A dynamic architecture linking autonomic activity to emotional dimensions: Real-time estimation from photoplethysmography

A dynamic architecture linking autonomic activity to emotional dimensions: Real-time estimation from photoplethysmography | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 27 January 2026 V1 Latest version Share on A dynamic architecture linking autonomic activity to emotional dimensions: Real-time estimation from photoplethysmography Authors : Shane R. McClafferty 0009-0009-3354-1130 and Bruce Friedman 0000-0003-0814-0393 [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.176951752.28031887/v1 220 views 96 downloads Contents Abstract Abstract Psychological Monitoring: Detecting Real-Time Emotional Changes Method Results Activation Valence Motivation Discrete Emotion Detection Discussion Conclusion Tables and Figures Intercorrelations Discrete-Dimensional Structure of Emotion Mean and 95% Confidence Interval of Predicted Dimensions per Film Mean Discrete Emotion Probabilities Appendix A Descriptive Subject Statistics of Continuous Variables Menstrual Phase Comfort With Being Emotionally Monitored Appendix B Appendix C References Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract The present study introduces a novel, real-time emotion (or generalized behavioral) tracking model capable of independently estimating valence, motivation (approach-avoidance), and activation (arousal) through estimations of parasympathetic (PNS), α-adrenergic (α-SNS), and β-adrenergic (β-SNS) sympathetic nervous system activity. These separate autonomic systems correspond to distinct emotional dimensions: valence to the PNS, motivation (approach-avoidance) to the α-SNS, and activation to the β-SNS. This framework enables continuous, real-time, and interpretable estimation of emotion from a single, wearable-compatible signal through photoplethysmography (PPG). The tested model utilizes inter-beat interval (IBI) and pulse wave or pulse volume amplitude (PVA), which are tracked using an extended Kalman filter (EKF) to extract frequencies (VLF-UHF) and standard heart rate variability metrics (HRV). These features were mapped onto emotional dimensions using supervised partial least squares (PLS) regressions from behavioral validation measures: facial electromyograph (EMG) for valence, a joystick for motivation, and eye-tracking (activation). The dimensional predictions reached within-subject accuracy levels comparable to those of traditional physiological models. Additionally, these emotional dimensions can be used to produce reasonable, discrete emotion probabilities based solely on theory (without requiring training data). These findings support a new model of emotion based on separate autonomic systems and dimensions that functionally define emotions in real time. Such an approach enables dynamic emotional inference or generalized behavior across various contexts, including experimental design, clinical monitoring, and ambulatory assessment, utilizing low-cost, wearable technology. Department of Psychology, Virginia Tech Author Note Shane R. McClafferty https://orcid.org/0009-0009-3354-1130Bruce H. Friedman https://orcid.org/0000-0003-0814-0393 This study used patent-pending technology (US 19/227,299). The patent does not restrict academic use or replication of the reported analyses. The authors would like to thank Casey Kozan, William Carlson, Shirin Mohammadian, Priya Mittal, Matt Summerson, Xinzhu Chai, Maddeline Netto, Anoushka Jawdekar, Sophia Barraza, Wanangwa Nyirongo, Clare O’Herron, and Jason Pittman, for their assistance with data collection and recruitment. Correspondence concerning this article should be addressed to Bruce H. Friedman, Department of Psychology (0436), 109 Williams Hall, Blacksburg VA 24061. Email: [email protected] Abstract The present study introduces a novel, real-time emotion (or generalized behavioral) tracking model capable of independently estimating valence, motivation (approach-avoidance), and activation (arousal) through estimations of parasympathetic (PNS), α-adrenergic (α-SNS), and β-adrenergic (β-SNS) sympathetic nervous system activity. These separate autonomic systems correspond to distinct emotional dimensions: valence to the PNS, motivation (approach-avoidance) to the α-SNS, and activation to the β-SNS. This framework enables continuous, real-time, and interpretable estimation of emotion from a single, wearable-compatible signal through photoplethysmography (PPG). The tested model utilizes inter-beat interval (IBI) and pulse wave or pulse volume amplitude (PVA), which are tracked using an extended Kalman filter (EKF) to extract frequencies (VLF-UHF) and standard heart rate variability metrics (HRV). These features were mapped onto emotional dimensions using supervised partial least squares (PLS) regressions from behavioral validation measures: facial electromyograph (EMG) for valence, a joystick for motivation, and eye-tracking (activation). The dimensional predictions reached within-subject accuracy levels comparable to those of traditional physiological models. Additionally, these emotional dimensions can be used to produce reasonable, discrete emotion probabilities based solely on theory (without requiring training data). These findings support a new model of emotion based on separate autonomic systems and dimensions that functionally define emotions in real time. Such an approach enables dynamic emotional inference or generalized behavior across various contexts, including experimental design, clinical monitoring, and ambulatory assessment, utilizing low-cost, wearable technology. Keywords: photoplethysmography (PPG), heart rate variability (HRV), pulse volume amplitude (PVA), emotion dimensions, real-time emotion tracking, Psychological Monitoring: Detecting Real-Time Emotional Changes Despite advances in data collection and modeling, behavioral science struggles with anticipating when, why, and how people act. This is evident in high-stakes domains such as suicide or crisis prevention (Asarnow et al., 2017; Goldstein, 1991; Lekkas et al., 2021) and violence risk assessment (Catchpole & Gretton, 2003; Coid et al., 2009; Rossegger et al., 2011), in which outcomes are objectively defined, yet prediction models are often only slightly better than chance despite known risk factors. These limitations also cast doubt on lower-stakes practices in which outcomes are less certain and typically rely on self-report, group trends, proxy variables, or retrospective interpretation. The core issue is a lack of high-resolution data (especially temporal resolution) and a dimensional structure to determine the direction of behavioral changes, which is particularly important in capturing the structure or cause of behavior. Good behavioral prediction requires representations that are relevant to adaptation, temporally continuous, and capable of capturing the direction and rate of change over time (Amer et al., 2020; Azorin-Lopez et al., 2014; Sarker et al., 2020). Existing emotion trackers typically lack at least one of the following major resolutions: continuous/scaled (versus categorical) or temporal/constant (versus explicitly defined time points or long intervals) (Amer et al., 2020; Kragel & LaBar, 2013; Stephens et al., 2010). The model architecture investigated in this study is based on three core emotional dimensions: activation (arousal), valence (pleasantness/unpleasantness), and motivation (approach/avoidance) (Christie & Friedman, 2004; Fontaine et al., 2007). When the dimensions are tracked over time they can combine to describe adaptive behavioral directions such as to improve (positive approach), destroy (negative approach), relax (positive avoidance), or escape (negative avoidance), with activation modulating the likelihood of action. These dimensional interactions provide a functional basis for predicting behaviorally relevant emotions like anger, fear, joy, or sadness. Despite the vast literature on valence arousal models, the simultaneous estimation of valence and motivation (without self reports) is rare. Preliminary work attempted to do so, motivating the present study (McClafferty & Friedman, 2024). Physiological measurement devices, such as photoplethysmography (PPG) would be well suited for such dimensional estimation since it can measure both cardiac timing and peripheral resistance from a single non-invasive sensor (Castaneda et al., 2018). The present study tested whether the three main emotional dimensions could be measured using autonomic nervous system (ANS) indices derived from PPG, with a model that describes how physiological measures can be combined with emotional dimensions. Emotion Emotions have adaptive functions that give situation-relevant signals to the brain and body (Friedman & Thayer, 2024; Schmidt & Cohn, 2008). Specific emotions are identified for their adaptiveness in unique situations and their recognizability across human societies/cultures, which include happiness/joy, contentment, sadness, fear, and anger. These adaptive emotions also fall along emotional dimensions. Most dimensional models rely on some combination of valence and arousal/activation to form a circumplex (Ascheid et al., 2019; Barrett & Russell, 1999; Davidson, 1993, 2005; Panayiotou, 2008; Simhon et al., 2025). However, the circumplex is weak in the functional separation of discrete emotions (notably anger and fear; see Table 1). In contrast, adding a third dimension reflecting behavioral direction helps to distinguish specific emotions, particularly aggression and fear, which have similar valence and arousal (Ascheid et al., 2019; Davidson, 1993, 2005; Hoofs et al., 2019; Panayiotou, 2008). Unlike a model with only valence and arousal, aggression (approach) and fear (avoidance) fall at opposite ends of the motivation dimension, allowing specific emotions to be clearly distinguishable using all three emotional dimensions (compare Tables 1 and 2). The differences in physiology for these specific emotions also match their differences in emotional dimensions (Kreibig, 2010). Therefore, it may be possible to measure these emotional dimensions, especially in the context of specific emotions, using physiological changes. Autonomic Nervous System The ANS is uniquely relevant to emotion due to its influence both on and from the Central Nervous System (CNS) and its variability during emotional changes (Kreibig, 2010). The ANS is typically separated into two main branches: the Sympathetic Nervous System (SNS) and the Parasympathetic Nervous System (PNS) (Langley, 1921). Higher stress is thought to indicate SNS activation and PNS inhibition (Berntson et al., 1994). However, there is differing evidence showing increased PNS activity for some high-activation emotions (increased SNS activity) and decreased PNS activity for some low-activation emotions (decreased SNS activity), indicating that some emotions have coactivation or coinhibition of the SNS and PNS (Berntson et al., 1994; Kreibig, 2010). Coactivation is observed primarily in high-activation emotions with positive valence and in negative conditions after habituation. In contrast, emotions with coinhibition are characterized by low activation and negative valence. If considered separately, then, PNS activity generally initiates digestive functions, slows breathing, and reduces heart rate. Overall, the PNS appears to induce relaxation or satiety in the body. The SNS can be distinguished by receptor subtypes, which vary by function and nerve locations, primarily between α (alpha) and β (beta) receptors (Wehrwein et al., 2016). Activation of the α-SNS helps to increase hemostasis (blood clotting; Aslam et al., 2023) and prevent hemorrhaging by increasing vasoconstriction (Wehrwein et al., 2016). It also increases constriction of blood vessels of the skin, muscle, and mucosal membranes, which naturally reduces absorption of particles in the skin, nose, and lungs for protection. Alternatively, activation of the β-SNS increases the rate and contractility of the heart, expands the bronchioles of the lungs to enhance blood flow and oxygenation, and increases energy availability in the body. These functions prepare the body for more demanding behavioral or mental activities. When the SNS is separated by α-SNS and β-SNS activity, only β-SNS activity continues to correlate with high-activation emotions (Kreibig, 