Aug 16, 2013 - oximetry [8-9], EEG [10], ECG [4, 11-13], thoracic and abdominal ..... developed by using Eclipse IDE and the Android SDK. We conducted ...
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a real-time mobile monitoring system Giovanna Sannino1,2, Ivanoe De Falco1 and Giuseppe De Pietro1 1 Institute of High Performance Computing and Networking, CNR 2 University of Naples Parthenope, Department of Technology {giovanna.sannino, ivanoe.defalco, giuseppe.depietro}@na.icar.cnr.it Abstract Real-time Obstructive Sleep Apnea (OSA) detection and monitoring are important for the society in terms of improvement in citizens' health conditions and of reduction in mortality and healthcare costs. This paper proposes an easy, cheap, and portable approach for monitoring patients with OSA. It is based on singlechannel ECG data, and on the automatic offline extraction, from a database containing ECG information about the monitored patient, of explicit knowledge under the form of a set of IF…THEN rules containing typical parameters derived from HRV analysis. This set of rules can be exploited in our real-time mobile monitoring system: ECG data is gathered by a wearable sensor and sent to a mobile device, where it is processed in real time, HRV-related parameters are computed from it, and, if their values activate some of the rules describing occurrence of OSA, an alarm is automatically produced. The approach has been tested on a well-known literature database of OSA patients. Rules are obtained which are specific for each patient. Numerical results have shown the effectiveness of the approach, and the achieved sets of rules evidence its userfriendliness. Furthermore, the method has been compared against other well-known classifiers. Keywords: Obstructive Sleep Apnea; Knowledge extraction; IF…THEN rules; Real-time monitoring system; Wearable sensors
1. Introduction Obstructive sleep apnea (OSA) [1] is a breathing disorder taking place during sleep, caused by the partial or complete constriction of the patient’s upper airway. This work has been partly supported by the project “Sistema avanzato per l’interpretazione e la condivisione della conoscenza in ambito sanitario A.S.K. – Health” (PON01_00850).
It causes hypoxemia, asphyxia and awakenings, increased heart rate and high blood pressure, and has long-term consequences as extreme fatigue, poor concentration, acompromised immune system, slower reaction times and cardio/cerebrovascular problems [2]. About 4% of the general population suffers from it to some extent, and it is estimated that less than 25% of OSA sufferers know they actually are [3]. These undiagnosed sufferers are believed to cause in the USA 70 billion dollars loss, 11.1 billions in damages, and 980 deaths each year [4]. Real-time OSA detection and monitoring are important in terms of improvement in citizens' health conditions and of reduction in mortality and healthcare costs. They are critical indeed in several situations as checking people suffering from OSA during activities which could worsen gravity of the disorder, e.g. during the intake of some drugs and in perioperative conditions where anesthesia is required [5]. Also, feedback for instantaneous and continuous CPAP pressure adjustments is another issue. Currently, to diagnose OSA patients undergo a sleep study, known as polysomnography (PSG). This examination has several drawbacks [6]: is quite complicate since needs many sensors, restrains the patients to a hospital environment for at least one night, sometimes two, may produce stresses that may influence the OSA pattern itself, is costly (national average price in the USA is about 2,600 USD [7]), and there are few places to undergo PSG, meaning long waiting times. For these reasons, simpler, cheaper, and home-based methods are highly welcome, and are actually being developed to diagnose or monitor OSA. The recent literature shows several papers dealing with detection and monitoring of OSA without making use of PSG. Several approaches have been proposed, based on the evaluation of different (sets of) vital parameters: oximetry [8-9], EEG [10], ECG [4, 11-13], thoracic and abdominal signals [14], or combinations of them [15-16]. This paper proposes an easy, cheap, and portable approach for monitoring patients with OSA, which gathers single-channel ECG data only. It is based on the automatic extraction, from a database containing
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
information about the monitored patient, of explicit knowledge under the form of a set of IF…THEN rules containing typical parameters derived from HRV analysis. This set of rules, generated offline, can be exploited in the real-time mobile monitoring system [17], developed at iHealthLab (ICAR-CNR)[18], which can monitor a large number of vital parameters for a patient by means of wearable sensors. Namely, for OSA ECG data is gathered by a wearable sensor and sent to a mobile device, where it is processed in real time each minute, and HRV-related parameters are computed from it. If their values activate some of the rules describing occurrence of OSA, the system can immediately wake up the patient and/or send an alert to medical personnel. The main features of our proposed system are: clear explanation of motivations for detection of each OSA episode due to the set of rules, ease of use (it just needs one ECG lead), real-time detection of occurrence of an OSA episode and consequent real-time action, execution of the whole processing on a mobile, personalized healthcare since different sets of rules will be found for different patients. Although several other proposals also use mobile devices for OSA, each of them lacks some of the features of our system. For example, [3, 19] use the mobile just to send data to a hospital server where analysis is performed. The systems proposed in [9, 20, 21], instead, lack an Action Layer allowing performing immediate actions. Moreover, they all detect apneas based on general knowledge rather than on rules specific for each patient. Our approach is tested on a database of OSA patients, and is compared against other well-known classifiers.
