Sleep Medicine Reviews 13 (2009) 333–343

Contents lists available at ScienceDirect

Sleep Medicine Reviews

journal homepage: www.elsevier.com/locate/smrv

PHYSIOLOGICAL REVIEW

Understanding sleep-disordered breathing through mathematical modelling

Tero Aittokallio a, b, *, Arho Virkki a, b, Olli Polo b, c
a
Biomathematics Research Group, Department of Mathematics, University of Turku, FI-20014 Turku, Finland
b
Sleep Research Unit, Department of Physiology, University of Turku, FI-20520 Turku, Finland
c
Department of Pulmonary Medicine, Tampere University Hospital, FI-33521 Tampere, Finland

s u m m a r y

Keywords: Recent studies have uncovered high prevalence of undiagnosed sleep-disordered breathing, and its
Sleep-disordered breathing linkage to metabolic or cardiovascular disorders which represent increasing health hazard. However, the
Obstructive sleep apnea mechanistic links behind these disorders as well as their contribution to the experimental observations
Flow limitation
and treatment responses remain poorly understood. Therefore, the screening of clinical measurements
Transcutaneous carbon dioxide
still relies upon relatively simple diagnostic features, such as signal averages or event frequencies, which
Non-invasive measurement
Mathematical modelling may represent suboptimal or surrogate markers of the underlying abnormality. Consequently, most
Predictive modelling patients are being treated with general therapies regardless of the cause of their key dysfunction.
Metabolic disorder Combining experimental measurements with mathematical modelling has the potential to provide
Vascular impairment mechanistic insights into the individual factors underlying the disease progression, which may finally
enable tailored treatment alternatives for each patient. This review depicts a number of modelling
approaches to elucidate sleep-related dysfunctions of the human respiratory system, and how these
models are being used to translate the measurements first into new ideas and then into testable
hypotheses. Such model-based investigations can provide systematic strategies towards better under-
standing, predicting or even preventing these dysfunctions. Along with the brief description of the
modelling approaches, we discuss their relative merits and potential implications especially for clinical
research.
Ó 2008 Elsevier Ltd. All rights reserved.

Introduction
Obstructive sleep-disordered breathing (SDB) is a common
condition, which manifests in a spectrum of conditions ranging
from simple snoring and airflow limitation through partial upper
airway collapsibility and hypopnea to the obstructive sleep apnea
syndrome (OSAS). A number of clinical conditions related to SDB
are nowadays occurring with greater frequency as the population
ages and becomes more obese in many countries. While it was
earlier estimated that 5% of the adults in Western countries suffer
from the symptoms of OSAS, such as excessive daytime sleepiness,1
the recent reports suggest that as many as one in four American
Abbreviations: AHI, apnea–hypopnea index; CO2, carbon dioxide; CPAP,
continuous positive airway pressure; EtCO2, end-tidal partial pressure of CO2;
HbA1c, glycosylated hemoglobin A1c; HDL-C, high-density lipoprotein cholesterol;
IMT, carotid artery intima–media thickness; NMD, nitroglycerin-mediated dilata-
tion; O2, oxygen; ODI, oxyhemoglobin desaturation index; OSAS, obstructive sleep
apnea syndrome; PaCO2, arterial partial pressure of CO2; SaO2, oxyhemoglobin
saturation; SDB, sleep-disordered breathing; TcCO2, transcutaneous partial pressure
of CO2.
* Corresponding author. Department of Mathematics, University of Turku, FI-
20014 Turku, Finland. Tel.: þ358 2 333 6030; fax: þ358 2 333 6595.
E-mail address: tero.aittokallio@utu.fi (T. Aittokallio).
adults are at high risk for OSAS.2 In addition to the patients having
symptoms of OSAS, recent studies have linked also the milder forms
of SDB with increased likelihood of hypertension and other
cardiovascular consequences.3–5 This cluster of conditions shares
also many risk factors with the metabolic syndrome.6–8 Moreover,
childhood SDB is associated with neurophysiological deficits of
memory, learning and executive function.9 Considering these
serious and lifelong health consequences, there is a critical need for
novel strategies to increase our understanding of the pathophysi-
ology of SDB, and its links to the related disorders, enabling their
early detection and effective treatment.
The development of SDB depends on the complex interplay of
multiple anatomic and physiological factors but the mechanisms
behind these factors or their impacts on the type or degree of the
disorder, experimental observations or on the treatment responses
remain poorly understood. While the key regulatory mechanisms
involved in the respiratory control are well-established,10 as well as
the main causes leading to SDB,11,12 what is lacking is an experi-
mental-analytic framework, consisting of a simple measurement
setup and formal decision rules, that could efficiently employ this
body of knowledge in transforming the experimental measure-
ments into clinical information. Although current experimental
techniques enable overnight monitoring of multiple signals in