2010; Stemmler et al., 2007). However, α-SNS activity positively correlates with motivation, where high approach tendencies are associated with higher α-SNS activity. These findings suggest that the ANS systems can be distinguished by emotional dimensions (see Table 1). Mapping ANS Systems to Emotional Dimensions The amalgamation of findings across studies enables the conceptualization of equivalence between emotional dimensions and separate ANS activities (Christie & Friedman, 2004; Kreibig, 2010; Stemmler et al., 2007). Kreibig (2010) found that only the β-SNS differed between high and low activation emotions, and that most positive emotions were associated with increases in PNS, and vice versa. In addition, α-SNS may be the closest estimator of motivation; as shown by avoidance versus approach in both anger and fear being best defined by PVA (Stemmler et al., 2007; see Table 2 or Figure 1 for this representation). Christie and Friedman (2004) found this relationship more directly through a discriminant analysis of discrete emotions. Their model generated three discriminant factors to differentiate the discrete emotions, two of which were associated with emotional dimensions. The activation factor was associated with skin conductance, systolic blood pressure, and heart rate variability (mean successive differences), indicating the influence on β-SNS and cholinergic-SNS/sweat gland pathways. For motivation, mean arterial pressure, diastolic blood pressure, and decreasing IBI were related to withdrawal. Reducing IBI (i.e., increasing heart rate) is crucial for maintaining blood pressure during vasoconstriction, which may explain why it was also correlated with this factor (Wehrwein et al., 2016). In addition, while PVA is a more direct measure of peripheral resistance diastolic pressure and mean arterial pressure provide a close estimate (motivation), but systolic pressure is more influenced by heart contractility or stroke volume (activation). These findings support the conclusion that Christie and Friedman’s (2004) motivation function follows the α-SNS activation, in which higher avoidance is related to higher α-SNS activity, and their activation measure followed β-SNS and the cholinergic sweat response. These dimensions should be interpreted as dominant autonomic biases rather than exclusive physiological sources. Effective Real-Time Measurement of ANS Systems Many of these studies relied on multiple measures of physiological activity which are more resource and time intensive. More importantly, simpler physiological devices and setup will increase the use of physiology in other studies. Studies attempting to estimate ANS activity typically use heartbeat detectors such as PPG (especially for MRI) or electrocardiography (ECG), often combined with respiration. However, these approaches primarily target PNS (vagal) activity via respiratory sinus arrhythmia (RSA) and frequently conflate PNS and SNS influences (often only referencing IBI/heart rate). Studies that aim to dissociate multiple ANS branches often incorporate additional sensors such as electrodermal activity (EDA), blood pressure, electromyography (EMG), impedance cardiography (ICG), and skin temperature, all of which increase cost and complexity. Yet, recent research suggests that the three primary ANS branches can be separately estimated using cardiovascular features: RSA or high frequency (HF) IBI for PNS, PVA for α-SNS, and very low frequency (VLF) IBI or IBI for β-SNS (Kreibig, 2010; Stemmler et al., 2007; Valenza et al., 2018). While ECG can capture PNS and potentially β-SNS activity via heart rate variability (HRV) and IBI, it lacks access to vascular signals that help to estimate α-SNS. In contrast, PPG provides a low-cost, non-intrusive method for capturing both heartbeat timing and vasculature simultaneously (Castaneda et al., 2018). It enables estimation of PNS through HF IBI, α-SNS through PVA, and β-SNS through UHF or VLF IBI variability. These distinctions reflect core emotional dimensions: PNS activity is preferentially indexed by valence, α-SNS activity by motivational tendency (positively correlating with PVA), and β-SNS activity by activation (see Table 2 and Figure 1). The present study investigated whether PPG, as a single, low-cost sensor, can serve as a real-time window into an individual’s emotional state, capturing distinct contributions from valence, motivation, and activation. If successful, this approach would provide a scalable, noninvasive solution for real-time emotion inference using only wearable PPG sensors. Estimating ANS The activation levels of the PNS, α-sympathetic, and β-sympathetic systems can be estimated separately using frequency-specific features of cardiovascular signals. Whereas both the PNS and β-SNS regulate IBI, they exhibit different temporal dynamics since β-SNS has a more delayed effect than the PNS, due to its action through catecholamine diffusion and receptor binding (Wehrwein et al., 2016). These separate latencies enable the two systems to be functionally dissociated in frequency domains. Specifically, the PNS produces smaller rhythmic IBI variability, aligned with respiration, referred to as RSA, which is primarily regulated by PNS (i.e., vagus nerve) activity (Shaffer & Ginsberg, 2017). This effect can be captured by the high frequency (HF; 0.15-0.4 Hz) IBI variability or by relating IBI to the respiration phase. In contrast, β-SNS, although slower to activate, results in larger and more irregular fluctuations in IBI variability (Valenza et al., 2018). Thus, while the PNS contributes to smooth, cyclic variability, β-SNS activity introduces sharper, more chaotic bursts or consistent increases once active, enabling spectral differentiation between the systems. Finally, α-SNS activity can be distinguished through its primary role in peripheral vasoconstriction, where β-SNS has no direct effect and PNS has only an indirect effect (mostly through central vasoconstriction) (Wehrwein et al., 2016). Therefore, a measure of blood volume at the extremities, specifically pulse volume/wave (PVA), calculated as the systolic minus diastolic peak, most reflects the α-SNS (Kreibig, 2010; Stemmler et al., 2007). However these separate ANS estimations do not operate in systematic isolation. For instance, α-SNS-driven vasoconstriction increases peripheral resistance, which elevates blood pressure and can trigger a baroreflex-mediated parasympathetic inhibition to preserve blood flow to critical regions (Fredericks & Moore, 1974). This results in an increase in heart rate even in the absence of direct vagal activation from superior cortical structures. This indirect pathway means that α-SNS activation can induce short-latency cardiac changes, potentially producing signal components within the UHF band that might otherwise be attributed to β-SNS. Additionally, α-SNS activity can produce stable, low-frequency (LF) rhythms in heart rate (Shaffer & Ginsberg, 2017). Likewise, while PVA is primarily an index of α-SNS activity, it may also reflect minor PNS contributions to vascular tone via central regulatory circuits or through vasodilators in muscles or other regions which justifies breaking down into frequencies (Wehrwein et al., 2016). Finally, UHF, HF, LF, and VLF are adjacent to each other, so some overlap is inevitable. These dynamics underscore the importance of interpreting cardiovascular features as biases towards ANS influence rather than isolated system estimation. Hypotheses In the present study, specific emotions were manipulated with unique differences in the three emotional dimensions (valence, activation, and motivation) to produce unique variance. The emotion manipulations targeted basic emotions through joy, amusement, contentment, fear, sadness, and high-approach anger (aggression), which were designed to induce opposite directions on each of the dimensions (see Table 2 and Figure 1). Each emotion occupies a distinct position across the three emotional dimensions. Validation measures for each dimension were compared with the measured changes in PNS, α-SNS, and β-SNS activity. We predicted that (1) α-SNS activity would predict motivation, with higher activity for approach; (2). PNS activity would predict valence, with higher activity for positive/pleasant feelings, and (3) β-SNS activity would predict activation, with greater activity indicating stronger arousal. Method Subjects Subjects were recruited based on convenience sampling from psychology undergraduate students. Attempts were made to recruit diverse samples (Kissel & Friedman, 2023). Recruitment was redirected to other majors, including biology and neuroscience, in an attempt to achieve an even sample of both sexes to avoid sex-based biases in HRV (Koenig & Thayer, 2016). Nonetheless, the final sample was primarily female (see below). Subjects were awarded class credit for participation. Eligibility criteria included English fluency and no history of seizures, cardiovascular disorders, facial Botox, or hypertension medication. All procedures were approved by the Virginia Tech IRB. Subjects included in the analyses were those who had fully completed the film induction ( n =45) with an average age of 19 (SD=1.9) and a majority female (60% female; Table 4), all of whom also completed the remaining emotion questionnaires. No subjects reported a gender different than their sex assigned at birth. All characteristics from the demographics, continuous variables, and scales are in Appendix A Table 1. The majority of subjects were white (51%) or mixed race (20%), with the remaining ethnicities represented in an even manner (see Table 4). Within-subjects modeling mitigates demographic influence. Measures Two types of measures were taken. Physiological measures were collected through a BIOPAC finger clip pulse oximeter (OXY100C) and used to predict the emotional dimensions. The other type involved well-established measures for the emotional dimensions unrelated to ANS activity (as opposed to using heart rate or EDA for activation). The measures of affect (outcomes) were the target for prediction by the physiological measures (predictors). Physiological Measures The present study used a finger clip PPG device (BIOPAC OXY100C; at 2000Hz; AcqKnowledge). The pulse wave was used to determine the time of each heartbeat (systolic peak), which was used to calculate the IBI in seconds and PVA. Then, relevant features were calculated, including time domains, variability calculations, and nonlinear transformations (see Preprocessing). These measures were then compared to affect measures to understand their viability for replacement. Affect Measures Measures of affect focused on established differences in emotional dimensions expressed through the somatic nervous system. These were continuously variable and constant measures that are highly reliable for the given emotional dimension. Valence was measured using facial expression by facial electromyography (EMG; Cacioppo & Tassinary, 1990), by measuring two key features: one being the distance between the eyebrows as measured by the corrugator muscle activity, which indicates negative valence, and the other being the amount of smiling as measured by the zygomatic muscle (major and minor) activity, which indicates positive valence (Wingenbach, 2023). Motivational tendency also has a unique, simple measurement method using a joystick. The subjects maintained constant contact with the joystick in their dominant hand. Without instruction, subjects will tend to push the joystick away during feelings of avoidance and pull it towards them during feelings of approach (Hoofs et al., 2019; Phaf et al., 2014). Subjects were also given brief directional instructions (“follow what feels natural”) to enhance their joystick measurements. Activation was more complicated to estimate because it measures the intensity of emotion; therefore, it is integrally linked to the other two dimensions. However, eye tracking, by measuring eye aspect ratio (EAR; the degree of eye openness: Dewi et al., 2022), is a reliable measure of activation due to its relation to increased alertness or wakefulness. Eye tracking was performed to capture EAR using a webcam and the OpenFace package (Baltrusaitis et al., 2018). Scales Emotion questionnaires were used to assess potential accuracy factors. For example, depression and alexithymia may inhibit self-reports