Vision-based Respiratory Movement Analysis for Monitoring the Sleep Apnea Severity Using Computer Vision and Machine Learning
Any opinions, findings, conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of UKDiss.com.
Abstract
Background: Sleep apnea is a chronic respiratory disorder which occurs due to the repetitive complete or partial collapses of the pharyngeal airway during sleep. Sleep apnea is a common respiratory disorder in 10% of the population. The gold standard method for diagnosis of sleep apnea is overnight in-laboratory sleep with polysomnography (PSG). However, the inconvenience and immobility of the PSG, as well as its high cost, have encouraged researchers to develop simple and portable devices to detect sleep apnea. The goal of this research is to develop a convenient and non-contact technique for sleep apnea diagnosis based on analyzing video recordings during sleep.
Methods: A full night sleep study will be performed and sleep apnea severity of the subjects will be assessed by the gold standard PSG. Simultaneously with PSG, an infrared camera will record the patient’s head and torso. Visual points in images will be identified within frames and will be tracked over time. The relationship between chest and abdomen movements and PSG during respiratory events will be investigated. On the basis of this comparison, the vision-based monitoring of the respiratory movements will be used to develop a novel machine learning system to estimate sleep apnea severity.
Results: This project will result in a set of features that are highly correlated with sleep apnea. Machine learning models will be developed to estimate sleep apnea severity; and they will be validated against the gold standard PSG.
Conclusion: This system will provide insight into the usefulness of state-of-the-art computer vision and machine learning methods to assess sleep apnea severity by monitoring respiratory movement. Successful implementation of this system would facilitate more frequent assessments of sleep apnea severity, allowing patients and their sleep physicians to make more informed treatment decisions.
Keywords: computer vision, sleep apnea, apnea hypopnea index, machine learning, physiological monitoring
Introduction
Sleep Apnea is a common respiratory disorder occurring in 10% of the population [1]. Sleep apnea is caused by complete (apnea) or partial (hypopnea) collapse of the pharyngeal airway during sleep (see Figure 1) [2]. Sleep apnea is associated with sleep fragmentation and increased daytime sleepiness [3], increased hypertension [4], increased risk of heart disease [5], and increased risk of stroke [6].
The gold standard method for the diagnosis of sleep apnea is overnight in-laboratory sleep with polysomnography (PSG) [7], [8]. PSG requires more than 20 sensors connected to the body, which is inconvenient [9]. Furthermore, a trained technician must analyze the PSG recordings, which is expensive [10] and time-consuming [11]. As a result of these challenges, up to 90% of individuals who may have sleep apnea are not diagnosed [12], [13]. Higher rates of diagnosis would be possible if there were portable monitoring devices to detect sleep apnea.
Existing portable sleep apnea monitoring devices differ in the number of channels connected to the body [14]; Type 1 is attended PSG with minimum of seven channels, type 2 is unattended PSG with minimum seven channels, type 3 is unattended with minimum four channels, and type 4 is with minimum one channel. Single channel devices suffer from low accuracy. Multi-channel devices suffer from high failure rates due to data loss in unattended home settings [15]. Thus, there is a trade-off between simplicity and accuracy. Therefore, there is a need to develop accurate, convenient, and affordable monitoring devices to diagnose sleep apnea.
In recent years, non-contact monitoring methods have been developed to diagnose sleep apnea. Respiratory movements caused by inhalation and exhalation and subtle movements caused by cardiac activity are used mostly as predictors of sleep apnea. Computer vision-based monitoring devices detect respiratory rates by identifying those movements or tracking changes in skin colour. The main challenges of vision-based monitoring methods are selecting the right region of interest (ROI), eliminating noise, working in low-light, and working in real-time. In this research, we will develop a convenient and non-contact algorithm for sleep apnea diagnosis based on analyzing video recordings during sleep in low-light conditions.
Literature Review:
Previous studies have shown an effort to diagnose sleep apnea based on visual features of respiratory and cardiopulmonary movements. Computer vision-based monitoring systems to detect physiological signals are categorized into two groups. Colour-based methods track the changes in colour or brightness of the skin caused by blood perfusion, and motion-based methods identify the motions associated with physiological signals.