1087-0792/$ – see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.smrv.2008.09.008
334 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343
a research setting, the complexity and cost of these methods and
their possible impact on sleep and the system may limit their usage
in clinical routine. Moreover, the conventional diagnosis screening
is based on the identification of respiratory-related events in these
measurements, such as episodes of apnea, hypopnea or arousal, but
such indices alone can provide only limited information about the
various mechanisms responsible for producing these events.
Accordingly, the emphasis of the diagnostic procedures should
move from complicated measurement of the secondary events
towards determination of the primary dysfunction in a given
individual using minimally invasive measurement techniques.
Mathematical models provide us a quantitative means for
understanding the dynamic behaviour of complex physiological
controlling systems, facilitating the mechanistic interpretation of
the experimental measurements and the prediction of clinically
important events in the systems function. Mathematical modelling
has successfully been used, for instance, in explaining the occur-
rence of periodic breathing in patients with cardiac symptoms as
a result of instability of the respiratory control system.13,14 Besides
providing mechanistic insights into the disease pathogenesis,
modelling can be used to suggest novel hypotheses, predict clini-
cally relevant outcome variables, or even to test the efficiency of
different treatment settings.15,16 This may lead to novel and exciting
physiological paradigms which can in turn contribute to the
development of efficient treatment alternatives to prevent these
conditions. For diagnostic testing, models can also help us to choose
optimal experimental setup by revealing the various disturbances
underlying the development of the medical conditions and by
suggesting the most effective measurements for their early detec-
tion. Given the rapidly growing incidence of SDB, there is an
increasing need for such automated tools to facilitate clinical
decision-making that could provide a cost-effective support to
conventional experimental studies.
This review describes how mathematical modelling can be used
to study the operation of the human respiratory control system in
health and disease. Our particular focus is on the dynamic behav-
iour of the system during transition from wakefulness to sleep, and
on those system dysfunctions that are implicated in the develop-
ment of SDB and its related conditions. As substantial efforts have
been devoted to the topic over the past 40 years, only representa-
tive examples of different modelling approaches can be surveyed
here, with an emphasis on those model-based studies that target at
addressing clinical questions, rather than mathematical or
computational issues only. After having introduced the necessary
modelling terminology and techniques in the context of modelling
the respiratory control system during sleep, two concrete case
studies are given that demonstrate how models can identify
important features from overnight experimental measurements,
including non-invasive recording of respiratory airflow and carbon
dioxide levels, and how these features can be used to predict
respiratory, metabolic and cardiovascular abnormalities. Potential
treatment options based on the model predictions will also be
discussed in the course of these two particular case studies. Finally,
we highlight some challenges in the field and offer our personal
view of the key future trends and developments in the model-
based analysis of SDB.
Modelling the human respiratory control system
during sleep
A mathematical modela can be broadly regarded as a quantita-
tive description of an observed system or its particular process,
a
Underlined terms are further defined in a Glossary of terms at the end of the
text.
which mimics those aspects of the real behaviour that enable useful
predictions to be made. Mathematical modelling as a systematic
tool to test new ideas and make hypotheses about the real word
phenomena has a long and successful record especially in the
physical sciences. Recently, modelling has increasingly being used
also in many fields of the biomedical sciences, especially in
biomedical engineering and systems biology. Due to the complexity
of many biological processes and our improved possibilities to
make high-throughput measurements at multiple levels of
a system, models have become indispensable tools in the integra-
tion, analysis and interpretation of the experimental measure-
ments. At the same time, powerful computing capacity has made it
possible to build realistic models of biological systems that can be
run on standard computers to simulate these processes in quanti-
tative terms. This is illustrated by the possibility of building meta-
models, in which different model components are coupled together
to predict macro-level behaviour. Beyond computer simulations,
model predictions can lead – sometimes after a great deal of effort –
to a hypothetical mechanism that is able to explain the observed
behaviour and can be tested with subsequent experimentation. In
addition to providing better understanding of the system’s prop-
erties, e.g., its many dysfunctions, perhaps the most exciting
potential comes from the possibility to design rational strategies to
modify a pathologic behaviour of the system back to its normal
state.
Due to its high inherent complexity and medical importance,
the many properties of the human respiratory control system have
been a topic of a large number of modelling studies, varying both in
their abstraction level and modelling approach. These modelling
choices are typically made on the basis of the specific physiological
application, that is, whether one is interested in the disease path-
ogenesis, its diagnosis or treatment. Of the various modelling
approaches, one should choose the approach that can best address
the question under analysis, and often the most useful models are
not necessarily those involving most mathematical sophistication
or computational complexity. So-called descriptive models aim to
reveal relationships among the experimental measurements and
other information (referred globally to as features), and their
contribution to a particular clinical outcome. Typical examples
include using stepwise linear regression modelling to select an
independent subset of anatomic and other features that could
account for most of the individual variability in SDB, e.g., in terms of
the observed apnea–hypopnea index (AHI).17 A related modelling
approach determines those features only that can predict the
correct AHI-class of a new subject as accurately as possible by
exploiting possible non-linear relationships between the features
in predictive models.18 These models and features can then be used,
e.g., for prioritizing patients for polysomnographic sleep studies.19
More information on the descriptive and predictive modelling
approaches can be found from related books.20
Compared to descriptive modelling, mechanistic models have
typically a rather different objective of providing mechanistic
insights into the behaviour of the respiratory system under
different physiological or pathophysiological conditions. Tradi-
tional applications include finding conditions under which the
respiration becomes stable or unstable, and the most important
factors affecting the occurrence and type of the periodic
breathing.21,22 These early modelling works have revealed, e.g., that
the control system instability may originate from prolonged
circulatory times or disturbed chemosensitivity to hypercapnia and
hypoxia. Such findings can have direct implications for medical
problems, such as congestive heart failure,23,24 or sudden infant
death syndrome.25,26 Model-based approaches to understanding
the nature of periodic breathing, especially in the context of central
SDB, have extensively been described in a recent review.14 In
general, the mechanistic modelling approach benefits from the
 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343
more direct linkage between the model and the real system,
allowing, e.g., hypothesis testing through physiologically relevant
model parameters, which is often lacking in the descriptive models.
On the other hand, the mechanistic interpretations rely more
profoundly upon a priori knowledge of the factors most relevant for
the condition of interest. A number of books have been devoted to
the techniques useful for building, solving and analysing mecha-
nistic models.27
After the model has been formulated, it is solved using analytical
or numerical techniques, and the results are then carefully analysed
for both mathematical soundness and physiological relevance. The
most fruitful analysis technique depends on the chosen description
level of a model, that is, whether the model tries to capture the
behaviour of the entire respiratory system or some of its sub-
systems only, and whether the chosen components are modelled in
detail or using a coarser-level representation. While detailed
system-level models can enable a realistic representation of the
respiratory system in its entirety, these models often become
intractable to be solved analytically. In those cases, computational
modelling can be used to numerically simulate the system behav-
iour to predict its response to different perturbations. Another
challenge involves choosing decent values for the model parame-
ters as the models increase in complexity. Alternative approach is to
reduce the complexity of the model to allow convenient mathe-
matical treatment, such as deriving analytical properties of its
solutions, or even model identification, which often involves esti-
mation of a part of the system parameters directly from the
measurements. While analytical calculations from the reduced
models can be mathematically attractive, the match between the
model and the real system may be significantly limited due to the
simplifying assumptions.
Regardless of the modelling approach chosen, it is important to
realize that all models are only idealizations of the real system and
are always limited in their applicability. Therefore, an essential
phase in any model building process is model validation, which can
be regarded as the process of gaining confidence that the model is
valid and useful for its intended purpose. While predictive
modelling can be very powerful for the purpose of diagnostic
classification of subjects into different clinical groups based on the
experimental measurements, these models cannot provide much
insights into the mechanisms leading, e.g., to the different facets of
SDB. On the other hand, mechanistic models have the potential to
reveal dynamic interactions involved in the pathophysiology of
SDB, but detailed system-level description can render them
prohibitively complicated to be useful in clinical practice. The
following two case studies demonstrate how combining the good
properties of these two approaches with experimental studies can
lead to efficient procedures with potential clinical applications. In
particular, we demonstrate how modelling can be used in trans-
lating the hidden information from non-invasive measurements
into physiologically meaningful parameters, thus providing novel
hypotheses for future experiments, targeted features for diagnostic
prediction and possible treatment outcomes when testing the
existing or developing new therapies.
Case study 1 – modelling upper airway dynamics of sleeping
subjects
One of the most critical components of SDB involves the
increased collapsibility of the upper airway after sleep onset, which
may predispose to significant pharyngeal resistance and airflow
limitation.12,28 There are a number of methodologies that enable
accurate measurement of the mechanical properties of the
collapsible airway segment in a research setting, and heuristic
mathematical forms of varying complexity have been applied to
reveal the upper airway pressure–area–flow relationships from
335
these measurements in patients with SDB.29–31 Reduced models
have successfully summarized the upper airway component into
parameters, such as the critical pressure index of upper airway
collapsibility, which has widely been used, e.g., to differentiate
subjects with various degrees of SDB and to evaluate treatment
responses in patients.32 Besides requiring rather heavy instru-
mentation, however, the critical pressure is conventionally esti-
mated under static, stable and/or passive conditions. Consequently,
this linear modelling approach was not initially designed for
dynamic description of the various flow limiting kinetic events that
are known to occur in naturally sleeping subjects. An inherent
limitation of any reduced model is their incapability to dissect the
multiple factors affecting the flow-pressure properties during an
upper airway collapse.33 With some further modifications of the
model and simplifications in the experimental setups, the tech-
nique is becoming more useful and streamlined for more routine
clinical use.34,35
Non-invasive measurement of respiratory airflow with nasal-
prongs has gained much popularity in routine sleep studies as
a simple and sensitive experimental setup for detecting respira-
tory-related events.36,37 Although much of the clinical attention has
been directed towards a rather limited spectrum of the events (e.g.,
no airflow, reduced airflow, normal airflow), the shape of the
airflow curve can provide additional information on the upper
airway behaviour throughout the respiratory cycle. Descriptive
flow shape analysis has successfully been exploited to automati-
cally adjust and test therapeutic devices such as the CPAP titra-
tion,38 distinguish subject with different degrees of SDB,39 or to
monitor the treatment effects on the upper airway dynamics
during sleep.40 Each particular flow profile is a result of a composite
of forces that either promote a collapse or support the airway
(Fig. 1A). To suggest physiological correlates behind the various
flow shape profiles, we developed a mechanistic model that
enables predictions of the relative contribution of various static and
dynamic upper airway forces to the inspiratory flow kinetics.41
Compared with the earlier dynamic models,42,43 our model intro-
duces anatomically more compatible pharyngeal description by
including also the longitudinal dimension of the upper airway
(Fig. 1B). The model predictions can be presented in the form of
nasal airflow, which allows us to compare its simulations with flow
measurements of sleeping subjects (Fig. 1C). By specifying indi-
vidual parameters settings, which characterise a number of physi-
ological and anatomic properties, the model is able to reproduce
and explain the flow characteristics that are also clinically observed
in subjects with SDB as well as their response to treatment options,
such as the CPAP therapy.44
As an illustration, we show here how the upper airway model
can be used to explain the rather distinct flow characteristics
observed not only between healthy individuals and subjects with
SDB, but also between the healthy men and women (Fig. 2A).
Through the adjustable model parameters, one can introduce
baseline settings (‘virtual subjects’) and their specific modifications
(‘intervention effects’), the impact of which on the inspiratory flow
shape can be monitored by means of computer simulations. The
model predicts, e.g., that physiological sleep effects on the upper
airway muscle tone can explain the flow-limited shapes that are
characteristic to the male subjects, and also suggest a potential
intervention to normalize these shape abnormalities (Fig. 2B).
Similarly, the model is not only able to explain the distinct flow
shape distribution observed in the female subjects, but also the
distinct effects of progesterone therapy on the flow shape dynamics
as a joint contribution from two of its known modes of action
(Fig. 2C). These flow shape responses in virtual subjects are
consistent to those observed also in real subjects.40 The model
predictions suggest that increased stimulation of respiration may
lead to even worsening of flow limitation in some subjects, whereas
336 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343