of emotion measurement (Christie & Friedman, 2004; Friedman et al., 2014), so subjects took the patient health questionnaire depression scale (PHQ-8; Kroenke et al., 2009) and the Toronto alexithymia scale (TAS; Bagby, Parker, et al., 1994; Bagby, Taylor, et al., 1994). Personality questionnaires included the big five aspects scale (DeYoung et al., 2007) and risk propensity scale (Meertens & Lion, 2008). Manipulations Emotion manipulations focused on variations in emotional dimensions and were aimed to establish the precedence of emotional triggers over physiological changes. The combination of emotion manipulations was designed to evoke specific emotions with unique differences in expected motivation, activation, and valence. The discrete emotion film clips targeted happiness, amusement, contentment, fear, depression/sadness, and high-approach anger (aggression), as seen in Table 2 and Figure 1. All emotions were induced through film clips on a computer monitor; the use of films was based on the recommendations of Siedlecka and Denson (2019), who found that visual stimuli were effective and reliable across studies. Film clip selection was guided by Wang et al.’s (2023) hybrid discrete-dimensional model of emotion, with a focus on identifying videos described by discrete emotions that also induce changes in the emotional dimensions (see Table 2 and Figure 1). The film clips for each category are presented in Table 3, which follows the relevant discrete emotion (Gross & Levenson, 1995; Hewig et al., 2005; Kaczmarek et al., 2021; Reynaud et al., 2012; Schaefer et al., 2010). Order effects of the videos were controlled by a Latin square design, where subjects were assigned their order based on gender, with the sequence resetting after completing a cycle of orders. Procedure Subjects were asked to abstain from caffeine for 6 hours, alcohol for 12 hours, and food and exercise for 3 hours before the session (Grant et al., 2023). Subjects were seated in front of a monitor with the joystick and a keyboard on opposite sides, with the joystick positioned to their right (the joystick was replaced with a mouse during the surveys). EMG electrode sites on the cheek and eyebrow were wiped with alcohol pads and dried using cotton balls. Reusable electrodes were placed with electrode gel and adhesives. The lead was taped to the face to increase comfort and security. The PPG device was placed on the left pointer finger. Recording of all measures was started for a baseline period and continued until the end of the film clips. Subjects were given about 10-15 minutes to acclimate while receiving instructions and performing the next tasks. Subjects were instructed to place their hand flat on the table and not to move it. Subjects completed the demographics and the Mind-Body lab health questionnaire (see Appendix C). Subjects answered scales about their current valence, motivation, and activation. When ready, the sequence of emotional videos was presented in the subject’s Latin square order, and further instruction was given on the screen. Subjects were informed of the target emotion being induced before each manipulation to avoid confusion and elicit more consistent emotional responses. After each emotionally charged film, a neutral film (The Lover; Schaefer et al., 2010) was presented to eliminate artifacts from the previous manipulation. Another survey was then provided, featuring emotion questionnaires related to the quality of the data and personality tests. The experimental manipulation series lasted approximately 22 minutes, and the entire experiment took about 90 minutes to complete. Analysis All subject data (except survey data) were saved to a single file at the end of each session, such that all timestamps of the manipulation start and stop, OpenFace, joystick, and BioPac data were synchronized. This process included an initial feature extraction, allowing the data file to have a row for each heartbeat, including calculated data such as IBI and PVA. Preprocessing Hypothesis testing utilized PPG data from the duration of emotional manipulations, including the label, emotion films, and neutral film periods. The order was not at all segmented or rearranged; instead, all physiological and affect data from the start of the films to the end were used in hypothesis testing. For the emotion validation measures, the valence dimension was calculated as zygomatic muscle activity subtracted from corrugator muscle activity, resulting in a scale in which positive emotions had positive values and negative emotions had negative values. For motivation, the joystick was calibrated such that the neutral position (the joystick straight up) was set to zero, with the distance toward the subject being positive (approach) and the distance away being negative (avoidance). Finally, activation was measured by the number of saccades per second or the EAR by eye tracking. Feature Extraction The feature extraction used only the PPG without any validator interference. The PPG data were run through a systolic peak detection algorithm using the Elgendi method form Neurokit2 to determine the timestamp of each heartbeat (Makowski et al., 2021). The timestamp of each heartbeat (systolic peak) was used to calculate IBI, and PVA was calculated by measuring the amplitude (change in light resistance) between the systolic and diastolic peaks which was divided by the sum of both peaks to ensure normalized data. The diastolic peak time was determined as the time at which the current value was the minimum between the current and the next peak (the detection was one heartbeat behind). The IBIs were limited to not be greater than 3 seconds or smaller than 0.3 seconds and defaulted to 1 second if conditions were met. For frequency domain extraction, a time-varying Extended Kalman Filter (EKF) was employed due to its ability to estimate instantaneous frequency in real-time, even in the presence of irregular sampling, irregular wave patterns, or signal dropout (McNames & Aboy, 2008). This method was preferred over more common methods (such as the Fast Fourier Transform) due to its real-time estimation capabilities, time-frequency resolution, ability to handle irregular sampling, and flexibility in accommodating phase and amplitude shifting. Additionally, the EKF was preferred due to its simplicity, particularly in lightweight or embedded devices. The EKF was applied to both PVA and IBI to extract UHF, HF, LF, and VLF, resulting in eight measures: IBI-UHF, IBI-HF, IBI-LF, IBI-VLF, PVA-UHF, PVA-HF, PVA-LF, and PVA-VLF. The amplitude output of each band reflects the engagement of its associated autonomic system, while the method’s high temporal resolution enables alignment with the dynamic time course of emotional states. In addition, the raw PVA and IBI values were used as separate features, and IBI calculations of root mean squared successive deviations (IBI-RMSSD) of the last ten seconds and standard deviation of normal to normal beats (IBI-SDNN) of the last 60 seconds were also added (Shaffer & Ginsberg, 2017). Normalization All predictors and dimensional validation measures were normalized individually for each subject before between-subject demeaning. To remove slower postural changes, the joy stick was linearly detrended (Virtanen et al., 2020). Dimensional validation measures (e.g., valence, motivation, and arousal) were further smoothed using exponential smoothing (McKinney, 2010). Initial burn-in and terminal edge effects were excluded. Signals were scaled to match the standard deviation of the joystick and stabilized using a rolling mean over the past twenty heartbeats. To avoid overfitting the model to idle, neutral, or low-amplitude fluctuations (noise), a variance mask was applied to the validation measures. The mask identified emotionally active time points using a mean absolute difference (MAD) threshold. If the threshold was never exceeded for that subject (flat or inexpressive), the time series was retained by default. Only time points exceeding the threshold in at least one validation dimension were kept. All physiological predictors were z-scored and then rescaled to match the standard deviation of the motivation validation, ensuring comparable variance across predictors. Finally, all features and outcomes were demeaned within each subject to remove between-subject differences in mean level. Although the measures were scaled during model development, the only requirement for using the following model in a new subject would be z-scaling between predictors to prevent individual predictors from over-biasing the dimension due to higher variability. Results Descriptive Statistics Of the 49 collected subjects, 2 were removed for missing joystick data and 2 for the films being muted. All 45 remaining subjects were included. No subjects reported consuming alcohol within the preceding twelve hours or caffeine in the preceding six hours, 1 had used tobacco or nicotine products, and 2 were undergoing treatment with selective serotonin reuptake inhibitors (SSRIs). 1 participant had exercised in the last two hours, and all but 4 exercised at least once per week (see Appendix A Table 2). See Appendix A for further subject data or Appendix B Table 2 for cross correlations. No subjects report medical disorders that effect autonomic activity. For mental health diagnoses, 6 reported diagnoses of ADHD (13%), 2 for ASD (4%), 1 for panic disorder (2%), 2 for OCD, 1 for social anxiety disorder, 8 for anxiety disorder (18%), 5 for depressive disorder (11%), and 1 for bipolar or seasonal affective disorder. Supervised Partial Least Squares Regressions The overall goal of this study was to predict emotional dimensions at each heartbeat using real time physiological measures. Partial Least Squares (PLS) regression was chosen to produce interpretable models by extracting the latent components of the predictors that are primarily relevant to the dimensions (Krishnan et al., 2011). Unlike standard regression, PLS detects patterns across the predictors, rather than single predictor-to-outcome relationships. Principal component analysis also identifies patterns across the predictors but does so without regard for a target variable. These functional differences enable PLS to avoid instability despite multicollinearity allowing for the extraction of components that align with the dimensions and avoids overfitting. All extracted physiological features were used as predictors, including SDNN, RMSSD, IBI, PVA, and frequency domains of IBI and PVA, with of the squared amplitudes of the frequency domains added. The model included every row of data (every detected heartbeat) for each participant, regardless of any noise or confounding events. Instead, the present study relied on the EKF algorithms to filter noise, which will at worst only limit the effect size and make it more relatable to real world signals. The variable importance projections were used to determine the contributors for each dimension. Because validation and predictor signals were on arbitrary and subject-specific scales, absolute reconstruction metrics were not directly interpretable across participants. Temporal correspondence was therefore assessed using within-subject functional data analysis (FDA). Predicted and observed dimensions were represented as B-spline functions on a normalized time axis, and functional correlation was computed via the L2 inner product, which quantifies the similarity of their temporal trajectories. Summary statistics were based on the median functional correlation. For each subject, the validation time series was circularly time-shifted to disrupt temporal alignment (2000 permutations), which preserves the within-subject distribution and temporal smoothness while generating a null model. Models were thus tested using time shifted nulls rather than shuffled data, yielding a conservative bias. In addition, the root mean squared errors (RMSE; including calibrated and clamped) are reported. However, since latent constructs were used (rather than direct predictors), the RMSE should be interpreted more generously. The results of the PLS regressions are presented in Table 5. Average intercorrelations between the predicted dimensions were low (max of 0.43 see Table 6) and even better in the real-time model (max absolute value of -0.23 see Appendix A Table 15). The full tables of