• Low frequency: LF • High frequency: HF • Power of the signal: P The parameters in the time domain are: • Average value of NN intervals: ANN • The 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 Those related to non-linear methods are: • approximate entropy: AE • fractal dimension: FD Each instance in the database related to that patient is constituted by those 12 values, together with the class of the instance as known from the annotations related to that recording. These latter will be represented by a 1 for a non-apnea minute and by a 2 for an apnea minute. Since apnea-ECG database recordings have different lengths, due to different duration of the sleep for the different patients, so will be the thereby extracted databases. Their length varies from 428 to 576 items. It should be noted that apnea periods, as well as nonapnea ones, can last even hours, therefore each database presents long sequences of consecutive items making reference to a same class. To avoid that this can negatively influence learning and generalization, the items in each of the 35 databases are shuffled. As an example, Figure 1 shows the final database for patient #4.
2. The Database Starting from the apnea-ECG database [22] we have to create a specific database for each patient. That database consists of 70 recordings, 35 of which with annotations about apneas for each 1-minute segment. Therefore we make reference to those 35 recordings only. Among them twenty (a1-a20) are known to be related to people definitely suffering from OSA, (b1-b5) five are borderline, and ten (c1-c10) are people with no OSA at all or very low values. For each of these we create a database. Namely, we take the whole patient recording over the whole night, and for each 1-minute segment we compute the values of a set of twelve typical parameters related to HRV by using GHRV software 1.2 [23]. Some of them are in the frequency domain, other in the time domain, and other make reference to non-linear methods. The parameters in the frequency domain are: • Low frequency/high frequency ratio: LF/HF • Ultra-low frequency: ULF • Very low frequency :VLF
Figure 1: the final database for patient #4.
3. The Experiments 3.1. Knowledge Extraction
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
To automatically extract from each of the 35 obtained databases a set of explicit IF-THEN rules for OSA detection, our DEREx tool [24] has been used. It is based on Differential Evolution [25], a fast and effective evolutionary algorithm specifically devised to tackle realvalued multivariable optimization problems. DEREx uses a 10-fold cross-validation to select the set of rules that maximize the correct classification rate over unseen examples. It is impossible here to describe with sufficient details the way the tool works, the interested reader can refer to [24]. We run DEREX 25 times because it is not deterministic, rather its execution depends on an initial random seed. We set DEREx so that over each database it should search for groups of at most six rules. Table I reports the overall discriminating ability of the system over the 35 cases in terms of accuracy, sensitivity, and specificity. Each value represents the average of the parameter over the 35 patient databases. For example, accuracy over the testing set means that the 35 best sets of rules found allow classificating correctly 92.26% of all the instances over the 35 databases. Table I: Discriminating Ability of the Best Sets of Rules Found.
