Contactless Sleep Apnea Detection on Smartphones
Rajalakshmi Nandakumar Shyamnath Gollakota Nathaniel Watson M.D.
Computer Science and Engineering Computer Science and Engineering UW Medicine Sleep Center
University of Washington University of Washington University of Washington
rajaln@uw.edu gshyam@uw.edu nwatson@uw.edu
Abstract – We present a contactless solution for detecting sleep
apnea events on smartphones. To achieve this, we introduce a novel
system that monitors the minute chest and abdomen movements
caused by breathing on smartphones. Our system works with the
phone away from the subject and can simultaneously identify and
track the fine-grained breathing movements from multiple subjects.
We do this by transforming the phone into an active sonar system
that emits frequency-modulated sound signals and listens to their
reflections; our design monitors the minute changes to these reflec-
tions to extract the chest movements. Results from a home bedroom
environment shows that our design operates efficiently at distances
of up to a meter and works even with the subject under a blanket.
Building on the above system, we develop algorithms that iden-
tify various sleep apnea events including obstructive apnea, central
apnea, and hypopnea from the sonar reflections. We deploy our sys-
tem at the UW Medicine Sleep Center at Harborview and perform
a clinical study with 37 patients for a total of 296 hours. Our study
demonstrates that the number of respiratory events identified by our
system is highly correlated with the ground truth and has a corre-
lation coefficient of 0.9957, 0.9860, and 0.9533 for central apnea,
obstructive apnea and hypopnea respectively. Furthermore, the av-
erage error in computing of rate of apnea and hypopnea events is as
low as 1.9 events/hr.
CATEGORIES AND SUBJECT DESCRIPTORS
J.3 [Computer Applications]: Life and Medical Sciences
GENERAL TERMS
Design; Human Factors; Algorithms
KEYWORDS
Mobile Health; Sleep Apnea; Phone Sonar; Contactless Breath-
ing Monitoring
1. INTRODUCTION
Sleep apnea is a common medical disorder that occurs when
breathing is disrupted during sleep. It is estimated to affect more
than 18 million American adults [9, 41] and is linked to attention
deficit/hyperactivity disorder, high blood pressure, diabetes, heart
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
MobiSys’15, May 18–22, 2015, Florence, Italy.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-3494-5/15/05 ...$15.00.
http://dx.doi.org/10.1145/2742647.2742674 .
Figure 1—Sensors used in the polysomnography test. The figure
shows all the sensors used in the test along with the data collection
unit. Polysomnography is used to diagnose various sleep disorders
including sleep apnea. Our goal is to use a smartphone to detect
sleep apnea events without any sensors on the human body.
attack, stroke, and increased motor vehicle accidents [6, 39]. Diag-
nosing sleep apnea in the clinic requires the polysomnography test
which is an expensive, time-consuming and labor-intensive process.
It requires a trained technician to attach and monitor various sensors
on the patient for the sleep duration and is typically associated with
long waiting lists [21]. While portable recording systems are being
developed for use in home settings, they require instrumenting ei-
ther the patient [25, 20, 20] or the bed [30] with various sensors
and most still require a trained technician to setup the recording
system [30].
In this paper we ask the following question: Can we leverage
smartphones to detect sleep apnea events without the need for sen-
sor instrumentation? The key challenge is that detecting sleep apnea
events requires tracking the fine-grained abdomen and chest move-
ments due to breathing [18]. While the iPhone Respiratory app [4]
can track the breathing movements, it requires placing the phone on
the body between the ribcage and the stomach and hence is intru-
sive. Vision-based solutions [43] can track these movements with-
out instrumenting users, but are limited to line-of-sight and good
lighting conditions and hence are not applicable to the sleep envi-
ronment, i.e., in the dark or under a blanket.
We introduce a novel contactless system that tracks the chest and
abdomen movements on smartphones and works in the sleep en-
vironment. It operates with the phone away from the user and can
concurrently track the breathing movements from multiple users.
Using this design, we build ApneaApp, a smartphone-based solu-
tion for detecting sleep-related respiratory events reported in a clin-
ical polysomnography test including hypopnea (when the subject’s
breathing becomes shallow), obstructive apnea (a complete or par-
tial blockage of the subject’s airway) and central apnea (when the
subject holds his or her breath).