loadings, VIPs, and weights are available in Appendix A Tables 5-7. Activation For the activation dimensions (validated by EAR), the PLS regression yielded significant temporal correspondence especially in the FDA correlation ( r =0.26, p <0.001, r 2 =0.07, median functional correlation = 0.14, permutation p = 0.008, RMSE=1.06). The key features were the linear PVA-LF (x-loading=-0.25, VIP=1.67), IBI (x-loading=.54, VIP=2.14), RMSSD (x-loading=0.65, VIP=2.58), and SDNN (x-loading=0.44, VIP=1.68). Valence For the valence dimensions (validated by EMG), the PLS regression was successful and had the highest FDA correlation ( r =-0.21, p <0.001, r 2 =0.04; median functional correlation = 0.20, permutation p = 0.018, RMSE=102). The main features were PVA-HF power (x-loading=0.58, VIP=2.44), followed by PVA-UHF power (x-loading=-0.36, VIP=2.21) and PVA-VLF (x-loading=0.52, VIP=1.62). Motivation For the motivation dimensions (validated by the joystick), the PLS regression was successful though the FDA permutation only trended towards significance ( r =-0.22, p <0.001, r 2 =0.05, median functional correlation = 0.12, permutation p = 0.058, RMSE=1.02). The main features were IBI-UHF (x-loading=0.53, VIP=2.06), PVA (x-loadings=0.23, VIP=1.69), and RMSSD (x-loadings=-0.37, VIP=1.58). Finally, almost all PVA features had moderate VIPs (just below 1 or above) and medium weights (>0.15), which generally indicate that PVA was used in a general manner, without biasing a particular frequency. There was some support from the IBI-LF. Discrete Emotion Detection To explore the practicality of the three-dimensional emotion model as a theoretical proof of concept, the present study produced a basic discrete emotion probability algorithm. To preserve sensitivity to rapid directional changes, the unsmoothed predictors were used to compute the dimensions by applying the PLS regression equations above (see Figure 2 for the dimensions over time for each film). The first derivative of each dimension was computed to capture changes over time. The fit for each emotional dimension was based on whether the movement of the dimensional direction and relative dimensional contribution aligned (e.g., if valence increased, then the predictions for joy, amusement, and contentment increased, and negative emotions decreased; see Table 2 and Figure 1 for directional details). At each new data row, the probability of each emotion was calculated based on its alignment with the dimensional directions. To smooth the probability time series, temporal accumulation was applied. A softmax filter was then used to produce probabilities for each emotion. The emotion probabilities over time were compared within each film clip based on emotionally relevant time points. The discrete emotion probabilities over time are presented in Figure 3 to show the average of each over time. It is clear from Figure 3 that fear (or anxiety) and contentment acted as the base states, in which most subjects felt either a low level of anxiety or contentment most of the time, likely when they did not experience any particular (discrete) emotion. Despite this, fear reached the max during the apex of the haunted house. Contentment was high during the low engagement periods and peaking during the contentment film. Anger peaked when the fight occurred. Sadness was high when the love interest died. Amusement showed increases during humorous moments. Discussion The dimensional loadings, weights, and VIPs supported the hypotheses of the present study, which posited that valence was associated with PNS indicators, motivation was linked to α-SNS activity, and activation was related to β-SNS measures. The valence dimension showed an association with HF variability in both heart rate and especially peripheral resistance. These HF fluctuations have been linked to PNS-related respiratory activity (Shaffer & Ginsberg, 2017; Valenza et al., 2018). The motivation dimension showed an association with most peripheral resistance domains (Stemmler et al., 2007) and LF variations, all of which relate to α-SNS activity (Shaffer & Ginsberg, 2017), and the PVA-UHF was likely associated with the rapid cardiac responses to sudden vasoconstriction (Fredericks & Moore, 1974). The activation dimension showed an association with general heart rate (particularly RMSSD which is biased by major fluctuations) and peripheral resistance responses to large changes in heart rate. These findings support the prediction that activation was influenced by β-SNS activity (Shaffer & Ginsberg, 2017). While individual physiological predictors often reflect input from multiple autonomic branches, the latent dimension emerged composed primarily of predictors associated with the hypothesized system. The PLS regressions yielded modest but reliable FDA correlations across dimensions. Given that noise is inherent in both physiological and behavioral measures, effect sizes of this magnitude are expected for continuous emotional trajectories rather than averaged outcomes. Whereas prior studies inferred these associations from averaged physiological responses to discrete emotions (e.g., Kreibig, 2010; Stemmler et al., 2007), the present study directly mapped each dimension onto latent physiological components extracted from real-time, continuous PPG. These findings were further supported by emotion induction studies based on discriminant functions (e.g., Christie & Friedman, 2004; Nyklicek et al., 1997; Stephens et al., 2010), where the motivation function was primarily explained by diastolic and mean arterial blood pressure, as well as a faster heart rate. In contrast, mean successive differences of heart rate and systolic blood pressure predicted activation. The present study contributes to and expands upon the context of these findings. In particular, we did not measure blood pressure in the present study; however, differences in frequency were observed, with activation predominantly occurring at LF frequencies and motivation primarily associated with PVA frequencies. The present study extended these findings by providing frequency domain separation, isolating valence, and enabling real-time predictions, especially from time points of unknown emotional relevance. Alternatively, other discriminant analysis studies used self-reports of both somatic and psychological changes,and were able to identify similar discriminant functions (Fontaine et al., 2007; Mohammadi & Vuilleumier, 2020). These shared findings across Christie and Friedman (2004), Mohammadi and Vuilleumier, and Fontaine et al. indicates a dimensional structure underlying discrete emotions. The PLS regressions yielded interpretable and accurate predictions that can be reproduced in any context with relevant sensors. Each of these dimensions had only low intercorrelations and distinct physiological signatures, requiring no baseline recordings or calibrations. PNS, α-SNS, and β-SNS activity presented as functionally dissociable autonomic influences. Against expectations, the estimated α-SNS activity (motivation) had a reverse correlation with the estimated β-SNS activity (see Appendix B Table 1). These results imply that the SNS does not vary uniformly, and the SNS and PNS are not perfectly reciprocal (Berntson et al., 1994; Jänig & Häbler, 2000). Further, the separation of these autonomic systems provided clear interpretations of emotional direction. Finally, previous studies have attempted to input as many physiological features as possible to discriminate emotional states (Christie & Friedman, 2004; Kragel & LaBar, 2013; Stephens et al., 2010). However, the present study provides evidence that only three estimations (computed from a single measuring device) are necessary. Implications Methodological Among the various implications of the present study, the most important is the methodological advancement in estimating the emotion-related signals sent by the CNS and ANS feedback. Thus, the present methodology has implications in two parts. The first is the ability to estimate latent factors associated with functionally dissociable autonomic influences, such as PNS, α-SNS, and β-SNS (or cholinergic SNS) activity. The second is that these latent factors are related to CNS-linked emotional dimensions. Therefore, not only does the ANS operate through three systems, but these systems can also be used to estimate emotions or predict general behavioral directions. Finally, all the estimations made can be directly interpreted and meaningfully tracked to their measurement core. In other words, unlike neural networks or “black box” artificial intelligence systems (Saganowski et al., 2023), all parts of the calculation are known and interpretable. In addition, the present method uses EKFs, SDNN (past 30), RMSSD (past 10s), and raw signals or transformations to allow for real-time estimation and estimation at high temporal resolution, potentially with accuracy at every heartbeat (though current validation measures for that do not exist) (McNames & Aboy, 2008; Shaffer & Ginsberg, 2017). In addition, these measures describe a method that only requires PPG. PPG is MRI compatible, cost-effective, scalable, and can be worn and measured with minimal interference, allowing for easy use in a wide variety of experimental settings. PPG is also widely available in consumer and commercial electronics such as smartwatches, fitness trackers, medical monitors, infrared cameras, and even measurable on phones (Castaneda et al., 2018), which means that a researcher capable of connecting to such devices can collect full behavioral data (only missing context/environment) from people in daily life. Ultimately, these emotional dimensions can be applied broadly. For example, the classic social psychology experiment using the Prisoner’s Dilemma could incorporate PPG to derive real-time dimensional signals throughout each round (Axelrod, 1980). While subjects would ultimately choose to cooperate or defect, their physiological data offer a continuous window into internal states leading up to the decision. High valence and high motivation, for instance, could predict a greater likelihood of cooperation. However, when an experimental condition is introduced, such as time pressure, perceived unfairness, or prior betrayal, the final behavioral choice may appear unchanged or statistically weak. In these cases, moment-by-moment shifts can reveal the impact of manipulation, even when the behavior remains unchanged. A significant drop in valence or spikes in avoidance may signal a breach of trust or moral discomfort that traditional binary outcomes miss. Thus, what might appear as a marginal or null behavioral effect can gain clarity through the added dimensional and temporal resolution of continuous physiological monitoring. Behavioral Applications Similarly, basic measures related to the fundamentals of behavior and emotion can be applied across various predictive contexts. These dimensions can provide the basic building blocks for prediction models, serving either as inputs (predictors) or as generalized outcomes (validations). For example, the discrete emotion detection in the present study utilized three dimensions as interpretable input methods and generated emotion probabilities using a simple model without requiring training data. In other behavioral models, these dimensions could be utilized for various mathematical algorithms or neural networks (or other such “black box” models) to produce more complex and accurate predictions. In addition, the general sentiment of such predictive models is “garbage in, garbage out” (Kilkenny & Robinson, 2018). The present study provides a method to avoid or filter out messy measures for higher relevance (like precise vectors in physics), meaning that the measures can be incorporated into a model as either the input or validator to eliminate noise. Naturally, these improvements could be applied to various modeling attempts (Asarnow et al., 2017; Coid et al., 2009; Rossegger et al., 2011). Limitations A significant limitation of this study is that both the ANS measures and the validation measures used are inherently noisy and susceptible to inaccuracies. While the physiological data were mostly denoised (through the EKFs and moving averages) and the final dimensional signal was based on a composite (reducing noise from any one signal), there will still be problems in the final product. For