Towards a cyber physical system for personalised and automatic OSA treatment

Abstract

Obstructive sleep apnoea (OSA) is a breathing disorder that takes place in the course of the sleep and is produced by a complete or a partial obstruction of the upper airway that manifests itself as frequent breathing stops and starts during the sleep. The real‐time evaluation of whether or not a patient is undergoing OSA episode is a very important task in medicine in many scenarios, as for example for making instantaneous pressure adjustments that should take place when Automatic Positive Airway Pressure devices are used during the treatment of OSA. Here, the design of a possible Cyber Physical System (CPS) suited to real‐time monitoring of OSA is described, and its software architecture and possible hardware sensing components are detailed. It should be emphasised here that this study does not deal with a full CPS, rather with a software part of it under a set of assumptions on the environment. The study also reports some preliminary experiments about the cognitive and learning capabilities of the designed CPS involving its use on a publicly available sleep apnoea database.

1 Introduction

Currently, healthcare is a very important task for the reduction in the world health costs that are continuously increasing due to the aging population and to the diffusion of chronic diseases [1]. As reported by the US National Commission on Sleep Disorders Research, about 38,000 deaths occurring each year are due to cardiovascular problems somehow linked to sleep apnoea. Obstructive sleep apnoea (OSA) is a breathing disorder that takes place in the course of the sleep and is produced by a complete or a partial obstruction of the upper airway that manifests itself as frequent breathing stops and starts during the sleep [2]. Statistics report that over 18 million Americans have been diagnosed with OSA, and that about 10 million US citizens suffering from OSA do not get proper diagnosis and treatment [3, 4]. In the UK, there are about 25 million or 40% of the population that are affected by OSA, as reported in NHS Choices [5], and 100 million people around the world are believed to suffer from OSA, according to the World Health Organization's report [6] on chronic respiratory diseases.

In general, the task of aiming at the evaluation of the quality of sleep for a subject and at investigating the presence of OSA is highly important in order to ameliorate health conditions for citizens while at the same time reducing both mortality and healthcare-related costs. It should be remarked here that this disease results in problems as hypoxaemia, asphyxia, and awakenings, and often has, on the one hand, immediate consequences as increased heart rate or high blood pressure, and on the other hand may yield long-term symptoms that negatively influence life quality. Among these latter, we can recall here extreme fatigue, poor concentration, a compromised immune system, slower reaction times, and cardio/cerebrovascular problems [7, 8].

Whenever people have been already diagnosed with OSA and have to undergo some particular medical therapies or to receive some specific drugs, the monitoring of their OSA is of high importance. In fact, in these situations, it becomes crucial to perform a suitable analysis of the side effects, as for instance sleep or breathing disturbances. Moreover, such a monitoring activity becomes pivotal in peri-operative situations [9]. Indeed, a very important task for the health of a patient suffering from OSA when she/he is undergoing surgery, as well as after this latter has ended, lies in the real-time evaluation of whether or not the patient is undergoing OSA episodes, and, in the positive case, which are the intensity and the duration of these episodes. Another matter where OSA monitoring in real time is very desirable lies in the evaluation of a suitable feedback for instantaneous pressure adjustments that should take place when Automatic Positive Airway Pressure (APAP) devices are used, instead of the Continuous Positive Airway Pressure (CPAP) device considered the gold standard for OSA treatment. This is also important for other medical cares that are related to the treatment of OSA disease.

The advantages of using APAP devices in the therapies of subjects suffering from OSA are reported in [10]. In this very recent paper (2017), Oldenburg et al. evidenced the effects of the use of APAPs on quality of life, cardiac function, and outcome in OSA patients.

To get the diagnosis about OSA and its monitoring and treatment, nowadays a sleep study, called PolySomnoGraphy (PSG), is used. The problem with PSG is that it has many drawbacks [11]. Firstly, it is quite complicate, because many recordings are to be carried out at the same time, among which electroencephalogram, electrooculogram, electromyogram, oro-nasal airflow, chest wall and abdominal wall movements, oxygen saturation (OS), and ElectroCardioGram (ECG). A negative consequence of the high number of these recordings is that, given the large amount of involved leads, wires, pipes, and so on, a patient, when undergoing PSG, is required to stay in the same position all night long, which is, of course, far from easy and comfortable. A further negative issue related to PSG lies in the fact that these tests, given their complexity, can only be performed in hospital environments, so that patients should stay in a hospital or a clinic centre for one or more nights. This unfamiliar environment can generate stress that may have a negative influence on the OSA pattern itself, as well as on the outcome of the test. As a further important issue, a PSG is quite expensive, for example, its average price in the USA is about 2925 USD [12]. Finally, given the need of undertaking PSG tests in hospitals, this implies that few places exist where to undergo the test, which results in queues and waiting times that can be long.

If on the one hand the diagnosis should compulsorily be done by using the PSG, on the other hand, healthcare researchers and companies are searching for innovative approaches to provide quality healthcare services for the OSA monitoring and treatment [1, 13]. In this way, citizen-centred solutions, and in particular human cyber physical systems (CPSs), are good candidates because they allow practitioners, patients, and their families (if suitable) to set the basis for partnerships so as to make sure that the needs and the preferences of the patients are complied with by procedures and decisions.