A
Promotion of airway collapse Promotion of airway patency

Negative pressure on inspiration Pharyngeal dilator

muscle contraction
Extralumenal positive pressure (genioglossus)
Fat decomposition
Small mandible Lung volume
(longitudinal traction)

B ∂u ∂u ∂p
ρ +u = − − R(A, u)u
∂t ∂x ∂x
unsteady and pressure viscous
convective gradient resistance
inertial terms
C

0 10 20 30 40 50 60
Time (sec)

Fig. 1. Analogy between the upper airway model parameters and their physiological counterparts governing the airflow kinetics. (A) The various static and dynamic upper airway
forces along with the factors that either promote the upper airway collapse or its patency. The collapsing forces include negative pressure during inspiration (dynamic force) and
decreased tonic muscle activity after sleep onset (static force), whereas dynamic supporting forces are the pharyngeal dilator muscle activation during inspiration and tracheal
traction at end-inspiration (at high lung volumes). Figure reproduced with permission from Elsevier.12 (B) The partial differential equation model of a collapsible upper airway is
based on the one-dimensional approximation of the Navier-Stokes equations that describe the dynamic interplay of the variables of interest: airflow velocity u and the upper airway
area A and pressure p. The fluid dynamic equations include the Euler’s equation for the conservation of mass and the Newton’s law for the conservation of momentum, whereas the
elastic properties of the upper airway are modelled using the so-called tube law.41 The model includes five key parameters which correspond to specific anatomic or physiological
properties: Pdrive, magnitude of the respiratory effort; Ptract, influence of obesity on the tracheal traction; Ptiss, influence of soft tissues on the airway collapsibility; Pphas, phasic
muscle activity during inspiration; and Ptone, tonic muscle activity (decreases at sleep onset). (C) A non-invasive recording of airflow with nasal-prongs, which shows a gradually
developing inspiratory flow limitation until an arousal normalizes the breathing pattern (the vertical dotted line). The collapsible upper airway model was designed to provide
plausible interpretations of the upper airway mechanisms and properties explaining the experimentally observed flow changes during sleep.

augmented stimulation of the upper airway dilators can effectively
stabilize the upper airway and consequently improve their airflow.
In this way, the model-based analysis allows a priori testing of how
individual properties can affect the mediating mechanisms and
outcome results of different treatment alternatives. These subtle
flow shape differences consistently observed in the normal subjects
may also explain the somewhat contradictory results about gender
differences in SDB obtained when using different indices of upper
airway collapsibility or resistance.45–49
In addition to the conventional treatment options, the feasibility
of novel treatment alternatives for SDB can be tested with the
model, e.g., before going into more elaborated human or animal
studies. For instance, the model prediction of the beneficial effect of
lowering the density of the breathing gas is consistent with the
observed changes in the optimal CPAP pressure in patients
breathing different gas mixtures.50 Moreover, while there currently
are no simple interventions that could modify the collapsible
pharyngeal length in human subjects, the model could prognosti-
cate its importance to the SDB predisposition.44 Similar observa-
tions supporting our model predictions have previously been made
using an animal model,51 and in retrospect using a more detailed
computational modelling approach.52 Computational fluid
dynamics models of the upper airway behaviour can be made
anatomically more accurate by allowing their dynamic state to vary
not only over the longitudinal dimension but over two or even
three spatial dimensions.53,54 Such computational modelling
enable, e.g., simulations of the mechanical properties of the
collapsible airway in SDB or in response to various therapies.55–57
While solving numerically such complex models do not pose
problems to today’s computers, strategies that could avoid per-
forming an upper airway 3D reconstruction for each individual
separately, or using a representative geometrical structure aver-
aged from a set of MRI or CT scans, would make the computational
models more useful in clinical setting or when planning individu-
alised therapies for SDB.
Case study 2 – modelling the chemoreflex control of
respiration
While the upper airway collapsibility is perhaps the single most
important factor behind SDB, isolating its effect from the rest of the
respiration system prevents making predictions about the influ-
ence of the compensatory mechanisms mediated, e.g., by the
feedback information from both the central and peripheral
chemoreceptors. Differentiating the contribution of these two
receptors to the ventilatory response would be of major biomedical
importance not only for better understanding of their relative roles
in physiological and pathophysiological conditions, but also for
discovering novel diagnostic measures that could reflect early
problems in gas exchange during sleep. While non-invasive
measurement of respiratory airflow and oxyhemoglobin saturation
(SaO2) are routinely performed in many sleep studies, non-invasive
 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343 337

A %

Percentage of breaths
50 Male Patients
40 Male Controls
Female Controls
30
20
10
0

1 2 3 4 5 6 7 8

B Wake Sleep Interventions

UPPP

Ptone ↓

Ptiss ↓
Ptract ↓
Overweight
Pdrive ↑, Ptiss ↑

C
MPA effect
Pdrive ↑ on effort

Ptone ↓

MPA effect
Pphas ↑ on dilators
Pdrive ↓, Ptiss ↓

Fig. 2. Examples of observed flow shape distributions and their explanation by the predictions from the upper airway model. (A) Distributions of automatically identified flow
shape classes (the numbered shapes) and their relative incidences (the histograms) in a study material consisting of 5 healthy men and 5 healthy women (AHI 0-1 events per hour)
as well as 5 men with mild-to-moderate degree of SDB (AHI 1 – 22 events per hour). Figure modified with permission from Elsevier.88 (B) The ‘male setting’ for the collapsible upper
airway model was characterised by a high respiratory drive and increased collapsibility of the upper airway soft tissues. The sleep state was modelled as a reduced tonic activity of
the pharyngeal muscles, as compared to the wake state, resulting in a slightly limited flow waveform similar to that of the observed shape class 2, which was characteristic to the
male subjects. By subsequently decreasing the vital capacity and thereby further increasing the upper airway collapsibility through tracheal traction, one can reproduce the late-
plateau shape class 4, which is frequently observed in obese men with SDB. In our study material, the male patients were indeed generally overweight (average BMI 31 kg/m2), as
compared to the controls (average BMI 21 kg/m2). The degree of flow limitation could be increased by making the virtual subjects more obese (data not shown). In this setting, the
model predicts that pharyngeal surgery, such as uvulopalatopharyngoplasty (UPPP), induces a flow shape with mid-plateau and increased volume. Similar shape and volume
responses were also observed in the recordings in obese men with OSAS treated with UPPP.39 (C) The ‘female setting’ was characterised by a lower drive and collapsibility. In this
setting, the model predicts that falling asleep results in a flow-limited shape with a peak at end-inspiration. Such a late-peak inspiratory flow shape was not frequent only in our
female control group, as exemplified by the shape class 7, but also in obese postmenopausal women with partial upper airway obstruction during sleep,39 suggesting that there may
be some non-symptomatic subjects in this population that are more prone to develop SDB later in life or during challenge. As a potential treatment alternative for such cases, the
model prognosticates that respiratory drive stimulation with a progesterone analogue, such as medroxyprogesterone acetate (MPA), can transform the flow-limited shape into one
consisting of two peaks with increased initial slope and inspiratory volume. This change in the flow profile is consistent with our results from treating obese postmenopausal
women with MPA, where the major treatment effect was indeed late-peak shapes turning into twin-peaked shapes.40 The observed secondary effects of MPA, including increased
inspiratory slopes and volumes, were also predicted by the model. The alternative treatment effect, observed decrease in inspiratory duration on MPA,40 can be reproduced by
postulating that in some subjects MPA can affect the upper airway muscle activity as well.89 This modification can also restore the sinusoidal inspiratory flow shape. The modelling
exercise demonstrates how one can make predictions about the effectiveness of treatments in different subject groups by means of simulations.

