Standardization of Chinese speech tasks for heart failure voice analysis: a cross-sectional diagnostic classification study.

Background: Voice analysis, a non-invasive and cost-effective approach, shows potential for improving heart failure (HF) management. Previous studies mainly focused on non-Chinese voice features, and speech tasks tailored to Chinese HF patients remain lacking. This study aimed to design and optimize speech tasks for Chinese HF patients and establish a foundation for a Chinese HF voice database. Methods: A total of 203 participants (101 HF patients and 102 healthy controls) were recruited from The Affiliated Taizhou People's Hospital of Nanjing Medical University. Based on Chinese phonetics, 12 speech tasks were developed, including vowels, sentences, numbers, and the paragraph. Voice recordings were obtained for all participants. Binary logistic regression was used to analyze factors associated with task completion. Differences in voice features between groups were assessed using independent t-tests (P<0.05), and classification performance was evaluated using a fully connected neural network (FNN) based on the ComParE 2016 and eGeMAPSv02 feature sets. Results: Among all participants, 105 (51.7%) completed all tasks (Group A), whereas 98 (48.3%) failed due to reading (75.5%) or execution problems (67.3%). Regression analysis revealed that age [middle-aged, odds ratio (OR) =3.716; young, OR =27.28], education (college or higher, OR =6.919), and task difficulty (easy, OR =14.185) significantly influenced completion rates (P<0.05). After removing six low-completion tasks (tasks 2, 3, 6, 8, 10, and 12), an optimized six-task set (Group B) increased the completion rate to 92.1%, particularly for participants with lower education (6.9% to 35.5%) and older age (15.3% to 50.7%). The ComParE 2016 feature set extracted 1,617 (task "i") to 3,039 (task "mm") features, while the eGeMAPSv02 feature set extracted 19 (task "i") to 53 (task "mm") features. Tasks "pg", "mm", and "1" demonstrated greater discriminative capacity (F-values are 67.5, 54.4, and 49.4, respectively) and superior classification performance (F1 scores are 0.817, 0.813, and 0.774, respectively). FNN models based on ComParE 2016 achieved higher performance than those based on eGeMAPSv02 (P<0.05), but there was no significant difference in the FNN models based on two datasets after demographic adjustment (P=0.16). Conclusions: This study developed standardized Chinese speech tasks for HF patients, identified factors influencing task completion, and demonstrated the feasibility of voice-based classification using both ComParE 2016 and eGeMAPSv02 feature sets. These findings support the development of large-scale voice databases and artificial intelligence models for early detection and management of HF in Chinese populations.