Here, we present a design of a logical scheme of a CPS particularised for OSA treatment, an overview of a multi-layered architecture for the CPS for OSA, and a description of possible hardware to use in the system. It should be emphasised here that this paper does not deal with a full CPS, rather with a software part of it under a set of assumptions on the environment. In fact, it should be remarked here that no open-source Automatic Positive Airway Pressure (APAP) devices are available on the market, and only some proprietary, closed ones exist. For these reasons, we have only been able to identify the hardware components related to sensing. When open-source APAPs will be available, the system could be implemented and made available to hospitals for a real-world testing. With respect to the literature, the proposed logical scheme realising the CPS is less invasive and more accurate. In fact, although other papers, as for example, [14] propose very invasive systems for diagnosing and treating patients with OSA, our approach is easier, cheaper, and portable, and it just uses data gathered by a single-channel ECG. The collected data could be processed and analysed on a mobile phone, where heart rate variability (HRV) parameters are computed and simple IF … THEN rules are applied. Subjects may find it comfortable, as there is the need of just one wearable sensor, so they are not forced to stay motionless during the night as it happens for PSG.

The paper is structured in the sections mentioned below. A review on CPSs for Healthcare is reported in Section 2. Our view on how to personalise the treatment of OSA is given in Section 3 in terms of logical scheme, possible hardware, personalisation of cognitive capabilities, and the process of knowledge extraction. Section 4 reports on the preliminary experiments performed by starting from a publicly available database, by suitably modifying it to fulfil our needs, and on the results obtained by running our system for automatic knowledge extraction. In the same section, the numerical results obtained are compared against those provided by several state-of-the-art classifiers. Finally, our Conclusions are given in Section 5 together with the foreseen future work.

2 CPSs for healthcare: state of the art

CPSs can be used in different areas of research. In particular, they are becoming very appealing for applications in the healthcare field, due to the latest progress made in issues as sensors for medicine, Wireless Sensor Networks, and Cloud Computing. These applications involve, among others, the care for patients both in hospitals and at home, as documented also in [15]. As far as health is considered, their presence is notable in distributed robotics (tele-presence, telemedicine), high-precision systems and devices for medicine, e-health and tele-health devices and systems, and in ubiquitous health (u-health) systems. A CPS for e-Health should be able to provide real-time responses to complex and time-changing circumstances while conserving privacy, safety, reliability, and Quality of Service across all levels.

Some very good reviews on the application of CPSs in healthcare are provided in the papers by Liu et al. [16] and by Haque et al. [17]. In [16], an analysis of the state of the art about research being made is presented, together with a discussion about the developments of such systems from different points of view, i.e. those related to models, technologies for data process, and design of the software. Finally, an overview on current problems and main future research is given. An added value of paper [17] is that in it a taxonomy is given to characterise and classify the several modules and procedures specifically needed when healthcare domain is dealt with.

One more paper presenting the state of the art in the use of CPSs for healthcare is that by Skorobogatjko et al. [18], in which also the issue of the compliance with the standards is provided, as well as ideas for the design of future CPSs.

A further paper reviewing the development of CPSs in the industry related to healthcare is that by Ariani and Soegijoko [15]. In it, healthcare tasks are grouped as personal emergency response systems, automated fall detection, telehealthcare, medical and medication intake management, and for each group examples of disease are given together with the corresponding CPSs available.

After this list of general review papers about CPSs for healthcare, in the remainder of this section, we focus our attention on those papers specifically related to the monitoring and the treatment of respiratory diseases and sleep problems, which constitute the target task of the present paper. Oddly enough, in spite of the relevance of this kind of diseases, as of February 2018, very few papers seem to exist in the scientific literature describing the application of CPSs to them. In the following, a description is provided for each of those papers.

A proposal for a system suited to the monitoring of respiration was advanced in Cao et al. [19]. Their approach is based on the use of a micro thermal flow sensor to keep under surveillance respiratory airflow, a triaxial micro-accelerometer to record subject's posture, and a micro-photoelectric sensor suited for the monitoring of patient's blood OS. A PC or a mobile phone allows transmitting the gathered data to a remote server via a bluetooth wireless connection. Then an analysis of the data is performed. To take a decision about the presence of sleep apnoea, two parameters are considered: apnoea–hypopnea index (AHI) and OS. Sleep apnoea is present if the following two conditions are satisfied at the same time: $\text{AHI}>5$ and $\text{OS}<90\%$. Two different thresholds are employed: the forced expiratory volume after 1 s (FEV1) and the forced vital capacity (FVC). A subject is diagnosed by the system as suffering from chronic obstructive pulmonary disease (COPD) or asthma if $\text{FEV}1/\text{FVC}<70\%$.

Another system to diagnose and treat subjects suffering from OSA was proposed by Corral et al. [14]. Their system is composed by a polygraph, a webcam, and a PC. For each subject many data are collected, among which oral/nasal flow, a pressure cannula, chest and abdomen movements, heart rate, body position, snoring, and OS (SO2). Once gathered, the data are then sent to a remote server where they are analysed. Twenty subjects were given medical care following a conventional consultation approach, whereas tele-consultation was used for other twenty. The outcome of this investigation was the discovery of the presence of OSA in 35 subjects out of the 40 participating in the study. Sixteen from among these subjects began a treatment relying on the use of CPAP devices. The correlation coefficients found in the study between data transmitted in real time and data stored in the polygraph were very high. The compliance rates for CPAP treatment with both a conventional clinic and tele-consultation resulted equal to 85 and 75%, respectively.

A further system for the monitoring of sleep pattern was proposed by Ni et al. [20]. The system incorporated a pressure sensor matrix. The corresponding trial took place thanks to the recruitment of ten young subjects. These subjects had to perform a total of three different kinds of typical sleep postures, namely left-lateral sleep, right-lateral sleep, and supine sleep. A total of 150 datasets were gathered during the trial. The next step consisted in gaining knowledge about sleep behaviour in bed for elder subjects. To this aim, an analysis of the data was performed by means of the use of naive Bayes (NB) and random forest, two widely used classification methods. Moreover, the validation phase took place in two different ways, i.e. ten-fold and leave-one-out. The combination of random forest and leave-one-out cross-validation allowed obtaining the highest average accuracy, equal to 87.33%.

