Prediction of state anxiety by machine learning applied to photoplethysmography data

Abstract

Background: As the human behavior is influenced by both cognition and emotion, affective computing plays a central role in human-machine interaction. Algorithms for emotions recognition are usually based on behavioral analysis or on physiological measurements (e.g., heart rate, blood pressure). Among these physiological signals, pulse wave propagation in the circulatory tree can be assessed through photoplethysmography (PPG), a non-invasive optical technique. Since pulse wave characteristics are influenced by the cardiovascular status, which is affected by the autonomic nervous activity and hence by the psychophysiological state, PPG might encode information about emotional conditions. The capability of a multivariate data-driven approach to estimate state anxiety (SA) of healthy participants from PPG features acquired on the brachial and radial artery was investigated.

Methods: The machine learning method was based on General Linear Model and supervised learning. PPG was measured employing a custom-made system and SA of the participants was assessed through the State-Trait Anxiety Inventory (STAI-Y) test.

Results: A leave-one-out cross-validation framework showed a good correlation between STAI-Y score and the SA predicted by the machine learning algorithm (r = 0.81; p = 1.87∙10^-9). The preliminary results suggested that PPG can be a promising tool for emotions recognition, convenient for human-machine interaction applications.

Keywords: Affective computing; General linear model (GLM); Machine learning; Photoplethysmography (PPG); State-trait anxiety inventory (STAI-Y).

Figure 1. (A) PPG system employed; (B) PPG probe were made of a LED at 940 nm of wavelength and a silicon photomultiplier (SiPM).

Figure 2. Schematic representation of the preprocessing steps of PPG signals and an example of the pulse average PPG with associated standard errors (shaded areas).

Figure 3. Histogram of the distribution of SA estimated by means of STAI-Y.

Figure 4. Histogram of the distribution of correlation coefficients estimated with a bootstrap procedure.

Figure 5. (A) Correlation and (B) Bland–Altman plot of SA and SAp for the GLM based estimation.

Figure 6. (A) Histogram reporting the distribution of AUC obtained with a Bootstrap procedure; (B) ROC of SAp in assessing high contingent anxiety (SA > 40).