measurement of CO2 has not become popular in sleep clinics. The
CO2 content can be estimated at the alveolar CO2 level from the
expired end-tidal partial pressure using nasal-prongs (EtCO2), or at
the peripheral level as the amount of CO2 that diffuses through the
skin using transcutaneous sensor (TcCO2). Comparative studies of
these two non-invasive techniques have demonstrated that TcCO2
enables accurate estimation of the arterial partial pressure of CO2
(PaCO2).58,59 Because of its non-invasive nature, TcCO2 measure-
ments have been used, e.g., for nocturnal monitoring of patients.60
Because the heated transcutaneous sensor reflects not only the
level of PaCO2, but there is an additional CO2 input from the skin
metabolism, TcCO2 variation has also been associated with over-
night changes in cardiac function and circulation delay.61 However,
the full clinical potential of the TcCO2 measurement, relative to that
of EtCO2, remains to be shown.
We have shown that menopause is associated with cardiore-
spiratory changes that can be observed in increased levels of TcCO2
from wakefulness to sleep, in the absence of changes in EtCO2.62 To
provide plausible explanation for this observation, a mechanistic
model was developed for predicting sleep-induced changes in the
338 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343

A B
Ventilation V˙
51 49 47 46 45.5 45

8
Alveolar 35
air
37

7
Blood Fast
flow Q˙ diffusion CO
2
Et

Blood flow (l/min)

6
Pulmonary blood Premenopausals 33
Venous Arterial Tc

5
CO
delay Lv delay La 2
31

4
Tissue capillary
Postmenopausals
blood 29
Diffusion

3
27
25

2
pH-receptors Feedback
Tissue mechanism 23

1
d 5.0 10 15 20
y = SKyR + V −1 Dy − diag ( D Tey) + ( Q ° L) y − diag ( Q Te)y + s(y) ×109
dt Controller gain (l/min/mol/l)
chemical diffusion terms blood flow terms sink and
reactions (with delay) source terms

C
Wake
REM
S1
S2
S3
S4
8
PTcCO2 (kPa)
7
6
5

23:00 23:30 00:00 00:30 01:00 01:30 02:00 02:30 03:00 03:30 04:00 04:30 05:00 05:30 06:00 06:30
Time

Fig. 3. A mechanistic model of the respiratory control system and its numerical predictions for nocturnal CO2 levels. (A) A schematic illustration of the multi-compartment
respiratory control model together with its governing equation. Multiple compartments for gas storage and exchange are combined with various transport and control mechanisms
(blood circulation, CO2 diffusion, conservation of mass, chemical reactions, and ventilatory feedback). In the differential equation, the variable y is a 23-dimensional state vector, and
S, K, R, D and Q are stoichiometric, kinetic, reaction, diffusion and flow matrices, respectively. Constant matrix V contains the volumes of the compartments on its diagonal.63 (B) The
diagram describes the model steady-state solutions as a function of the velocity of blood flow and the gain of the feedback function. The model predicts that the observed
differences between end-tidal (EtCO2) and transcutaneous (TcCO2) carbon dioxide levels can be explained by variations in the respiratory controller loop gain and blood flow. The
arrows illustrate typical changes in the EtCO2 and TcCO2 levels between wakefulness and sleep which are consistent with the measurements in pre- and postmenopausal women.62
Figure reproduced with permission from Springer Science and Business Media.63 (C) A representative result of an overnight TcCO2 recording with hypnogram (the top graph). The
dotted line indicates the slope at sleep onset (referred to as initial slope), which has been used as a target feature when detecting metabolic and vascular disturbances. This
application shows how mechanistic models can be used in explaining experimental observations and suggesting optimal targets for predictive modelling studies.

chemical regulation of ventilation.63 The model combines
multiple compartments for gas storage and exchange with various
transport and control mechanisms (Fig. 3A). Simulation results
supported our idea that the contradictory trends between the
EtCO2 and TcCO2 measurements across sleep onset could be
explained by menopausal state-related adjustments in the respi-
ratory parameters (Fig. 3B). Such adjustments result in an
observed difference between the peripheral tissue pH levels (as
reflected by TcCO2) and the alveolar blood CO2 levels (as reflected
by EtCO2). Modelling the interplay of CO2 levels at the two ends of
the system provides not only mechanistic insights into the
disturbances in the intracellular pH control that are known to be
involved in several major diseases, including diabetes and
atherosclerosis, but also better understanding of how these
disturbances are reflected in the non-invasive CO2 measurements
in a given patient. However, estimation of the control system
parameters using individual measurements is still complicated by
many experimental and theoretical challenges.64,65 While sensi-
tivity analysis makes it possible to identify the key respiratory,
metabolic and circulatory parameters from the non-invasive CO2
measurements,66 the limited amounts of data can still result in
many candidate parameter estimates, indicating that more
measured variables will be needed for accurate system identifi-
cation enabling clinical applications.
Besides the central chemoreflex, there is also a complementary
component of the ventilatory response associated with the
peripheral chemoreceptor in the carotid bodies that respond to
changes in the arterial O2, CO2 and pH. The precise mechanisms of
how these different signals are integrated into a neural output and
what is the importance of the peripheral chemoreceptor in health
and disease is still poorly understood.67,68 Using a simple flow-
sensitive model of the carotid body chemoreception,69 it was
 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343 339

A PaO2 (kPa)
B ) PaO2 (kPa)
18
) (kPa
(kPa 16
PaCO 2 PaCO 2
14
15 15
12
Impulses/s

Impulses/s
10 10
10
8
5 5
6
6.5 6.5 4
0 0
5.5 5.5
8 8 10 2
10 4.5 12 4.5
12 14
14 16
16 0

C D
16

7.0
14

SaO2 98

6.5
hypoxic + hypercapnic
12

6.0
10

SaO2 95

PaCO2 (kPa)
Impulses/s

5.5
8

only hypoxic SaO2 85

6

5.0
only hypercapnic
4

4.5

normal
2

4.0
0

1/8 1/2 3/4 1 4/3 3/2 40 60 80 100 120

Blood flow (fraction) Blood flow ( of the reference)

Fig. 4. Model-based predictions on how blood flow limitation can alter the neural response of the carotid body O2 and CO2 chemoreceptors. (A) The combined neural response as
a function of the arterial O2 and CO2 partial pressures (PaO2 and PaCO2). The neural response to hypoxia increases markedly as PaCO2 falls below 10 kPa. (B) The upper surface shows
the relative effect of a severe blood flow limitation (1/8 of the normal value). At high arterial O2 levels, the sensitivity to CO2 increases relative to the normal situation as the removal
of CO2 from the chemoreceptor cells is compromised. At lower O2 levels, in contrast, the relative sensitivity to CO2 decreases and the chemoreceptor response becomes saturated. (C)
Another view of the chemoreceptor neural stimulus in response to changes in blood flow through the carotid bodies. The best sensitivity and the maximal difference between the
‘normal’ and the ‘hypoxic þ hypercapnic’ stimuli is obtained at the level of maximal flow. In this example, the values SaO2 ¼ 85% and PaCO2 ¼ 6.4 kPa represented the hypoxic and
hypercapnic responses, respectively. (D) An example showing how increased blood flow can compensate some degree of hypercapnia. Each prediction curve corresponds to
a constant neural stimulus produced by a fixed arterial O2 saturation saturation level (85%, 95% or 98%). The results suggest that an increase in the blood flow (the horizontal arrow)
can fully compensate a 1.5 kPa increase in the arterial PaCO2 (the vertical arrow), provided that O2 level is relatively high. In contrast, only small increases in PaCO2 can be
compensated during hypoxia (SaO2 ¼ 85%). Taken together, the model predicts that the O2 response is affected even by moderate changes in the blood flow, whereas the CO2
response is not altered until blood flow becomes severely limited. This type of modelling application is aimed at computational testing of the feasibility of new ideas before
formulating specific hypotheses for subsequent human or animal experiments. Figures reproduced with permission from Elsevier.69