In 2017, Ling et al. proposed SleepSense [21], a non-contact and cost-effective sleep monitoring system aimed at recognising in a continuous way the different sleep conditions, comprising issues as on-bed movement, bed exit, and breathing section. The system is made up of three components: a Doppler radar-based sensor, an automated radar demodulation module, and a sleep status recognition framework. A set of features from both time and frequency domains were extracted to carry out the task of recognising sleep. Testing of the system was performed in two different experiments. In a short-term one, the prototype obtained 95.1% of accuracy rate in the classification of different sleep conditions. In the latter, consisting in a 75-min sleep case, the system showed that it could be reliably used in a real-world situation.

In 2017 too, Manfredi described in a general way CPS applications for healthcare [22], among which sleep monitoring. Rather than presenting a specific system, he considered a general scenario of systems for the remote monitoring of patients’ health through the use of wireless heterogeneous networks. He applied the consensus-based algorithm realised through hop-by-hop mechanism in the case of wireless technology. The tuning of the controller gains was performed at each of the nodes. The algorithm was implemented at the network layer. He claimed his methodology can be used also if background traffic arriving from the neighbours is present, thanks to the presence of a feedback term.

3 Our way for making OSA treatment personalised

3.1 Proposed logical scheme of a CPS for OSA treatment

A CPS involves the combination of sensors, actuators, and computation modules to solve issues that lie across the physical and computational areas. In recent years, great interest has grown about CPSs that are regarded as emerging technologies [17]. A CPS combines computation and communication capabilities with the physical world.

A logical scheme of a CPS, with some details and examples referred to the specific OSA treatment problem discussed here, is reported in Fig. 1. Our idea of a CPS for OSA works exactly like this: it is aware of the physical environment by realising an effective feedback loop between sensing and actuation through cognitive and learning capabilities.

[IMAGE OMITTED. SEE PDF]

From the software implementation point of view, the logical scheme in Fig. 1 could be mapped on a multilayer architecture, as for example the one reported in Fig. 2. The software architecture could be logically divided into three layers: the Data Layer, the Communication Layer, and the Decision Layer. Each of them could be composed of software modules (or components), each performing its particular duty and exchanging information among themselves via suitable interfaces.

[IMAGE OMITTED. SEE PDF]

As shown in Fig. 2, the CPS could offer functionalities and services for data processing in the Decision Layer combined with other functionalities and serviced related to sensing and actuating in the Data Layer, and to communication in the Communication Layer to be able to analyse complex situations and to take autonomous decisions.

The Data Layer could be in charge of managing data arriving from sensors, as a wearable ECG sensor, and for sending data to actuators, such as an air pump of an APAP. Additionally, this layer could also supply ways to produce alarm messages and to send them to physicians, or to alert the patient through sound and/or visual warnings.

The Communication Layer could be in charge of guaranteeing highly reliable transmission of the received original data coming from the sensors to the Decision Layer, and meanwhile of ensuring an effective communication between the Decision Layer and the actuators. To ensure syntactic interoperability, the set of monitored data should be represented and structured in conformance to standardised models (HL7 Reference Information Model for clinical data [23]). The semantic interoperability should be supported by adopting international and standardised coding systems, as LOINC (Logical Observation Identifiers Names and Codes [24]) and SNOMED-CT (Systematized Nomenclature of Medicine – Clinical Terms [25]), that are widely used to code clinical data in primary and secondary care settings.

To take any decision, the CPS should be able to process the acquired data in the computing and cognition phase in order to, for example, recognise specific patterns, and detect anomalous and dangerous situations and/or phenomena. The Decision Layer should be responsible for filtering data and performing parameter estimation, i.e. denoising of the ECG signal, detection of the peak of QRS, computing of the values of the HRV features, and it should also contain the CPS ‘intelligent’ core. Here, the data are elaborated in real time on the basis of the rules personalised for the subject, so deciding whether an action should be performed and, if so, its kind (e.g. the activation of an APAP, an alarm or a suggestion to the patient). The choice of dealing with personalised knowledge is proved effective in several papers in the literature, as for example [26, 27].

In the following subsections, all the phases of the logical scheme of CPS for OSA treatment, i.e. the sensing and communication phases, and the computing and cognition phases, will be detailed. Unfortunately, due to the unavailability of an open-source airflow pump, it was not possible to include it into a possible CPS prototype for OSA treatment, so the actuating phase is not detailed. For this reason, it should be emphasised here that this paper has not dealt with a full CPS, rather with a software part of it under a set of assumptions on the environment.

3.2 Sensing and communication phases

To deal with OSA treatment, in our idea of CPS for OSA, we identified one wearable device to be used, i.e. the Zephyr BioHarness BH3, an advanced physiological monitoring device in which a one-lead ECG sensor is embedded.

This sensor, shown in Fig. 3, is very small and provides a medical-grade ECG, as well as heart rate, breathing rate, temperature, and three-axis accelerometer for monitoring subject posture and activity. Data are transmitted by bluetooth and this allows physiological data to be monitored using any suitably configured mobile device with bluetooth technology, such as a laptop, a mobile phone, or a Personal Digital Assistant (PDA).