Specificity
Training Set
Testing Set
Whole Database
83.78
100.00
84.62
Instead, the best set of rules for patient a19 is: IF (LF ≥ 3052.02) AND ((mean_value_HR < OR (mean_value_HR > 85.83)) THEN no_apnea IF (standard_deviation_HR ≥ 2.74) THEN apnea IF (HF < 8334.66) AND (standard_deviation_HR ≤ 2.80) THEN no_apnea and its discriminating ability is shown in Table III. A comparison between these two sets or rules confirms our hypothesis that the most relevant parameters for OSA discrimination depend on the specific patient being monitored: a set of rules which is very good for one could be unsuitable for other patients. Table III: Discriminating Ability of the Best Set of Rules Found for Patient a19. Training Set
Testing Set
Whole Database
Accuracy
88.69
98.00
89.62
Training Set
Testing Set
Whole Database
Sensitivity
94.44
95.83
94.61
Accuracy
85.04
92.26
85.76
Specificity
84.88
100.00
86.20
Sensitivity
65.43
82.14
65.82
Specificity
65.43
79.40
66.03
A very interesting remark here is that in the testing set values for sensitivity are quite higher than those for specificity. This is a positive feature of our system, as in monitoring apneas it is of course better to avoid missing OSA segments than to generate unnecessary alarms. At the end of this experimental phase the system provides users with 35 sets of rules. As an example, that for patient a04, is the following: IF ((HF < 3321.48) OR (HF > 5685.89) AND (mean_value_HR > 71.44) THEN no_apnea IF (mean_value_HR < 73.32) THEN apnea
On the one hand, some parameters such as the mean value of HR are often present in the 35 sets of rules, yet, on the other hand, each patient requires some very specific parameters. If, instead, the 35 patient databases are gathered in one global database, the best set of rules found on it is: IF ((power < 114958.95) OR (power > 2360972.72)) AND (pnn50 < 5.13) THEN no_apnea IF (fractal_dimension ≥ 1.62) THEN apnea IF (ULF ≥ 101363.00) THEN apnea which achieves the performance shown in Table IV.
This set perfectly takes all cases over testing set, as it is shown in Table II. Furthermore, also performance over the training set and over the whole database are really satisfactory. Table II: Discriminating Ability of the Best Set of Rules Found for Patient a04. Training Set
Testing Set
Whole Database
Accuracy
98.42
100.00
98.57
Sensitivity
99.75
100.00
99.78
Table IV: Discriminating Ability of the Best Set of Rules Found for the global database. Training Set
Testing Set
Whole Database
Accuracy
70.51
72.00
70.66
Sensitivity
91.33
91.20
91.31
Specificity
33.45
35.49
33.65
The parameters involved now are not those best suited for the patients shown above. Also, global performance is worse, especially in terms of specificity. This is a very important conclusion: a general database of people
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
suffering from OSA, as typical of other approaches, is not well suited for taking care of a specific patient.
3.2 Comparison With Other Classifiers To evaluate the goodness of the results in terms of accuracy, sensitivity, and specificity, other four wellknown classifiers have been used for comparison. Therefore, the Waikato Environment for Knowledge Analysis (WEKA) system release 3.4 [26] has been used: it contains a large number of such techniques, divided into groups depending on the their basic working principles. From each such group a representative has been chosen. Among the Bayesian, the Bayes Net (BN) [27] has been considered, and, among the function-based, the MultiLayer Perceptron Artificial Neural Network (MLP) [28] has been selected. The KStar (KS) [29] has been considered as a representative of the lazy methods, while, among the tree-based, the J48 [30] has been chosen. Similarly to what was done for DEREx, no preliminary parameter tuning has been carried out for all of the above techniques as well, so the parameter values used for each such method are those set as default in WEKA. Furthermore, also for them 25 runs have been performed. Also for all of these tools 10-fold cross-validation has been carried out in each run. Table V shows the results in terms of the average accuracy A over the 35 databases, the standard deviation StD, and the maximum and the minimum values. Table V: Results Achieved by the Tested Classifiers. BN
MLP
KS
J48
DEREx
A
85.53
87.59
82.32
87.20
92.26
StD
9.40
8.62
11.37
8.83
7.25
Max
99.79
99.79
99.79
99.79
100.00
Min
65.67
70.82
65.82
72.10
73.91
For each parameter in Table V, the best value obtained by all algorithms is reported in bold. The average percentage of correct classification provided by DEREX over the 35 databases is the highest, much higher than the other ones by about 5%. Moreover, the standard deviation is the lowest, meaning that the algorithm is quite insensitive to the different initial random seeds. DEREx also achieves the highest maximum value, and is the only one to correctly classify 100% over five databases, i.e. those related to patients a01, a04, c05, c06, and c07. It also obtains the highest among the minimum values. Table VI contains the comparison of the algorithms in terms of average values of sensitivity over the 35 databases, and of the related standard deviation as well. Table Vi: Average Results for Sensitivity.
