Predicting extremely low body weight from 12-lead electrocardiograms using a deep neural network
Ken Kurisu1, Tadahiro Yamazaki1, Kazuhiro Yoshiuchi2
1Department of Stress Sciences and Psychosomatic Medicine, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan.
Scientific Reports
|February 27, 2024
Summary
Deep learning can predict extremely low body weight using 12-lead electrocardiograms (ECGs). The model identified QRS waves and lower QRS voltage as key indicators in patients with anorexia nervosa.
Area of Science:
- Cardiology
- Artificial Intelligence
- Medical Informatics
Background:
- Previous research successfully used deep learning to predict overweight status from 12-lead ECGs.
- Predictive models for underweight status using ECG data have not been explored.
- Extremely low body weight is a critical health concern, particularly in conditions like anorexia nervosa.
Purpose of the Study:
- To assess the feasibility of deep learning for predicting extremely low body weight (BMI ≤ 12.6 kg/m²) using 12-lead ECGs.
- To investigate the rationale behind ECG-based predictions by identifying relevant ECG features.
- To explore the association between ECG characteristics and extremely low body weight in patients with anorexia nervosa.
Main Methods:
- A convolutional neural network (CNN) was trained on 12-lead ECG data from inpatients, primarily those with anorexia nervosa.
- The CNN was designed to output a binary prediction of whether a patient's body mass index (BMI) was ≤ 12.6 kg/m².
- Gradient-weighted class activation mapping (Grad-CAM) was used to visualize the ECG regions influencing the model's predictions.
Main Results:
- The CNN model achieved an area under the receiver operating characteristic curve (AUC) of 0.807 on the test dataset, indicating good predictive performance.
- Gradient-weighted class activation mapping highlighted the QRS waves as the primary focus of the CNN's prediction.
- A negative correlation was observed between QRS voltage and the model's prediction scores, suggesting lower QRS voltage is associated with extremely low body weight.
Conclusions:
- Deep learning models are feasible for predicting extremely low body weight using standard 12-lead ECGs.
- Specific ECG features, notably lower QRS voltage, are associated with extremely low body weight in patients with anorexia nervosa.
- This approach offers a potential non-invasive method for identifying individuals at risk of severe underweight conditions.