[IMAGE OMITTED. SEE PDF]

In our laboratory, tests on the acquisition and the exchanging of ECG data from the BioHarness BH3 sensor were done by using two different devices with two different Android versions: a Samsung GT-i9505 (S4) with the Android 5.0.1 version and a HUAWEI GRA-L09 (P8) with the Android 6.1 version.

On these mobile devices, a prototypal java app was installed, where the Data Layer receives the ECG signal and is responsible to represent it in compliance with the HL7 Reference Information Model for clinical data [23]. This mechanism allows a correct syntactic and semantic interoperability of the monitored information with other frameworks, application, and/or services, as for example the services offered by the Decisional Layer.

3.3 Computing phase

This phase is carried out in the Decisional Layer where the acquired signal is processed by executing: (i) the denoising of the signal by using [28, 29]; (ii) the detection of the QRS complex of the ECG signal by using the Pan and Tompkins algorithm [30]; (iii) the calculation of the peak of QRS and hence that of the heart rate; and (iv) the estimation of the HRV parameters.

The extracted HRV values could be used in the cognition phase in personalised IF … THEN rules to determine if some action has to be executed and, in this case, the type of such an action (i.e. alarm generation, a suggestion to the patient, and/or the managing of the automatic pump).

In detail, the complete set of the estimated HRV parameters can be found in Table 1.

Table 1 Complete set of the estimated HRV parameters

Parameter	Symbol
frequency domain
low frequency/high frequency ratio	LF/HF
ultra-low frequency	ULF
very low frequency	VLF
low frequency	LF
high frequency	HF
power of the signal	P
time domain
average value of NN intervals	ANN
standard deviation of the average NN intervals	SDANN
proportion of NN50 divided by total number of NNs (NN50 is the number of pairs of successive NNs that differ by more than 50 ms)	pNN50
the square root of the mean squared difference of successive NNs	rMSSD
non-linear methods
approximate entropy	AE
fractal dimension	FD

3.4 Cognition phase

3.4.1 Methodology to personalise the cognitive capabilities of designed CPS for OSA

For the particular case study, the OSA treatment, we have designed an ad-hoc methodology to extract explicit knowledge composed by a set of IF … THEN rules personalised for each patient to insert in the Decisional Layer for the cognition phase of the CPS for OSA. The methodology, designed to guarantee the best performance in terms of correctness in detecting the OSA episodes, is divided into two phases (Fig. 4).

[IMAGE OMITTED. SEE PDF]

The first one is about the data gathering essential to extract the explicit knowledge for that patient. In this phase, the patient goes to a hospital and receives from the doctor an ECG sensor to wear for one night and a mobile device onto which the acquired data are recorded.

The second phase consists of the patient returning the hardware to the medical personnel, and in data annotation, diagnosis, and offline knowledge extraction. Namely, physicians annotate the acquired ECG recording in connection with the appearance of OSA events. Moreover, the ECG signal is analysed in terms of 1-min time intervals through HRV, so that from the original ECG the values of several typical HRV parameters are computed for each interval, and a database is created. As a result of the annotations, each 1-min time interval of the recording can be classified as either OSA or no-OSA in a supervised way. In this way, personalised knowledge on the subject can be extracted offline in an automatic manner from the database as a set of explicit IF … THEN rules comprising HRV-related parameters. We realise this through our DEREx tool relying on a Differential Evolution (DE) algorithm [31]. An example of this knowledge is reported in Fig. 5.

[IMAGE OMITTED. SEE PDF]

It should be remarked here that the lack of a cooperation agreement with the medical staff from some University Hospital has prevented us from carrying out both of the above-described phases in a real-world environment. Consequently, we have simulated this methodology by using a set of publicly available data.

3.4.2 Process of knowledge extraction

The heuristic algorithm used here to automatically extract knowledge from the available database is DEREx [31], an acronym for DE for Rule Extraction. As the acronym reveals, DEREx is based on DE, a stochastic heuristic optimisation algorithm [32–34] designed to face multivariable optimisation problems and belonging to the family of Evolutionary Algorithms [35]. DEREx is a general-purpose algorithm that can be used without any modification to automatically extract knowledge from any database coming from several application areas; an example is given in [36]. All is needed is to prepare the database to be given as input to DEREx tool. For the experiments in this paper, the database has been described in the following Section 4.1. The knowledge is represented in DEREx under the form of a set of IF–THEN rules.

A thorough description of DE and of DEREx would be too lengthy here, and outside the scope of the paper, hence the interested reader is addressed to [31]. What is important to say here is that, given a database consisting in a number of $N_{\mathrm{A}}$ database attributes and divided into $N_{\mathrm{C}}$ classes, by means of ten-fold supervised classification DEREx is able to extract a set of explicit IF–THEN rules that allow maximising the classification accuracy of the items belonging to the test set. Before algorithm execution, the user can specify the maximum number $N{R}_{max}$ of rules composing the set to be extracted, and can also suitably set the values of two DEREx parameters Rule_active and Literal_Active so as to drive the search, respectively, towards sets with higher or lower numbers of rules, and towards rules having a large or a low number of database attributes.

Each of these rules is composed by an antecedent part in which literals are joined through logical AND connectors. A generic literal has the form:$({\mathrm{var}}_i\text{OP}{c}_1[,c_2])$ where $\text{va}{\mathrm{r}}_i$ is the ith database attribute, OP is a relation operator belonging to the set $\{<,\le ,=,\ge ,>,\text{IN},\text{OUT}\}$, whereas $c_1$ and $c_2$ are two constant values in the admissible range for the attribute ${\mathrm{var}}_i$ contained in the literal under account. The operators IN and OUT mean ‘within the range’ and ‘outside the range’, respectively, hence they need two constant values, $c_1$ and $c_2$, to represent to extremes of the range. The other five operators, instead, need one constant value only, i.e. $c_1$.

