A photoplethysmography-based system for talking detection in bedridden patients

Background and objectives: Verbal interaction may help bedridden patients to manage or prevent frustration, anxiety, and depression caused by the restrictions they find when performing daily living activities. In this re-gard, automatic monitoring of how long and often bedridden patients talk could help to identify who is at risk. A considerable body of work has focused on using sensing devices to capture and quantify speech events. However, such approaches may raise privacy concerns and produce discomfort. This study introduces a non-invasive, easy -to-deploy, and privacy-protective system based on photoplethysmography (PPG) to detect talking in bedridden patients. Method: Raw finger PPG signals were acquired from 36 participants who were lying in a bed for six minutes within which they were allowed to talk. We averaged six features extracted from PPG records and investigated statistically significant differences and effect sizes between silence and talking periods. Features showing sta-tistically significant differences and moderate-to-high effect sizes were normalized to train a single perceptron and a binomial logistic regression. Results: The absolute amplitude, the pulse amplitude, and the interpulse interval of PPG waveforms decreased significantly with talking and showed moderate-to-high effect sizes. Using the abovementioned features, the perceptron and the logistic regression achieved classification accuracies of 88.89% and 94.12%, respectively. Conclusions: Results showed that it is possible to detect speech events in individuals with restricted mobility by tracking changes in the PPG signal's contour. Future work should aim to discriminate talking-driven effects on PPG signals during physical activity and establish validation criteria for correctly identifying speech events.