Our key insight is to transform the phone into an active sonar
system. At a high level, we transmit 18-20 kHz sound waves from
the phone speaker and listen to their reflections at the microphone.
The chest and abdomen motion due to breathing creates changes
to the reflected sound waves. These changes, however, are minute
and extracting them reliably from other environmental reflections is
challenging. To overcome this, we employ FMCW (frequency mod-
ulated continuous wave) transmissions that allow us to separate re-
flections arriving at different times by mapping time differences to
shifts in the carrier frequency. Specifically, the reflections from the
human body arrive at a specific time depending on the distance from
the phone speaker. Thus, focusing on the corresponding frequency
allows us to reliably extract the amplitude changes due to breathing,
in the presence of all other environmental reflections. Further, since
reflections from multiple subjects would arrive at different times,
the corresponding frequencies provide us with the ability to simul-
taneously track multiple breathing signals. Finally, non-breathing
body motion creates reflection patterns distinct from breathing, en-
abling us to distinguish between them.
We implement our design on off-the-shelf smartphones and run
benchmark experiments with five healthy participants in a bedroom
environment using the Vernier respiratory belt as a baseline. Our
results show the following:
• Our system estimates the coarse-grained breathing frequency
1
to within 99.2% of the baseline at distances of up to a meter
from the subject. This translated to an error of less than 0.11
breaths/min. These accuracies remain this high even when the
subjects use blankets.
• The above accuracies remain unaffected by audible noise in the
environment from vehicles on a nearby street as well as human
conversations. This is because, we use a high-pass filter to filter
out audible signals below 18 kHz.
• It can separate and concurrently track the breathing movements
of two subjects on the bed separated by 20 cm.
Building on the above system, we design algorithms to compute
the number of central, obstructive and hypopnea events as well as
the apnea-hypopnea index which is the average rate of apnea and
hypopnea events during the sleep duration. We achieve this by pro-
cessing both the fine- and coarse- grained changes due to the chest
and abdomen movements as well as non-breathing body motion.
We deploy ApneaApp at the UW Medicine Sleep Center at Har-
borview and perform a clinical study with 37 patients for a total of
296 hours. The patients in our study were ordered by their physi-
cians to undergo the polysomnography (PSG) test. Our study was
done concurrently with the PSG test and we consider the sensor
data and diagnosis from the latter as the ground truth for evaluating
our system. Our study shows the following:
• Across patients, the number of central apnea, hypopnea and ob-
structive apnea events detected by our system is highly corre-
lated with the ground truth. Specifically, the intra-class correla-
tion coefficient between PSG and ApneaApp, is 0.9957, 0.9533
and 0.9860 for central apnea, hypopnea and obstructive apnea
respectively.
1
Detecting sleep apnea events requires tracking the fine-grained ab-
domen and chest motion variations in addition to the coarse-grained
breathing frequency. We evaluate ApneaApp’s ability to track these
variations in our clinical study.
Figure 2—Snapshot of a Clinical PSG Report. It summarizes
the number of obstructive, central and hypopnea events along with
the apneas-hypopneas index (AHI). An AHI value between 0–5 is
classified as no-apnea, values between 5–15 are classified as mild-
apnea, AHI values between 15–30 are classified as moderate-apnea,
and higher AHIs are severe apnea conditions.
• The average error in computing the rate of apnea and hypopnea
events is 1.9 events/hr; this is a clinically acceptable value [30].
• Our system accurately classifies 32 out of 37 patients between
four sleep apnea levels (no apnea, mild, moderate, and severe
apnea). The five misclassifications occur between no-apnea and
mild-apnea; four of them happen right at the boundary between
the two levels with an error less than 1 event/hr. These bound-
ary cases are handled separately by physicians depending on the
patient preferences, symptoms, and insurance; thus, effectively
reducing the number of misclassifications to one.
• We ran an audibility test with 87 sleep apnea patients (ages be-
tween 23 and 93 with a mean age of 50) and 57 healthy under-
graduate students at UW CSE. None of the 87 sleep apnea pa-
tients reported any audible sounds from ApneaApp. Only two
of the 57 undergraduates reported hearing audible sounds. This
demonstrates that ApneaApp is inaudible for most of the adult
population.