Just to show an example, given a database with four attributes denoted with $a_1$, $a_2$, $a_3$, and $a_4$, and divided into two classes represented as $\text{clas}{\mathrm{s}}_1$ and $\text{clas}{\mathrm{s}}_2$, and setting $N{R}_{max}=3$, a set of rules that could be provided by DEREx could have the following form:$\text{IF}(a_1\le 3.07)\text{AND}(a_3\ge 6.82)\text{THEN}\text{class}=\text{clas}{\mathrm{s}}_1$$\text{IF}\left(a_3<6.24\right)\text{AND}\left(a_4\text{IN}\left(12.74,19.05\right)\right)\text{THEN}\text{class}=\text{clas}{\mathrm{s}}_2$ This example shows that some database attributes could be not contained in the ruleset, as it is the case here for $a_2$. This would mean that such attributes are not discriminant for classification purposes. Hence, DEREx also performs the automatic selection of the database attributes that are the most important for the classification.

Moreover, the example also shows that the actual number of rules contained in the best set of rules does not necessarily coincide with $N{R}_{max}$, since it could also be lower than this latter. This holds true especially if the values for Rule_Active are kept relatively low.

An interesting issue is that of the indeterminate cases. This can mean two different things. On the one hand, it can represent database items that during the training phase are taken by more rules so that the item would be assigned to more than one class (yes–yes indeterminate). On the other hand, it can refer to items that are taken by no rule assigning it to a class (no–no indeterminate). In both cases, these items are considered as incorrectly classified in the training phase. In the test phase, nonetheless, DEREx is provided with a suitable recovery mechanism to assign each item to one and only one class. Consequently, no item can exist that is assigned to more classes or to no class during test. This is very important because the quality of each set of rules proposed by DEREx during its evolution is represented by its classification accuracy over the train set, yet the best set of rules found is the one with the highest correct classification rate over the previously unseen items making up the test set.

4 Preliminary experiments about the cognitive capabilities of designed CPS for OSA

4.1 Database

To test the effectiveness of the approach, since to present we do not have real-world data available, we have carried out some preliminary experiments by using a publicly available dataset, i.e. the apnoea-ECG database [37]. This is composed by 70 recordings, out of which 35 have been annotated with the presence of OSA episodes by considering segments lasting 1 min. We have made reference to the 35 annotated recordings only. Among them, 20, denoted as A followed by a number, makes reference to people actually suffering from OSA, 10, represented by C plus a number, refers to either healthy people or subjects showing a very low degree of the disease, and 5 more subjects, identified by B plus a number, are considered borderline. From the raw data related to each such subject, a dataset has been created by us by considering the complete overnight recording, and by taking the computing steps mentioned in @@@Section 3.3.

The complete set of the estimated HRV parameters can be found in Table 1, and we have computed the values for each segment lasting 1 min.

This has allowed us to obtain one database for each of the 35 subjects. Each item in it makes reference to a 1-min segment. More specifically, in each item we have the values computed for those 12 parameters in the minute of reference, followed by the class associated with the specific item. This latter is based on the annotations made by medical experts to that 1-min segment and is equal to 1 for minutes in which no apnoea episode has been detected by the doctors, and is equal to 2 in the opposite case. Since the durations of the recordings are different for the different subjects, so are the corresponding databases. In fact, the shortest of them has 428 items, whereas the longest one 575.

This database is then passed on to DEREx. By doing so, the problem of knowledge extraction becomes the problem of classifying correctly the items, by means of a set of IF–THEN rules each suggesting a class. Hopefully, DEREx will be also able to find out the subset of the HRV parameters that are the most helpful in correctly classifying the items for the subject under the account, i.e. to carry out feature selection from among the 12 parameters. These latter should be present in the rules making up the best set in terms of correct classification accuracy.

4.2 Numerical findings

We have carried out experiments for each subject, i.e. a set of 35 separate experiments over the 35 data sets. As a consequence, we have been able to extract personalised knowledge. As already stated, DEREx works in supervised mode, namely it is run in ten-fold classification. This means that the database we are working on is divided into ten parts, called folds. Then, for the generic i-fold, the remaining nine folds are used for training, whereas the items in this ith fold are kept for the test.

As for the parameters, we have decided to search for sets composed by at most six rules, i.e. $N{R}_{max}=6$. Moreover, for Active_Rule and Active_Literal we have used values equal to 0.80 and 0.80 so that sets with few rules, each of them containing just few parameters, should be preferred. No preliminary tuning phase has been carried out for all the values of all the parameters pertaining to DEREx algorithm, rather we have used this setting because we have noticed that on a wide set of databases coming from different application domains it provides good results. Of course, the search for a better setting could result in better results than those reported in the following.

In order to evaluate the goodness of the rule sets achieved for the different subjects, we have considered accuracy (Acc), sensitivity (Sens), specificity (Spec), and the area under the ROC curve (AUC). The definitions for these parameters are reported in [13].

Table 2 shows the average results obtained over the 20 data sets for the 20 people suffering from OSA. Namely, the average value over the 20 subjects (aver), the related standard deviation (std. dev), the maximum (max) and the minimum (min) values over the train set, the test set, and over the whole database are reported.

