Automatic Estimation of the Apnea-hypopnea Index Using Neural Networks to Detect Sleep Apnea

Automatic Estimation of Apnea-hypopnea Index (AHI) Using Neural Networks to Assist in the Diagnosis of Sleep Apnea-hypopnea Syndrome (SAHS)

The sleep apnea hypopnea syndrome (SAHS) is a respiratory disorder characterized by frequent breathing cessations (apneas) or partial collapses (hypopneas) during sleep. These respiratory events lead to deep oxygen desaturations, blood pressure and heart rate acute changes, increased sympathetic activity and cortical arousals. The gold standard method for SAHS diagnosis is in-hospital, technician-attended overnight polysomnography (PSG). However, this methodology is labor-intensive, expensive and time-consuming, which has led to large waiting lists, delaying diagnosis and treatment. Blood oxygen saturation (SpO2) from nocturnal pulse oximetry (NPO) provides relevant information to detect apneas, it can be easily recorded ambulatory and it is less expensive and highly reliable. The investigators hypothesize that an automatic analysis of single oximetric recordings at home could provide essential information on the diagnosis of SAHS. The aim of this study is two-fold: firstly, the research focuses on assessing the reliability and usefulness of NPO carried out at patient's home in the context of SAHS detection and, secondly, the study aims at assessing the performance of an automatic regression model of the AHI by means of neural networks using information from NPO recordings. To achieve this goal, both PSG and NPO studies are carried out. A polysomnography equipment (E-Series, Compumedics) is used for standard in-hospital PSG studies, whereas a portable pulseoximeter (WristOX2 3150, Nonin) is used for ambulatory NPO. NPO is carried out the day immediately before or after the PSG at patient's home. Patients are assigned to carry out the NPO study before or after the in-hospital PSG randomly. In addition, in-hospital attended oximetry is also performed simultaneously to the PSG using the portable pulseoximeter.

Study Overview

• Status: Unknown
• Conditions: Sleep Apnea/Hypopnea Syndrome

Detailed Description

Subjects under study are recruited from the sleep unit of the "Hospital Universitario Río Hortega" (HURH) from Valladolid (Spain). All subjects are derived to the sleep unit due to suspicion of suffering from SAHS. The whole population set is subsequently divided into training set and test set. The training set is used to compose the regression model, whereas the test set is used to further assess its performance.

The standard apnea-hypopnea index (AHI) from PSG is used to diagnose SAHS. According to the American Academy of Sleep Medicine (AASM) rules, apnea is defined as a drop in the airflow signal greater than or equal to 90% from baseline lasting at least 10s, whereas hypopnea is defined as a drop greater than or equal to 50% during at least 10 s accompanied by a desaturation greater than or equal to 3% and/or an arousal. Subjects with an AHI >= 10 events per hour (e/h) are diagnosed as suffering from SAHS.

A portable pulseoximeter (WristOX2 3150, Nonin) is used for ambulatory NPO. NPO is carried out the day immediately before or after the PSG at patient's home. Patients are assigned to carry out the NPO study before or after in-hospital PSG randomly. In addition, oximetry is also performed simultaneously to the PSG by means of the portable pulseoximeter. Therefore, every patient has 3 oximetric recordings: (i) SpO2 from unattended portable monitoring at home, (ii) SpO2 from attended in-hospital portable monitoring and (iii) SpO2 from attended in-hospital standard PSG.

SpO2 is recorded at a sampling rate of 1 Hz. All SpO2 recordings are saved to separate files and process offline. An automatic signal pre-processing stage is carried out to remove artifacts.

Our methodology is divided into two stages: feature extraction and pattern recognition. Oximetric recordings are parameterized by means of 16 features from four feature subsets to compose the initial feature set from oximetry: time domain statistics, frequency domain statistics, conventional spectral measures and nonlinear features. All features are computed for each whole overnight recording.