recently suggested that the huge perfusion capacity of the carotid investigated the extent to which overnight features in the TcCO2
body arteries enables an effective gas monitoring apparatus when measurements could predict metabolic outcome variables in
fully vasodilated, and that limiting the perfusion through vaso- subjects with SDB.71 Predictive modelling demonstrated that TcCO2
constriction may also offer a highly dynamic and fast response indeed has a substantial role, e.g., in predicting the levels of high-
mechanism to drive breathing beyond metabolic needs (Fig. 4). density lipoprotein cholesterol (HDL-C) and of glycosylated
Such overshoot is observed, for instance, before physical exercise in hemoglobin A1c (HbA1c). Because of the well-established associ-
wakefulness or before central apnea in periodic breathing.70 The ation between low HDL-C and coronary heart disease, as well as
model also predicts that the increase in CO2 observed during between elevated HbA1c and type 2 diabetes, this finding may open
transition from wakefulness to sleep, which is conventionally a novel functional perspective into the stationary markers for these
attributed to the withdrawal of the ‘wakefulness stimulus’, could at diseases. To further evaluate the contribution of peripheral CO2 to
least partly be mediated through the vasodynamic changes that the development of vascular impairment, we used TcCO2 features
occur when falling asleep. Adjusting ventilation through control- to predict early changes in vascular function.72 It was found out that
ling the carotid body perfusion is a new concept, but if experi- the initial slope of TcCO2 was an important predictor of carotid
mentally verified, it can have a number of important medical and artery intima–media thickness (IMT), as well as of nitroglycerin-
clinical implications. These predictions can lead, e.g., to effective mediated dilatation (NMD) of brachial artery. As both IMT and NMD
therapeutic targets to manipulate symptoms such as dyspnoea or are established markers of sub-clinical atherosclerosis, these
hyperventilation. For instance, one of the model-predicted features, model-based findings suggest a novel link between peripheral CO2
the relative increase in the TcCO2 levels across sleep onset (the measurements and early signs of vascular impairment. Interest-
initial slope in Fig. 3C), is being evaluated as a non-invasive indi- ingly, oxyhemoglobin desaturation index (ODI) or other features in
cator of the vascular function. the SaO2 measurements were not among the key predictors, indi-
To validate the prediction that metabolic disturbances should cating that it is not sleep apnea or hypoxemia per se but other
result in changes of peripheral perfusion, we systematically nocturnal events that are better reflected in the TcCO2 signal and
340 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343
enable identification of individuals at higher risk for metabolic and
cardiovascular abnormalities. In this way, models can be used for
designing cost-effective setups in terms of experimental tech-
niques, diagnostic features, and analytical methods that can be
optimized for each clinical application.
Model limitations and future research directions
As with any mathematical model of a physiological system,
simplifying assumptions of different importance were made in the
above modelling studies. Important mechanisms omitted in the
upper airway model (case study 1), but which have been consid-
ered in other models, included, e.g., neuromechanical interac-
tions,73 chemoreceptor feedbacks,74 and respiratory pattern
generators.75 When the dynamic interplay between such neuro-
chemical control mechanisms with the upper airway collapsibility
is taken into consideration in the models, the predicted breathing
pattern can properly reflect the continuous adjustments due to
feedback information from prior breaths.74,76 The chemoreflex
control model (case study 2), on the other hand, lacks, e.g., an
integrated carotid body and central chemoreceptor feedback
functions as well as the cardiac autonomic and baroreceptor reflex
control of heart rate, blood pressure and arousals from
sleep.22,24,77,78 The future modelling studies could use these and
other models of particular sub-systems as building blocks when
constructing larger meta-models to provide a more realistic
representation of the respiratory control system.
By coupling the existing and emerging models into a more
comprehensive modelling and simulation framework, one could
start looking for inter-relationships between the multitude of
factors contributing to the severity of SDB,79 as well as mechanistic
or even causal links to the related diseases, such as hypertension,80
atherosclerosis,81 cardiac failure,82 stroke,83 and diabetes.84 Such
bottom-up building of meta-models would provide a public test-
bench for educational purposes in a wide range of experimentally
inconvenient or even inaccessible situations, as well as for testing
and improving the conventional therapeutic strategies or devel-
oping novel and potentially feasible treatment alternatives. The
availability of such test-bench could substitute – or prioritize –
some of the human or animal experiments in the future. For
instance, we have made available a simulator server for the
collapsible upper airway model at our website (http://www.math.
utu.fi/en/research/groups/bio/projects/sleep.html). The simulator
enables user to predict the dynamic changes in nasal flow signal in
response to specific parameter settings.
As an alternative approach to the bottom-up model building
strategy, it would be appealing to start with a global, coarse-
grained model, and then gradually focus on the specific sub-
systems of interest by adjusting the description level as more
experimental data becomes available. Such a top-down modelling
approach is increasingly being used in many applications of
systems biology, as it is better targeted at finding novel regulatory
mechanisms, such as those involving genetic factors,85,86 rather
than relying upon a priori knowledge only. Both the bottom-up and
top-down modelling approaches enable an iterative model building
process, in which model simulations and predictions are closely
coupled with the new experiments, thus providing increasingly
more realistic descriptions of the system and more accurate
predictions for the subsequent experiments. We expect that one of
the most promising directions in the development of customized
medical applications will come from those modelling approaches
that can efficiently use several types of complementary information
gathered from the existing and emerging experimental
technologies.87
An advantage of the mechanistic mathematical models is that
the model parameters are individual-specific. In theory, the model
parameters fully characterise the condition of each subject, and
therefore the individual parameter vector can be used to distin-
guish between healthy subjects and patients with different
degrees of anatomic or physiological abnormalities, as well as to
predict their responses to interventions. In practice, however,
certain parameters can be extremely difficult or even impossible
to measure directly. While parameter estimation can in principle
be done through any identifiable model, accurate estimation of all
the model parameters from limited measurements can be infea-
sible especially in clinical practice. Simpler linear models are
easier to estimate using the current measurement technologies,
but their power of expression can be limited. At the other extreme,
there are the computational fluid dynamics models of the upper
airway behaviour that rely on averaged 2D or 3D anatomical
templates. The use of generic anatomical or physiological
parameters may, however, limit the accuracy of the model
predictions.
Conclusion
Advances in physiological knowledge, clinical measurement
techniques, computing power, and numerical methods are making
it possible to develop models that incorporate a greater number of
interacting control mechanisms, thus giving models the potential
to reflect more accurately the complex behaviour of the cardiore-
spiratory control system. Given the high incidence of clinical
conditions related to the dysfunction of these control mechanisms,
such as SDB, that are occurring with increasing frequency as the
population ages and becomes more obese, transforming these
advances into medical knowledge through mathematical model-
ling will be of major clinical importance for addressing the growing
health consequences in the near future. Since the models allow