Table 2 Results over the subjects suffering from OSA

	Train set	Test set	Whole dataset
Aver	Std. dev	Max	Min	Aver	Std. dev	Max	Min	Aver	Std. dev	Max	Min
accuracy	85.09	10.32	98.86	68.30	93.51	6.12	100.00	81.63	85.92	9.83	98.98	69.62
sensitivity	90.99	13.56	99.76	51.74	94.11	10.47	100.00	64.71	91.32	13.27	99.79	52.91
specificity	65.84	19.08	87.55	30.64	89.74	14.94	100.00	53.33	67.81	18.33	88.89	33.69
AUC	65.87	18.35	87.74	31.46	89.31	14.75	100.00	53.95	67.82	17.71	88.90	34.39

As a general comment, results in terms of accuracy are satisfactory, and those about sensitivity are promising indeed. Specificity, instead, could be improved, and their values yield that also those for the AUC are not as high as hoped. It should be remarked here that sensitivity is related to the false negative errors, the higher the values for Sens the lower the number of those errors. False negatives are more dangerous than false positives, because the former ones imply that an apnoea event has not been taken, whereas the latter says that a normal situation has been erroneously taken as an OSA episode.

4.3 Knowledge extracted

In this subsection, we show the knowledge obtained about some of the subjects suffering from OSA.

The knowledge obtained for subject A18 is represented by the two following rules:$\text{IF}\left(\text{VLF}<500.074\right)\text{AND}(\text{ANN}\ge 63.934)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnoea}$$\text{IF}(\text{ANN}\le 64.930)\text{THEN}\text{class}=\text{apnoea}$ and the numerical results are reported in Table 3.

Table 3 Results over the subject A18 suffering from OSA

	Train set	Test set	Whole dataset
accuracy	95.88	100.00	96.29
sensitivity	99.23	100.00	99.31
specificity	66.67	100.00	68.75
AUC	66.79	100.00	68.85

As it can be seen, the two above rules, although simple and straightforward, allow obtaining explicit knowledge that is very helpful for the correct identification of OSA episodes. They perfectly fit the test set and are of high performance over the whole database related to subject A18. They say that for subject A18 the most important parameter is ANN which is contained in both rules, followed by the parameter VLF.

As a further example, let us take into account the results obtained for subject A3, suffering from OSA as well.

The best set of rules for this subject has turned out to be:$\text{IF}\left(\text{ANN}\text{OUT}\left(82.998,89.815\right)\right)\text{AND}(\text{SDANN}\le 7.341)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnea}$$\text{IF}\left(\text{SDANN}>7.480\right)\text{THEN}\text{class}=\text{apnea}$$\text{IF}\left(\text{ULF}<94885.049\right)\text{AND}\left(\text{pNN}50\text{OUT}\left(5.75825.648\right)\right)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnea}$ and the related numerical results are shown in Table 4.

Table 4 Results over the subject A03 suffering from OSA

	Train set	Test set	Whole dataset
accuracy	89.03	96.08	91.87
sensitivity	91.12	94.74	87.78
specificity	87.25	94.77	87.94
AUC	87.42	89.73	83.30

Unlike the previous case, here three rules are used, and the most discriminating parameter is SDANN, followed by ANN (useful in the previous case too), ULF, and pNN50. The numerical results too are a bit different from the previous case, in that classification over the test set is very good yet not perfect. Moreover, and very importantly, the numerical values for accuracy and sensitivity are lower than for A18, but those for sensitivity and AUC are much higher than before, namely the results for sensitivity and specificity are much more similar each other than they are for subject A18.

As an example of the behaviour when a perfectly healthy subject is monitored, let us consider C08. The best set of rules for him are$\text{IF}(\text{pNN}50\le 8.067)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnoea}$$\text{IF}\left(\mathrm{P}\text{IN}\left(927137.821,955297.313\right)\right)\text{THEN}\text{class}=\text{apnoea}$$\text{IF}\left(\text{SDANN}\text{IN}\left(8.825,42.157\right)\right)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnoea}$$\text{IF}(P\ge 520347.163)\text{AND}(\text{SDANN}\ge 18.292)\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnoea}$$\text{IF}\text{THEN}\text{class}=\text{no}\mathrm{\_}\text{apnoea}$ Here, the last rule, with empty antecedent, works as an ELSE driving towards a diagnosis of the non-OSA episode. The ruleset contains five rules, four suggesting non-apnoea and one apnoea. This latter rule covers no item in the dataset related to the subject C08, as it can be seen in Table 5 where the numerical results are shown.

Table 5 Results over the healthy subject C08

	Train set	Test set	Whole dataset
acc.	100.00	100.00	100.00
sens.	100.00	100.00	100.00
spec.	100.00	100.00	100.00
AUC	100.00	100.00	100.00

The three values for the specificity in table say that there is no 1-min segment that is erroneously taken as an OSA episode, so no false positive is present. The values for accuracy and for sensitivity reveal that no false negative errors take place, i.e. no-OSA episode is considered as a non-OSA segment. This is what should be expected, because subject C08 does not suffer at all from OSA, so his recordings do not contain any 1-min segment labelled by doctors as an OSA episode.

As a general comment, although these results are very preliminary, have been obtained over a publicly available database and without any preliminary tuning phase for the parameters of DEREx algorithm, they are promising. Very importantly, the different rulesets obtained for the different OSA subjects show that a personalised approach is very helpful in efficiently assessing whether or not an apnoea episode has taken place.