Colour-based methods rely on the properties of the blood (hemoglobin) which absorbs light more than the surrounding tissue [16]. Takano et al. proposed a method to extract PPG data with ambient light [17]. Bousefsaf et al. were able to reduce the effects of illumination variations by using spectral skin analysis [18]. Feng et al. used an adaptive green/red difference to extract the signal from moving subjects [19]. Xu et al. developed a remote measuring system to extract PPG signal under changing light conditions using partial least squares and multivariate empirical mode decomposition [20].
There are several challenges with colour-based methods. The ROIs need to be clearly detected to extract the cardiorespiratory signal. PPG waveforms vary due to different ROIs, artificial illumination variations, and skin tone. Also, it is difficult to track the changes of skin colour in the dark.
Motion-based methods use motion or optical flow information of consecutive images to extract respiratory and cardiac motion from video data. Nakajima et al. used a camera to extract the respiratory activity by detecting the optical flow of the movement resulting from respiratory rhythms [21]. Li et al. used infrared camera to record the video images of the subjects while sleeping to monitor cardiopulmonary activities. They extract respiratory and heart rate by tracking several distinctive points in video and selecting top ranked components using principal components analysis [22]. Cattani et al. proposed a method to estimate the periodicity of pathological movements using Eulerian video magnification and Maximum Likelihood (ML) criterion for periodicity detection [23]. He et al. developed a system to estimate respiratory and heart rate using infrared night camera in order to monitor physiological movements during the night. They used Eulerian video magnification on manually selected ROI and used a butterworth low-pass filter to isolate the respiratory signal [24].
Many researchers used depth image sensing camera like time of flight and Kinect sensor to monitor respiratory movements [25], [26], [27], [28], [29], [30], [31], [32], [33]. However, the major challenges of these methods are short detection ranges and the need for specific hardware.
Comparing the motion-based computer vision methods with the colour-based methods, motion-based methods have several advantages. They work effectively even when skin is not visible or even the subject is covered by a blanket or a mask. Moreover, as they rely on motion detection, they are not affected by artificial illumination variations or skin tone.
In this study, a motion-based computer vision method to assess sleep apnea severity is proposed. To the author’s knowledge, motion-based computer vision methods have not been used to assess sleep apnea severity.
Research Objectives
Research Goal: The goal of this research is to develop a clinically reliable non-contact algorithm to estimate sleep apnea severity assessed by Apnea-Hypopnea Index (AHI) by analyzing visual features of respiratory movements captured by infrared cameras during sleep.
Research Question: Can computer vision and machine learning techniques measure the severity of sleep apnea?
Research Hypothesis: We hypothesize that by applying computer vision and machine learning techniques sleep apnea severity could be estimated from overnight infrared video with more than 80% accuracy (compared to the PSG) for clinical diagnosis.
Aim 1:
Research Question: Can infrared video images be used to find visual features correlated with respiratory events during sleep?
Hypothesis: If features like respiratory frequency and amplitude will change with the apnea-hypopnea, then these features can be used to estimate sleep apnea severity.
The first goal of aim 1 is extracting visual features using computer vision techniques from the infrared video which correlate significantly with apnea and hypopnea. The second goal is implementing a machine learning classifier to classify events into normal breading, central, and obstructive apnea hypopnea. To address our aim, we will recruit 20 subjects and perform overnight sleep study with gold standard PSG. Simultaneously, video images will be recorded using the infrared camera.
Methodology:
Sleep study and Polysomnography: Before starting this research, we will get the required approvals from the research ethics board. A consent form will be prepared to read and sign by participants. Participants will come to the sleep lab of Toronto Rehabilitation Institute (TRI) around 8 PM, and they will undergo an overnight sleep study for 7-8 hours. At the beginning, participant’s medical history including history of heart attack, any respiratory disorder such as cough and flu, and smoking history will be determined. Overnight PSG will then be performed. Using standard techniques and criteria, a qualified PSG technician will mark the sleep stages and arousals as well as the apneas and hypopneas.