instance, movement could cause certain heartbeats to be missed (though EKFs correct it; See Appendix A “Noise Robustness”), or, for the PVA, environmental lighting or sensor movement could change how the measured values relate to actual pulse volume (a limitation that does not exist for blood pressure) (Charlton et al., 2023). In addition, PPG is not the most precise measure of cardiac timing (ECG is generally preferred; (Quigley et al., 2024), but the focus of this study was for more limited applications. In addition, the validation measures are likely to be equally noisy and potentially less accurate than the predicted values. This is especially true for measures more related to conscious control, in which motivation (arm movement) is likely the least accurate, followed by valence (facial movement), with activation likely the most accurate, since people rarely think about how open their eyes are. However, if PLS had produced only the reliable estimate (by focusing on predictor variability and averaging across subjects ), then the physiological prediction may be equivalent or more accurate. Another limitation is that the film clips themselves do not necessarily evoke the same emotional response in all individuals (McGinley & Friedman, 2017). Instead, subjects may feel a different emotion or no emotion at all (although self-reports tend to support emotional feelings; actual feelings may vary). The more often this occurs, the harder it would be to separate each dimension (as each dimension correlates differently for each emotion), and thus the less accurate the prediction may be. Including the time series data from all subjects might help to reduce this problem. The fact that individual differences in responses exist, however, may allow for the discovery of between-subject differences, such as personality or mood. Future Research Much of the present study, including the three-dimensional model, the use of EKFs for emotional prediction, and providing high temporal accuracy, is novel. Motivation is rarely considered in addition to valence and arousal (Gasper & Middlewood, 2014; Hoofs et al., 2019; Kaczmarek et al., 2021), and it is often presented as an alternative to valence in the circumplex space (Christie & Friedman, 2004; Harmon-Jones & Allen, 1998). To use all three measured as a time series or as the core for predicting discrete emotions has not been previously tested. While there were predictions regarding which ANS systems would correlate with each dimension, there were few assumptions about which measures would correlate with each dimension. Therefore, replication is clearly needed. Furthermore, the present study can only be as predictive as the validation measures. Therefore, new techniques for validation should be sought. These include validation through measures of the emotional dimensions or the ANS systems (such as through fMRI or electrical activity of the ANS nerves). Similarly, these dimensions should also be tested for neural correlations (such as through EEG or fMRI), particularly in the regions known to relate to the dimensions. In addition, studies should test using manipulations that are more temporally trackable, such as the presentation of emotional pictures, to show the more immediate and time-relevant predictability of the calculated dimensions. Finally, the current and other options for measuring these dimensions (such as facial tracking, which is not yet available for motivation) should be implemented to enable wider usability and applications. The discrete emotion prediction models can be improved upon based on the findings of the present study. First, validation measures of discrete emotions could be used to determine prediction accuracy. Many other additions can also be made to improve either accuracy or usefulness. On that point, although the present study showed that training was not required, models would likely still benefit from such methods. The introduction of a neutral state or cool down period may provide estimations that are more applicable to avoiding over-attributing emotions during arbitrary states. Finally, temporal combinations of dimensional directions could elicit more complex emotions, such as disgust from fear directionality, with greater PNS variability for the faint or vomiting response. However, to date, no models have demonstrated interpretability for these combinations. Another way to provide these tests (possibly more applicable to real applications) is an ecological momentary assessment approach (Shiffman et al., 2008) where subjects using a PPG wearable device in their daily lives are occasionally prompted on emotional scales such as the positive and negative affect schedule (PANAS; Watson et al., 1988). This method would not only enable the prediction of emotions but also provide the data necessary to determine normalization capabilities in the long term and across new environments. Future studies should also attempt to relate these dimensions to mood by validating longer-term measures and relating the dimensions to personality profiles by comparing the emotional responses between subjects. Personality prediction may require contextual information about the current situation (such as a social situation for extroversion). The collection of contextual information will also be important for the basis of predictive models or broader applications. Overall, the present study provides a wide range of possibilities for the production and improvement of predictive models. Conclusion Emotions may provide effective estimations or predictors of future emotions or behaviors, particularly core behavior types such as improving, destroying, relaxing, or escaping. These would be more effectively determined when emotions or behaviors are measured by their functional differences, such as motivation (approach-avoidance), valence (positive-negative), and activation (arousal), and when they are constantly measured (high temporal variability) and continuously variable rather than strict categories. The present study offers a novel method for estimating such emotional dimensions, capable of real-time recording using PPG, a technology commonly employed in smartwatches, fitness trackers, and medical monitors. It was found that valence could be estimated primarily from indicators of PNS activity, motivation primarily from α-SNS indicators, and activation from β-SNS indicators. These three dimensions may be combined for a rudimentary but effective estimation of discrete emotions. Such results indicate that emotions are more than states, but rather dynamic trajectories along dimensions related to functionally dissociable autonomic influences. Cheap, interpretable, and non-invasive emotional and behavioral predictions are now within reach, and a significant portion of the public already wears the necessary device and is gaining popularity. Tables and Figures Differentiation of Emotional Dimensions From SNS and PNS Activity Positive (PNS Activation) Amusement/Joy* Contentment* Negative (PNS Inhibition) Fear/anger* Depression/Sadness* Note: Columns represent differences in activation, with high activation on the left and low activation on the right. Rows represent differences in valence, with positive valence at the top and negative valence at the bottom. SNS = Sympathetic Nervous System; PNS = Parasympathetic Nervous System. * Evidence available in (Kreibig, 2010). Differentiation of The Three Emotional Dimensions Using ANS Activity Activation Approach (α-SNS activation) Avoidance (α-SNS Inhibition) Approach (α-SNS activation) Avoidance (α-SNS Inhibition) Positive (PNS Activation) Joy* Amusement* Contentment* Negative (PNS Inhibition) Anger* Fear* Sadness* Note: The top columns represent differences in valence, with positive valence on the left and negative valence on the right. The lower columns represent differences in activation, with high activation on the left and low activation on the right. Low activation refers to either a decrease in β-SNS activity or a level lower than its counterpart under other conditions held constant. Rows represent motivation, with avoidance on top and approach at the bottom. SNS = Sympathetic Nervous System; PNS = Parasympathetic Nervous System. * Evidence available in (Kreibig, 2010). Emotion Manipulations Activation Approach (α-SNS activation) Avoidance (α-SNS Inhibition) Approach (α-SNS activation) Avoidance (α-SNS Inhibition) Positive (PNS Activation) Benny & Joone 1 There Is Something About Mary (1) 1 The Dead Poets Society (2) 1,2 Negative (PNS Inhibition) In the Name of the Father 1 The Blair Witch Project 1 City of Angels 1 Note: The top columns represent differences in valence, with positive valence on the left and negative valence on the right. The lower columns represent differences in activation, with high activation on the left and low activation on the right. Low activation refers to either a decrease in β-SNS activity or a level lower than its counterpart under other conditions held constant. Rows represent motivation, with avoidance on top and approach at the bottom. SNS = Sympathetic Nervous System; PNS = Parasympathetic Nervous System. 1 Evidence available (Schaefer et al., 2010). https://sites.uclouvain.be/ipsp/FilmStim/film.htm 2 Evidence available (Kaczmarek et al., 2021). Gender and Ethnicity White or European 23 (51.1) 8 (44.4) 15 (55.6) Black or African 2 (4.4) 0 2 (7.4) East Asian 2 (4.4) 0 2 (7.4) Hispanic or Latino 3 (6.7) 3 (16.7) 0 Native American or Indigenous 1 (2.2) 0 1 (3.7) South Asian 4 (8.9) 5 (71.43) 4 (14.8) Middle Eastern or North African 1 (2.2) 1 (5.6) 0 Multiracial or Mixed Ethnicity 9 (20.0) 6 (33.3) 3 (11.1) Total (Gender %) 45 18 (40.0) 27 (60.0) Note. The reported ethnicity subjects. PLS Regression Outcomes Activation 0.07 0.0004 0.26 <0.001 Valence 0.04 0.0008 -0.21 <0.001 Motivation 0.05 0.0008 -0.22 <0.001 Note. Each row represents a separate PLS regression reporting the variance explained for the predictors (X) and the validators (Y). PLS = partial least squares. Intercorrelations Activation 1 Valence -0.16 1 Motivation 0.43 -0.01 1 Note. The intercorrelations of the calculated dimensions. Discrete-Dimensional Structure of Emotion Note. The black arrows represent the emotional dimensions, which are described by the colored axis labels. The points represent the discrete emotional outcomes of the dimensional changes. PNS = parasympathetic nervous system, α-SNS = alpha adrenergic sympathetic nervous system, β-SNS = beta sympathetic nervous system. Mean and 95% Confidence Interval of Predicted Dimensions per Film Note. Average valence, motivation, and activation dimensions over time for each film clip. Ninety-five percent confidence intervals are presented through the color fields. Mean Discrete Emotion Probabilities Note. Average emotion probabilities over time from the start to finish of each emotional induction (including the beginning label, emotional film, and the following neutral clip). Appendix A The following appendix provides additional, data, analyses, and application of the method to derive real-time emotional dimensions from physiological signals. The PLS model loadings, weights, and variable importance projections (VIPs) are found in Tables 1-3. Most descriptive statistics are presented in Table 4 and physical activity in Table 5. Sobolev FDA results: motivation ( fdcor =0.003, p =0.552), valence ( fdcor =0.031, p =0.204), and activation ( fdcor =0.015, p =0.373). Additional Subject Data 12 female subjects reported using hormonal contraception. Among those not using contraception, 4 reported being in the early phase of their menstrual cycle, 8 were in the mid-cycle phase, and 1 at the end of the cycle (see Table 6). The remaining 3 were unknown. No subjects with medical disorders were identified, except 1 for lazy eye and 7 with visual deficiency (20%). Cronbach’s alphas for the emotion questionnaires are reported for PHQ-8 ( α =0.73) and TAS ( α =0.86), as well as for the factors of difficulty identifying feelings (DIF; α =0.85), difficulty describing feelings (DDF; α =0.89), and externally oriented thinking (EOT; α =0.52). Results from the patient health questionnaire (see Table 7) showed that most subjects’ depression symptoms were rated as moderately severe (n=20, 44%), followed by severe (n=12, 27%), and moderate (n=10, 22%). However, only 3 subjects reported minor symptoms (7%). The TAS scale showed 6 subjects rated borderline and 6 as above the cut-off for alexithymia. 