computational testing of ideas and planning of experiments in an
unlimited number of simulated conditions, such physiology-driven
modelling frameworks could significantly enhance the produc-
tivity of medical research. Model-based strategies for designing
experiments may eventually challenge the common practice of
producing millions of bits and pieces for the human jigsaw-puzzle
at a much faster rate than anybody can test whether they fit
together, not to talk about building unifying paradigms or overall
understanding.
Although numerous modelling approaches have been applied
in the context of SDB, they are best understood according to their
ultimate goals. Three main medical applications of modelling
were identified that either try to improve our 1) pathophysio-
logical knowledge (e.g., carotid body model); 2) diagnostic
capabilities (e.g., transcutaneous CO2 features); or 3) therapeutic
options (e.g., individual pharyngeal properties). The modelling
approach or their combination should be chosen on the basis of
the specific application and measurements available. In the
typical case of limited clinical data, rigorously reduced models are
likely to be most successful at estimating individual parameters of
the respiratory system, such as the critical pressure for pharyn-
geal collapsibility or the loop gain for ventilatory control.
However, while these single parameters do represent critical
aspects of the system, the simplified models cannot reflect the full
dynamic range of the factors involved in SDB. Detailed models can
dissect the multiple factors behind experimental observations
under different conditions, and predict those features in the
measurements that are most important for distinguishing the
conditions. Clinical success of the modelling approaches is likely
to be driven by the parallel improvements in the experimental
setups that enable reliable identification of more complex models
as well.
Successful collaboration between clinical and mathematical
researchers requires that the different scientific goals of the two
 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343
disciplines are sufficiently met. From the medical point of view,
mathematics is a supporting science that helps to outline the
medical outcomes. In contrast, the mathematical scientists
consider the medical problem as a practical test-bench for their
theories and applications. Building together mathematical models
for concrete medical applications is a powerful tool to handle this
collaboration. Perhaps the most important lesson learned so far
from the mathematical modelling of SDB is that a multitude of
different outcomes of the respiratory system can be predicted by
changing only a few model parameters. This implies also that
certain medical therapies that are effective in some patients may
prove ineffective in others. The eventual aim of the model-based
approaches is therefore to facilitate the prediction of the optimal
therapy for each patient. By making the mathematical models
accessible to the medical community will eventually help to
underpin the full clinical potential of mathematical modelling in
establishing a new era of truly predictive, preventive and person-
alized medicine for SDB and other complex disorders.
Practice points
Mathematical modelling of sleep-disordered breathing is
a useful tool to:
1. explain, simulate and predict the system responses to
various clinical conditions that may be infeasible to
study in practice;
2. enable public test-benches for clinicians and students
that can be used for educational, scientific and clinical
purposes;
3. provide valuable checkpoint for novel hypotheses and
pinpoint flaws in the experimental designs and
measurement setups;
4. suggest specific targets for diagnostic procedures that
are optimal at distinguishing normal and abnormal
behaviour;
5. assess and improve conventional treatment strategies
and find those combinations that are most effective for
a given patient.
Research agenda
A model-based strategy to predict and prevent SDB involves
the following steps:
1. improve and combine the existing models of the respi-
ratory control system in conjunction with experimental
observations and physiological information;
2. discern the factors responsible for the experimentally
observed abnormal behaviour such as flow limitation,
snoring, sleep apnea, or periodic breathing;
3. reduce the complexity of the models to allow accurate
identification of those sub-systems that can explain and
accurately predict the abnormal behaviour;
4. use model predictions to establish a simple experi-
mental setup for sensitive screening of sleeping subjects
and their classification into different groups;
5. choose an existing therapy or their combination, or
develop a novel therapy, that can manipulate the indi-
vidual parameters back into their normal range.
341
Acknowledgements
The work was supported by the Academy of Finland (TA), and by
Tampere Tuberculosis Foundation (OP).
Glossary of terms
Mathematical model is a mathematical structure together with
an interpretation rule. The structure is typically a set of equations
describing the interplay between model variables and parameters,
and the interpretation rule tells the modeller how these quantities
are connected to the properties of the system or process being
modelled. A good mathematical model should at least be able to
quantitatively describe and predict important properties, such as
behaviour or state, of the real system.
Descriptive model presents and summarizes the important
aspects of the measurements in a mathematical form. Such data-
driven modelling effort involves choosing the functional form of
the model, and using measurements to determine the adjustable
parameters of the model. While typically it is sufficient that
a descriptive model provides a reasonable fit with the data without
being over-complicated, predictive model has the additional
requirement that it must carry also generalization power to unseen
data or subjects as well.
Mechanistic model is based on the underlying laws and mech-
anisms of the real world system from which the governing equa-
tions are formulated. The nature of these equations can be either
exact (like the law of mass action), empirical (like the Newton’s
laws of motion), or more hypothetical (like the response function to
hypercapnia). Compared to descriptive models, the interpretation
of the explanatory models is more closely connected to the real
system parameters, thus enabling mechanistic explanation for the
observed behaviour.
Model identification techniques deal with the estimation of
model parameters from the given experimental measurements of
an individual subject. This problem is complicated by the fact that
the models may become highly non-linear and that the data
available is typically partial and noisy. Sensitivity analysis can be
used to prioritize the parameters that are most important for the
model predictions and should – at the minimum – be estimated
from the data.
Computational model is a computer program that tries to
emulate the behaviour of a particular system. Computer simula-
tions have become useful in many natural systems in physics,
chemistry, biology and physiology, to gain insight into the opera-
tion of those systems or to observe their malfunctions using tech-
niques of numerical analysis. For instance, finite element method is
a typical algorithm for solving computational fluid dynamic models
over complex domains (like the upper airway anatomy).
Critical pressure is a measure of pharyngeal collapsibility based
on the representation of upper airway as a simple tube with
a collapsible segment. In this model, airflow cannot occur until the
pressure upstream of the segment exceeds the surrounding pres-
sure (Pcrit). In normal subjects, Pcrit is negative, whereas in snoring
subjects Pcrit becomes less negative, and in patients with OSAS,
Pcrit is typically positive enabling total occlusion of the upper
airway. Pcrit can be defined in practice as the level of nasal CPAP
mask pressure needed to induce a cessation of airflow
Model validation is a process of checking the soundness and
usefulness of a model. Depending on the modelling approaches,
part or all of the following questions can be addressed: Does the
model yield new relevant insight into the system? Are the under-
lying structure and assumptions of the model feasible? Does the
model fit the empirical observations and measurements available?
Do the model parameters fall into a plausible range consistent with
342 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343