The Infrared Camera: While sleeping, the room will be illuminated by an IR light and respiratory movements will be recorded using an infrared camera which will be placed approximately 1.5 m above the bed, to capture the person’s head and torso. Video data will be transferred from the camera to a computer for further analysis.
Data Analysis: Changes in respiratory rate are useful predictors of sleep apnea. We will look for these changes in the video images using computer vision techniques. Respiratory rate and respiratory amplitude will be extracted by tracking the trajectories of specified visual points in subjects’ chest and abdomen. During the apnea or hypopnea, pauses in breading (central apnea) will change the respiratory rate, and shallow breading (obstructive apnea) will change the amplitude of the respiratory movements. These changes should last 10 seconds or more to be considered as apnea or hypopnea. Thus, by looking for these changes in a sliding window of time over the entire video we will be able to mark the events.
We will implement a supervised machine learning algorithm to be able to classify the events. Detected events and extracted features will be used as inputs of our implemented supervised machine learning algorithm. The actual events that marked by PSG technician will be used as ground truth to train our model. The classifier will classify the events to normal breading, central apnea-hypopnea, and obstructive apnea-hypopnea. In order to gain higher accuracy, we will tune the implemented classifier by dividing our data into train, validation, and test and using several cross-validation methods. Finally, by comparing the results we will choose the best cross-validation method for our classifier.
Methods:
Histogram of oriented gradients (HOG) and histogram of oriented optical flow [34]: Will be implemented and assessed to extract features from recorded infrared videos.
Harris corner detection [35]: Will be used to identify the visual points.
Support vector machine and support vector regression [36]: Will be used for our classification goal.
Cross-validation [37]: Will be used to tune the machine learning algorithm.
Expected Results: We expect that, features that are highly correlated with sleep apnea severity such as respiratory frequency and respiratory amplitude will be extracted in this aim. By incorporating these features in our machine learning algorithm, we will then be able to classify the events to normal breading, central and obstructive apnea-hypopnea. Results of the implemented algorithm will be assessed by comparison with PSG in overnight experiments with n=20 participants. By measuring the true positive, true negative, false positive, and false negative, the sensitivity, specificity, and precision of the method will be investigated using confusion matrix.
Aim 2:
Research Question: Can deep learning methods be used to estimate sleep apnea severity?
Hypothesis: We hypothesize by using deep learning techniques sleep apnea severity could be estimated from the overnight infrared video for clinical diagnosis.
The goal of aim 2 is to develop and validate a deep learning algorithm to estimate sleep apnea severity. For this aim, machine learning models that consider the temporal dependency of video data will be used.
Methodology: We will implement a deep learning method for spatial-temporal analysis of recorded video data. Deep learning is an unsupervised learning method and is one of the cutting-edge technologies in the field of machine learning. The reason we choose deep learning methods is that these methods are capable of extracting the features and doing the classification. The algorithm will provide us with a set of features extracted from video images and will assign a weight to each feature. We will choose the features with higher weights and will retrain our algorithm with regard to those selected features. At the end, we will have a trained model which is able to detect apnea-hypopnea events. By implementing a classifier, we will classify the events to normal breading, central and obstructive apnea-hypopnea. Finally, we will compare the performance and accuracy of this algorithm with the algorithm developed in aim 1 in estimating sleep apnea severity.
Methods:
Support vector machine and support vector regression [36]: Will be used for our classification goal.
Cross-validation [37]: Will be used to tune the machine learning algorithm.
Recurrent neural networks (RNN) [38]: Will be used as deep learning method. It considers temporal information of video frames and will let us to analyze the video images to extract weighted features.
Expected results: We expect that a deep learning model trained to estimate sleep apnea severity will be developed in this aim. The concurrent validity of the estimated results of the algorithm will be assessed against the ground truth PSG in overnight experiments with n=20 participants. By measuring the parameters needed for confusion matrix, the sensitivity, specificity, and precision of the method will be investigated.
