Further thanks to Prof Carina and Sr Adele from the HART institute for providing a sphygmomanometer for testing and Michelle from Cachet Park for providing the location for data collection. Finally, I thank all 133 participants who allowed me to measure their blood pressure and use their data in my research.
Background
This shows that the high prevalence of MetS may be due to a lack of awareness. This shows that an app that provides adequate feedback can help guide users to better choices regarding the treatment or prevention of MetS.
Literature review
- Metabolic Syndrome screening methods
- Photoplethysmography
- Classication techniques
- What is a screening test?
This is further demonstrated by Kim [16] in a meta-analysis of observational studies used to assess the association between the consumption of different types of meat and the risk of MetS. This correlation found between heart rate and MetS suggests that heart rate may be a valuable measurement to perform for MetS screening.
Research question
To keep the app as cheap as possible, thereby also increasing the accessibility, will be a requirement to exclude any external sensors. Is it necessary to determine all the characteristic risk factors for MetS (abdominal obesity, fasting blood sugars, serum triglyceride levels and blood pressure) to diagnose metabolic syndrome.
Methodology
Can a smartphone be used as a plethysmograph in order to accu-
These figures illustrate the disadvantages of both the techniques that can be used to determine heart rate. Each of these conditions will be discussed in more detail to determine if they can be determined in an inexpensive and non-invasive manner.
Would an application of this kind be easy enough to use for it to
Relative efficiency can be calculated by also using the time of the tasks that were not performed. The results from the questionnaire are then compared with the SUS in Figure 1.5, which was made on the basis of information from [50], to draw a conclusion about the user's satisfaction.
Chapter overview
- Chapter 2: Photoplethysmogrophy measurements
- Chapter 3: Metabolic Syndrome Classication
- Chapter 4: Experimental procedure
- Chapter 5: Results
- Chapter 6: Conclusion
- Heart rate
- Blood Pressure
ANN is used to determine blood pressure from features extracted from the PPG waveform. Finally, ANN is used to determine blood pressure from the features extracted from the PPG waveform.
PPG on a smartphone
Colour selection
It is immediately clear from Figure 2.1 that the color that appears most from the skin is red. Blue light, however, is almost completely absorbed, barely penetrates the skin and therefore receives minimal signal. Considering that red light penetrates the skin much better than green light, red would be a better option for obtaining the PPG signal in cases where the light source is not.
Signal verication
However, blue and green channels were still used as references to determine if the user was operating the application correctly, which is described in the next section. The first validation step for each frame will be to determine if the user has placed their nger correctly on the camera lens. Deviating from the calibrated values will alert the application of motion or other issues during the measurement, which will cause the application to enter an error state and prompt the user to restart the measurement.
Heart rate
Something to consider when determining the heart rate is the measurement resolution of the method used. The measurement resolution of the FFT-based method also depends on the sampling frequency and is further dependent on the number of samples (Ns). For a 10 s measurement, the resolution for the FFT method is 3 BPM, which is usable, but still much higher than the 0.2 BPM for the time-based method.
Determining blood pressure
- Hydraulic analogy
- Data analysis and preprocessing
- Feature extraction
- ANN design
As we saw in Section 2.4.1, the relationship between blood pressure and PPG can be seen in the frequency response of the PPG. The frequency response was the first 14 points (10 Hz) of the magnitude and phase response as shown in Figure 2.10. with 70% of the dataset (approximately 6650 samples) used for training and 15% (approximately 1400 samples) used for validation and testing.
Chapter summary
This chapter discusses the design of the ANN to be used to screen for MetS. The study also used a linear regression model, which did not have the high accuracy of the ANN, but was more useful than the ANN in determining which variables correlate the most [47]. There are some deviations from the existing research based on the traits available in the training data, and the addition of other lifestyle-based traits is also being tested to determine if this would have a significant impact on the model's accuracy.