4.4 Comparison

To evaluate the effectiveness of DEREx in facing this problem, we have compared its results against those provided by many classification algorithms coming from the literature. Namely, we have used the Waikato Environment for Knowledge Analysis (WEKA) system [38] in which many such algorithms are included. Actually, these algorithms are grouped in sets on the basis of their working principles: Bayesian, function-based, lazy, tree-based, rule-based. We have picked some algorithm from each of these groups. Namely, the NB has been chosen from among the Bayesian methods, whereas the MultiLayer Perceptron Artificial Neural Network (MLP) and the Radial Basis Function Artificial Neural Network (RBF) have been picked as the representatives for the methods based on functions. Going to the group consisting of the lazy algorithms, we have chosen the the IB1, while the J48 and the Naive Bayes Tree (NBTree) have been considered as exemplary for the methods based on trees. Finally, since DEREx is a rule-based method, we have had a particular interest in the group of the algorithms based on rules, from among which three techniques, i.e. the OneR, the Part and the Ripple Down Rule (Ridor) have been chosen.

In the same way as for DE, for all of these algorithms, we have not performed any preliminary tuning phase aiming at finding good values for their parameters. Rather, we have used for these parameters the default values that are found in WEKA.

Moreover, as also these methods must be evaluated in terms of average values over 25 runs as it is for DEREx, for each of them we have varied either the values of the starting seeds or those of some parameters.

The results over the test set are reported in Table 6 for all the algorithms. For each of them, both average final value and standard deviation are shown for the four parameters.

Table 6 Comparison of the results over the test set for the subjects suffering from OSA

		NB	MLP	RBF	IB1	J48	NBTree	OneR	Part	Ridor	DEREx
accuracy	aver	76.37	85.63	81.04	80.99	84.80	84.88	82.11	84.49	84.45	93.51
std. dev	12.97	8.14	10.89	10.46	8.16	8.13	10.78	8.61	8.37	6.12
sensitivity	aver	81.77	87,87	84.37	78.75	87.60	87.49	87.72	88.96	87.66	94.11
std. dev	16.07	10.10	17.84	14.14	9.48	9.54	10.14	10.14	11.96	10.47
specificity	aver	63.15	71.05	61.64	69.50	69.07	69.99	61.65	66.02	65.76	89.74
std. dev	29.13	15.92	23.44	20.18	16.68	16.20	18.26	17.44	19.46	14.94
AUC	aver	81.99	86.46	79.88	74.13	78.59	83.50	74.67	82.43	76.71	89.31
std. dev	10.08	7.29	8.26	9.32	9.90	8.61	9.68	10.03	8.16	14.75

The table reports in bold the best result obtained in terms of highest average value and lowest standard deviation. As it can be observed, DEREx always obtains the highest average value, independently of which specific parameter is considered out of the four investigated here. As concerns the standard deviations, instead, DEREx is the best for accuracy and specificity, whereas J48 is the best for sensitivity and MLP for AUC.

As we are especially interested in algorithms providing sets of rules, we take a closer look at the comparison of DEREx against the other three rule-based algorithms. This shows its superiority, Part being the runner-up, although at a large distance, followed by Ridor and by OneR, respectively.

5 Conclusions and future works

In this paper, we have presented a design of a logical scheme of a CPS particularised for OSA treatment, an overview of a multi-layered architecture for the CPS for OSA, and a description of possible sensing hardware to use in the system. Due to open-source APAPs unavailability, it has not been possible to realise a real prototype implementing completely the logical scheme of a CPS for OSA described in this paper. For this reason, it should be emphasised here that this paper has not dealt with a full CPS, rather with a software part of it under a set of assumptions on the environment.

Some preliminary experiments about the cognitive capabilities of our designed CPS for OSA have been carried out on a publicly available sleep apnoea database, leading to promising results.

Future work will investigate two issues. The former makes reference to an implementation of the system by simulating the APAP at a software level. The latter, instead, consists in a cooperation agreement with the Faculty of Medicine of the University of Naples ‘Federico II’, aiming at starting a trial study in which a set of subjects suffering from OSA will participate. The result of this latter activity will be a real testing of the CPS proposed here.

Finally, due to the fact that the data annotation process required in the second phase of our methodology is very laborious, we are also investigating the possibility to apply a deep learning approach to help the medical personnel in this phase.

References

Nikita, K.S., Lin, J.C., Fotiadis, D.I. et al.: ‘Special issue on mobile and wireless technologies for healthcare delivery’, IEEE Trans. Biomed. Eng., 2012, 59, (11), pp. 3083–3089

Park, J.G., Ramar, K., Olson, E.J.: ‘Updates on definition, consequences, and management of obstructive sleep apnea’, Mayo Clin. Proc., 2011, 86, pp. 549–555

Lee, W., Nagubadi, S., Kryger, M.H. et al.: ‘Epidemiology of obstructive sleep apnea: a population‐based perspective’, Expert Rev. Respir. Med., 2008, 2, (3), pp. 349–364

Young, T., Peppard, P.E., Gottlieb, D.J.: ‘Epidemiology of obstructive sleep apnea: a population health perspective’, Am. J. Respir. Crit. Care Med., 2002, 165, (9), pp. 1217–1239

NHS Choices: ‘Key obstructive sleep apnoea statistics’. Available at http://www.NHS.uk

World Health Organisatio: ‘Chronic respiratory diseases’. Available at http://www.who.int/gard/publications/chronic_respiratory_diseases.pdf

Koskenvuo, M., Kaprio, J., Telakivi, T. et al.: ‘Snoring as a risk factor for ischaemic heart disease and stroke in men’, Br. Med. J., Clin. Res. Ed., 1987, 294, (6563), pp. 16–19

Melillo, P., Castaldo, R., Sannino, G. et al.: ‘Wearable technology and ECG processing for fall risk assessment, prevention and detection’. 2015 37th Annual Int. Conf. of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 2015, pp. 7740–7743

Den Herder, C., Schmeck, J., Appelboom, D.J. et al.: ‘Risks of general anaesthesia in people with obstructive sleep apnoea’, Br. Med. J., 2004, 329, (7472), p. 955