BN
MLP
KS
J48
DEREx
A
70.01
69.61
61.60
68.13
82.14
StD
35.58
35.60
34.61
36.23
31.75
The best values are reported in bold. DEREx achieves by far the highest value of sensitivity, so it has a much lower number of false negatives, i.e. apneas that are erroneously seen as non-apneas. This is very important for a system of this kind, as it is better able to generate alarms only when necessary. Table VII shows the same data for the specificity. Table Vii: Average Results for Specificity. BN
MLP
KS
J48
DEREx
A
76.39
76.29
69.18
75.90
79.40
StD
17.95
20.23
25.00
20.22
31.26
Here too DEREx provides the highest value. This means that DEREx has a lower number of false positives, resulting in a lower number of unnecessary alarms. This comparison makes us confident that the set of rules shown is capable of discriminating apnea episodes much more accurately than these other artificial intelligence tools can. Moreover, all these other classifiers do not extract rules, so they would be useless for the creation of the knowledge base needed by our system. We believe this user-friendliness of our system is very helpful to doctors.
4. Mobile Real-Time Monitoring System This method, shown effective on the OSA ECG database used, should now be tested in a real-world case. Therefore, it has been included in our mobile health monitoring system developed at iHealthLab, and experiments will be carried out in next weeks with volunteers suffering from OSA. Each of them will be asked to wear sensors of the system for a few hours, and the resulting ECG trace will be annotated by a doctor. So a set of rules can be achieved for that patient. These rules will be used to monitor the patient in the following nights. Software and hardware details about our mobile health monitoring system are given below.
4.1. Software Details The realized system relies on a multi-layer architecture [17] designed to provide the system with flexibility capacity through which algorithms and sensors can be adapted to various applications and new sensors can be easily added. As shown in figure 2, each layer is independent of the upper one and of the lower one. The Data Layer provides user interfaces and mechanisms to manage sensors data and patient
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
information that will be processed by the Decisional Layer. This layer has to collect data from BioHarness device, which will be described in next subsection, and is responsible for computing complex parameters such as the peak of QRS to estimate Heart Rate and its variability represented by the parameters discussed in section 2.
diagnosing with a complete representation of the electrical activity of the heart, including the heart rate (HR), which is interpreted as the R-to-R Interval, as shown in figure 3.
Figure 3: ECG Signal
Figure 2: Multi-layer Architecture of the Realized Monitoring System.
This layer has also the task to collect data from an accelerometer embedded into the BioHarness device so as to recognize the patient's posture and activity. Finally, it manages the saving of data in a specific data format, the European Data Format (EDF). The choice of EDF format has been made because it is the de-facto standard for EEG and PSG recordings in commercial equipment and multicenter research projects. It is a simple and flexible format for exchange and storage of multichannel biological and physical signals. In addition many freeware EDF viewers and EDF analyzers can be found. The Decisional Layer is the intelligent core of the system and includes the rule engine described in [31]. Data coming from the Data Layer are elaborated here. Thanks to the presence of a set of rules representing the formalization of experts’ knowledge about anomalies, this layer by means of a Decision Support System recognizes critical situations and determines the most suitable actions to be performed by the Action Layer. Finally, the Action Layer executes the actions inferred by the Decisional Layer by implementing mechanisms which produce reactions like the generation of alarms.
The device used in our system is the BioHarness™ 3, an advanced physiological monitoring device that uses Bluetooth technology to transmit data. It is small and provides a medical-grade ECG, as well as heart rate, breathing rate, and 3-axis accelerometery. The monitor could be used with the BioHarness™ strap, a lightweight elasticized component which incorporates Zephyr Smart Fabric ECG and Breathing Rate sensors. Data is transmitted by Bluetooth, and this allows physiological data to be monitored using any mobile device with bluetooth technology, as a laptop, phone or PDA.
4.3. Implementation Details All layers are implemented for resource-limited mobile devices, such as PDA and smart phone, using the java programming language, but the system could be used also to build desktop applications, except for the user interfaces. In particular, for this use case the system was developed by using Eclipse IDE and the Android SDK. We conducted some preliminary tests using an Android-based tablet, an ASUS Eee Pad Transformer Prime TF201 model, as shown in figure 4.