• Features 1 to 4. First to fourth-order moments (M1t - M4t) in the time domain: arithmetic mean (M1t), variance (M2t), skewness (M3t) and kurtosis (M4t) are applied to quantify central tendency, amount of dispersion, asymmetry and peakedness, respectively.
• Features 5 to 8. First to fourth-order moments (M1f - M4f) in the frequency domain.
• Feature 9. Median frequency (MF), which is defined as the component which comprises 50% of signal power.
• Feature 10. Spectral entropy (SE), which is a disorder quantifier related to the flatness of the spectrum.
• Feature 11. Total spectral power (PT), which is computed as the total area under the PSD.
• Feature 12. Peak amplitude (PA) in the apnea frequency band, which is the local maximum of the spectral content in the frequency range 0.014 - 0.033 Hz.
• Feature 13. Relative power (PR), which is the ratio of the area enclosed under the PSD in the apnea frequency band to the total signal power.
• Feature 14. Sample entropy (SampEn), which quantifies irregularity in time series, with larger values corresponding to more irregular data.
• Feature 15. Central tendency measure (CTM), which provides a variability measure from second order difference plots.
• Feature 16. Lempel - Ziv complexity (LZC), which is a measure of complexity linked with the rate of new subsequences and their repetition along the signal.

The second stage corresponds to regression analysis, which aims to provide an analytical expression for the AHI as a function of the extracted features. A multilayer perceptron (MLP) neural network is used. MLP networks are models for expressing knowledge using a connectionist paradigm inspired in the human brain. They are composed of multiple simple units or neurons known as perceptrons. Perceptrons are arranged in several interconnected layers. Each network connection between two of them is associated with a network adaptive parameter or weight. MLP networks with a single hidden layer composed of nonlinear perceptrons (i.e., with a nonlinear activation function) are implemented since they are capable of universal approximation. The proposed regression task aims to approximate a 1-D continuous variable representing the AHI. Thus, a single output unit with a linear activation function is required.

• Study Type: Observational
• Enrollment (Actual): 322

Contacts and Locations

Study Locations

• Spain
  • Valladolid, Spain, 47012
    • Hospital Universitario Rio Hortega

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: 18 years and older (Adult, Older Adult)
• Accepts Healthy Volunteers: No
• Genders Eligible for Study: All

• Sampling Method: Probability Sample

Study Population

Subjects derived to the reference sleep unit from primary care with suspicion of suffering from SAHS due to daytime hypersomnolence, loud snoring, nocturnal choking and awakenings, and/or apnoeic events

Description

Inclusion Criteria (consecutive patient sampling):

• Men and women over 18 years old
• Subjects submitted to the sleep unit due to previous symptoms of sleep apnea (daytime hypersomnolence, loud snoring, nocturnal choking and awakenings, and/or apnoeic events)
• Written informed consent signed

Exclusion Criteria:

• Subjects under 18 years old
• Subjects not signing the informed consent
• Presence of any previously diagnosed sleep disorders: narcolepsy, insomnia, chronic sleep deprivation, regular use of hypnotic or sedative medications and restless leg syndrome
• Patients with chronic diseases: congestive heart failure, renal failure, neuromuscular diseases, chronic respiratory failure
• Patients with > 50% of central apneas or the presence of Cheyne-Stokes respiration
• Previous CPAP treatment for SAHS diagnosis
• A medical history that may interfere with the study objectives or, in the opinion of the investigator, compromise the conclusions

Study Plan

How is the study designed?