the current knowledge? How well does the model predict new 26. Batzel JJ, Tran HT. Modeling instability in the control system for human
respiration: applications to infant non-REM sleep. Appl Math Comput
observations and events?
2000;110:1–51.
Systems biology is an approach to study the complex interac- 27. Fowler AC. Mathematical models in the applied sciences. Cambridge: Cambridge
tions between the components of biological systems, and how University Press; 1997.
these interactions give rise to the function of the system as a whole. 28. Sériès F. Upper airway muscles awake and asleep. Sleep Med Rev 2002;6:229–42.
29. Isono S, Feroah TR, Hajduk EA, Brant R, Whitelaw WA, Remmers JE. Interaction
Such studies are enabled by new biotechnologies that can measure of cross-sectional area, driving pressure, and airflow of passive velopharynx.
cellular components and their interactions on a large-scale. Inte- J Appl Physiol 1997;83:851–9.
gration of the multiple data sources using mathematical models 30. Tamisier R, Pepin JL, Wuyam B, Smith R, Argod J, Levy P. Characterization of
pharyngeal resistance during sleep in a spectrum of sleep-disordered
can offer systematic means to understand the dynamic behaviour breathing. J Appl Physiol 2000;89:120–30.
of biological processes at systems-level, rather than investigating 31. Mansour KF, Rowley JA, Meshenish AA, Shkoukani MA, Badr MS. A mathe-
snapshots of individual components separately. matical model to detect inspiratory flow limitation during sleep. J Appl Physiol
2002;93:1084–92.
*32. Gold AR, Schwartz AR. The pharyngeal critical pressure. The whys and hows of
using nasal continuous positive airway pressure diagnostically. Chest
References 1996;110:1077–88.
33. Sforza E, Petiau C, Weiss T, Thibault A, Krieger J. Pharyngeal critical pressure in
patients with obstructive sleep apnea syndrome. Clinical implications. Am J
1. Young T, Peppard PE, Gottlieb DJ. Epidemiology of obstructive sleep apnea:
Respir Crit Care Med 1999;159:149–57.
a population health perspective. Am J Respir Crit Care Med 2002;165:1217–39.
34. Patil SP, Punjabi NM, Schneider H, O’Donnell CP, Smith PL, Schwartz AR. A
2. Hiestand DM, Britz P, Goldman M, Phillips B. Prevalence of symptoms and risk
simplified method for measuring critical pressures during sleep in the clinical
of sleep apnea in the US population: results from the national sleep founda-
setting. Am J Respir Crit Care Med 2004;170:86–93.
tion sleep in America 2005 poll. Chest 2006;130:780–6.
35. Patil SP, Schneider H, Marx JJ, Gladmon E, Schwartz AR, Smith PL. Neuro-
3. Shahar E, Whitney CW, Redline S, Lee ET, Newman AB, Javier Nieto F, et al.
mechanical control of upper airway patency during sleep. J Appl Physiol
Sleep-disordered breathing and cardiovascular disease: cross-sectional results
2007;102:547–56.
of the Sleep Heart Health Study. Am J Respir Crit Care Med 2001;163:19–25.
36. Montserrat JM, Farré R, Navajas D. How to use the nasal pressure in clinical
4. Marin JM, Carrizo SJ, Vicente E, Agusti AG. Long-term cardiovascular outcomes
practice. Sleep Med 2003;4:381–3.
in men with obstructive sleep apnoea-hypopnoea with or without treatment
37. Ayappa I, Norman RG, Suryadevara M, Rapoport DM. Comparison of limited
with continuous positive airway pressure: an observational study. Lancet
monitoring using a nasal-cannula flow signal to full polysomnography in
2005;365:1046–53.
sleep-disordered breathing. Sleep 2004;27:1171–9.
5. Gozal D, Kheirandish-Gozal L. Cardiovascular morbidity in obstructive sleep
38. Farré R, Montserrat JM, Rigau J, Trepat X, Pinto P, Navajas D. Response of
apnea: oxidative stress, inflammation, and much more. Am J Respir Crit Care
automatic continuous positive airway pressure devices to different sleep
Med 2008;177:369–75.
breathing patterns: a bench study. Am J Respir Crit Care Med 2002;166:469–73.
6. Vgontzas AN, Bixler EO, Chrousos GP. Sleep apnea is a manifestation of the
39. Aittokallio T, Saaresranta T, Polo-Kantola P, Nevalainen O, Polo O. Analysis of
metabolic syndrome. Sleep Med Rev 2005;9:211–24.
inspiratory flow shapes in patients with partial upper-airway obstruction
7. Redline S, Storfer-Isser A, Rosen CL, Johnson NL, Kirchner HL, Emancipator J,
during sleep. Chest 2001;119:37–44.
et al. Association between metabolic syndrome and sleep-disordered
40. Saaresranta T, Aittokallio T, Polo-Kantola P, Helenius H, Polo O. Effect of
breathing in adolescents. Am J Respir Crit Care Med 2007;176:401–8.
medroxyprogesterone on inspiratory flow shapes during sleep in post-
8. Seicean S, Kirchner HL, Gottlieb DJ, Punjabi NM, Resnick H, Sanders M, et al.
menopausal women. Respir Physiol Neurobiol 2003;134:131–43.
Sleep-disordered breathing and impaired glucose metabolism in normal-
41. Aittokallio T, Gyllenberg M, Polo O. A model of a snorer’s upper airway. Math
weight and overweight/obese individuals: the Sleep Heart Health Study.
Biosci 2001;170:79–90.
Diabetes Care 2008;31:1001–6.
42. Gavriely N, Jensen O. Theory and measurements of snores. J Appl Physiol
9. Halbower AC, Mahone EM. Neuropsychological morbidity linked to childhood
1993;74:2828–37.
sleep-disordered breathing. Sleep Med Rev 2006;10:97–107.
43. Fodil R, Ribreau C, Louis B, Lofaso F, Isabey D. Interaction between steady flow
10. Mitchell GS, Douse MA, Foley KT. Receptor interactions in modulating venti-
and individualised compliant segments: application to upper airways. Med
latory activity. Am J Physiol 1990;259:R911–20.
Biol Eng Comput 1997;35:638–48.
*11. Cherniack NS. Sleep apnea and its causes. J Clin Invest 1984;73:1501–6.
*44. Aittokallio T, Gyllenberg M, Saaresranta T, Polo O. Prediction of inspiratory
*12. Malhotra A, White DP. Obstructive sleep apnoea. Lancet 2002;360:237–45.
flow shapes during sleep with a mathematic model of upper airway forces.
13. Cherniack NS. Apnea and periodic breathing during sleep. N Engl J Med
Sleep 2003;26:857–63.
1999;341:985–7.
45. Trinder J, Kay A, Kleiman J, Dunai J. Gender differences in airway resistance
*14. Cherniack NS, Longobardo GS. Mathematical models of periodic breathing and
during sleep. J Appl Physiol 1997;83:1986–97.
their usefulness in understanding cardiovascular and respiratory disorders.
46. Thurnheer R, Wraith PK, Douglas NJ. Influence of age and gender on upper
Exp Physiol 2006;91:295–305.
airway resistance in NREM and REM sleep. J Appl Physiol 2001;90:981–8.
15. Marini JJ, Crooke PS. A general mathematical model for respiratory dynamics
47. Rowley JA, Zhou X, Vergine I, Shkoukani MA, Badr MS. Influence of gender on
relevant to the clinical setting. Am Rev Respir Dis 1993;147:14–24.
upper airway mechanics: upper airway resistance and Pcrit. J Appl Physiol
16. Hota S, Crooke PS, Adams AB, Hotchkiss JR. Optimal phasic tracheal gas
2001;91:2248–54.
insufflation timing: an experimental and mathematical analysis. Crit Care Med
48. Jordan AS, Wellman A, Edwards JK, Schory K, Dover L, MacDonald M, et al.
2006;34:1408–14.
Respiratory control stability and upper airway collapsibility in men and
17. Dempsey JA, Skatrud JB, Jacques AJ, Ewanowski SJ, Woodson BT, Hanson PR,
women with obstructive sleep apnea. J Appl Physiol 2005;99:2020–7.
et al. Anatomic determinants of sleep-disordered breathing across the spec-
49. Anttalainen U, Saaresranta T, Kalleinen N, Aittokallio J, Vahlberg T, Polo O.
trum of clinical and nonclinical male subjects. Chest 2002;122:840–51.
Gender differences in age and BMI distributions in partial upper airway
18. El-Solh AA, Mador MJ, Ten-Brock E, Shucard DW, Abul-Khoudoud M, Grant BJ.
obstruction during sleep. Respir Physiol Neurobiol 2007;159:219–26.
Validity of neural network in sleep apnea. Sleep 1999;22:105–11.
50. Farré R, Rigau J, Montserrat JM, Buscemi L, Ballester E, Navajas D. Static and
19. Tsai WH, Remmers JE, Brant R, Flemons WW, Davies J, Macarthur C. A decision
dynamic upper airway obstruction in sleep apnea: role of the breathing gas
rule for diagnostic testing in obstructive sleep apnea. Am J Respir Crit Care Med
properties. Am J Respir Crit Care Med 2003;168:659–63.
2003;167:1427–32.
