Development of a mobile application for metabolic syndrome screening

Abstract

This dissertation was created in pursuit of determining the viability of a smartphone based technique for metabolic syndrome screening. This screening method consists of two artificial neural networks. The first network uses a photoplethysmography (PPG) waveform measured using a smartphone camera and its FFT analysis to determine blood pressure. The second classifier uses the heart rate and blood pressure determined from the PPG measurement and user data, including sex, age, BMI, waist-to-height ratio, smoking, drinking habits, medical history and activity level, to determine whether a user may be at risk of metabolic syndrome. The blood pressure network is initially trained and tested using PPG data, measured with a blood oximeter, from the MIMIC III online waveform dataset. The algorithm is then tested with PPG data obtained from participants with a smartphone camera. In real world scenarios the BP (blood pressure) algorithm has a mean absolute error of 7.7 mmHg (±6:9) for SBP (systolic blood pressure) and 8.0 mmHg (±6:3) for DBP (diastolic blood pressure). Multiple metabolic syndrome algorithms are tested, with and without lifestyle related factors, and with a simulated error being added to blood pressure inputs based on results from participant tests. It was found that the inclusion of lifestyle factors improved the performance of the model by approximately 2%, from 75.2% to 77.2%, when assessed using 10-fold cross validation. When including simulated HR and BP errors, the network including lifestyle factors achieved an average accuracy of 74.9%. An approach similar to bootstrapping was followed where the 10-cross validation process is executed 10 times, randomly shuffling the data with each cycle, resulting in a total of 100 testing sets. The average accuracy across the 100 sets for the first model was 73.4%, with a sensitivity of 73.1% and a specificity of 73.6%. For the model with lifestyle factors the accuracy was 76.3%, with a sensitivity of 74.9% and a specificity of 77.4%. A t-test was done to show that the accuracy difference between the models was statistically significant (t = -2.48; p < 0.014). A bias in the BP algorithm was found for the real world testing, causing the mean SBP to be 5 mmHg too low. Due to a strong dependence on DBP, this caused the MetS algorithm to perform less optimally. This highlights the necessity of an accurate BP algorithm for the MetS algorithm to provide reliable results. The accuracy obtained across the 100 testing sets when including HR and BP errors was 73.1%, with a sensitivity of 64.9% and a specificity of 79.9%. In order to determine if the output from the model could be interpreted as a risk indicator instead of just a binary prediction, the models were assessed at different confidence levels. The models achieved 86.8% to 88.4% accuracy for confidence levels over 80% and 92.7% to 94.1% accuracy for confidence over 90%. This seems to indicate that there is definite value in using this model as a risk indicator, which a user could potentially use to determine whether a full diagnosis is warranted. In that sense, it falls perfectly in line with what is expected from a screening solution.

The results show that smartphone based metabolic syndrome screening is a promising venture, that can be expanded upon in further research and considered for real world screening solutions. However, further research would be required to test whether this solution achieves similar results across multiple devices before it could be rolled out as a final product.