Design Details

• Observational Models: Case-Control
• Time Perspectives: Prospective

Number of groups / cohorts

2

Cohorts and Interventions

Group / Cohort
SAHS negative; Subjects derived to the sleep unit due to suspicion of suffering from sleep apnea which finally do not have the disease according to standard PSG
SAHS positive; Subjects derived to the sleep unit due to suspicion of suffering from sleep apnea which finally have the disease according to standard PSG

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Correlation between our estimated AHI from oximetry and real AHI from gold standard PSG	A measure of correlation between our estimation of the AHI and the real AHI derived from conventional in-hospital PSG will be measured by means of the intra-class correlation coefficient (ICC). This measure shows how similar are both indexes (estimated AHI and real AHI) in order to assess the severity of SAHS using our estimated AHI. In addition, Bland and Altman plots of agreement between NPO-based estimated AHI and PSG-based standard AHI will be drawn in order to assess under/over-estimation along the whole range of AHI values.	12 months after the inclusion of the last patient
Percentage of patients correctly classified	Percentage of patients correctly classified by the optimum portable NPO-based algorithm using the NPO-based estimated AHI. PSG is used as the reference gold standard method. Subjects with an AHI >= 10 event per hour (e/h) are considered as suffering from SAHS.	12 months after the inclusion of the last patient

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Prevalence of SAHS	Prevalence of SAHS in patients derived to the sleep unit	12 months (inclusion period)
Severity of SAHS	Severity of SAHS patients in terms of the AHI	12 months (inclusion period)
Demographic and anthropometric characteristics	Demographic and anthropometric characteristics of the study population (mean +/- standard deviation): age, gender, body mass index, neck circumference, waist circumference, blood pressure.	12 months (inclusion period)
Clinical characteristics of the study population	Clinical characteristics of the study population: previous symptoms of suffering from SAHS and additional conditions (hypertension and chronic obstructive pulmonary disease) co-occurring with SAHS according to standard definitions.	12 months (inclusion period)
Patients' lifestyle	Patients' lifestyle derived from questionnaires on sleep (Epworth Sleepiness Scale, ESS), smoking and alcoholism (Test EuroQol, EQ-5D)	12 months (inclusion period)
PSG-derived variables	PSG-derived variables (AHI; apnea index (AI); hypopnea index (HI); percentage of time in phase I, II, III, IV and REM sleep; percentage of time in supine position; arousal index; sleep efficiency)	12 months (inclusion period)
Portable NPO-derived variables	Portable NPO-derived variables (oxygen desaturation index of 3% (ODI3) and 4% (ODI4), cumulative time with a saturation value below 90% (CT90), minimum saturation, average saturation)	12 months (inclusion period)
Compliance with portable device	Compliance with the portable NPO recording device	12 months (inclusion period)
Physiological interpretation	Physiological interpretation of features included in the regression model	24 months
Cost-effectiveness	Cost-effectiveness study of the proposed model for SAHS screening based on AHI estimation from NPO	24 months

Collaborators and Investigators

• Sponsor: Sociedad Española de Neumología y Cirugía Torácica
• Investigators: Principal Investigator: Felix Del Campo, PhD, MD, Hospital Universitario Río Hortega, University of Valladolid

Publications and helpful links

Helpful Links

• Biomedical Engineering Group (GIB) Website

Study record dates

Study Major Dates

• Study Start: July 1, 2013
• Primary Completion (Actual): March 1, 2014
• Study Completion (Anticipated): July 1, 2014

Study Registration Dates

• First Submitted: May 15, 2014
• First Submitted That Met QC Criteria: May 20, 2014
• First Posted (Estimate): May 21, 2014

Study Record Updates

• Last Update Posted (Estimate): May 21, 2014
• Last Update Submitted That Met QC Criteria: May 20, 2014
• Last Verified: May 1, 2014

More Information

Terms related to this study

• Keywords: Pulse Oximetry; Sleep Apnea/Hypopnea Syndrome; Nonlinear Dynamics; Fourier Analysis; Neural Networks (Computer)
• Additional Relevant MeSH Terms: Pathologic Processes; Nervous System Diseases; Respiratory Tract Diseases; Respiration Disorders; Sleep Disorders, Intrinsic; Dyssomnias; Sleep Wake Disorders; Disease; Signs and Symptoms, Respiratory; Sleep Apnea Syndromes; Sleep Apnea, Obstructive; Syndrome; Apnea
• Other Study ID Numbers: SEPAR-265/2012