A simplified three-feature Pre-CPD index model (MEF25 %pred, FEV1/FVC, and EF) achieved 89.03% predictive accuracy for cardiopulmonary events, outperforming existing clinical concepts.
Cohort (n=2,370)
Does a machine learning-derived Pre-Cardiopulmonary Dysfunction (Pre-CPD) index accurately predict cardiopulmonary events in asymptomatic individuals without baseline disease?
A machine learning-derived Pre-Cardiopulmonary Dysfunction (Pre-CPD) index using three functional features accurately identifies high-risk asymptomatic individuals, offering a potential tool for early prevention.
Abstract Background Cardiovascular and respiratory diseases are leading global causes of death worldwide, and functional decline in these systems often begins decades before clinical diagnosis. Thus, intervention during this subclinical stage is critical. However, existing early identification concepts are largely limited to single organ/system, overlooking the interaction and frequent co-impairment of cardiopulmonary function. Therefore, there is an urgent need to characterize pre-cardiopulmonary dysfunction (Pre-CPD). Methods This study utilized unsupervised clustering and supervised learning models on 2370 individuals without baseline cardiopulmonary disease from the Xiangya Cardiopulmonary Health and Disease Cohort (XY-CPHDC). Constrained k-means clustering was applied to 3,080 combinations of features across four modules (pulmonary function, echocardiography, electrocardiogram, and serum lipids) to identify clinically heterogeneous phenotypes. Cluster labels were used as supervised targets to build logistic regression models of varying complexity, thereby developing the Pre-CPD Index. The model was validated in a cardiopulmonary event-driven test set (n = 237) and compared with existing clinical concepts. Findings Among 2,133 participants in the train and validation sets, systematic evaluation of multidimensional physiological indices identified an optimal feature set comprising seven key indicators and revealed four heterogeneous cardiopulmonary phenotypes:Cluster 1:Preserved cardiopulmonary function (event rate 1.5%);Cluster 2: High lung volume with preserved cardiac function (event rate 2.49%);Cluster 3: Isolated small airway dysfunction with compensatory cardiac enhancement (event rate 3.32%);Cluster 4: Combined cardiopulmonary impairment (event rate 5.69%).Cluster 4 showed the lowest overall functional metrics (e.g., EF 60.0%, FVC %pred 102.1%), whereas Cluster 3 demonstrated markedly reduced small airway function (MEF25 %pred 63.3%) but elevated cardiac performance (EF 70.0%). Based on these phenotypes, multiple Pre-CPD index models were developed and evaluated. The simplified three-feature model, consisting of MEF25 %pred, FEV1/FVC, and EF, achieved the highest predictive accuracy (89.03%) in the test set (n = 237), substantially outperforming existing clinical concepts. Interpretation We build a concise and quantitative tool, pre-CPD index, to quantify high-risk asymptomatic individuals with cardiopulmonary function deline through machine learning approach, which has the potention to be utilized for early prevention and precision intervention in cardiopulmonary diseases. This abstract is funded by: This work was supported by the National Key Research and Development Program (No. 2022YFC3601001) and National Multidisciplinary Cooperative Diagnosis and Treatment Capacity Building Project for Major Diseases (Lung Cancer) (No. z027002)
Zhang et al. (Fri,) conducted a cohort in Pre-cardiopulmonary dysfunction (n=2,370). Pre-CPD Index (simplified three-feature model) vs. Existing clinical concepts was evaluated on Predictive accuracy in the cardiopulmonary event-driven test set. A simplified three-feature Pre-CPD index model (MEF25 %pred, FEV1/FVC, and EF) achieved 89.03% predictive accuracy for cardiopulmonary events, outperforming existing clinical concepts.