Cardiac Evidence Backtracking for Eating Behavior Monitoring using Collocative Electrocardiogram Imagining

TL;DR

The paper tackles non-invasive, continuous eating monitoring using 24-hour ECG data. It introduces a collocative learning framework that transforms 1D ECG into 2D collocative tensors, incorporates periodic attention regulation via Coached Attention Gates, and enables CAM-based evidence backtracking to produce human-interpretable cardiac evidence. The approach achieves superior performance on the largest ECG eating-behavior dataset and provides readable, clinician-friendly representations through decision trees/forests, supported by qualitative saliency analyses. This work advances interpretable, continuous eating monitoring and offers practical pathways for clinical adoption and cardiology research.

Abstract

Eating monitoring has remained an open challenge in medical research for years due to the lack of non-invasive sensors for continuous monitoring and the reliable methods for automatic behavior detection. In this paper, we present a pilot study using the wearable 24-hour ECG for sensing and tailoring the sophisticated deep learning for ad-hoc and interpretable detection. This is accomplished using a collocative learning framework in which 1) we construct collocative tensors as pseudo-images from 1D ECG signals to improve the feasibility of 2D image-based deep models; 2) we formulate the cardiac logic of analyzing the ECG data in a comparative way as periodic attention regulators so as to guide the deep inference to collect evidence in a human comprehensible manner; and 3) we improve the interpretability of the framework by enabling the backtracking of evidence with a set of methods designed for Class Activation Mapping (CAM) decoding and decision tree/forest generation. The effectiveness of the proposed framework has been validated on the largest ECG dataset of eating behavior with superior performance over conventional models, and its capacity of cardiac evidence mining has also been verified through the consistency of the evidence it backtracked and that of the previous medical studies.

Figures (7)

• Figure 1: Illustration of ECG signals and the typical intervals for cardiac analysis. An ECG period includes 6 waves that are related to different parts of the heart where the electrical impulse initiation, conduction, and depolarization have been performed (e.g., P wave indicates the impulse conduction from the sino-atrial node to the atrio-ventricular node). However, in cardiology, the interpretation of ECG has been built more on the comparative features like the morphology and intervals of these waves (e.g., PP, QRS) than the values of a single wave or time point, because a heart event is usually a dynamic process that includes a set of consecutive movements of several heart parts. This makes the classical features in signal processing (e.g., zero-crossing rates) play a less effective role than the comparative features. In other words, the cardiology way of interpretation focuses on the inter-wave binary relation rather than the unary relation (of single wave).
• Figure 2: The collocative framework for modeling the comparative relations and cardiac evidence backtracking in Electrocardiograms. It consists of three parts: 1) the transformation of 1D ECG signals into 2D views to be adapted to the conventional deep models; 2) the periodic attention coach which guides the deep inference with human logic of comparative analysis; 3) the evidence backtracking based on saliency maps and the generation of human comprehensible representations of the evidence and inference logic.
• Figure 3: The comparative relations among segments have already been encoded as "backslashes" into the collocative tensors. In the figure, several backslashes for the QRS complex are shown. It can be used to model the comparative relations of Q, R, and S waves. The duration has also been encoded as the length of backslashes. The period is encoded into the distances among backslashes.
• Figure 4: The coached periodic attention gates (CAGs). The network of CAGs consists of three layers. The first layer is a single node fixed value 1. The second layer is with $4$ parameter nodes of $(\alpha,\beta,\gamma,T)$. The third layer is a parameterized (by the parameters of the second layer) and diagonalized cosine attention mask $\boldsymbol{\Omega}$ which will guide the learning to focus on the diagonal "backslash" patterns and sense the period of the signals at the same time.
• Figure 5: Examples of saliency maps generated by the Base Model (using the ECG signals as 1D inputs directly) and collocative learning (using the 2D-based collocative tensors as inputs). Maps by the collocative learning are generated at wave- and tensor-levels. Results of the collocative models with and without the periodic CAG regulations have been demonstrated respectively. It is easy to see that the collocative with CAGs has captured the periodic nature of the signals at either wave or tensor level.