Estimating actigraphy from motion artifacts in ECG and respiratory effort signals

P. Fonseca, R.M. Aarts, X. Long, J. Rolink, S. Leonhardt

Research output: Contribution to journalArticleAcademicpeer-review

5 Citations (Scopus)
94 Downloads (Pure)

Abstract

Recent work in unobtrusive sleep/wake classification has shown that cardiac and respiratory features can help improve classification performance. Nevertheless, actigraphy remains the single most discriminative modality for this task. Unfortunately, it requires the use of dedicated devices in addition to the sensors used to measure electrocardiogram (ECG) or respiratory effort. This paper proposes a method to estimate actigraphy from the body movement artifacts present in the ECG and respiratory inductance plethysmography (RIP) based on the time-frequency analysis of those signals. Using a continuous wavelet transform to analyze RIP, and ECG and RIP combined, it provides a surrogate measure of actigraphy with moderate correlation (for ECG+RIP, $\rho =0.74$ , p  <  0.001) and agreement (mean bias ratio of 0.94 and 95% agreement ratios of 0.11 and 8.45) with reference actigraphy. More important, it can be used as a replacement of actigraphy in sleep/wake classification: after cross-validation with a data set comprising polysomnographic (PSG) recordings of 15 healthy subjects and 25 insomniacs annotated by an external sleep technician, it achieves a statistically non-inferior classification performance when used together with respiratory features (average κ of 0.64 for 15 healthy subjects, and 0.50 for a dataset with 40 healthy and insomniac subjects), and when used together with respiratory and cardiac features (average κ of 0.66 for 15 healthy subjects, and 0.56 for 40 healthy and insomniac subjects). Since this method eliminates the need for a dedicated actigraphy device, it reduces the number of sensors needed for sleep/wake classification to a single sensor when using respiratory features, and to two sensors when using respiratory and cardiac features without any loss in performance. It offers a major benefit in terms of comfort for long-term home monitoring and is immediately applicable for legacy ECG and RIP monitoring devices already used in clinical practice and which do not have an accelerometer built-in.
Original languageEnglish
Pages (from-to)67-82
Number of pages16
JournalPhysiological Measurement
Volume37
Issue number1
DOIs
Publication statusPublished - 1 Jan 2016

Fingerprint

Plethysmography
Actigraphy
Electrocardiography
Inductance
Artifacts
Healthy Volunteers
Sleep
Sensors
Equipment and Supplies
Monitoring
Wavelet Analysis
Accelerometers
Wavelet transforms

Cite this

@article{403a9debbe0f47d9820db7d144c5ae0b,
title = "Estimating actigraphy from motion artifacts in ECG and respiratory effort signals",
abstract = "Recent work in unobtrusive sleep/wake classification has shown that cardiac and respiratory features can help improve classification performance. Nevertheless, actigraphy remains the single most discriminative modality for this task. Unfortunately, it requires the use of dedicated devices in addition to the sensors used to measure electrocardiogram (ECG) or respiratory effort. This paper proposes a method to estimate actigraphy from the body movement artifacts present in the ECG and respiratory inductance plethysmography (RIP) based on the time-frequency analysis of those signals. Using a continuous wavelet transform to analyze RIP, and ECG and RIP combined, it provides a surrogate measure of actigraphy with moderate correlation (for ECG+RIP, $\rho =0.74$ , p  <  0.001) and agreement (mean bias ratio of 0.94 and 95{\%} agreement ratios of 0.11 and 8.45) with reference actigraphy. More important, it can be used as a replacement of actigraphy in sleep/wake classification: after cross-validation with a data set comprising polysomnographic (PSG) recordings of 15 healthy subjects and 25 insomniacs annotated by an external sleep technician, it achieves a statistically non-inferior classification performance when used together with respiratory features (average κ of 0.64 for 15 healthy subjects, and 0.50 for a dataset with 40 healthy and insomniac subjects), and when used together with respiratory and cardiac features (average κ of 0.66 for 15 healthy subjects, and 0.56 for 40 healthy and insomniac subjects). Since this method eliminates the need for a dedicated actigraphy device, it reduces the number of sensors needed for sleep/wake classification to a single sensor when using respiratory features, and to two sensors when using respiratory and cardiac features without any loss in performance. It offers a major benefit in terms of comfort for long-term home monitoring and is immediately applicable for legacy ECG and RIP monitoring devices already used in clinical practice and which do not have an accelerometer built-in.",
author = "P. Fonseca and R.M. Aarts and X. Long and J. Rolink and S. Leonhardt",
year = "2016",
month = "1",
day = "1",
doi = "10.1088/0967-3334/37/1/67",
language = "English",
volume = "37",
pages = "67--82",
journal = "Physiological Measurement",
issn = "0967-3334",
publisher = "Institute of Physics",
number = "1",

}

Estimating actigraphy from motion artifacts in ECG and respiratory effort signals. / Fonseca, P.; Aarts, R.M.; Long, X.; Rolink, J.; Leonhardt, S.

