Core Concepts
Explainable AI techniques reveal age-related changes in electrocardiogram features and patterns, providing insights beyond traditional feature-based approaches.
Abstract
The study investigates age-related changes in electrocardiogram (ECG) data from a dataset of 1,095 healthy individuals using two different models: an XGBoost model operating on expert-engineered ECG features and an XResNet50 model working with raw ECG signals.
Key findings:
The XGBoost model reveals consistent age-related trends in several ECG features:
Decreasing pNN20, MCVNN, and breathing rate with age
Increasing alpha-fluctuations, P-wave amplitude, and PAS with age
Higher SDANN5 values associated with older individuals
The XResNet50 model consistently focuses on the P-wave, highlighting potential changes in the distribution of different P-wave types with age. It also places relevance on the Q-peak, S-peak, and TP-interval.
The findings from both models align with existing literature on age-related ECG changes, validating the insights. Additionally, the models uncover new relationships, such as the decline in breathing rate and the significance of high SDANN5 values for older individuals.
The use of explainable AI techniques enables a comprehensive understanding of the age-related patterns in ECG data, going beyond traditional feature-based approaches.
Stats
Breathing rate and signal decrease with age.
SDANN5 (average standard deviation of normal-to-normal RR-intervals within a 5-minute interval) values increase with age, with notably high values being more indicative of elderly individuals.
Quotes
"Our model reveals interesting insights regarding the SDANN5 feature. Contrary to established research showing a general decline in SDANN with age, the explainability analysis suggests that high SDANN5-values contribute positively towards the age prediction of older individuals."
"The XResNet50 model consistently demonstrates a predilection for the entire P-wave as individuals age. It specifically focuses on the offset, with some onset in early age groups. These variations may reflect different P-wave types."