Timeline: The proposed timeline (see Table 1) is included to demonstrate our strategy for tackling this project. Here, we consider Sept. 01st, 2017 as the start date with the target end date in Aug. 31st, 2019.
Dissemination Plan: This project has the potential to yield one peer-reviewed publication on the sleep apnea severity detection using infrared videos. This paper could potentially be published in the following journals:
- IEEE journal of biomedical and health informatics
- Journal of Clinical Sleep Medicine
Conclusion and Significance:
This study will demonstrate the potential effectiveness of measuring of visual respiratory movement features to monitor and estimate the severity of sleep apnea. Once validated in a larger clinical population, the results of this study will be used to develop cost-effective and convenient non-contact technologies based on video images to assess sleep apnea severity.
References:
1. Peppard PE, Young T, Barnet JH, Palta M, Hagen EW, Hla KM. Increased Prevalence of Sleep-Disordered Breathing in Adults. Am J Epidemiol. 2013 May 1;177(9):1006–14.
2. Quan S, Gillin JC, Littner MR, Shepard JW. Sleep-related breathing disorders in adults: Recommendations for syndrome definition and measurement techniques in clinical research. Editorials. Vol. 22. 1999. 662 p.
3. Rakel RE. Clinical and Societal Consequences of Obstructive Sleep Apnea and Excessive Daytime Sleepiness. Postgrad Med. 2009 Jan;121(1):86–95.
4. Malhotra A, White DP. Obstructive sleep apnoea. The lancet. 2002;360(9328):237–245.
5. Takama N, Kurabayashi M. Influence of Untreated Sleep-Disordered Breathing on the Long-Term Prognosis of Patients With Cardiovascular Disease. Am J Cardiol. 2009 Mar;103(5):730–4.
6. Calvin AD, Somers VK. Obstructive sleep apnea and risk of stroke: time for a trial. Nat Clin Pract Cardiovasc Med. 2009 Feb;6(2):90–1.
7. Pack AI, Gurubhagavatula I. Economic implications of the diagnosis of obstructive sleep apnea. Ann Intern Med. 1999;130(6):533–534.
8. Kushida CA, Littner MR, Morgenthaler T, Alessi CA, Bailey D, Coleman Jr J, et al. Practice parameters for the indications for polysomnography and related procedures: an update for 2005. Sleep. 2005;28(4):499–523.
9. Chesson AL, Ferber RA, Fry JM, Grigg-Damberger M, Hartse KM, Hurwitz TD, et al. Practice parameters for the indications for polysomnography and related procedures. Sleep. 1997;20(6):406–422.
10. Sleep Study CPT codes list 95806, 95810, 95811, 95807 | Medicare Fee, Payment, Procedure code, ICD, Denial [Internet]. [cited 2017 Oct 10]. Available from: http://www.medicarepaymentandreimbursement.com/2011/08/polysomnography-and-sleep-studies-cpt.html
11. Flemons WW, Douglas NJ, Kuna ST, Rodenstein DO, Wheatley J. Access to Diagnosis and Treatment of Patients with Suspected Sleep Apnea. Am J Respir Crit Care Med. 2004 Mar 15;169(6):668–72.
12. Kapur V, Strohl KP, Redline S, Iber C, O’Connor G, Nieto J. Underdiagnosis of Sleep Apnea Syndrome in U.S. Communities. Sleep Breath. 2002 Apr 1;6(2):49–54.
13. Young T, Evans L, Finn L, Palta M. Estimation of the clinically diagnosed proportion of sleep apnea syndrome in middle-aged men and women. Sleep. 1997;20(9):705–706.
14. Collop NA, Anderson WM, Boehlecke B, Claman D, Goldberg R, Gottlieb DJ, et al. Clinical guidelines for the use of unattended portable monitors in the diagnosis of obstructive sleep apnea in adult patients. Portable Monitoring Task Force of the American Academy of Sleep Medicine. J Clin Sleep Med JCSM Off Publ Am Acad Sleep Med. 2007;3(7):737–747.