51. Thut DC, Schwartz AR, Roach D, Wise RA, Permutt S, Smith PL. Tracheal and
20. Hand D, Mannila H, Smyth P. Principles of data mining. London: The MIT Press;
neck position influence upper airway airflow dynamics by altering airway
2001.
length. J Appl Physiol 1993;75:2084–90.
*21. Longobardo GS, Gothe B, Goldman MD, Cherniack NS. Sleep apnea considered
*52. Huang Y, White DP, Malhotra A. The impact of anatomic manipulations on
as a control system instability. Respir Physiol 1982;50:311–33.
pharyngeal collapse: results from a computational model of the normal
*22. Khoo MC, Gottschalk A, Pack AI. Sleep-induced periodic breathing and apnea:
human upper airway. Chest 2005;128:1324–30.
a theoretical study. J Appl Physiol 1991;70:2014–24.
53. Huang Y, Malhotra A, White DP. Computational simulation of human upper
23. Batzel JJ, Kappel F, Timischl-Teschl S. A cardiovascular-respiratory control
airway collapse using a pressure-/state-dependent model of genioglossal
system model including state delay with application to congestive heart
muscle contraction under laminar flow conditions. J Appl Physiol
failure in humans. J Math Biol 2005;50:293–335.
2005;99:1138–48.
24. Topor ZL, Vasilakos K, Younes M, Remmers JE. Model based analysis of sleep
54. Xu C, Sin S, McDonough JM, Udupa JK, Guez A, Arens R, et al. Computational
disordered breathing in congestive heart failure. Respir Physiol Neurobiol
fluid dynamics modeling of the upper airway of children with obstructive
2007;155:82–92.
sleep apnea syndrome in steady flow. J Biomech 2006;39:2043–54.
25. Tehrani FT. A model study of periodic breathing, stability of the neonatal
55. Shome B, Wang LP, Santare MH, Prasad AK, Szeri AZ, Roberts D. Modeling of
respiratory system, and causes of sudden infant death syndrome. Med Eng
airflow in the pharynx with application to sleep apnea. J Biomech Eng
Phys 1997;19:547–55.
1998;120:416–22.
56. Huang Y, White DP, Malhotra A. Use of computational modeling to predict
responses to upper airway surgery in obstructive sleep apnea. Laryngoscope
2007;117:648–53.
* The most important references are denoted by an asterisk.
 T. Aittokallio et al. / Sleep Medicine Reviews 13 (2009) 333–343
57. De Backer JW, Vanderveken OM, Vos WG, Devolder A, Verhulst SL,
Verbraecken JA, et al. Functional imaging using computational fluid dynamics
to predict treatment success of mandibular advancement devices in sleep-
disordered breathing. J Biomech 2007;40:3708–14.
58. Casati A, Squicciarini G, Malagutti G, Baciarello M, Putzu M, Fanelli A. Trans-
cutaneous monitoring of partial pressure of carbon dioxide in the elderly
patient: a prospective, clinical comparison with end-tidal monitoring. J Clin
Anesth 2006;18:436–40.
59. Parker SM, Gibson GJ. Evaluation of a transcutaneous carbon dioxide monitor
("TOSCA") in adult patients in routine respiratory practice. Respir Med
2007;101:261–4.
60. Ward S, Chatwin M, Heather S, Simonds AK. Randomised controlled trial of
non-invasive ventilation (NIV) for nocturnal hypoventilation in neuromus-
cular and chest wall disease patients with daytime normocapnia. Thorax
2005;60:1019–24.
61. Tkacova R, Niroumand M, LorenziFilho G, Bradley TD. Overnight shift from
obstructive to central apneas in patients with heart failure: the role of PCO2
and circulatory delay. Circulation 2001;103:238–43.
62. Aittokallio J, Virkki A, Aittokallio T, Saaresranta T, Polo-Kantola P, Polo O.
Non-invasive respiratory monitoring during wakefulness and sleep in pre-
and postmenopausal women. Respir Physiol Neurobiol 2006;150:
66–74.
63. Aittokallio T, Gyllenberg M, Polo O, Toivonen J, Virkki A. Model-based analysis
of mechanisms responsible for sleep-induced carbon dioxide differences. Bull
Math Biol 2006;68:315–41.
64. Zenker S, Rubin J, Clermont G. From inverse problems in mathematical phys-
iology to quantitative differential diagnoses. PLoS Comput Biol 2007;3:e204.
65. Fink M, Batzel JJ, Tran H. A respiratory system model: parameter estimation
and sensitivity analysis. Cardiovasc Eng 2008;8:120–34.
66. Aittokallio T, Gyllenberg M, Polo O, Virkki A. Parameter estimation of a respi-
ratory control model from noninvasive carbon dioxide measurements during
sleep. Math Med Biol 2007;24:225–49.
67. Prabhakar NR, Peng YJ. Peripheral chemoreceptors in health and disease. J Appl
Physiol 2004;96:359–66.
68. Nattie E. Why do we have both peripheral and central chemoreceptors? J Appl
Physiol 2006;100:9–10.
*69. Virkki A, Polo O, Gyllenberg M, Aittokallio T. Can carotid body perfusion act as
a respiratory controller? J Theor Biol 2007;249:737–48.
70. Smith CA, Chenuel BJ, Henderson KS, Dempsey JA. The apneic threshold
during non-REM sleep in dogs: sensitivity of carotid body vs. central
chemoreceptors. J Appl Physiol 2007;103:578–86.
71. Virkki A, Polo O, Saaresranta T, Laapotti-Salo A, Gyllenberg M, Aittokallio T.
Overnight features of transcutaneous carbon dioxide measurement as
predictors of metabolic status. Artif Intell Med 2008;42:55–65.
*72. Aittokallio J, Polo O, Hiissa J, Virkki A, Toikka J, Raitakari O, et al. Overnight
variability in transcutaneous carbon dioxide predicts vascular impairment in
women. Exp Physiol 2008;93:880–91.
343
73. Huang L, Williams JE. Neuromechanical interaction in human snoring and
upper airway obstruction. J Appl Physiol 1999;86:1759–63.
74. Aittokallio T, Gyllenberg M, Polo O. Adjustment of the human respiratory
system to increased upper airway resistance during sleep. Bull Math Biol
2002;64:3–28.
75. Longobardo G, Evangelisti CJ, Cherniack NS. Introduction of respiratory
pattern generators into models of respiratory control. Respir Physiol Neurobiol
2005;148:285–301.
76. Longobardo GS, Evangelisti CJ, Cherniack NS. Analysis of the interplay between
neurochemical control of respiration and upper airway mechanics producing
upper airway obstruction during sleep in humans. Exp Physiol 2008;93:271–87.
77. Belozeroff V, Berry RB, Khoo MC. Model-based assessment of autonomic
control in obstructive sleep apnea syndrome. Sleep 2003;26:65–73.
78. Fowler AC, McGuinness MJ. A delay recruitment model of the cardiovascular
control system. J Math Biol 2005;51:508–26.
79. Younes M, Ostrowski M, Atkar R, Laprairie J, Siemens A, Hanly P. Mechanisms
of breathing instability in patients with obstructive sleep apnea. J Appl Physiol
2007;103:1929–41.
80. Cooper VL, Pearson SB, Bowker CM, Elliott MW, Hainsworth R. Interaction of
chemoreceptor and baroreceptor reflexes by hypoxia and hypercapnia:
a mechanism for promoting hypertension in obstructive sleep apnoea. J
Physiol 2005;568:677–87.
81. Lorenzi-Filho G, Drager LF. Obstructive sleep apnea and atherosclerosis: a new
paradigm. Am J Respir Crit Care Med 2007;175:1219–21.
82. Caples SM, Wolk R, Somers VK. Influence of cardiac function and failure on
sleep-disordered breathing: evidence for a causative role. J Appl Physiol
2005;99:2433–9.
83. Arzt M, Young T, Finn L, Skatrud JB, Bradley TD. Association of sleep-disor-
dered breathing and the occurrence of stroke. Am J Respir Crit Care Med
2005;172:1447–51.
84. West SD, Nicoll DJ, Stradling JR. Prevalence of obstructive sleep apnoea in men
with type 2 diabetes. Thorax 2006;61:945–50.
85. Riha RL, Diefenbach K, Jennum P, McNicholas WT. For the Management
Committee of the COST B26 Action on Sleep Apnoea Syndrome Genetic
aspects of hypertension and metabolic disease in the obstructive sleep
apnoea-hypopnoea syndrome. Sleep Med Rev 2008;12:49–63.
86. Lee JC, Weissglas-Volkov D, Kyttälä M, Dastani Z, Cantor RM, Sobel EM, et al.
WW-domain-containing oxidoreductase is associated with low plasma HDL-C
levels. Am J Hum Genet 2008;83:180–92.
87. Eberhard P. The design, use, and results of transcutaneous carbon dioxide
analysis: current and future directions. Anesth Analg 2007;105:S48–52.
88. Aittokallio T, Malminen JS, Pahikkala T, Polo O, Nevalainen OS.
Inspiratory flow shape clustering: an automated method to monitor upper
airway performance during sleep. Comput Methods Programs Biomed
2007;85:8–18.
89. Popovic RM, White DP. Upper airway muscle activity in normal women:
influence of hormonal status. J Appl Physiol 1998;84:1055–62.