8 people reported using devices such as fitness trackers, smartwatches, or similar devices. Refer to Figure 1 for their reported comfort levels. No subjects reported a gender deviating from sex assigned at birth. Alternate Implementations To aid future applications, two additional models were tested. The first was a model that did not use PVA. The intention was to demonstrate the validity of using physiological sensors that do not provide estimates of peripheral resistance (such as ECG, phonography, or Doppler radar techniques). If possible, this may also produce a model that requires less normalization to environmental light or sensor placement, since time is a non-arbitrary value. IBI Only The analysis without PVA (see Table 8) resulted in slightly worse prediction outcomes, although motivation suffered the most. The reconstruction r 2 s are reported (activation: r 2 =0.04, valence: r 2 =0.03, motivation: r 2 =0.01). The FDA permutations showed only valence significant, activation trending significant, and motivation insignificant. These finding are to be expected given how much each dimension relies on PVA. Some of the reductions may be the natural outcome of reducing predictors and conscious noise in the validation measure. The parameters used are available in Tables 9-11. The RMSEs were generally higher for activation (RMSE=1.07), valence (RMSE=1.03), and motivation (RMSE=1.21). Average intercorrelations of the predicted dimensions are in Table 12. Real-Time Implementation To verify the use of these dimensions in real-time (potentially small-device) implementations, C code was generated to mimic the Python code. However, scaling and normalization are based on the overall data and cannot be properly calculated in real-time, so no scaling was applied. However, some weights will reverse sign and generally not reflect the actual relationship. The PLS regression was still successful (See Table 13). Model weights for each dimension are shown in Table 14. The RMSEs were similar for activation (RMSE=1.12), valence (RMSE=0.95), and motivation (RMSE=1.11). Average intercorrelations of the predicted dimensions are in intercorrelations in Table 15. IBI Only Real-time measures without PVA were also tested (see Table 16). The weights are available in Table 17. The RMSEs are reported for activation (RMSE=1.15), valence (RMSE=0.97), and motivation (RMSE=1.13). Average intercorrelations of the predicted dimensions are in Table 18. PLS Regression Activation X-Loadings, Weights, and VIP IBI 0.54 0.48 2.14 IBI-UHF -0.03 -0.04 0.19 IBI-UHF 2 -0.06 -0.03 0.14 IBI-HF -0.05 -0.08 0.38 IBI-HF 2 -0.01 -0.08 0.35 IBI-LF -0.07 0.12 0.56 IBI-LF 2 -0.06 -0.03 0.13 IBI-VLF -0.02 -0.06 0.27 IBI-VLF 2 -0.02 -0.07 0.30 IBI-RMSSD 0.65 0.58 2.58 IBI-SDNN 0.44 0.38 1.68 PVA -0.21 -0.25 1.11 PVA-UHF 0.30 0.18 0.81 PVA-UHF 2 -0.01 0.00 0.00 PVA-HF 0.09 0.05 0.24 PVA-HF 2 0.06 -0.13 0.58 PVA-LF -0.25 -0.37 1.67 PVA-LF 2 0.11 0.00 0.02 PVA-VLF 0.07 -0.01 0.04 PVA-VLF 2 0.06 -0.03 0.15 Note. The x-loading, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. PLS Regression Valence X-Loadings, Weights, and VIP IBI -0.06 -0.15 0.66 IBI-UHF -0.31 -0.25 1.11 IBI-UHF 2 -0.36 -0.49 2.21 IBI-HF -0.27 -0.10 0.44 IBI-HF 2 -0.25 0.04 0.19 IBI-LF -0.26 0.09 0.39 IBI-LF 2 -0.29 -0.22 0.98 IBI-VLF -0.25 0.04 0.17 IBI-VLF 2 -0.24 0.08 0.38 IBI-RMSSD 0.16 0.19 0.85 IBI-SDNN 0.06 0.04 0.17 PVA -0.06 -0.17 0.76 PVA-UHF 0.31 0.03 0.12 PVA-UHF 2 0.23 0.02 0.10 PVA-HF 0.42 0.00 0.01 PVA-HF 2 0.58 0.55 2.44 PVA-LF -0.10 -0.01 0.03 PVA-LF 2 0.38 0.07 0.29 PVA-VLF 0.52 0.37 1.67 PVA-VLF 2 0.38 0.29 1.31 Note. The x-loading, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. PLS Regression Motivation X-Loadings, Weights, and VIP IBI -0.29 -0.17 0.74 IBI-UHF -0.29 -0.15 0.68 IBI-UHF 2 -0.42 -0.22 1.00 IBI-HF -0.36 0.06 0.26 IBI-HF 2 -0.51 -0.07 0.29 IBI-LF -0.40 -0.24 1.09 IBI-LF 2 -0.42 -0.03 0.15 IBI-VLF -0.51 -0.17 0.78 IBI-VLF 2 -0.52 -0.17 0.75 IBI-RMSSD -0.37 -0.35 1.58 IBI-SDNN -0.33 -0.27 1.23 PVA 0.23 0.38 1.69 PVA-UHF 0.18 0.48 2.14 PVA-UHF 2 -0.12 -0.09 0.39 PVA-HF -0.21 -0.31 1.37 PVA-HF 2 -0.12 -0.18 0.81 PVA-LF -0.15 -0.14 0.62 PVA-LF 2 0.07 -0.02 0.09 PVA-VLF -0.11 -0.15 0.69 PVA-VLF 2 -0.08 0.15 0.67 Note. The x-loading, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Descriptive Subject Statistics of Continuous Variables Age (years) 45 19.16 18.71 19.6 1.492 18 25 Height (inches) 45 66.76 65.57 67.94 3.938 55 74 Weight (lbs) 45 148.5 141 156 24.87 98 210 BMI 45 2.221 2.123 2.318 0.325 1.571 2.937 Hours since caffeine 7 7.143 6.505 7.781 0.69 6 8 Hours since ate 45 7.611 6.137 9.085 4.906 1 18 Average sleep 45 6.962 6.656 7.268 1.018 5 9 PHQ8 45 16.71 15.28 18.14 4.756 8 26 TAS 45 47.18 43.68 50.68 11.65 25 74 TAS DIF 45 15.22 13.58 16.86 5.456 7 31 TAS DDF 45 14.69 13.07 16.3 5.376 5 25 TAS EOT 45 17.27 16.09 18.44 3.922 10 25 Note. The emotional dimensions were self-reports on a 0-10 scale. CI = confidence interval, SD = standard deviation, Caffeine time = hours since last caffeinated beverage, Last ate (hrs) = hours since last food consumption, Hours asleep = hours slept last night, Avg sleep = average nightly sleep, PHQ-8 = patient health questionnaire version 8, TAS = Toronto alexithymia scale, DIF = difficulty identifying feelings, DDF = difficulty describing feelings, EOT = externally oriented thinking. Physical Activity Never 1 2.2 Rarely 2 4.4 One to two times per month 1 2.2 One to two times per week 15 33.3 Three to four days per week 6 13.3 Five to six days per week 15 33.3 Seven days per week 5 11.1 Note. The subjects self-reported physical activity for the past month. Menstrual Phase Beginning 4 8.9 Middle 8 17.8 End 1 2.2 Unknown 5 11.1 N/A 22 48.9 Missing 5 11.1 Note. N/A and missing values were from males or females on birth control. PHQ-8 Severity Mild 3 6.7 Moderate 10 22.2 Moderately severe 20 44.4 Severe 12 26.7 Note. The standard severity ratings for PHQ-8. PHQ-8 = patient health questionnaire version 8. PLS Regression Outcomes Without PVA Activation 0.05 0.0003 0.23 <0.001 0.09 0.077 0.004 0.403 Valence 0.03 0.0003 0.17 <0.001 0.05 0.031 -0.003 0.576 Motivation 0.01 0.0006 0.11 <0.001 0.08 0.200 -0.002 0.506 Note. Each row represents a separate PLS regression reporting the variance explained for the predictors (X) and the validators (Y). PLS = partial least squares, L2-r=functional correlation, L2-p=functional permutation significance, S-r=Sobolev functional derivative correlation, S-p=Sobolev permutation significance. Activation Univariate PLS Regression Without PVA IBI 0.56 0.56 1.84 IBI-UHF -0.02 -0.05 0.17 IBI-UHF 2 -0.05 -0.04 0.12 IBI-HF -0.05 -0.10 0.33 IBI-HF 2 0.00 -0.09 0.30 IBI-LF -0.07 0.14 0.48 IBI-LF 2 -0.05 -0.03 0.11 IBI-VLF 0.00 -0.07 0.24 IBI-VLF 2 0.00 -0.08 0.26 IBI-RMSSD 0.70 0.67 2.22 IBI-SDNN 0.51 0.44 1.45 Note. The x-loadings, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Valence Univariate PLS Regression Without PVA IBI 0.24 0.22 0.74 IBI-UHF 0.63 0.37 1.24 IBI-UHF 2 0.79 0.74 2.47 IBI-HF 0.45 0.15 0.49 IBI-HF 2 0.48 -0.06 0.21 IBI-LF 0.42 -0.13 0.43 IBI-LF 2 0.54 0.33 1.10 IBI-VLF 0.48 -0.06 0.20 IBI-VLF 2 0.48 -0.13 0.42 IBI-RMSSD -0.17 -0.29 0.95 IBI-SDNN -0.13 -0.06 0.19 Note. The x-loadings, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Motivation Univariate PLS Regression Without PVA IBI 0.31 0.25 0.84 IBI-UHF 0.28 0.23 0.77 IBI-UHF 2 0.38 0.34 1.14 IBI-HF 0.33 -0.09 0.30 IBI-HF 2 0.47 0.10 0.33 IBI-LF 0.37 0.37 1.24 IBI-LF 2 0.38 0.05 0.18 IBI-VLF 0.47 0.27 0.88 IBI-VLF 2 0.48 0.25 0.85 IBI-RMSSD 0.32 0.54 1.79 IBI-SDNN 0.30 0.42 1.39 Note. The x-loadings, model weights, and VIP were used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Intercorrelations Without PVA Activation 1 Valence -0.24 1 Motivation 0.51 0.40 1 Note. The intercorrelations of the calculated dimensions. Real-Time PLS Regression Outcomes Activation 0.06 0.0001 -0.25 <0.001 0.21 0.005 0.026 0.232 Valence 0.14 0.0002 0.38 <0.001 0.25 0.003 0.091 0.006 Motivation 0.07 0.0002 0.27 <0.001 0.06 0.343 0.061 0.033 Note. Each row represents a separate PLS regression reporting the variance explained for the predictors (X) and the validators (Y). PLS = partial least squares, L2-r=functional correlation, L2-p=functional permutation significance, S-r=Sobolev functional derivative correlation, S-p=Sobolev permutation significance. Real-Time PLS Regression Weights IBI -0.09 -0.19 -0.01 IBI-UHF -0.17 0.46 0.06 IBI-UHF 2 0.08 0.16 0.12 IBI-HF 0.00 0.23 0.02 IBI-HF 2 0.05 0.21 0.12 IBI-LF 0.02 -0.30 0.06 IBI-LF 2 0.01 0.19 0.06 IBI-VLF -0.37 0.44 -0.42 IBI-VLF 2 0.31 -0.17 0.36 IBI-RMSSD 0.27 0.18 0.14 IBI-SDNN 0.04 0.24 0.12 PVA 0.20 0.07 -0.05 PVA-UHF 0.08 0.02 0.05 PVA-UHF 2 -0.39 -0.20 0.01 PVA-HF -0.25 0.15 0.27 PVA-HF 2 -0.25 -0.17 0.28 PVA-LF 0.19 -0.06 0.13 PVA-LF 2 0.40 -0.04 0.23 PVA-VLF 0.22 0.16 0.40 PVA-VLF 2 0.27 -0.26 0.48 Note. The model weights used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Real-Time Intercorrelations Activation 1 Valence -0.23 1 Motivation -0.10 -0.14 1 Note. The intercorrelations of the calculated dimensions. Real-Time PLS Regression Outcomes Without PVA Activation 0.02 0.0001 0.16 <0.001 0.04 0.249 0.045 0.117 Valence 0.11 0.0001 0.32 <0.001 0.05 0.245 -0.003 0.546 Motivation 0.04 0.0001 -0.21 <0.001 0.05 0.228 0.073 0.0165 Note. Each row represents a separate PLS regression reporting the variance explained for the predictors (X) and the validators (Y). PLS = partial least squares, L2-r=functional correlation, L2-p=functional permutation significance, S-r=Sobolev functional derivative correlation, S-p=Sobolev permutation significance. Real-Time PLS Regression Weights Without PVA IBI 0.15 -0.22 0.02 IBI-UHF 0.28 0.51 -0.10 IBI-UHF 2 -0.13 0.18 -0.20 IBI-HF 0.00 0.26 -0.03 IBI-HF 2 -0.08 0.23 -0.19 IBI-LF -0.04 -0.34 -0.10 IBI-LF 2 -0.01 0.22 -0.09 IBI-VLF 0.63 0.49 0.68 IBI-VLF 2 -0.52 -0.19 -0.58 IBI-RMSSD -0.45 0.20 -0.23 IBI-SDNN -0.07 0.27 -0.20 Note. The model weights used for the PLS regression. PLS = partial least squares, VIP = variable importance projection, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Real-Time Intercorrelations without PVA Activation 1 Valence -0.12 1 Motivation -0.72 0.212 1 Note. The intercorrelations of the calculated dimensions. Comfort With Being Emotionally Monitored Note. Subjects’ self-reported comfort with having their emotions monitored by a wearable device. Appendix B Within-Subject Average Correlation Table Valence (EMG) 1 0.053 0.097 -0.011 0.006 0.009 -0.033 -0.073 -0.078 -0.02 0.03 0.046 0.049 -0.017 -0.075 0.151 -0.012 -0.073 Activation (EAR) 0.053 1 0.04 0.114 -0.066 -0.092 -0.082 -0.11 0.081 0.084 -0.036 -0.032 -0.167 -0.24 -0.118 0.066 0.153 -0.033 Motivation (Joystick) 0.097 0.04 1 0.051 -0.029 -0.144 -0.096 -0.111 0.126 0.126 -0.027 -0.149 0.015 -0.03 -0.012 -0.164 0.104 0.05 IBI -0.011 0.114 0.051 1 0.137 0.157 0.121 0.113 0.436 0.136 -0.125 0.042 0.023 0.068 0.002 0.13 0.602 0.427 IBI-UHF 0.006 -0.066 -0.029 0.137 1 0.385 0.248 0.304 0.008 -0.03 -0.016 -0.06 0.02 0.114 -0.097 0.324 -0.022 0.272 IBI-HF 0.009 -0.092 -0.144 0.157 0.385 1 0.681 0.766 -0.009 0.005 -0.008 -0.004 0.187 0.128 0.055 0.293 -0.076 0.404 IBI-LF -0.033 -0.082 -0.096 0.121 0.248 0.681 1 0.798 -0.005 0.058 0.031 0.002 0.159 0.102 0.065 0.257 -0.066 0.507 IBI-VLF -0.073 -0.11 -0.111 0.113 0.304 0.766 0.798 1 0.035 0.088 0.035 -0.015 0.134 0.076 0.038 0.301 -0.059 0.524 IBI-RMSSD -0.078 0.081 0.126 0.436 0.008 -0.009 -0.005 0.035 1 0.512 -0.155 0.026 0.047 0.024 0.055 -0.281 0.752 0.473 IBI-SDNN -0.02 0.084 0.126 0.136 -0.03 0.005 0.058 0.088 0.512 1 -0.012 -0.088 -0.006 0.021 0.068 -0.154 0.547 0.443 PVA 0.03 -0.036 -0.027 -0.125 -0.016 -0.008 0.031 0.035 -0.155 -0.012 1 0.007 0.037 0.006 0.039 0.162 -0.365 -0.318 PVA-UHF 0.046 -0.032 -0.149 0.042 -0.06 -0.004 0.002 -0.015 0.026 -0.088 0.007 1 0.228 -0.081 0.35 -0.159 0.118 -0.146 PVA-HF 0.049 -0.167 0.015 0.023 0.02 0.187 0.159 0.134 0.047 -0.006 0.037 0.228 1 0.262 0.228 -0.257 -0.023 0.212 PVA-LF -0.017 -0.24 -0.03 0.068 0.114 0.128 0.102 0.076 0.024 0.021 0.006 -0.081 0.262 1 0.248 -0.184 -0.171 0.235 PVA-VLF -0.075 -0.118 -0.012 0.002 -0.097 0.055 0.065 0.038 0.055 0.068 0.039 0.35 0.228 0.248 1 -0.506 -0.029 0.176 Valence (Predicted) 0.151 0.066 -0.164 0.13 0.324 0.293 0.257 0.301 -0.281 -0.154 0.162 -0.159 -0.257 -0.184 -0.506 1 -0.156 0.009 Activation (Predicted) -0.012 0.153 0.104 0.602 -0.022 -0.076 -0.066 -0.059 0.752 0.547 -0.365 0.118 -0.023 -0.171 -0.029 -0.156 1 0.431 Motivation (Predicted) -0.073 -0.033 0.05 0.427 0.272 0.404 0.507 0.524 0.473 0.443 -0.318 -0.146 0.212 0.235 0.176 0.009 0.431 1 Note. The uncalibrated average (per subject) correlation of the validation measures and PPG predicted measures and their squares. The correlations of the features are also included. EMG=electromyography, EAR=eye aspect ratio, IBI = inter-beat interval, PVA = pulse volume amplitude, UHF = ultra-high frequency, HF = high frequency, LF = low frequency, VLF = very low frequency, RMSSD = root mean squared successive differences, SDNN = standard deviation normal to normal beat. Between-Subject Correlation Table Age 1 Height -0.11 1 Weight 0.15 0.52 1 BMI 0.22 0.19 0.94 1 Gender 0.178 -0.73 -0.51 -0.288 1 Avg Sleep -0.35 -0.08 0.007 0.031 -0.04 1 Physical activity 0.036 -0.092 0.058 0.094 0.136 -0.014 1 ADHD -0.05 0.059 0.074 0.051 -0.19 0.083 0.081 1 Anxiety disorder 0.088 -0.074 0.011 0.044 -0.024 -0.046 -0.015 0.16 1 Depressive disorder 0.133 -0.167 -0.005 0.058 0.144 -0.048 -0.011 0.069 0.39 1 Monitoring acceptance -0.29 0.054 -0.134 -0.175 -0.058 0.052 -0.027 0.089 0.167 -0.109 1 Withdrawal 0.364 0.117 0.062 0.035 -0.01 -0.33 -0.36 -0.02 -0.12 0.023 -0.12 1 Volatility 0.222 -0.17 -0.15 -0.095 0.302 -0.31 -0.23 -0.44 0.229 0.262 0.074 0.353 1 Compassion -0.15 0.042 0.058 0.065 0.009 0.025 0.026 -0.13 -0.08 -0.1 -0.03 -0.17 -0.03 1 Politeness 0.111 -0.23 0.028 0.138 0.198 -0.01 0.003 -0.03 -0.3 -0.22 -0.22 0.218 -0.21 0.277 1 Industriousness -0.035 -0.145 -0.04 0.011 0.001 0.214 0.47 0.16 -0.164 -0.259 -0.079 -0.515 -0.458 0.056 0.187 1 Orderliness 0.15 -0.145 -0.121 -0.076 0.109 -0.025 0.248 0.059 -0.169 -0.15 -0.047 -0.117 0.038 -0.033 0.179 0.613 1 Enthusiasm 0.06 -0.37 -0.211 -0.086 0.474 -0.036 0.409 -0.065 0.311 0.34 0.044 -0.286 0.25 0.324 -0.028 0.227 0.134 1 Assertiveness -0.04 -0.04 -0.11 -0.104 0.113 0.04 0.248 0.105 0.251 0.043 0.086 -0.39 0.095 0.23 -0.18 0.233 0.036 0.458 1 Intellect 0.022 -0.052 0.168 0.215 -0.051 0.076 0.289 -0.07 -0.247 -0.372 0.029 -0.285 -0.382 0.091 0.015 0.528 0.314 -0.024 0.152 1 Openness -0.02 0.216 0.076 -46.24 -0.07 -0.04 -0.1 -0.24 -0.29 -0.16 -0.29 0.066 -0.15 0.193 -0.02 -0.15 -0.16 -0.16 0.116 0.406 1 Neuroticism 0.344 -0.051 -0.072 -0.046 0.199 -0.384 -0.349 -0.31 0.093 0.19 -0.013 0.772 0.867 -0.109 -0.024 -0.585 -0.036 0.017 -0.141 -0.411 -0.067 1 Agreeableness -0.047 -0.098 0.056 0.121 0.114 0.012 0.02 -0.105 -0.22 -0.193 -0.143 0.002 -0.133 0.851 0.74 0.141 0.075 0.212 0.062 0.071 0.126 -0.089 1 Conscientiousness 0.067 -0.162 -0.091 -0.037 0.063 0.102 0.396 0.12 -0.186 -0.226 -0.07 -0.346 -0.227 0.011 0.204 0.893 0.903 0.2 0.147 0.466 -0.172 -0.339 0.119 1 Extraversion 0.007 -0.23 -0.19 -0.112 0.332 0.005 0.379 0.029 0.327 0.214 0.078 -0.4 0.196 0.322 -0.13 0.269 0.097 0.835 0.871 0.081 -0.02 -0.08 0.155 0.201 1 Openness/Intellect 0.003 0.089 0.149 0.135 -0.072 0.025 0.128 -0.181 -0.32 -0.324 -0.146 -0.142 -0.324 0.165 -20.66 0.249 0.107 -0.107 0.161 0.857 0.818 -0.296 0.116 0.196 0.04 1 Riskiness -0.16 0.176 -0.04 -0.124 -0.01 0.091 0.269 0.041 0.126 -0.05 -0.03 -0.31 -0.26 0.053 -0.32 0.002 -0.47 0.187 0.423 0.016 0.279 -0.34 -0.14 -0.26 0.365 0.167 1 PHQ8 0.199 0.028 -0.07 -0.114 -0.17 -0.36 -0.25 -0.15 -0.03 -0.14 -0.23 0.339 0.07 -0.12 0.047 -0.36 -0.17 -0.38 -0.37 -0.15 0.229 0.228 -0.06 -0.29 -0.44 0.035 -0.1 1 TAS DIF 0.146 0.016 -0.07 -0.094 -0.16 -0.24 -0.03 0.053 0.116 -0.2 -0.06 0.24 0.031 -0.36 -0.12 -0.27 -0.08 -0.23 -0.02 -0.24 -0.01 0.149 -0.32 -0.19 -0.14 -0.16 0.122 0.618 1 TAS DDF -0.05 0.093 0.053 -0.003 -0.33 -0.15 -0.02 0.112 -0.02 -0.25 -0.05 0.072 -0.28 -0.32 -0.08 -0.09 -0.18 -0.47 -0.26 0.056 0.019 -0.15 -0.27 -0.15 -0.42 0.046 0.085 0.615 0.577 1 TAS EOT -0.08 -0.12 0.047 0.095 -0.04 0.053 0.113 -0.13 0.242 0.134 -0.14 0.028 0.124 -0.21 0.071 -0.07 -0.13 -0.03 -0.12 -0.39 -0.55 0.099 -0.11 -0.11 -0.09 -0.55 -0.14 0.15 0.301 0.355 1 TAS 0.018 0.012 0.009 -0.014 -0.24 -0.16 0.012 0.034 0.128 -0.16 -0.1 0.155 -0.07 -0.39 -0.07 -0.19 -0.16 -0.33 -0.17 -0.22 -0.18 0.033 -0.31 -0.2 -0.29 -0.24 0.05 0.624 0.836 0.851 0.642 1 Note. The between-subjects correlations. For personality, subjects completed the big five personality aspects test (DeYoung et al., 2007) and the risk propensity scale (Highhouse et al., 2022; Meertens & Lion, 2008). Cronbach’s alphas for the personality scales are reported for neuroticism ( α =0.87) and sub-traits withdrawal ( α =0.76) and volatility ( α =0.86), agreeableness ( α =0.70) and sub-traits compassion ( α =0.78) and politeness ( α =0.28), conscientiousness ( α =0.91) and sub traits industriousness ( α =0.87) and orderliness ( α =0.86), extraversion ( α =0.89) and sub traits enthusiasm ( α =0.77) and assertiveness ( α =0.85), openness ( α =0.87) and sub traits intellect ( α =0.81) and aesthetic openness ( α =0.83), and risk propensity ( α =0.68). Gender (1 = Male, 2 = Female), Contraception (1 = uses hormonal contraception, 2 = does not), for Visual deficiency, Anxiety disorder, Depressive disorder (1 = diagnosed, 2 = not diagnosed). BMI = body mass index, Hours asleep = hours slept last night, Avg sleep = average nightly sleep, PHQ-8 = patient health questionnaire version 8, TAS = Toronto alexithymia scale, DIF = difficulty identifying feelings, DDF = difficulty describing feelings, EOT = externally oriented thinking. Appendix C Mind-Body Health Questionnaire How many hours ago was the last time that you had an alcoholic beverage? (If more than 12 hours, type ”-”.) How many hours has it been since you had a caffeinated beverage? (If more than 8 hours, type ”-”.) Do you use tobacco or nicotine products of any kind? Yes No Do you currently take selective serotonin reuptake inhibitors (SSRIs)? (Some examples: Prozac, Paxil, Zoloft, Luvox, Lexapro, Celexa). Yes No How many hours has it been since you last ate? Do you currently take hormonally-based contraception (including birth control pills, skin patches, or vaginal rings)? Yes No (FEMALE PARTICIPANTS ONLY IF NOT ON HORMONE-BASED CONTRACEPTIVE) What phase of the menstrual cycle are you currently in? Beginning Middle End Unknown N/A How many hours of sleep did you get last night? < 2 hours 2-4 hours 4-6 hours 6-8 hours 8+ hours How many hours of sleep do you average per night? Did you engage in vigorous exercise within the last 2 hours? Yes No On average, how often do you engage in physical activity for at least 30-minute sessions? Never Rarely One to two times per month One to two times per week Three to four days per week Five to six days per week Seven days per week Dominant hand Right Left Do you have amblyopia (lazy eye)? Yes No If yes, which eye is slower? Right Left N/A Do you currently have, or have you ever had any of the following? Arthritis Auditory deficiency that noticeably affects your ability to hear Visual deficiency that noticeably affects your ability to see Hypertension (high blood pressure) Hypotension (low blood pressure) Have you ever been diagnosed as having: Attention deficit/hyperactivity disorder Autism spectrum disorder or Asperger syndrome (ASD) Phobia of any type Post-traumatic stress disorder Panic disorder Obsessive Compulsive Disorder Social Anxiety Disorder Anxiety disorder of any type Depressive disorder of any type Bipolar disorder or Seasonal affective disorder Affective disorder of any type References 1. Amer, A. Y. A., Wouters, F., Vranken, J., De Korte-de Boer, D., Smit-Fun, V., Duflot, P., Beaupain, M.-H., Vandervoort, P., Luca, S., Aerts, J.-M., & Vanrumste, B. (2020). Vital signs prediction and early warning score calculation based on continuous monitoring of hospitalised patients using wearable technology. Sensors , 20 (22), 6593. https://doi.org/10.3390/s20226593Asarnow, J. R., Berk, M., Zhang, L., Wang, P., & Tang, L. (2017). Emergency department youth patients with suicidal ideation or attempts: Predicting suicide attempts through 18 months of follow‐up. Suicide and Life-Threatening Behavior , 47 (5), 551–566. https://doi.org/10.1111/sltb.12309Ascheid, S., Wessa, M., & Linke, J. O. (2019). Effects of valence and arousal on implicit approach/ avoidance tendencies: A fMRI study. Neuropsychologia , 131 , 333–341. https://doi.org/10.1016/j.neuropsychologia.2019.05.028Aslam, R. D., Liew, J., & Besi, E. (2023). Is 1:1000 adrenaline as a topical haemostat an effective alternative to control bleeding in dentistry and oral surgery? British Dental Journal , 235 (1), 29–34. https://doi.org/10.1038/s41415-023-6010-7Axelrod, R. (1980). Effective choice in the prisoner’s dilemma. Journal of Conflict Resolution , 24 (1), 3–25. https://doi.org/10.1177/002200278002400101Azorin-Lopez, J., Saval-Calvo, M., Fuster-Guillo, A., & Oliver-Albert, A. (2014). A predictive model for recognizing human behaviour based on trajectory representation. 2014 International Joint Conference on Neural Networks (IJCNN) , 1494–1501. https://doi.org/10.1109/IJCNN.2014.6889883Bagby, R. M., Parker, J. D. A., & Taylor, G. J. (1994). The twenty-item Toronto alexithymia scale—I. Item selection and cross-validation of the factor structure. Journal of Psychosomatic Research , 38 (1), 23–32. https://doi.org/10.1016/0022-3999(94)90005-1Bagby, R. M., Taylor, G. J., & Parker, J. D. A. (1994). The twenty-item Toronto alexithymia scale—II. Convergent, discriminant, and concurrent validity. Journal of Psychosomatic Research , 38 (1), 33–40. https://doi.org/10.1016/0022-3999(94)90006-XBaltrusaitis, T., Zadeh, A., Lim, Y. C., & Morency, L.-P. (2018). OpenFace 2.0: Facial behavior analysis toolkit. 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018) , 59–66. https://doi.org/10.1109/FG.2018.00019Barrett, L. F., & Russell, J. A. (1999). The structure of current affect: Controversies and emerging consensus. Current Directions in Psychological Science , 8 (1), 10–14. https://doi.org/10.1111/1467-8721.00003Berntson, G. G., Cacioppo, J. T., Quigley, K. S., & Fabro, V. T. (1994). Autonomic space and psychophysiological response. Psychophysiology , 31 , 44–61.Cacioppo, J. T., & Tassinary, L. G. (1990). Inferring psychological significance from physiological signals. American Psychologist .Castaneda, M., Esparza, A., Ghamari, M., Soltanpur, C., & Nazeran, H. (2018). A review on wearable photoplethysmography sensors and their potential future applications in health care. International Journal of Biosensors & Bioelectronics , 4 (4). https://doi.org/10.15406/ijbsbe.2018.04.00125Catchpole, R. E. H., & Gretton, H. M. (2003). The predictive validity of risk assessment with violent young offenders: A 1-year examination of criminal outcome. Criminal Justice and Behavior , 30 (6), 688–708. https://doi.org/10.1177/0093854803256455Charlton, P. H., Allen, J., Bailón, R., Baker, S., Behar, J. A., Chen, F., Clifford, G. D., Clifton, D. A., Davies, H. J., Ding, C., Ding, X., Dunn, J., Elgendi, M., Ferdoushi, M., Franklin, D., Gil, E., Hassan, M. F., Hernesniemi, J., Hu, X., … Zhu, T. (2023). The 2023 wearable photoplethysmography roadmap. Physiological Measurement , 44 (11), 111001. https://doi.org/10.1088/1361-6579/acead2Christie, I. C., & Friedman, B. H. (2004). Autonomic specificity of discrete emotion and dimensions of affective space: A multivariate approach. International Journal of Psychophysiology , 51 (2), 143–153. https://doi.org/10.1016/j.ijpsycho.2003.08.002Coid, J., Yang, M., Ullrich, S., Zhang, T., Sizmur, S., Roberts, C., Farrington, D. P., & Rogers, R. D. (2009). Gender differences in structured risk assessment: Comparing the accuracy of five instruments. Journal of Consulting and Clinical Psychology , 77 (2), 337–348. https://doi.org/10.1037/a0015155Davidson, R. J. (1993). The neuropsychology of emotion and affective style. In Biological and Neuropsychological Approaches .Davidson, R. J. (2005). Neuropsychological perspectives on affective styles and their cognitive consequences. In T. Dalgleish & M. J. Power (Eds.), Handbook of Cognition and Emotion (pp. 103–123). John Wiley & Sons, Ltd. https://doi.org/10.1002/0470013494.ch6Dewi, C., Chen, R.-C., Chang, C.-W., Wu, S.