In: Physiological Measurement, Vol. 37, No. 1, 01.01.2016, p. 67-82.

Research output: Contribution to journalArticleAcademicpeer-review

TY - JOUR

T1 - Estimating actigraphy from motion artifacts in ECG and respiratory effort signals

AU - Fonseca, P.

AU - Aarts, R.M.

AU - Long, X.

AU - Rolink, J.

AU - Leonhardt, S.

PY - 2016/1/1

Y1 - 2016/1/1

N2 - Recent work in unobtrusive sleep/wake classification has shown that cardiac and respiratory features can help improve classification performance. Nevertheless, actigraphy remains the single most discriminative modality for this task. Unfortunately, it requires the use of dedicated devices in addition to the sensors used to measure electrocardiogram (ECG) or respiratory effort. This paper proposes a method to estimate actigraphy from the body movement artifacts present in the ECG and respiratory inductance plethysmography (RIP) based on the time-frequency analysis of those signals. Using a continuous wavelet transform to analyze RIP, and ECG and RIP combined, it provides a surrogate measure of actigraphy with moderate correlation (for ECG+RIP, $\rho =0.74$ , p  <  0.001) and agreement (mean bias ratio of 0.94 and 95% agreement ratios of 0.11 and 8.45) with reference actigraphy. More important, it can be used as a replacement of actigraphy in sleep/wake classification: after cross-validation with a data set comprising polysomnographic (PSG) recordings of 15 healthy subjects and 25 insomniacs annotated by an external sleep technician, it achieves a statistically non-inferior classification performance when used together with respiratory features (average κ of 0.64 for 15 healthy subjects, and 0.50 for a dataset with 40 healthy and insomniac subjects), and when used together with respiratory and cardiac features (average κ of 0.66 for 15 healthy subjects, and 0.56 for 40 healthy and insomniac subjects). Since this method eliminates the need for a dedicated actigraphy device, it reduces the number of sensors needed for sleep/wake classification to a single sensor when using respiratory features, and to two sensors when using respiratory and cardiac features without any loss in performance. It offers a major benefit in terms of comfort for long-term home monitoring and is immediately applicable for legacy ECG and RIP monitoring devices already used in clinical practice and which do not have an accelerometer built-in.

AB - Recent work in unobtrusive sleep/wake classification has shown that cardiac and respiratory features can help improve classification performance. Nevertheless, actigraphy remains the single most discriminative modality for this task. Unfortunately, it requires the use of dedicated devices in addition to the sensors used to measure electrocardiogram (ECG) or respiratory effort. This paper proposes a method to estimate actigraphy from the body movement artifacts present in the ECG and respiratory inductance plethysmography (RIP) based on the time-frequency analysis of those signals. Using a continuous wavelet transform to analyze RIP, and ECG and RIP combined, it provides a surrogate measure of actigraphy with moderate correlation (for ECG+RIP, $\rho =0.74$ , p  <  0.001) and agreement (mean bias ratio of 0.94 and 95% agreement ratios of 0.11 and 8.45) with reference actigraphy. More important, it can be used as a replacement of actigraphy in sleep/wake classification: after cross-validation with a data set comprising polysomnographic (PSG) recordings of 15 healthy subjects and 25 insomniacs annotated by an external sleep technician, it achieves a statistically non-inferior classification performance when used together with respiratory features (average κ of 0.64 for 15 healthy subjects, and 0.50 for a dataset with 40 healthy and insomniac subjects), and when used together with respiratory and cardiac features (average κ of 0.66 for 15 healthy subjects, and 0.56 for 40 healthy and insomniac subjects). Since this method eliminates the need for a dedicated actigraphy device, it reduces the number of sensors needed for sleep/wake classification to a single sensor when using respiratory features, and to two sensors when using respiratory and cardiac features without any loss in performance. It offers a major benefit in terms of comfort for long-term home monitoring and is immediately applicable for legacy ECG and RIP monitoring devices already used in clinical practice and which do not have an accelerometer built-in.

U2 - 10.1088/0967-3334/37/1/67

DO - 10.1088/0967-3334/37/1/67

M3 - Article

C2 - 26641863

VL - 37

SP - 67

EP - 82

JO - Physiological Measurement

JF - Physiological Measurement

SN - 0967-3334

IS - 1

ER -