Contributions: We make four key contributions: (1) We introduce
a novel contactless technique for tracking chest and abdomen move-
ments due to breathing on smartphones. We achieve this by analyz-
ing the reflections from FMCW sonar transmissions. (2) We design
algorithms to detect central apnea, obstructive apnea, and hypopnea
as well as estimate the apnea-hypopnea index from the sonar reflec-
tions. (3) We implement our design on off-the-shelf smartphones
and demonstrate the ability to concurrently track breathing move-
ments from multiple subjects. (4) We perform a clinical study with
37 patients demonstrating the feasibility of using our system to ac-
curately compute the number of central, obstructive, and hypopnea
events as well as the apnea-hypopnea index.
2. POLYSOMNOGRAPHY OVERVIEW
The clinical polysomnography test (PSG) is traditionally used
to diagnose sleep apnea and other sleep disorders. It is conducted
overnight in a sleep laboratory where a trained technician monitors
the patient’s sleeping patterns. To do this, the technician attaches
the patient with a number of sensors including a chest and abdomen
belt to measure breathing movements, a nasal pressure transducer, a
snore microphone, a pulse oximeter to measure oxygen saturation,
a movement sensor on each leg to detect movements and five EEG
sensors to measure brain activity. The sensors are all connected us-
ing wires and the technician monitors the live data stream from the
sensors, throughout the sleep duration.
Fig. 2 shows a snapshot of a PSG report. The key metric used for
sleep apnea diagnosis is the AHI — the Apnea-Hypopnea Index —
that represents the rate at which apnea and hypopnea events occur
during the sleep period. Physicians classify the sleep apnea level
using these AHI values. Specifically, AHI values between 0–5 are
classified as no-apnea, those between 5–15 are classified as mild-
0 10 20 30 40 50 60
Time (in sec)
Nasal Chest
Pressure Motion
Central Apnea
(a) Central Apnea Event
0 20 40 60 80
Time (in sec)
Nasal Chest
Pressure Motion
Hypopnea
(b) Hypopnea Event
0 10 20 30 40 50 60
Time (in sec)
Nasal Chest
Pressure Motion
Spike
Obstructive Apnea
Obstructive Apnea
(c) Obstructive Apnea Event
Figure 3—American Academy of Sleep Medicine (AASM) Sig-
nal Characterization of the Apnea Events. The figures show the
chest motion and nasal pressure signals for the three apneas. A cen-
tral apnea event occurs when the subject holds her breath for a non-
negligible duration. A hypopnea event occurs when the subject’s
chest motion drops by more than 30% with an accompanying 3–
4% oxygen desaturation. Finally, an obstructive apnea event occurs
when the subject makes an increased effort to pull air into the lungs
but air does not reach the lungs due to blockage.
apnea, AHI values between 15–30 are classified as moderate-apnea,
and higher AHIs are severe apnea.
The apnea-hypopnea index is computed as follows:
AHI =
#central apnea + #hypopnea + #obstructive apnea
total sleep time
In the above equation, central apnea, hypopnea, and obstructive ap-
nea denote the various apnea conditions that are tracked during the
study. Mixed apneas are another class of apneas that are sometimes
included in the above equation. However, none of our PSG reports
showed non-zero mixed apneas and so we ignore them in our com-
putation.
To compute the above parameters, the eight-hour sensor data is
split into 30-second intervals called epochs. The scoring process
of analyzing these epochs involves two main steps. The first step
is staging, which identifies whether the patient is awake or asleep
in each epoch. This is achieved by examining the brain activity ob-
tained from the five EEG sensors. At the end of this step, each epoch
is marked as being in either a wake or sleep stage. The second step
involves identifying the number of central apnea, hypopnea, and
obstructive apnea events, using the AASM guidelines [18] outlined
below.
Identifying central apnea events. A central apnea event occurs
when the subject holds her breath for a non-negligible duration.
Fig. 3(a) shows the nasal pressure and chest motion signals during a
central apnea event. The figure shows that the chest movements are
flat indicating the absence of breathing effort; consequentially the
nasal pressure is also flat. If this persists for more than ten seconds,
it is marked as a central apnea event.
Identifying hypopnea events. A hypopnea event occurs when the
subject’s breathing becomes shallow. Fig. 3(b) plots the nasal pres-
sure and chest motion signals during a hypopnea event. The figure
shows that during a hypopnea event, the chest movements reduce
in amplitude. In particular, if this amplitude drops by more than
30% and has an accompanying 3–4% oxygen desaturation, then the
corresponding epoch is marked as hypopnea. We note that recent
clinical research [30] has shown that the 30% reduction alone can
be used for detecting hypopneas without a significant loss of accu-
racies.