Application interface
Another study that may have shown the most promise regarding lifestyle factors in information available is one by Jahandideh, who used an ANN to infer risk factors for cardiovascular disease (HDL cholesterol, total cholesterol, triglycerides) from factors such as to determine gender, age. , build, weight, marital status, individual's status in the family, physical activity, hours of sleep per day, smoking, type of tobacco and BMI [47]. The research by Ivanovi¢ [15], which used an ANN to determine MetS from only non-invasive features, is used as a starting point. Based on the promise shown in previous studies that determined MetS with only non-invasive measurements [15], a similar technique was used in this thesis.
Data analysis
- Main features
- Lifestyle factors
- Activity level
- Scaling
The correlation matrix of all the relevant variables in the SAPBA dataset can be seen in Figure 3.2. In Figure 3.3, the distribution of some of the characteristics for patients with and without MetS can be seen. The initial features selected to determine MetS were largely based on [ 15 ] and the features available in the SAPBA dataset.
ANN design
Network Optimisation
While the results from the other sessions were not so clear in this regard, the trend was expected to remain the same. The trend regarding the number of layers was not always as clear in the case of the model including lifestyle factors, as can be seen in Figure 3.5, but the single-layer network tended to have slightly more consistent performance. Finally, the number of neurons for the hidden layer was chosen as 30 in both cases.
Without lifestyle factors
With lifestyle factors
Chapter summary
Answering question 1: PPG based heart rate and blood pressure
There have been previous studies where blood pressure was determined from PPG data obtained from blood oximeters. However, there have been fewer studies showing comparable blood pressure results with PPG data obtained from a cell phone camera. There are also no known online datasets containing blood pressure and PPG data measured with a cell phone camera.
Answering questions 2 and 3: Metabolic Syndrome
There are already many heart rate monitoring apps available on Google Play and the Apple App Store on mobile phones, some of which have also been the subject of research. This made it necessary to obtain smartphone-measured data that could be used to test the proposed heart rate and blood pressure algorithms. This model was then integrated into an application that would be used to collect mobile phone-based BP, HR, and PPG data from participants.
Answering question 4: Usability
To further test the potential accuracy of a fully integrated application, the distribution of the errors from the BP and HR results were used as margins to add random errors to the test sets of the MetS model, including lifestyle factors.
Overview of the datasets for testing
MIMIC
The same standards used to validate the accuracy of standard cu-based shygmomanometers were used for the blood pressure algorithm on the test set. American National Standards Institute, Association for the Advancement of Medical Instrumentation and International Organization for Standardization. Society for the Advancement of Medical Instrumentation, European Society of Hypertension and International Organization for Standardization, 2018.
SABPA
These values indicate a standard variation that can be expected between diagnostic methods and will be used as a baseline in the coming chapters when discussing the effectiveness of the MetS algorithm.
Smartphone data
Data collection process
- Data collection methodology
- Study location
- Study population
- Study limitations
A commercial cu-based blood pressure monitor (Omron HEM-712C [70]) was used to measure participants' blood pressure. The measured value was entered into the application, which stored it in the cloud along with the rest of the data. Due to the focus on measuring blood pressure, other risk factors for metabolic syndrome were not measured.
Chapter summary
This chapter discusses the results obtained from the experimental procedure described in the previous chapter. The ME, MAE and MAPE obtained from the 133 participants' smartphone PPG HR measurements are shown in Table 5.1. This can be seen more clearly in Figure 5.1b, which is a residual plot showing the error for each PPG HR measurement.
Blood pressure
MIMIC Testing Data
As you can see in Figure 5.4, the majority of measurements have an error of less than 10 mmHg in both cases. a) Regression graph of smartphone PPG. This may be due to the fact that the target population of the MIMIC III dataset is ICU patients. Figure 5-7 shows two examples of the time-based PPG portion of the input (i.e., the first 45 samples), both of which had an error greater than 30 mmHg.