4.2. Hardware Details Electrocardiography –ECG– is a non-invasive technique based on interpretation of the electrical activity of the heart over time. These signals or activities are recorded by using skin electrodes. In our system, the ECG sensor is composed of three electrodes. The ECG module gets small voltages around about 1mV that normally appear on the skin that may help to monitor the cardiac activity in human body. The signals from the different leads provide help to cardiologist for
Figure 4: Mobile Monitoring System Interface
5. Conclusions
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
The real-time Obstructive Sleep Apnea (OSA) detection is critical for several reasons. This paper has proposed an approach, embedded in a real-time mobile monitoring system, for monitoring patients with OSA. It is quite simple, since gathers single-channel ECG data only, and is based on the automatic extraction of explicit knowledge from a database containing information about the monitored patient as a set of IF…THEN rules containing typical HRV parameters. It has been tested on a literature database of OSA patients. Numerical results have shown the effectiveness of the approach, and the achieved sets of rules evidence its user-friendliness, which is in our opinion a very helpful feature to doctors. An important conclusion is that a general database of people suffering from OSA is not well suited for taking care of a specific patient. Future works involve the testing of this system in a real-world situation by means of a group of volunteers suffering from OSA.
Acknowledgements We thank Dr. M. Esposito and Dr. A. Minutolo for allowing us to use their DSS in our system [31].
References [1] W. T. McNicholas and P. Levy, “Sleep-related breathing disorders: Definitions and measurements,” Eur. Respir. J., vol. 15, no. 6, pp. 988–989, 2000. [2] M. Koskenvuo, J. Kaprio, T. Telakivi, M. Partinen, K. Heikkila, and S. Sarna. Snoring as a risk factor for ischemic heart disease and stroke in men. Br. Med J., 294:9–16, 1987. [3] S. Alqassim, M. Ganesh, S. Khoja, M. Zaidi, F. Aloul, A. Sagahyroon, “Sleep apnea monitoring using mobile phones”, proceedings of e-Health Networking, Applications and Services (Healtcom) 2012 Conference, pp. 443-446, IEEE Press, 2012. [4] L. Almazaydeh, K. Elleithy, M. Faezipour, “Detection of Obstructive Sleep Apnea through ECG Signal Features”, IEEE International Conference on Electro/Information Technology, pp. 1-6, 2012. [5] C. den Herder, J. Schmech, D. Appelboom, and N. de Vries, “Risks of general anaesthesia in people with obstructive sleep Apnea,” Br. Med. J., (BMJ), vol. 329, pp. 955–959, 2004, 2011. [6] P. Ryan, M. Hilton, D. Boldy, A. Evans, S. Bradbury, S. Sapiano, K. Prowse, R. Cayton, “Validation of British Thoracic Society guidelines for the diagnosis of the sleep apnoea/hypopnoea syndrome: can polysomnography be avoided?” Thorax, vol. 50, pp. 972-975, 1995.
http://www.newchoicehealth.com/Directory/Procedure/51/Sleep %20Study%20%28Polysomnography%29 . [8] L. Almazaydeh, M. Faezipour, K. Elleithy,” Neural Network System for Detection of Obstructive Sleep Apnea Through SpO2 Signal Features”, International Journal of Advanced Computer Science and Applications, Vol. 3, No.5, 2012. [9] N. Oliver, F. Flores-Mangas, “HealthGear; Automatic Sleep Apnea detection and Monitoring with a Mobile Phone”, Journal of Communications, vol. 2, no. 2,pp. 1-9, Academy Publisher, 2007. [10] R. Lin, R. Lee, C. Tseng, H. Zhou, C. Chao, J. Jiang,“A New Approach for Identifying Sleep Apnea Syndrome Using Wavelet Transform and Neural Networks,” Biomedical Engineering: Applications, Basis & Communications, vol. 18, no. 3, pp. 138-143, 2006. [11] A. F. Quiceno-Manrique, J.B. Alonso-Hernandez, C. M. Travieso-Gonzalez, M. A. Ferrer-Ballester and G. CastellanosDominguez, “Detection of obstructive sleep apnea in ECG recordings using time-frequency distributions and dynamic features”, 31st International Conference of the IEEE EMBS, 5559-5562, 