Digital respiratory acoustic biomarkers for scalable detection and monitoring of pulmonary congestion in heart failure
D Kotak, D Enriquez-Vazquez, E Barge-Caballero, C Sheppard, M Antunez-Ballesteros, Z Grille-Cancela, Y Raykov, G Barge-Caballero, M J Paniagua-Martin, P Blanco-Canosa, M Padilla-Bautista, J Muniz-Garcia, J M Vazquez-Rodriguez, A Gratiot, M G Crespo-LeiroAbstract
Background
Implantable pulmonary artery pressure monitoring improves outcomes in selected patients with chronic heart failure (HF) but is invasive, costly, and largely restricted to advanced disease. Respiratory acoustic biomarkers provide a complementary, non-invasive assessment of lung mechanics and interstitial fluid changes that may precede haemodynamic decompensation.
Purpose
To evaluate whether smartphone-recorded expiratory acoustics can (i) derive a digital surrogate of NT-proBNP and (ii) enable automated detection of HF-associated acoustic signatures of pulmonary congestion, including sub-audible crackle-like events beyond standard auscultation.
Methods
This retrospective study analyzed single forced-expiratory recordings from a single-site cohort of participants with NYHA classe I–III HF of diverse etiologies (n= 93) and controls (n= 28); age 36–94; 92 male, 29 female). An independent cohort contributed 106 matched controls for training and evaluation. Recordings were acquired using a smartphone-based Eupnoos application. Reference standards included NT-proBNP, NYHA class, and clinician-annotated crackles. Acoustic features capturing time–frequency, spectral, and transient characteristics were extracted. Ridge regression estimated NT-proBNP, while logistic regression, random forest, and neural network models were evaluated for crackle detection and HF classification, with covariate adjustment and cohort reweighting. Performance was assessed by cross-validation and reported as mean (SD) across folds.
Results
Digital NT-proBNP was collected using point-of-care devices and regularized ridge regression of acoustic features derived from the Eupnoos tests, we obtained strong out-of-sample (leave k subjects out) association with laboratory values (n=93; RMSE 155.8 (41.5), MAE 110.4 (32.0), R² 0.9966 (0.0041), Spearman ρ 0.9888 (0.0058)). The acoustic features were also used to train supervised machine learning method to predict HF-associated crackles and achieved average out-of-sample AUC of 0.94 (0.07) with balanced accuracy 0.77 (0.17) (n=17) despite many events being non-audible on routine auscultation.
Acoustic features were also used to train an algorithm to classify HF versus control status, with adjustment for age, sex, and key comorbidities. The classifier achieved accuracy 0.76 (0.02) and AUC 0.61 (0.07) in a single-site cohort (n=121), improving to sensitivity 0.92 (0.06) and specificity 0.82 (0.09) when external controls were included (n=227).
Conclusions
Smartphone-derived respiratory acoustic biomarkers enable digital, scalable, non-invasive detection and monitoring of pulmonary congestion in HF, with applications in remote care and clinical trials.For image description, please refer to the figure legend and surrounding text.