-H., Jiang, X., & Yu, H. (2022). Eye aspect ratio for real-time drowsiness detection to improve driver safety. Electronics , 11 (19), 3183. https://doi.org/10.3390/electronics11193183DeYoung, C. G., Quilty, L. C., & Peterson, J. B. (2007). Between facets and domains: 10 aspects of the Big Five. Journal of Personality and Social Psychology , 93 (5), 880–896. https://doi.org/10.1037/0022-3514.93.5.880Fontaine, J. R. J., Scherer, K. R., Roesch, E. B., & Ellsworth, P. C. (2007). The world of emotions is not two-dimensional. Psychological Science , 18 (12), 1050–1057. https://doi.org/10.1111/j.1467-9280.2007.02024.xFredericks, A., & Moore, J. W. (1974). Selective autonomic blockade of conditioned and unconditioned heart rate changes in rabbits. Pharmacology Biochemistry and Behavior , 2 , 493–501.Friedman, B. H., Stephens, C. L., & Thayer, J. F. (2014). Redundancy analysis of autonomic and self-reported, responses to induced emotions. Biological Psychology , 98 , 19–28. https://doi.org/10.1016/j.biopsycho.2013.12.006Friedman, B. H., & Thayer, J. F. (2024). Is emotion physiology more compatible with discrete, dimensional, or appraisal accounts? In Emotion theory: The Routledge comprehensive guide: (Vol. 1, pp. 488–510). Routledge, Taylor & Francis.Goldstein, R. B. (1991). The prediction of suicide: Sensitivity, specificity, and predictive value of a multivariate model applied to suicide among 1906 patients with affective disorders. Archives of General Psychiatry , 48 (5), 418. https://doi.org/10.1001/archpsyc.1991.01810290030004Grant, S. S., Kim, K., & Friedman, B. H. (2023). How long is long enough? Controlling for acute caffeine intake in cardiovascular research. Brain Sciences , 13 (2), 224. https://doi.org/10.3390/brainsci13020224Gross, J. J., & Levenson, R. W. (1995). Emotion elicitation using films. Cognition & Emotion , 9 (1), 87–108. https://doi.org/10.1080/02699939508408966Harmon-Jones, E., & Allen, J. J. B. (1998). Anger and frontal brain activity: EEG asymmetry consistent with approach motivation despite negative affective valence. Journal of Personality and Social Psychology , 74 (5), 1310–1316. https://doi.org/10.1037/0022-3514.74.5.1310Hewig, J., Hagemann, D., Seifert, J., Gollwitzer, M., Naumann, E., & Bartussek, D. (2005). A revised film set for the induction of basic emotions. Cognition & Emotion , 19 (7), 1095–1109. https://doi.org/10.1080/02699930541000084Highhouse, S., Wang, Y., & Zhang, D. C. (2022). Is risk propensity unique from the big five factors of personality? A meta-analytic investigation. Journal of Research in Personality , 98 , 104206. https://doi.org/10.1016/j.jrp.2022.104206Hoofs, V., Carsten, T., Boehler, C. N., & Krebs, R. M. (2019). Interactions between incentive valence and action information in a cued approach–avoidance task. Psychological Research , 83 (1), 13–25. https://doi.org/10.1007/s00426-018-0975-xJänig, W., & Häbler, H.-J. (2000). Specificity in the organization of the autonomic nervous system: A basis for precise neural regulation of homeostatic and protective body functions. In Progress in Brain Research (Vol. 122, pp. 351–367). Elsevier. https://doi.org/10.1016/S0079-6123(08)62150-0Kaczmarek, L. D., Behnke, M., Enko, J., Kosakowski, M., Guzik, P., & Hughes, B. M. (2021). Splitting the affective atom: Divergence of valence and approach-avoidance motivation during a dynamic emotional experience. Current Psychology , 40 (7), 3272–3283. https://doi.org/10.1007/s12144-019-00264-3Kissel, H. A., & Friedman, B. H. (2023). Participant diversity in Psychophysiology . Psychophysiology , 60 (11), e14369. https://doi.org/10.1111/psyp.14369Koenig, J., & Thayer, J. F. (2016). Sex differences in healthy human heart rate variability: A meta-analysis. Neuroscience & Biobehavioral Reviews , 64 , 288–310. https://doi.org/10.1016/j.neubiorev.2016.03.007Kragel, P. A., & LaBar, K. S. (2013). Multivariate pattern classification reveals autonomic and experiential representations of discrete emotions. Emotion , 13 (4), 681–690. https://doi.org/10.1037/a0031820Kreibig, S. D. (2010). Autonomic nervous system activity in emotion: A review. Biological Psychology , 84 (3), 394–421. https://doi.org/10.1016/j.biopsycho.2010.03.010Krishnan, A., Williams, L. J., McIntosh, A. R., & Abdi, H. (2011). Partial least squares (PLS) methods for neuroimaging: A tutorial and review. NeuroImage , 56 (2), 455–475. https://doi.org/10.1016/j.neuroimage.2010.07.034Kroenke, K., Strine, T. W., Spitzer, R. L., Williams, J. B. W., Berry, J. T., & Mokdad, A. H. (2009). The PHQ-8 as a measure of current depression in the general population. Journal of Affective Disorders , 114 (1–3), 163–173. https://doi.org/10.1016/j.jad.2008.06.026Langley, J. N. (1921). Part 1. In The Autonomic Nervous System . Cambridge.Lekkas, D., Klein, R. J., & Jacobson, N. C. (2021). Predicting acute suicidal ideation on Instagram using ensemble machine learning models. Internet Interventions , 25 , 100424. https://doi.org/10.1016/j.invent.2021.100424Makowski, D., Pham, T., Lau, Z. J., Brammer, J. C., Lespinasse, F., Pham, H., Schölzel, C., & Chen, S. H. A. (2021). NeuroKit2: A Python toolbox for neurophysiological signal processing. Behavior Research Methods , 53 (4), 1689–1696. https://doi.org/10.3758/s13428-020-01516-yMcClafferty, S. R., & Friedman, B. H. (2024). A model for predicting emotional dimensions within the individual: Analysis of the POPANE dataset [abstract]. Psychophysiology , 61 (S1), S250.McGinley, J. J., & Friedman, B. H. (2017). Autonomic specificity in emotion: The induction method matters. International Journal of Psychophysiology , 118 , 48–57. https://doi.org/10.1016/j.ijpsycho.2017.06.002McKinney, W. (2010). Data structures for statistical computing in Python . 56–61. https://doi.org/10.25080/Majora-92bf1922-00aMcNames, J., & Aboy, M. (2008). Statistical modeling of cardiovascular signals and parameter estimation based on the extended Kalman filter. IEEE Transactions on Biomedical Engineering , 55 (1), 119–129. https://doi.org/10.1109/TBME.2007.910648Meertens, R. M., & Lion, R. (2008). Measuring an individual’s tendency to take risks: The risk propensity scale 1 . Journal of Applied Social Psychology , 38 (6), 1506–1520. https://doi.org/10.1111/j.1559-1816.2008.00357.xMohammadi, G., & Vuilleumier, P. (2020). A multi-componential approach to emotion recognition and the effect of personality. IEEE Transactions on Affective Computing , 13 (3), 127–1139. https://doi.org/10.48550/ARXIV.2010.11370Nyklicek, I., Thayer, J. F., & Van Doornen, L. J. P. (1997). Cardiorespiratory differentiation of musically-induced emotions. Journal of Psychophysiology , 11 , 304–321.Panayiotou, G. (2008). Emotional dimensions reflected in ratings of affective scripts. Personality and Individual Differences , 44 (8), 1795–1806. https://doi.org/10.1016/j.paid.2008.02.006Phaf, R. H., Mohr, S. E., Rotteveel, M., & Wicherts, J. M. (2014). Approach, avoidance, and affect: A meta-analysis of approach-avoidance tendencies in manual reaction time tasks. Frontiers in Psychology , 5 . https://doi.org/10.3389/fpsyg.2014.00378Quigley, K. S., Gianaros, P. J., Norman, G. J., Jennings, J. R., Berntson, G. G., & De Geus, E. J. C. (2024). Publication guidelines for human heart rate and heart rate variability studies in psychophysiology—Part 1: Physiological underpinnings and foundations of measurement. Psychophysiology , 61 (9), e14604. https://doi.org/10.1111/psyp.14604Reynaud, E., El Khoury-Malhame, M., Rossier, J., Blin, O., & Khalfa, S. (2012). Neuroticism modifies psychophysiological responses to fearful films. PLoS ONE , 7 (3), e32413. https://doi.org/10.1371/journal.pone.0032413Rossegger, A., Laubacher, A., Moskvitin, K., Villmar, T., Palermo, G. B., & Endrass, J. (2011). Risk assessment instruments in repeat offending: The usefulness of FOTRES. International Journal of Offender Therapy and Comparative Criminology , 55 (5), 716–731. https://doi.org/10.1177/0306624X09360662Saganowski, S., Perz, B., Polak, A. G., & Kazienko, P. (2023). Emotion recognition for everyday life using physiological signals from wearables: A systematic literature review. IEEE Transactions on Affective Computing , 14 (3), 1876–1897. https://doi.org/10.1109/TAFFC.2022.3176135Sarker, I. H., Colman, A., Han, J., Khan, A. I., Abushark, Y. B., & Salah, K. (2020). BehavDT: A behavioral decision tree learning to build user-centric context-aware predictive model. Mobile Networks and Applications , 25 (3), 1151–1161. https://doi.org/10.1007/s11036-019-01443-zSchaefer, A., Nils, F., Sanchez, X., & Philippot, P. (2010). Assessing the effectiveness of a large database of emotion-eliciting films: A new tool for emotion researchers. Cognition & Emotion , 24 (7), 1153–1172. https://doi.org/10.1080/02699930903274322Schmidt, K. L., & Cohn, J. F. (2008). Human facial expressions as adaptations: Evolutionary questions in facial expression research. American Journal of Physical Anthropology , 33 , 3–24.Shaffer, F., & Ginsberg, J. P. (2017). An overview of heart rate variability metrics and norms. Frontiers in Public Health , 5 , 258. https://doi.org/10.3389/fpubh.2017.00258Shiffman, S., Stone, A. A., & Hufford, M. R. (2008). Ecological momentary assessment. Annual Review of Clinical Psychology , 4 (1), 1–32. https://doi.org/10.1146/annurev.clinpsy.3.022806.091415Simhon, R., Meiran, N., Shvimmer, S., Amihai, L., Yitzhaky, Y., Rosenblatt, J., & Gilead, M. (2025). Beyond valence and arousal: Distributed facial patterns linked to specific emotions cannot be reduced to core affect. Affective Science . https://doi.org/10.1007/s42761-025-00311-2Stemmler, G., Aue, T., & Wacker, J. (2007). Anger and fear: Separable effects of emotion and motivational direction on somatovisceral responses. International Journal of Psychophysiology , 66 (2), 141–153. https://doi.org/10.1016/j.ijpsycho.2007.03.019Stephens, C. L., Christie, I. C., & Friedman, B. H. (2010). Autonomic specificity of basic emotions: Evidence from pattern classification and cluster analysis. Biological Psychology , 84 (3), 463–473. https://doi.org/10.1016/j.biopsycho.2010.03.014Valenza, G., Citi, L., Saul, J. P., & Barbieri, R. (2018). Measures of sympathetic and parasympathetic autonomic outflow from heartbeat dynamics. Journal of Applied Physiology , 125 (1), 19–39. https://doi.org/10.1152/japplphysiol.00842.2017Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D., Burovski, E., Peterson, P., Weckesser, W., Bright, J., Van Der Walt, S. J., Brett, M., Wilson, J., Millman, K. J., Mayorov, N., Nelson, A. R. J., Jones, E., Kern, R., Larson, E., … Vázquez-Baeza, Y. (2020). SciPy 1.0: Fundamental algorithms for scientific computing in Python. Nature Methods , 17 (3), 261–272. https://doi.org/10.1038/s41592-019-0686-2Watson, D., Anna, L., & Tellegen, A. (1988). Development and validation of brief measures of positive and negative affect: The PANAS scales. Journal of Personality and Social Psychology , 54 (6), 1063–1070.Wehrwein, E. A., Orer, H. S., & Barman, S. M. (2016). Overview of the anatomy, physiology, and pharmacology of the autonomic nervous system. In R. Terjung (Ed.), Comprehensive Physiology (1st ed., pp. 1239–1278). Wiley. https://doi.org/10.1002/cphy.c150037Wingenbach, T. S. H. (2023). Facial EMG – Investigating the interplay of facial muscles and emotions. In P. S. Boggio, T. S. H. Wingenbach, M. L. Da Silveira Coêlho, W. E. Comfort, L. Murrins Marques, & M. V. C. Alves (Eds.), Social and Affective Neuroscience of Everyday Human Interaction (pp. 283–300). Springer International Publishing. https://doi.org/10.1007/978-3-031-08651-9_17 Crossref Google Scholar Information & Authors Information Version history V1 Version 1 27 January 2026 Copyright This work is licensed under a Non Exclusive No Reuse License. Authors Affiliations Shane R. McClafferty 0009-0009-3354-1130 Virginia Polytechnic Institute and State University View all articles by this author Bruce Friedman 0000-0003-0814-0393 [email protected] Virginia Polytechnic Institute and State University View all articles by this author Metrics & Citations Metrics Article Usage 220 views 96 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Shane R. McClafferty, Bruce Friedman. A dynamic architecture linking autonomic activity to emotional dimensions: Real-time estimation from photoplethysmography. Authorea . 27 January 2026. DOI: https://doi.org/10.22541/au.176951752.28031887/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176951752.28031887/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fe3250a1e74aa64',t:'MTc3OTE5NDE1OQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();