Identifying obstructive apnea events. Obstructive apnea occurs
when there is a complete or partial blockage of the upper airway
during sleep. During an obstructive apnea event, the subject makes
an effort to pull air into the lungs, however air does not reach the
lungs because of blockage. Fig. 3(c) shows the signals where the
breathing effort can be seen in the chest band signals and the air
flow is flat in the nasal pressure sensor.
We note the following three points about PSG:
• The current procedure for sensor data collection and process-
ing is both labor and time intensive. Specifically, it takes about
an hour for the technician to fit each patient with these sensors.
Throughout a sleep duration of eight hours, the technician mon-
itors the sensors and ensures that they remain properly attached
to the patient’s body. The sensor data is then processed manually
to tag every epoch with the sleep apnea events.
• While portable sleep apnea testing is performed in the home,
it still require setting up the patient with chest and abdomen
belts, nasal pressure sensors, transducer and thermistors, EKG
and pulse oximetry. Home testing has a high failure rate of up to
33% due to the loss of signal resulting from detachment of wires
and cables [33].
• A PSG test is also used to diagnosis other sleep-related con-
ditions including upper airway resistance syndrome which in-
volve respiratory effort related arousals (RERA) that are shown
in Fig. 2. RERAs are sleep arousals that do not meet the above
definitions of apneas and hypopneas. While these are respiratory-
related and could be detected using our sonar-based system, ex-
ploring them in detail is not in the scope of this paper.
3. APNEAAPP
ApneaApp is a contactless system that enables detection of sleep
apnea events using smartphones. To understand how ApneaApp op-
erates, we first describe how we transform the phone into an active
sonar system that tracks the chest and abdomen movements due to
breathing. We then describe our algorithms to detect sleep apnea
events from these movements.
3.1 Transforming the Phone into an Active Sonar
An FMCW waveform is a chirp signal, as shown in Fig. 4,
where the transmitted frequency increases linearly with time be-
tween 18 kHz and 20 kHz. These signals reflect off the reflector
(e.g., human body) and arrive at the microphone after a time delay.
To determine this delay, at a high level, an FMCW receiver com-
Figure 4—Traditional FMCW Processing. The transmitter con-
tinuously transmits signals where the frequency increases linearly
with time between f
0
and f
1
. The reflections that arrive with a time
delay ∆t create a frequency shift ∆f . The receiver extracts this fre-
quency shift by performing an FFT over the chirp duration.
pares the frequencies of the transmitted and reflected signals. Since
the transmitted frequency increases linearly in time, time delays in
the reflected signals translate to frequency shifts in comparison to
the transmitted signals.
For instance, the red line is the transmitted signal from the phone
speaker and the green line is the reflected signal from a human body
that arrives with a time delay ∆t. This delay is given by
2d
v
sound
, where
d is the distance from the human body and v
sound
is the speed of
sound. Now, the frequency shift ∆f between the transmitted and
reflected signal is:
∆f =
f
1
− f
0
T
sweep
∆t
When we have multiple reflectors that are at different distances
from the receiver, their reflections translate to different frequency
shifts in the signal. An FMCW receiver can extract all these fre-
quency shifts by performing an Fourier transform over a chirp du-
ration as shown in Fig. 4. The chirp duration, T
sweep
, in practice is
picked so that the reflections from all points within the desired oper-
ational distance would start arriving before the chirp ends. Since our
operational distance is a meter we pick a chirp duration of 10.75 ms
in our implementation.
Challenge: The act of breathing creates minute chest and abdomen
motion that can be captured by monitoring the corresponding bin in
the Fourier transform as a function of time. The challenge however
is that breathing movements are minute and create a very small fre-
quency shift. Specifically, a 2 cm breathing displacement creates a
11.7 Hz shift.
2
This is problematic because given our sampling rate
and chirp duration, the width of each FFT bin is 93.75 Hz which is
much greater than the frequency shifts created due to breathing.