Oldenburg, O., Fox, H., Wellmann, B. et al.: ‘Automatic positive airway pressure for treatment of obstructive sleep apnea in heart failure’, Somnologie, 2017, 21, pp. 273–280

Ryan, P., Hilton, M., Boldy, D. et al.: ‘Validation of British Thoracic Society guidelines for the diagnosis of the sleep apnoea/hypopnoea syndrome: can polysomnography be avoided?’, Thorax, 1995, 50, (9), pp. 972–975

New Choice Health: ‘Sleep study (polysomnography) cost and procedure information’. Available at http://www.newchoicehealth.com/Directory/Procedure/51/Sleep%20Study%20%28Polysomnography%29

Sannino, G., De Falco, I., De Pietro, G.: ‘Monitoring obstructive sleep apnea by means of a real‐time mobile system based on the automatic extraction of sets of rules through differential evolution’, J. Biomed. Inf., 2014, 49, pp. 84–100

Coma‐del Corral, M.J., Alonso‐Álvarez, M.L., Allende, M. et al.: ‘Reliability of telemedicine in the diagnosis and treatment of sleep apnea syndrome’, Telemed. e‐Health, 2013, 19, (1), pp. 7–12

Ariani, A., Soegijoko, S.: ‘The development of cyber‐physical system in health care industry’, Computational intelligence for decision support in cyber‐physical systems’ (Springer, 2014), pp. 107–148

Liu, Y., Peng, Y., Wang, B. et al.: ‘Review on cyber‐physical systems’, IEEE/CAA J. Autom. Sin., 2017, 4, (1), pp. 27–40

Haque, S.A., Aziz, S.M., Rahman, M.: ‘Review of cyber‐physical system in healthcare’, Int. J. Distrib. Sens. Netw., 2014, 10, (4), p. 217415

Alyona, S., Andrejs, R., Nadezhda, K.: ‘State of the art in the healthcare cyberphysical systems’, Inf. Technol. Manag. Sci., 2014, 17, (1), pp. 126–131

Cao, Z., Zhu, R., Que, R.‐Y.: ‘A wireless portable system with microsensors for monitoring respiratory diseases’, IEEE Trans. Biomed. Eng., 2012, 59, (11), pp. 3110–3116

Ni, H., Abdulrazak, B., Zhang, D. et al.: ‘Unobtrusive sleep posture detection for elder‐care in smart home’. Aging Friendly Technology for Health and Independence, Seoul, South Korea, 2010, pp. 67–75

Lin, F., Zhuang, Y., Song, C. et al.: ‘SleepSense: a noncontact and cost‐effective sleep monitoring system’, IEEE Trans. Biomed. Circuits Syst., 2017, 11, (1), pp. 189–202

Manfredi, S.: ‘Application to cyber‐physical systems’ (Springer, 2017), pp. 99–134

‘Health Level 7’, Available at http://www.hl7.org

McDonald, C., Huff, S., Suico, J. et al.: ‘Logical observation identifiers names and codes (loinc®) users’ guide’ (Regenstrief Institute, Indianapolis, 2004)

C. Snomed: ‘Systematized nomenclature of medicine‐clinical terms’ (International Health Terminology Standards Development Organisation, 2011)

Forkan, A., Khalil, I., Ibaida, A. et al.: ‘BDCam: big data for context‐aware monitoring‐a personalized knowledge discovery framework for assisted healthcare’, IEEE Trans. Cloud Comput., 2015

Forkan, A.R.M., Khalil, I.: ‘A clinical decision‐making mechanism for context‐aware and patient‐specific remote monitoring systems using the correlations of multiple vital signs’, Comput. Methods Programs Biomed., 2017, 139, pp. 1–16

Cuomo, S., De Pietro, G., Farina, R. et al.: ‘A revised scheme for real time ECG signal denoising based on recursive filtering’, Biomed. Signal Proc. Control, 2016, 27, pp. 134–144

Cuomo, S., Galletti, A., Farina, R. et al.: ‘A framework for ECG denoising for mobile devices’. Proc. of the 8th ACM Int. Conf. on PErvasive Technologies Related to Assistive Environments, Corfu, Greece, 2015, p. 48

Pan, J., Tompkins, W.J.: ‘A real‐time QRS detection algorithm’, IEEE Trans. Biomed. Eng., 1985, (3), pp. 230–236

De Falco, I.: ‘Differential evolution for automatic rule extraction from medical databases’, Appl. Soft Comput., 2013, 13, (2), pp. 1265–1283

Price, K., Storn, R.: ‘Differential evolution: numerical optimization made easy’, Dr. Dobb's J., 1997, 220, pp. 18–24

Storn, R., Price, K.: ‘Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces’, J. Glob. Optim., 1997, 11, (4), pp. 341–359

Price, K., Storn, R.M., Lampinen, J.A.: ‘Differential evolution: a practical approach to global optimization’ (Springer Science & Business Media, 2006)

Baeck, T., Fogel, D., Michalewicz, Z.: ‘Handbook of evolutionary computation’ (Taylor & Francis, 1997)

Sannino, G., De Falco, I., De Pietro, G.: ‘A supervised approach to automatically extract a set of rules to support fall detection in an mHealth system’, Appl. Soft Comput., 2015, 34, pp. 205–216

Penzel, T.: ‘The apnea‐ECG database’. Computers in Cardiology, Cambridge, MA, USA, 2000, vol. 27, pp. 255–258

Frank, E., Hall, M.A., Witte, I.H.: ‘Data mining: practical machine learning tools and techniques’ (2016, 4th edn.)