better indicators of poor vigilance for female OSA patients than changes in PPG frequency domain features. The present stepwise regression analyses strongly support this conclusion (table 3), as female sex and comorbid depression were significant predictors of large within-test variation.

Elevated intermittent sympathetic activity leads to higher blood pressure [22] and further, to higher nocturnal heart rate [24] which associates to deteriorated vigilance especially in men via increased APF (figures 2 and 3). One of the reasons behind this association could be sympathetic overdrive and its effect on haemodynamics and cerebrospinal fluid oscillations in slow-wave (SW) sleep. Recent literature shows that SW oscillations in N2 and N3 sleep drive the oscillations in cerebral blood volume (CBV) and cerebrospinal fluid (CSF), enabling protein clearance from the brain via the glymphatic system [28, 29]. Under high blood pressure, the protein clearance process via oscillation of CBV and CSF can be disrupted and even reversed [28]. In addition, glymphatic propulsion associates with cardiac pulsations and breathing pulsations [30]. Therefore, untreated OSA as a chronic condition can contribute to the accumulation of proteins in the brain over time [31-33], and thus, together with hypoxaemia-induced neuronal brain damage, lead to cognitive decline [10]. In this study, only a minor difference between Q1 and Q4 in the amount of N3 sleep was found (table 2). This implies that higher APF, which is linked to elevated sympathetic tone and blood pressure, could be a marker for reduced protein clearance during slow-wave sleep (SWS) irrespective of SWS quantity. Moreover, higher heart rate [16], elevated blood pressure [34], elevated sympathetic tone [22] and moderate-to-severe OSA [2, 35] are associated with comorbid cerebral small vessel disease, which could also explain the poor vigilance in male patients having higher APF (figures 2 and 3). To test this hypothesis, a large prospective study combining current protocol with PPG analysis within different sleep stages, MR imaging and CSF protein analysis is warranted.

It is noteworthy that the presented method does not include the exact detection of pulse peaks in PPG for pulse detection. This is, however, advantageous as PPG waveforms are in part modulated by the functioning of the aortic valve and mechanical properties of the arteries that carry the pulse wave. In addition, the computation of spectrograms and spectrums is fast and convenient; the presented scheme can be easily implemented in existing diagnostic methods such as PSG or home sleep apnoea testing. These findings together with previous studies imply that the assessment of OSA severity and evaluation of related daytime dysfunctions could be conducted using pulse oximeter measurements and deep learning [9, 10, 36]. This would further facilitate referrals to in-depth examinations for those with the highest risk of severe consequences of OSA. To assess specifically the risk of deteriorated vigilance, we compiled univariate models for the three most important predictors (ESS score, t90% and median APF) of belonging to the lapses Q4 and also different combinations of the multivariate models (figure 6). The ROC analyses revealed that none of the three parameters are capable of producing meaningful predictions alone and suffer from a high number of false positives with high sensitivity. However, when median APF, ESS or t90% are combined with sex and age, models' sensitivity and specificity outperform the univariate models and the differentiation capability of the AHI (figure 6). However, to produce a more reliable model, similar analyses need to be conducted in various patient cohorts to optimise the used predictor variables and related β-coefficients.

This study has certain limitations. First, the median total sleep times were short, which is most likely due to a common first-night effect in PSG. It is acknowledged that for a more comprehensive analysis, nocturnal PPG recordings from multiple nights would be needed. However, the current protocol consisting of PVT prior to PSG mitigates the first-night effect to PVT results. Second, our data do not include a sleep diary or long-term actigraphy. Chronic sleep deprivation is shown to affect the PVT performance to a large extent [37]; therefore, the lack of these data is a limitation. Third, we could not consider motivational aspects affecting PVT performance such as having a driver's licence or being a professional driver. These factors also affect the ESS scores, and we acknowledge that this is a limitation. Fourth, a complete record of the patients' medication at the time of measurement was not available. It is acknowledged that, for example, psychoactive medication can affect the PVT results and medications such as β-blockers can affect the arterial pulsation waveforms. We minimised this effect in stepwise regression by adjusting for comorbidities, but we acknowledge that this is a limitation. Fifth, the existence of hypertension is based on medical records and earlier diagnosis. Thus, this cohort may contain patients with undiagnosed arterial hypertension. In addition, hypertension was treated as a dichotomous variable due to the lack of blood pressure follow-ups of the patients in this cohort. Sixth, besides PVT, other concentration and sleepiness tests such as the Oxford Sleep Resistance (OSLER) test and Maintenance of Wakefulness test exist. By combining these tests with the PVT, a more comprehensive overview of the patients’ daytime performance and symptoms would be achieved.

In conclusion, an APF shift to higher frequencies and increased median frequency of the PPG provide a PSG marker for poor psychomotor vigilance task performance in male OSA patients. These findings highlight the connection of sleep disorders to cardiorespiratory regulation. In addition, the presented method could help in detecting those OSA patients at the highest risk of deteriorated vigilance.