Avaliação cardíaca do padrão de responsividade à cafeína através de métodos de aprendizagem computacional

Abstract

Approximately 80% of the adult population worldwide consumes one caffeinated product on a daily basis and, as opposed to other drugs, the consumption happens in every socioeconomic level. In addition to the cost-benefit, caffeine acts as a powerful stimulant to the central nervous system, causing physiological effects that are very appealing to the consumer, specifically on a cognitive level and in terms of physical performance. The way in which this psychoactive substance is administered seems to influence the body´s response and, therefore, studies have been developed to find more efficient solutions. Caffeine delivery through Oral Films (OF) constitutes a great opportunity for researchers. In order to bridge this area´s scientific gap, advantage was taken from the impact that caffeine has on the modulation of the nervous system´s activity by tracking the variability of cardiac activity in healthy individuals that were subject to the consumption of different caffeine modalities: coffee, decaffeinated coffee, OF_caffeine and OF_placebo. Taking into consideration the scientific and technological age we live in, and in an attempt to maximize the discriminating capability of electrocardiographic signals (ECG), its dynamic nature was exploited using robust signal processing methods and Artificial Intelligence techniques in order to analyze the binary comparisons decaffeinated coffee/coffee and OF_placebo/OF_caffeine. Therefore, after subjecting the ECGs to a multi-band analysis through the use of the Discrete Wavelet Transform (DWT), metrics were extracted for all the 5 decomposition levels: energy, entropy, Lyapunov exponent, Hurst exponent and Higuchi´s fractal dimension. These metrics fed 23 classification models every minute, in a leave-one-out cross-validation process. In order to optimize the classification process, the metrics that better describe the signals were automatically selected, removing redundant information and avoiding overfitting, resorting to classifying training with the Principal Component Analysis (PCA) at 100%, 80%, 70%, 50%, 20%, 10%, 5% and 1%. In order to understand the development of the modalities’ physiological effect on the body over time, 50 min accuracy curves with 1 min resolution were outlined for every binary comparison, classifier and PCA training. In view of the high volume of data, the classifier´s decision that better specifies the modalities in each binary comparison was made based on the average accuracy of the accuracy curves. The individual analysis´ results showed the best average accuracy values for the decision tree classifiers in PCA training with PCA 100% (50.2%) and Fine Gaussian in training with PCA 95% (72%) for decaffeinated coffee/coffee and OF_placebo/OF_caffeine comparisons, respectively. Nevertheless, there are higher amplitude discriminatory peaks camouflaged by the average accuracy over time. In contrast with the decaffeinated coffee/coffee comparison, the accuracy curves showed significant differences between the OF_placebo and the OF_caffeine over time, pointing out that caffeine intake through OF shows effective capacity when compared to its placebo. Meanwhile, as far as the decaffeinated coffee/coffee comparison is concerned, the results point to the fact that decaffeinated coffee may not be a real coffee placebo. In this analysis, the main factor highlights that both modalities were prepared using the same machine and that it is possible that the decaffeinated beverage got caffeine residue.

Date of Award	13 Jan 2023
Original language	Portuguese
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Pedro Miguel Rodrigues (Supervisor) & Patrícia Oliveira-Silva (Co-Supervisor)