A Comparative Study of Deep Learning Algorithms for Detecting Food Intake

The choice of appropriate machine learning algorithms is crucial for classification problems. This study compares the performance of state-of-the-art time-series deep learning algorithms for classifying food intake using sensor signals. The sensor signals were collected with the help of a wearable sensor system (the Automatic Ingestion Monitor v2, or AIM-2). AIM-2 used an optical and 3-axis accelerometer sensor to capture temporalis muscle activation. Raw signals from those sensors were used to train five classifiers (multilayer perceptron (MLP), time Convolutional Neural Network (time-CNN), Fully Convolutional Neural Network (FCN), Residual Neural Network (ResNet), and Inception network) to differentiate food intake (eating and drinking) from other activities. Data were collected from 17 pilot subjects over the course of 23 days in free-living conditions. A leave one subject out cross-validation scheme was used for training and testing. Time-CNN, FCN, ResNet, and Inception achieved average balanced classification accuracy of 88.84%, 90.18%, 93.47%, and 92.15%, respectively. The results indicate that ResNet outperforms other state-of-the-art deep learning algorithms for this specific problem.