15. Golpe R, Jime´nez A, Carpizo R. Home Sleep Studies in the Assessment of Sleep Apnea/Hypopnea Syndrome. Chest. 2002 Oct 1;122(4):1156–61.
16. Meredith DJ, Clifton D, Charlton P, Brooks J, Pugh CW, Tarassenko L. Photoplethysmographic derivation of respiratory rate: a review of relevant physiology. J Med Eng Technol. 2012 Mar;36(1):1–7.
17. Takano C, Ohta Y. Heart rate measurement based on a time-lapse image. Med Eng Phys. 2007 Oct;29(8):853–7.
18. Bousefsaf F, Maaoui C, Pruski A. Continuous wavelet filtering on webcam photoplethysmographic signals to remotely assess the instantaneous heart rate. Biomed Signal Process Control. 2013 Nov;8(6):568–74.
19. Litong Feng, Lai-Man Po, Xuyuan Xu, Yuming Li, Ruiyi Ma. Motion-Resistant Remote Imaging Photoplethysmography Based on the Optical Properties of Skin. IEEE Trans Circuits Syst Video Technol. 2015 May;25(5):879–91.
20. Xu L, Cheng J, Chen X. Illumination variation interference suppression in remote PPG using PLS and MEMD. Electron Lett. 2017;53(4):216–218.
21. Nakajima K, Osa A, Miike H. A method for measuring respiration and physical activity in bed by optical flow analysis. In: Engineering in Medicine and Biology Society, 1997 Proceedings of the 19th Annual International Conference of the IEEE [Internet]. IEEE; 1997 [cited 2017 Oct 9]. p. 2054–2057. Available from: http://ieeexplore.ieee.org/abstract/document/758752/
22. Li MH, Yadollahi A, Taati B. Non-Contact Vision-Based Cardiopulmonary Monitoring in Different Sleeping Positions. IEEE J Biomed Health Inform. 2017;PP(99):1–1.
23. Cattani L, Alinovi D, Ferrari G, Raheli R, Pavlidis E, Spagnoli C, et al. Monitoring infants by automatic video processing: A unified approach to motion analysis. Comput Biol Med. 2017 Jan;80:158–65.
24. He X, Goubran R, Knoefel F. IR night vision video-based estimation of heart and respiration rates. In: 2017 IEEE Sensors Applications Symposium (SAS). 2017. p. 1–5.
25. Schaller C, Penne J, Hornegger J. Time-of-flight sensor for respiratory motion gating: Time-of-flight sensor for respiratory motion gating. Med Phys. 2008 Jun 13;35(7Part1):3090–3.
26. Penne J, Schaller C, Hornegger J, Kuwert T. Robust real-time 3D respiratory motion detection using time-of-flight cameras. Int J Comput Assist Radiol Surg. 2008 Nov;3(5):427–31.
27. Falie D, David L, Ichim M. Statistical algorithm for detection and screening sleep apnea. In: Signals, Circuits and Systems, 2009 ISSCS 2009 International Symposium on [Internet]. IEEE; 2009 [cited 2017 Oct 9]. p. 1–4. Available from: http://ieeexplore.ieee.org/abstract/document/5206206/
28. Yu M-C, Liou J-L, Kuo S-W, Lee M-S, Hung Y-P. Noncontact respiratory measurement of volume change using depth camera. In: Engineering in Medicine and Biology Society (EMBC), 2012 Annual International Conference of the IEEE [Internet]. IEEE; 2012 [cited 2017 Oct 9]. p. 2371–2374. Available from: http://ieeexplore.ieee.org/abstract/document/6346440/
29. Bernacchia N, Scalise L, Casacanditella L, Ercoli I, Marchionni P, Tomasini EP. Non contact measurement of heart and respiration rates based on KinectTM. In: Medical Measurements and Applications (MeMeA), 2014 IEEE International Symposium on [Internet]. IEEE; 2014 [cited 2017 Oct 9]. p. 1–5. Available from: http://ieeexplore.ieee.org/abstract/document/6860065/
30. Martinez M, Stiefelhagen R. Breathing Rate Monitoring during Sleep from a Depth Camera under Real-Life Conditions. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). 2017. p. 1168–76.