To address this problem, as shown in Fig. 5, the ApneaApp re-
ceiver performs a Fourier transform over N FMCW chirps. This is
in contrast to a traditional FMCW receiver that computes a Fourier
transform over the duration of a single FMCW chirp. Such an oper-
ation, decreases the width of each FFT bin by a factor of N. In our
implementation we set N to ten which results in an FFT bin width
of 9.37 Hz. This allows us to capture the minute frequency shifts re-
sulting from the breathing movements. We note that performing an
2
Given the speed of sound, a 48 kHz sampling rate translates to a
resolution of 0.71 cm per sample. Further, a 10.7 ms chirp duration
corresponds to 512 samples. With 18–20 kHz FMCW chirps, each
sample corresponds to a 3.9 Hz frequency shift. Thus, a displace-
ment of 0.71 cm translates to a 3.9 Hz change in the frequency
domain. Consequentially, a 2 cm breathing movement creates a
11.7 Hz frequency shift.
Figure 5—FMCW Processing in ApneaApp. To extract the
minute frequency shifts created by breathing motion, ApneaApp
performs an FFT over an integer number of chirp durations.
FFT over multiple chirp durations reduces our ability to track high-
frequency movements that occur during these chirps. However, in
our implementation, ten chirps correspond to a very short 107 ms, a
duration within which significant breathing movements are unlikely
to occur.
The final question is: how do we compute the distance of the sub-
ject from the phone? At a high level, we start at a distance of zero
and search for breathing movements at increasing distance values
up to the maximum distance of one meter. Specifically, we search
for breathing movements in the 58 Fourier bin corresponding to
18 kHz to 18.546 kHz.
3
In our implementation we reduce the com-
putation by searching in every alternate FFT bin. To search for these
breathing movements in each FFT bin, we perform another FFT
over a 30s duration and search for peaks in the typical breathing
frequencies of 0.2-0.3 Hz.
To summarize, the phone transmits FMCW signals in the 18-20
kHz range with a chirp duration of 10.7 ms from its speaker. The
microphones receive the reflected signals and process them to track
the breathing movements. Specifically, we first find the distance
to the human by searching for a periodic breathing signal starting
from the closed distance value to the maximum range of one meter.
Once we find this distance, we track the breathing movements by
performing in a shorter FFT over ten chirp durations and monitor
the reflected signals corresponding to the estimated distance value.
Note that the above procedure is repeated every time the user moves
their position, which we identify using the algorithm in §3.2.2. This
prevents the need for manually calibrating the distance between the
user and the phone.
We note the following points about our algorithm.
Computational Complexity. Our algorithm requires one 5120-point
FFT and between one to twenty nine 24000-point FFTs to success-
fully extract the breathing motion. We stop our search at the first
FFT bin that has the breathing movements.
Tracking breathing from multiple subjects. Reflections correspond-
ing to subjects at different distances arrive with different time de-
lays. Thus they create different frequency shifts with the FMCW
signal. Therefore, to track breathing from two subjects, we modify
the above algorithm to continue its search until it finds two FFT
bins with the breathing motion.
Leveraging angle-of-arrival algorithms. One could track breathing
movements from equidistant subjects who are at different angles,
by using multiple microphones and implementing angle-of-arrival
3
For
18-20 kHz FMCW transmissions, a distance of zero corre-
sponds to the FFT bin for 18-kHz. The operational distance of 1 m
corresponds to a frequency shift of 18.546 Hz.
algorithms. Evaluating this, however, is not in the scope of this pa-
per.
FMCW versus pulse-modulated transmissions. Pulse-modulated
transmissions use high-amplitude short pulses and are an alterna-
tive to continue-wave FMCW signals. In our experiment, however,
they created low frequency components in the 0-18 kHz range that
made them noticeably audible. FMCW transmissions, on the other
hand, have lower-amplitudes and are limited to 18-20 kHz, making
them inaudible for most of the adult population.
3.2 Sleep Apnea Detection Algorithm
As described in §2, diagnosing sleep apnea requires estimating
the Apnea-Hypopnea Index (AHI) which is the average rate of ap-
nea events during the sleep duration. This requires computing the
number of central, obstructive, and hypopneas as well the total sleep
time. In this section, we first describe our algorithms to compute the
number of apneas and then the total sleep time.
3.2.1 Estimating the Number of Apneas
ApneaApp detects a hypopnea event when the chest motion re-
duces below a threshold (30%).