Metabolic Syndrome
Without lifestyle factors
The next column in Table 5.5 is precision, which is simply the ability of the classier not to classify a sample that is negative as positive. The last column is the F1 score, which can be interpreted as a harmonic mean of the precision and recall. As the results in Table 5.7 show, there is a clear increase in performance as the confidence level of the output increases.
With lifestyle factors
Nevertheless, this is largely irrelevant because, as with the previous model, an overall increase in accuracy is still visible when looking at the higher confidence levels, indicating that the proposed screening method remains viable for this model as well.
Application data error simulation
The results of the 10-fold cross-validation done on the MetS model including lifestyle habits and including an added error in SBP, DBP and HR are given in Table 5.11. At first glance, when looking at accuracy, the results in Table 5.11 show a very slight drop in performance from the error-free BP en HR results. This bias in the MetS results is most likely due to the negative bias observed in the experimental DBP results earlier.
Usability
Eectiveness
Eciency
Most cases that took longer were those that had difficulty taking a successful measurement in a single attempt. To get a better idea of what the efficiency is for single measurements, we used equation 5.2, which included only the measurements taken with only a single attempt. This further reinforces the statement from the beginning of this section that on average less than half of the time used is on the actual measurement.
Satisfaction
Chapter summary
Based on the ISO 9241-11 standard, we measured the usability of the application in terms of efficiency, effectiveness and satisfaction. This was based on the abilities of the researchers and the inability to measure invasive risk factors for MetS. Stefanadis, Diet, Exercise and the Metabolic Syndrome, Journal of the Society of Biomedical Diabetes Research, vol.
Example of smartphone based PPG measurement
Visualisation of (a) PPG waveform measured using a smartphone and (b)
Order of operations from the user's point of view
Flow diagram of operations to be completed by the software
Grade rankings of SUS scores
Example images recorded using smartphone camera
Camera lenses for Samsung Galaxy A50
Example of PPG signal with low frequency interference with same signal
Comparison between PPG and ABP before and after time alignment
PPG waveform resampled to 30Hz and bandpass lter applied
Analogous electrical and hydraulic equations
Selection criteria used to eliminate abnormal BP samples [27]
ANN design for blood pressure estimation
Harris-Benedict equations as revised by Miin and St Jeor
Value ranges of metabolic syndrome features
ANN topology for Mets without lifestyle factors
ANN topology for Mets with lifestyle factors
Summary of the three datasets used for testing
Confusion matrices of IDF and WHO MetS diagnoses compared to the
Results from the HR algorithm on smartphone PPG measurements
ISO 81060-2 specication for blood pressure testing [48]
Results from the BP algorithm on on the MIMIC III test dataset
Results from the BP algorithm on smartphone PPG measurements
Cross validation results of the MetS algorithm, without lifestyle factors
Cross validation accuracy of the MetS algorithm, without lifestyle factors
Cross validation results of the MetS algorithm, with lifestyle factors on 10
Average confusion matrix of the MetS algorithm, with lifestyle factors
Cross validation accuracy of the MetS algorithm, with lifestyle factors for
Cross validation results of the MetS algorithm with BP error simulation
Average confusion matrix of the MetS algorithm, with lifestyle factors and
Number of participants to number of measurements taken
Number of participants to obtain dierent levels of eectiveness
Usability ratings provided by users for PPG measurement
Rescaled usability ratings with SUS Scores
Summary of real-world and error-simulated results across all variables
Summary of results from IDF and WHO MetS denitions and ANN models
Description of variables 1 to 21 from SABPA dataset applicable to metabolic
Description of variables 22 to 24 from SABPA dataset applicable to metabolic
![Figure 2.6: Cardiac cycle showing left atrial pressure, left ventricular pressure, aortic pressure ventricular volume, the electrocardiogram, and the phonocardiogram from [60]](https://thumb-ap.123doks.com/thumbv2/pubpdfnet/10396028.0/48.892.212.685.128.530/cardiac-pressure-ventricular-pressure-pressure-ventricular-electrocardiogram-phonocardiogram.webp)