2009. [12] B. Yilmaz, M. Asyali, E. Arikan, S. Yektin and F. Ozgen, “Sleep Stage and Obstructive Apneaic Epoch Classification Using Single-Lead ECG,” Biomedical Engineering Online, vol. 9, 2010. [13] M. Bsoul, H. Minn, L. Tamil, “Apnea MedAssist: real-time sleep apnea monitor using single-lead ECG”, IEEE Transactions on Information Technology in Biomedicine, vol. 15, no. 3, pp. 416-427, 2011. [14] A. Ng, J. Chung, M. Gohel, W. Yu, K. Fan and T. Wong, “Evaluation of the Performance of Using Mean Absolute Amplitude Analysis of Thoracic and Abdominal Signals for Immediate Indication of Sleep Apnoea Events,” Journal of Clinical Nursing, vol. 17, no. 17, pp. 2360-2366, Sep. 2008. [15] D. Alvarez, R. Hornero, J. Marcos, F. Campo and M. Lopez, “Spectral Analysis of Electroencephalogram and Oximetric Signals in Obstructive Sleep Apnea Diagnosis,” in Proceedings of the 31st IEEE International Conference on Engineering in Medicine and Biology Society (EMBS 2009), pp. 400-403, Sep. 2009. [16] B. Xie, H. Minn, “Real Time Sleep Apnea Detection by Classifier Combination,” in IEEE Transactions on Information Technology in Biomedicine, vol. 15 no. 3 pp. 469-477, 2012. [17] Sannino, G.; De Pietro, G., “An Intelligent Mobile System For Cardiac Monitoring”, In Proc. of IEEE Healthcom’10, Lyon, France, pp: 52-57, April 2010. [18] iHealthLab - http://ihealthlab.icar.cnr.it/
[19] R. Ishida, Y. Yonezawa, H. Maki, H. Ogawa, I. Ninomiya, K. Sada, S. Hamada, A. W. Hahn, W. M. Caldwell, “A Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479 [7]
New
Choice
Health:
medical
cost
comparison
2013 IEEE 14th International Conference on Information Reuse and Integration (IRI)
wearable, mobile phone-based respiration monitoring system for sleep apnea syndrome detection”, Biomedical Sciences Intrumentation, vol. 41 pp 289-293, Instrument Society of America, 2005. [20] D. Patil, V M Wadhai, S. Gujar, K. Surana, P. Devkate, S. Waghmare, “APNEA Detection on Smart Phone”, International Journal of Computer Applications 59(7):15-19, Foundation of Computer Science, New, USA, 2012. [21] M. Rofouei, M. Sinclair, R. Bittner, T. Blank, N. Saw, G. DeJean, J. Heffron, “A Non-invasive Wearable Neck-Cuff System for Real-Time Sleep Monitoring”, Proceedings of the 2011 International Conference on Body Sensor Networks, pp. 156-161, IEEE Computer Society, Washington, DC, USA, 2011. [22] T. Penzel, “The Apnea–ECG database,” Comput. Cardiol., vol. 27, pp. 255–258, 2000. [23] http://milegroup.github.io/ghrv/index.html [24] I. De Falco, “Differential Evolution for automatic rule extraction from medical databases”, Applied Soft Computing, vol. 13, pp. 1265-1283, Elsevier, 2013. [25] K. Price, R. Storn, J. Lampinen, “Differential Evolution: A Practical Approach to Global Optimization”, Springer, 2005. [26] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I.H. Witten, “The WEKA data mining software: an update”, SIGKDD Explorations vol. 11 n. 1, pp. 10–18, 2009. [27] F. Jensen, An Introduction to Bayesian Networks, UCL Press/Springer-Verlag, New York, NY, 1996. [28] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representation by back-propagation errors, Nature 323 (1986) 533–536. [29] J.G. Cleary, L.E. Trigg, “K*: an instance-based learner using an entropic distance measure”, 12th International Conference on Machine Learning, 1995, pp.108-114. [30] R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Publishers, San Mateo, CA, 1993. [31] A. Minutolo, M. Esposito and G. De Pietro “A Mobile Reasoning System for Supporting the Monitoring of Chronic Diseases”, in the 2nd International ICST Conference on Wireless Mobile Communication and Healthcare MobiHealth 2011, pp. 225-232.
Sannino, G.; De Falco, I.; De Pietro, G., "Automatic extraction of effective rule sets for Obstructive Sleep Apnea detection for a realtime mobile monitoring system," Information Reuse and Integration (IRI), 2013 IEEE 14th International Conference on , vol., no., pp.247,253, 14-16 Aug. 2013 - doi: 10.1109/IRI.2013.6642479