In this chapter, the design decisions that were made when developing the algorithms to screen for MetS were discussed. The focus was on achieving this with limited features that could be available to a smartphone user or could be measured using a smartphone.
User interaction and usability were major deciding factors for which features to keep and which to discard. A portion of the information required to determine MetS that would already be known by the individual were chosen, such as weight, height, etc. The remaining features, such as HR and BP, would be determined from the PPG measurements in the previous chapter.
The data used for training and optimisation was from the SAPBA dataset [46]. From this, the following features were selected:
1. Age 2. Gender 3. BMI
4. Waist-to-height ratio 5. SBP
6. DBP 7. Heart rate 8. Medical history 9. Alcohol use 10. Smoking 11. Activity level
Two models were developed, one excluding and one including lifestyle factors (features 8 to 11 in the list above). All the input features were scaled to be between 0 and 1, and an optimised ANN topology was determined for each model. The procedure followed when testing these models is discussed in Chapter 4, with the results from the testing procedure in Chapter 5.
4
EXPERIMENTAL PROCEDURE
In this chapter the experimental procedure used to validate the solutions designed in chapter 2 and 3 is discussed. The solutions are assessed according to how well they answer the research questions posed by the problem statement. The methods used to obtain data and how the data is used are also discussed.
4.1 Addressing the research questions
The following questions were formed in Chapter 1 to expand on the main research question:
1. Can a smartphone be used as a plethysmograph in order to accurately determine heart rate and blood pressure?
2. Is it necessary to determine all the characteristic risk factors for MetS (abdominal obesity, fasting blood sugars, serum triglyceride levels and blood pressure) in order to diagnose metabolic syndrome?
3. Can information from lifestyle factors, such as diet, exercise or smoking, be used to improve the accuracy of MetS predictions?
4. Would an application of this kind be easy enough to use for it to be a feasible solution?
These questions are used as a guide to answer the research question, and by answering them, we hope to adequately address the overall problem.
4.1.1 Answering question 1: PPG based heart rate and blood pressure
There have been previous studies that determined blood pressure from PPG data obtained from blood oximeters. There are also many heart rate monitoring cell phone applications already available on the Google Play Store as well as the Apple App Store, some of which have also been the subjects of research studies [21], [26], [37]. However, there have been fewer studies to show similar BP results with PPG data obtained from a cell phone camera. There are also no known online datasets available that contain blood pressure and PPG data, measured using a cell phone camera. For this reason it was necessary to obtain data measured with a smartphone that could be used to test the proposed heart rate and blood pressure algorithms.
For the initial training, testing and validation the MIMIC III waveform database was used [62]. This model was then integrated into an application that would be used to gather cellphone based BP, HR and PPG data from participants. The data collection process is discussed in Section 4.3
4.1.2 Answering questions 2 and 3: Metabolic Syndrome
As discussed in the previous chapter, to determine the impact of removing some elements from the consideration when determining MetS, the following features were chosen from the SABPA dataset to use in an initial model:
1. Age 2. Sex 3. BMI
4. Waist-to-height ratio 5. SBP
6. DBP 7. Heart rate
These features were chosen because they are all non-invasive measurements that can potentially be made by a smartphone or would already be known by a user.
A second model was built that included all the other information that was available in the SABPA dataset that may be useful for determining MetS. The following features were selected:
1. Medical history
2. Alcohol use 3. Smoking 4. Activity level
Both of these models were evaluated in the same manner, discussed in Chapter 5, and compared to other standard denitions of MetS. 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 testing sets of the MetS model, including lifestyle factors.
4.1.3 Answering question 4: Usability
Observable and quantiable metrics were determined based on the ISO 9241-11 standard, which denes usability as "the extent to which a product can be used by specied users to achieve specied goals with eectiveness, eciency and satisfaction in a specied context of use" [48]. The usability of the application was measured in terms of eectiveness, eciency and satisfaction.
There was, however, a major limiting factor with regards to determining the usability of the application. Since the data collected in the data gathering process only included PPG, HR and BP measurements, participants did not get a perspective of the entire application, but rather just a single aspect thereof, i.e. the PPG measurement. While this means that the results will only be limited to a single element of the application, considering that it is the most time consuming, involved and experimental aspect of the application (the rest simply entails lling in text elds), this should still provide a valuable indication of what the usability of a full MetS application would be.