JMIR Publications

ABSTRACT

Background:

Estimated pulmonary biological age (ePBA) has emerged as a more reliable indicator for disease progression and mortality than chronological age, with chest CT as a promising tool for calculating ePBA. However, the lack of models trained and validated with large-scale healthy adults hinders the generalizability of the CT-based ePBA.

Objective:

The aim of this study was to develop a biomarker of ageing—estimated pulmonary biological age (ePBA)—from chest CTs of multi-center healthy adults using deep learning models.

Methods:

11187 chest CTs of healthy adults from 3 health management centers were enrolled to develop a biomarker of ageing —ePBA based on a combination of convolutional neural network (CNN) and long short-term memory (LSTM) networks. Among them, 7726 chest CTs from one center (Institution A) were divided into the training dataset (n=5408), the tuning dataset (n=787) and the internal test dataset (n=1531) to construct the model. And data from another two institutions were used for external test datasets (Institution B: n=1506; Institution C: n=1955). The CNN architecture respectively used VGG11, ResNet18, and ConvNeXT frameworks. The performance of the regression model was evaluated using the correlation coefficient (r), coefficient of determination (R2), and mean absolute error (MAE).

Results:

The external test dataset of healthy individuals showed a strong correlation between ePBA derived from the VGG11-LSTM deep learning model and chronological age (Institution B: r = 0.97 [99% CI: 0.96-0.97]; Institution C: r = 0.98 [99% CI: 0.96-0.99] ) and good model performance (Institution B: R2= 0.93[99% CI: 0.92-0.94], MAE = 4.59 [99% CI: 4.34-4.83]; Institution C: R2= 0.93[99% CI: 0.84-0.94], MAE = 3.64 [99% CI: 3.47-3.81]). These results demonstrated higher correlation coefficient, coefficient of determination, and lower MAE values than those of the ResNet18-LSTM and ConvNeXT-LSTM deep learning models.

Conclusions:

This study developed and validated the biomarker of ageing—ePBA— with a CNN-LSTM deep learning model based on chest CT, in which VGG11-LSTM model demonstrated the best efficiency.