Artificial Intelligence-Enabled Spirometry for Early Detection of Right Heart Failure

TL;DR

A self-supervised representation learning approach combining spirogram time series and demographic data, demonstrating promising potential for early RHF detection in clinical practice is presented.

Abstract

Right heart failure (RHF) is a disease characterized by abnormalities in the structure or function of the right ventricle (RV), which is associated with high morbidity and mortality. Lung disease often causes increased right ventricular load, leading to RHF. Therefore, it is very important to screen out patients with cor pulmonale who develop RHF from people with underlying lung diseases. In this work, we propose a self-supervised representation learning method to early detecting RHF from patients with cor pulmonale, which uses spirogram time series to predict patients with RHF at an early stage. The proposed model is divided into two stages. The first stage is the self-supervised representation learning-based spirogram embedding (SLSE) network training process, where the encoder of the Variational autoencoder (VAE-encoder) learns a robust low-dimensional representation of the spirogram time series from the data-augmented unlabeled data. Second, this low-dimensional representation is fused with demographic information and fed into a CatBoost classifier for the downstream RHF prediction task. Trained and tested on a carefully selected subset of 26,617 individuals from the UK Biobank, our model achieved an AUROC of 0.7501 in detecting RHF, demonstrating strong population-level distinction ability. We further evaluated the model on high-risk clinical subgroups, achieving AUROC values of 0.8194 on a test set of 74 patients with chronic kidney disease (CKD) and 0.8413 on a set of 64 patients with valvular heart disease (VHD). These results highlight the model's potential utility in predicting RHF among clinically elevated-risk populations. In conclusion, this study presents a self-supervised representation learning approach combining spirogram time series and demographic data, demonstrating promising potential for early RHF detection in clinical practice.

Figures (6)

• Figure 1: Overall Design. Our model utilizes clinical data including demographic information and raw spirograms collected from pulmonary tests. The Volume-Time spirograms are then generate to standardized Volume-Flow spirograms. The methodology is structured as a two-stage learning process using the UK Biobank dataset partitioned into training, validation, and testing sets. The proposed model proceeds in two stages: first, the SLSE network training stage, followed by the downstream classification task. The weight set of the online network is $\theta$, and the weight set of the target network is another different set $\xi$. Feature extractor using VAE-encoder. The SLSE network training phase completes the robust representation of the spirogram encoding, and the downstream classification task phase uses the CatBoost classifier with heterogeneous feature fusion as input to complete the prediction of patients with right heart failure.
• Figure 2: Data augmentation method. (a) shows the original capacity-flow curve, (b) shows the curve after Gaussian noise addition, (c) shows the curve after a certain period of amplification after the peak, (d) shows the curve after vertical stretching, (e) shows the curve after horizontal stretching, and (f) shows the curve after downsampling.
• Figure 3: Subgroup analysis. Subgroup analysis is performed on the test set for different subgroups to demonstrate the accuracy of the model classification in different subsets in the population. The subgroups involves age group, sex group, COPD status group, LHF status group, smoke status group, obesity status group, hypertension status group, diabetes status group, chronic kidney disease (CKD) status group, coronary heart disease (CHD) status group, and Valvular Heart Disease (VHD) group. a AUROC performance on the whole test set. b Subgroup analysis on AUROC performance. c Subgroup analysis on RHF prediction probability distribution. "YES" means that the patient has the disease or the condition, and "NO" means that the patient does not have the disease or the condition.
• Figure 4: Ablation experiment. a Ablation experiment of different modules. Comparing the AUROC performance between the methods after deleting each module and the original method on the same test dataset. b Ablation experiment of augmentation methods. It shows the loss obtained by using different combined signal augmentation methods. It can be seen that SLSE method enables the VAE encoder to learn a robust low-dimensional representation of the lung volume map with good anti-interference ability.
• Figure 5: SHAP analysis. (a) Beeswarm diagram. The darker the color, the greater the positive impact of the feature on the model prediction results. (b) Analysis of the dependence of RHF prediction results on age characteristics. The figure shows how age affects the model's prediction of RHF, showing that the elderly are more likely to be predicted as RHF. (c) Analysis of the dependence of RHF prediction results on sex characteristics. The figure shows how sex affects the model's prediction of RHF, showing that males are more likely to be predicted as RHF than females.