4
A central apnea event is detected
when the subject holds her breath and as a result the amplitude of
the chest motion signal reduces to zero. Thus, ApneaApp identifies
hypopnea when the amplitude of the breathing motion decreases
below the threshold but the periodicity that is expected in a breath-
ing signal still exists. But to identify central apnea, we couple an
amplitude reduction in the chest motion signal with an absence of
the breathing periodicity.
An obstructive apnea event, on the other hand, occurs when there
is an obstruction in the airflow, i.e., the subject makes a breathing
effort but the airflow is obstructed by a tissue overgrowth in the
neck. In a clinical PSG study, obstructive apnea is detected using
the nasal pressure sensor that directly measures the airflow; this,
however, is not available in our system. We instead perform an anal-
ysis of the chest motion data that reveals that the subject usually
tends to increase her breathing effort in these scenarios. This re-
sults in a clear spike in the amplitude of the chest motion, as shown
in Fig. 3(c).
Thus, measuring the amplitude and periodicity in the chest mo-
tion signal is critical to detecting obstructive, central and hypop-
neas. Note that as shown in Fig. 3, the chest motion signal can be
approximated as a periodic sinusoidal wave. Hence the magnitude
of the peaks of these sinusoidal waves represents the amplitude and
their peak locations determine their periodicity. Thus, we design a
peak detection algorithm to compute the amplitude and periodicity
of the chest motion signal.
ApneaApp’s peak detection algorithm. Standard peak detection
algorithms identify the transition point at which the signal changes
from an increasing to a decreasing trend. In other words, for every
set of three points, if the middle point is the maximum, then it is
labeled as a peak. Such an algorithm, however, would result in a
number of erroneous peaks with our chest motion signal. As an
example, Fig. 6 plots a typical chest motion signal. Running the
standard peak detection algorithm on this signal results in a number
of unintended peaks as shown in the figure. To reduce the number
of such peaks, we introduce two key heuristics.
4
AASM and Medicare guidelines require the 30% reduction to be
accompanied with a 3% and 4% oxygen saturation respectively.
However, recent clinical research [30] has shown that the 30% re-
duction alone can be used for detecting hypopneas.
0 5 10 15 20 25 30
Chest Motion Signal
0 5 10 15 20 25 30
Traditional Peak Detection Algorithm
0 5 10 15 20 25 30
ApneaApp’s Peak Detection Algorithm
Time(in sec)
Figure 6—Understanding ApneaApp’s peak detection algo-
rithm. Standard peak detection algorithms identify the transition
point at which the signal changes from an increasing trend to a de-
creasing trend. Thus, they identify a number of unintended peaks in
our chest motion signal. ApneaApp’s peak detection algorithm, in
contrast, uses minimum peak distance and amplitude heuristics to
reduce these unintended peaks.
The first heuristic is to set a threshold on the minimum dis-
tance between two consecutive peaks. In particular, the breath-
ing frequency in an adult human typically varies between 12-18
breaths/min. Thus each breath takes 3.3 s at the maximum fre-
quency of 18 breaths/min. We set a conservative threshold of three
seconds in our implementation.
The second heuristic is to set a threshold on the minimum am-
plitude at which a peak is detected. To do this, we first run the
above peak detection algorithm with the minimum distance heuris-
tic on the chest motion data for the first hour to obtain an initial set
of peaks. We then compute the minimum amplitude threshold as
µ
peaks
− 2σ
peaks
, where µ
peaks
and σ
peaks
are the mean and standard
deviation of the peak amplitudes. Finally, we go back and apply
both the amplitude as well as the minimum distance thresholds on
the entire chest motion data for the eight-hour sleep duration to ob-
tain the actual set of peaks. Fig. 6 shows that our peak detection
algorithm identifies the correct peaks.
Central apnea estimation algorithm. We run the peak detection
algorithm to identify the locations of the peaks in the chest motion
signal. We then compute the distance between these consecutive
peaks. If this distance is greater than ten seconds, it means that the
subject holds her breath for a non-negligible period of time and
hence we declare it as a central apnea event.
Hypopnea estimation algorithm. We again use the peak detection
algorithm to detect the peaks. When the peak values reduce beyond
a threshold and still maintain their periodicity, we declare it as a
hypopnea event. To compute this threshold for our sonar data, we
perform a linear regression on the data from a single patient to max-