This study aims to estimate CPF from cough sounds which were measured by smartphone using artificial intelligence (AI) techniques such as artificial neural networks (NN).
The study included 65 participants without respiratory or cardiovascular diseases, approved by the Hiroshima Cosmopolitan University Ethics Committee (S20160310). CPFs were measured using a spirometer (AS-507), and cough sounds were measured on an iPhone 6 with a custom app, calculating the maximum sound pressure level (SPL). We developed a NN model to estimate CPF and compared it with random forest (RF) and multiple regression (Reg) models. Data were split into 80% training and 20% testing sets. Explanatory variables were optimized using ElasticNet, and hyperparameters of the NN and RF were tuned by the grid search. The dependent variable was CPF, while explanatory variables included age, gender, height, weight, and SPL. Estimation accuracy was assessed using the coefficient of determination (R²). Absolute errors were compared using repeated measures ANOVA and Tukey’s HSD test. Systematic errors were evaluated with Bland-Altman plots, Pearson correlation, and one-sample t-tests. All analyses were conducted with Python 3.0, with a significance level of 0.05. Results were presented as mean values with 95% confidence intervals.
The absolute error for CPF estimated by Reg was 42.9 (22.2 – 63.7) L/min, by RF was 69.9 (45.9 – 94.0) L/min, and by NN was 35.9 (18.9 – 52.9) L/min. Repeated measures ANOVA showed significant differences among models (p = 0.0089). The NN had significantly lower errors compared to the RF (p = 0.036), with no other significant differences noted. The R² values for the NN, Reg, and RF models were 0.918, 0.913, and 0.839, respectively. Bland-Altman plots indicated proportional errors for the RF and Reg models (r = -0.793, p = 0.001; r = -0.654, p = 0.015, respectively), but not for the NN model (r = -0.253, p = 0.404). One-sample t-tests confirmed no bias errors (p > 0.05).
The NN model estimates CPF with higher accuracy and without systematic errors compared to Reg and RF models. Incorporating the proposed model into a smartphone app enables easy and non-facial contact assessment of cough strength. Future studies should include larger, diverse populations to improve generalizability and examine the impact of comorbidities. Refining the model with additional variables or advanced AI techniques could enhance accuracy.
The proposed method provides a quick, non-facial contact tool for assessing cough strength, valuable in physiotherapy for early detection of reduced cough effectiveness, especially in patients with neuromuscular disorders, respiratory conditions, or the elderly. It would support remote assessments, aligns with telehealth, and enables clinicians to monitor patients without frequent visits, enhancing care in limited-contact settings.
Neural Networks
Smartphone
