In an exploratory retrospective study, researchers evaluated five supervised machine learning algorithms to screen for unrecognized heart failure with preserved ejection fraction among patients hospitalized with hypercapnic chronic obstructive pulmonary disease. The study analyzed clinical and echocardiographic data from 441 patients treated at Panzhihua Central Hospital, identifying concurrent heart failure with preserved ejection fraction in 123 individuals. By integrating routine physiological, demographic, and echocardiographic variables, an ensemble Random Forest model achieved the highest discriminatory accuracy among tested models, recording a testing area under the receiver operating characteristic curve of 0.940. The findings demonstrate how computational models using routinely collected clinical data can address diagnostic overlap between advanced respiratory disease and concealed cardiac impairment.
Chronic obstructive pulmonary disease represents a progressive respiratory disorder characterized by chronic airway inflammation and persistent airflow limitation. As lung function deteriorates, a substantial proportion of patients experience chronic hypercapnia, defined by elevated arterial carbon dioxide tensions. Elevated partial pressure of carbon dioxide reflects severe ventilatory compromise and associates with pulmonary vascular alterations, elevated pulmonary arterial pressures, and broader systemic manifestations across the cardiovascular system. Concurrently, heart failure with preserved ejection fraction frequently coexists with advanced pulmonary disease, yet its detection remains difficult because primary symptoms such as exertional dyspnea, fatigue, and exercise limitation overlap almost entirely with the baseline manifestations of severe chronic obstructive pulmonary disease.
Diagnostic adjudication of cardiac status in hypercapnic pulmonary disease faces structural and physiological obstacles during standard clinical workups. Pulmonary hyperinflation and severe thoracic distortion frequently degrade acoustic windows, impeding transthoracic echocardiographic visualization of left ventricular filling pressures and structural remodeling. Furthermore, standard single-parameter screening thresholds, such as resting early diastolic transmitral to mitral annular velocity ratios or baseline B-type natriuretic peptide cutoffs, often lose specificity in hypercapnic populations because chronic alveolar hypoxia and elevated carbon dioxide stimulate pulmonary vasoconstriction and right ventricular afterload, inducing mild natriuretic peptide elevations even in the absence of primary left ventricular failure. To overcome these diagnostic ambiguities, an automated triage framework using multidimensional clinical profiles was established and compared against consensus diagnostic standards.
The study cohort was drawn from an initial pool of 512 consecutive patients with hypercapnic chronic obstructive pulmonary disease retrospectively screened at a single medical center. After excluding 71 patients due to reduced or mildly reduced ejection fractions below 50 percent, severe valvular pathology, primary pulmonary arterial hypertension, advanced renal impairment, or incomplete diagnostic records, 441 eligible patients remained for complete-case evaluation. Reference diagnoses of concurrent heart failure with preserved ejection fraction were established by two independent cardiologists applying the multi-domain European Society of Cardiology HFA-PEFF algorithm, requiring preserved ejection fractions alongside diagnostic morphological, functional, biomarker, and diuretic response criteria. Baseline measures encompassed 16 demographic, arterial blood gas, spirometric, and echocardiographic parameters gathered during standardized clinical periods.
To develop predictive algorithms, the complete-case cohort was randomly partitioned into a 70 percent training subset comprising 308 individuals and an internal hold-out testing set of 133 individuals. To counter moderate class imbalance within the training data, investigators implemented the Random Over-Sampling Examples technique within cross-validation folds, generating synthetic minority observations through smoothed bootstrap kernel density estimation to prevent synthetic data leakage into evaluation folds. Five distinct supervised algorithms were trained and calibrated across repeated cross-validation frameworks: Random Forest, XGBoost, Support Vector Machines with radial basis function kernels, LightGBM, and standard Logistic Regression. Hyperparameter optimization across all algorithms was conducted via exhaustive grid search strictly confined to the repeated internal training folds.
Evaluation within the independent testing set demonstrated substantial performance variations across algorithmic architectures. The Random Forest model exhibited the strongest discrimination, recording an area under the receiver operating characteristic curve of 0.940, accompanied by a 95 percent confidence interval spanning 0.898 to 0.982, while maintaining an average cross-validation area under the curve of 0.928 across ten repetitions. XGBoost performed similarly with an area of 0.920, whereas Support Vector Machines and LightGBM registered areas of 0.812 and 0.782, respectively. Conventional Logistic Regression achieved a significantly lower area under the curve of 0.642. In precision-recall assessments designed for imbalanced cohorts, the Random Forest model maintained the top area under the precision-recall curve at 0.912, significantly outperforming linear modeling approaches.