31. Lin KY, Chen DY, Yang C, Chen KJ, Tsai WJ. Automatic Human Target Detection and Remote Respiratory Rate Monitoring. In: 2016 IEEE Second International Conference on Multimedia Big Data (BigMM). 2016. p. 354–6.
32. Procházka A, Schätz M, Centonze F, Kuchyňka J, Vyšata O, Vališ M. Extraction of breathing features using MS Kinect for sleep stage detection. Signal Image Video Process. 2016;10(7):1279–86.
33. Rihana S, Younes E, Visvikis D, Fayad H. Kinect2 #x2014; Respiratory movement detection study. In: 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). 2016. p. 3875–8.
34. Histogram of oriented gradients. In: Wikipedia [Internet]. 2017 [cited 2017 Oct 16]. Available from: https://en.wikipedia.org/w/index.php?title=Histogram_of_oriented_gradients&oldid=768396739
35. Corner detection. In: Wikipedia [Internet]. 2017 [cited 2017 Oct 16]. Available from: https://en.wikipedia.org/w/index.php?title=Corner_detection&oldid=801295857
36. Support vector machine. In: Wikipedia [Internet]. 2017 [cited 2017 Oct 16]. Available from: https://en.wikipedia.org/w/index.php?title=Support_vector_machine&oldid=805071956
37. Cross-validation (statistics). In: Wikipedia [Internet]. 2017 [cited 2017 Oct 26]. Available from: https://en.wikipedia.org/w/index.php?title=Cross-validation_(statistics)&oldid=807143385
38. Recurrent neural network. In: Wikipedia [Internet]. 2017 [cited 2017 Oct 16]. Available from: https://en.wikipedia.org/w/index.php?title=Recurrent_neural_network&oldid=802740929
39. Levitzky MG. Using the pathophysiology of obstructive sleep apnea to teach cardiopulmonary integration. Adv Physiol Educ. 2008 Sep 1;32(3):196–202.
Figure 1:
image from: http://advan.physiology.org/content/32/3/196
Using the pathophysiology of obstructive sleep apnea to teach cardiopulmonary integration [39]
Table 1:
Research Activities | Time (Months) | ||||||||||||
2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 | 20 | 22 | 24 | ||
Literature review | |||||||||||||
Defining the Project and Writing the Thesis Proposal | |||||||||||||
Aim 1 | Learning computer vision and machine learning techniques | ||||||||||||
Sleep study and Polysomnography | |||||||||||||
Data Analysis | |||||||||||||
Aim 2 | Learning the deep learning methods | ||||||||||||
Implement deep learning algorithm | |||||||||||||
Writing – Peer Review Publication | |||||||||||||
Writing – Thesis | |||||||||||||
Thesis Defense |
Response to Review:
Comments | Responses |
Does the introduction adequately explain the rational for proposing this research? (If so, how? Could there be improvements?) (If not, what is missing? What could be deleted/edited?)
|
|
In the introduction, I think that a figure demonstrating the collapse of pharyngeal airway causing apnea or hypopnea could be included. | It is a very good comment. Figure 1 is added into appendix section. |
Polysomnography is a specialized technique used to assess sleeping disorders. Its importance was mentioned in the introduction, but the details behind the technique were not included. | Good comment. The details required for the introducing the PSG is included. For more detail the reader could refer to the cited reference. |
There was a brief description of portable sleep apnea devices. Maziar could expand more on the channels, what they are, and what they measure. | Good comment. The brief details required for the introducing the portable sleep apnea devices is included. For more detail the reader could refer to the cited reference. |
Maziar could expand more on the technical details of colour- and motion-based computer vision technologies. | It is a very good comment. Brief details are added. |
Maziar could reduce some of the jargon included in his introduction, like some of the methods to estimate pathological movements in those suffering from sleep apnea. | I think they could help the reader to understand the method we use and why we chose motion-based method. |
More information could be included about machine learning and how it will fit into the study. | Very good comment. More details added. |
Are the research objectives clear? Address: 1) How clear are the stated goals of the proposal? Could they be improved?, 2) Give a suggestion how the research question can be shortened, and 3) Explain how each hypothesis is good or needs to be improved.