To evaluate whether the predictive capacity of the Random Forest model stemmed merely from repeating diagnostic echocardiographic markers or natriuretic peptide measurements, investigators conducted a sensitivity ablation analysis excluding both amino-terminal pro-B-type natriuretic peptide and the early diastolic velocity ratio from the input matrix. Within this restricted 14-variable parameter framework, the ablated Random Forest model retained an area under the curve of 0.846 in the testing cohort, confirming that algorithmic performance relied upon integrated cardiopulmonary interactions rather than isolated cardiac indicators. Model classification at an optimal out-of-fold probability cutoff of 0.58 demonstrated symmetrical diagnostic distributions in the test cohort, while decision curve analysis indicated positive net clinical benefit across a broad range of threshold probabilities relative to default clinical assumptions.
Model interpretability was interrogated using SHapley Additive exPlanations, translating algorithmic decision-making into quantified statistical feature contributions. Global importance assessments identified percent predicted forced expiratory volume in one second, arterial carbon dioxide tension, and amino-terminal pro-B-type natriuretic peptide as the primary mathematical contributors to elevated predictive probabilities. Local waterfall plots illustrated individualized risk attributions, while dependence plots revealed pronounced non-linear associations, including steep risk increments when carbon dioxide tensions exceeded 55 millimeters of mercury and natriuretic peptide concentrations surpassed 400 picograms per milliliter. Bivariate interaction evaluations demonstrated marked upward deviations in predictive values when elevated carbon dioxide levels coincided with higher body mass index values, reflecting a potential mathematical synergy characteristic of obesity-hypoventilation profiles.
Investigating cohort heterogeneity beyond binary diagnostic classifications, researchers performed unsupervised K-means consensus clustering across 1,000 resampling iterations using standardized pathophysiological biomarkers. Quantitative clustering criteria identified three distinct clinical sub-phenotypes across the 441 patients. Phenotype 1 represented 38.1 percent of the cohort, characterized by mild hypercapnia, normal-to-low body mass index, and baseline natriuretic peptide levels. Phenotype 2 encompassed 33.8 percent of individuals, displaying severe hypercapnia, moderate natriuretic peptide increases, and elevated body mass index values defining an obesity-hypoventilation dominant profile. Phenotype 3 comprised 28.1 percent of patients, presenting with extreme natriuretic peptide elevations, substantial hypercapnia, and lower body mass indices, consistent with severe cor pulmonale and cardiac dysfunction.
Prognostic relevance of the identified phenotypic clusters was verified through survival tracking and multivariable Cox proportional hazards modeling evaluating event-free survival. Multivariable regressions adjusting for age above 70 years, male sex, heavy smoking history, and severe baseline airflow limitation revealed significant divergence in clinical trajectories. Assignment to Phenotype 3 was independently associated with a 3.20-fold increase in the hazard of adverse clinical outcomes relative to Phenotype 1, while Phenotype 2 carried an adjusted hazard ratio of 1.85. The original clustering variables were excluded from regression specifications to prevent mathematical multicollinearity, confirming that unsupervised algorithmic phenotyping offers additive prognostic stratification independent of conventional demographic and spirometric markers.
Despite promising predictive metrics, the authors emphasized several methodological constraints requiring cautious interpretation. The investigation was conducted retrospectively at a single clinical institution with a moderate sample size of 441 patients, lacking external validation cohorts from independent healthcare environments to rule out model optimism and institutional overfitting. Furthermore, the complete-case inclusion criterion necessitated comprehensive baseline echocardiography and biomarker testing, potentially introducing selection bias toward more acutely symptomatic individuals with higher pre-test cardiac suspicion. The authors noted that until prospective, multicenter external validation and calibration studies are completed, these machine learning algorithms and phenotypic stratifications serve as hypothesis-generating investigative frameworks rather than clinically deployable automated decision systems.
Subject of Research: Machine learning detection and phenotypic clustering of concurrent heart failure with preserved ejection fraction in hypercapnic chronic obstructive pulmonary disease
Article Title: Interpretable machine learning for the identification and prognostic phenotyping of concurrent HFpEF in hypercapnic COPD
Article References: Deng, C., Du, Y., Liu, J., Liu, H., Min, P., Wang, J., & Cheng, X. (2026). Interpretable machine learning for the identification and prognostic phenotyping of concurrent HFpEF in hypercapnic COPD. Frontiers in Medicine, 13, Article 1990320. https://doi.org/10.3389/fmed.2026.1990320
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1990320
Keywords: chronic obstructive pulmonary disease, heart failure with preserved ejection fraction, hypercapnia, machine learning, Random Forest, SHAP analysis, phenotypic clustering, NT-proBNP, pulmonary medicine, cardiology, prognostic stratification, retrospective study, HFA-PEFF algorithm, body mass index, event-free survival, ventilatory failure
News Source: Teresa Odom. (November 17, 2026). Machine Learning Model Detects Hidden Heart Failure in Patients With Hypercapnic COPD. Scienmag.