|
|
The research question could be shortened as follows: Can machine learning measure the severity of sleep apnea? | Thanks for the comment. It is shortened. |
The hypothesis for Aim 1 could be made more specific. Maziar could explicitly state frequency and amplitude of respiration as the main metrics he plans to measure from his experiment. It could be made more explicit what he expects to see when he measures frequency and amplitude of visual movements, and how that will affect his measurements of sleep apnea severity. | Good point. I added some details of how they correlate with sleep apnea and why we want to monitor them. |
Maziar could alternatively state that he plans to measure several features, and that he expects to see one feature that will have a statistically significantly positive correlation with sleep apnea. | Already added. |
I understand that the results of Aim 1 will dictate some of the actions taken as part of Aim 2 in Maziar’s thesis. | This is a very good point. I changed both aims in order to make them independent to each other. |
However, it could be made more explicit what Maziar plans to measure with his machine learning algorithm. | Good comment. I tried to make it more clear by adding the methodology part to both aims. |
It is also not clear why 80% accuracy in diagnosis was chosen as the threshold for success of the algorithm. | The accuracy will be measured with regard to the gold standard PSG. The reason we mentioned 80% is that the accuracy of common vision-based methods is under 80% and we want to achieve a higher accuracy. |
How well described is the methodology and the methods?
|
|
The methodology behind the experiments is not completely explained. There are details about what the experiments will try to measure or determine. However, it is not fully explained why the experiments are being conducted. Maziar does not explain why he is using the techniques he mentions in his methods. | It is a very good comment. I added the methodology part into both aims and I tried to explain the methodology we want to follow in each aim.
I mentioned that the reason that we conduct multiple experiments is that we want to compare different approaches in solving our problem. |
He does not go into enough detail about how he will implement those techniques. A brief explanation would be useful. | Good point. I divided the aims into methodology and methods and provided details of what and how we want to use them in our project. |
How well are the implications communicated? Address: 1) Is it clear that the author addressed either or both of the following questions: “So What?”, “Why should I care?” and 2) Does your understanding of the wider implications of this research match the author’s descriptions?
|
|
Maziar addressed the questions “So What?” and “Why Should I care?”. Maziar is developing a cost effective diagnostic system to treat patients suffering from sleep apnea. Many people go undiagnosed because of the cost of the procedure to determine if patients have sleep apnea. His system would solve this problem. My understanding of the wider implications matches the author’s description. Maziar’s system results would act as a proof of concept as a non-contact system for diagnosing sleep apnea using machine learning. It would require further validation from a larger patient population to ensure its accuracy in diagnosing sleep apnea and hypopnea. | Thanks for the comments. |
If applicable, identify any other major issues and/or specific recommendations.
|
|
I understand that machine learning in computer vision requires large data sets to “train” computers to identify objects or patterns correctly in images. I was wondering if Maziar has considered whether his sample population of 20 individuals is too small to adequately train his algorithm to identify pathological movements associated with sleep apnea and hypopnea. | Some of the patients that will participate in our study have mild to severe apnea problem. The data that will be used for training the deep learning network is the video taken from patients with AHI index of ~10 to ~50. The video recorded from each patient will last around 7 to 8 hours. Thus, amount of useful data to train the algorithm would be between ~70 to ~400 per patient. |
I was wondering if Maziar was planning to process the same experimental data using the several implementations of the neural network? What are the benefits of choosing to conduct the experiment with different networks, when they each seem to have specific functions? | Good point. I changed the aim 2 having this comment in mind. |
Maziar mentioned that he planned to gain 80% accuracy in diagnosis using his system. I was wondering how he chose that value. I was also wondering what the margin of error would be on this figure. |