Abstract
This study proposes the application of a comprehensive signal processing framework, based on inhomogeneous pointprocess models of heartbeat dynamics, to instantaneously assess affective haptic perception using electrocardiogramderived information exclusively. The framework relies on inverseGaussian pointprocesses with Laguerre expansion of the nonlinear WienerVolterra kernels, accounting for the longterm information given by the past heartbeat events. Up to cubicorder nonlinearities allow for an instantaneous estimation of the dynamic spectrum and bispectrum of the considered cardiovascular dynamics, as well as for instantaneous measures of complexity, through Lyapunov exponents and entropy. Shortterm caresslike stimuli were administered for 4.3–25 seconds on the forearms of 32 healthy volunteers (16 females) through a wearable haptic device, by selectively superimposing two levels of force, 2 N and 6 N, and two levels of velocity, 9.4 mm/s and 65 mm/s. Results demonstrated that our instantaneous linear and nonlinear features were able to finely characterize the affective haptic perception, with a recognition accuracy of 69.79% along the force dimension, and 81.25% along the velocity dimension.
Introduction
Emotions such as anger, fear, and love can be communicated through touch to the forearm^{1,2}. Indeed, the human tactile system is very sensible to affective stimuli like caressing, stroking or patting, thanks to the activity of lowthreshold mechanoreceptors called CT fibers^{1,2}. Through this knowledge, the relationship between the physical characteristics of a haptic stimulus and its pleasantness was previously studied^{3,4,5}. Most pleasant stimuli were associated to the lowest applied force^{3}, with movements velocity in the range 1–10 cm/s (3 cm/s considered as the most pleasant)^{4}. Of note, these findings are consistent with stimuli administered in different parts of the body (forehead, arm, palm, thigh)^{6}.
The dynamics of hearth contractions, which is modulated by the Autonomic Nervous System (ANS) activity, is significantly affected by emotion elicitation and by passive touch^{7,8,9,10,11,12,13,14,15}. Specifically, linear and nonlinear analysis of Heart Rate Variability (HRV)^{16} revealed parasympatetic changes after massage^{13,17,18}. Of note, the use of nonlinear quantifiers is justified by the fact that heartbeat dynamics shows exemplary nonlinear behavior, mainly generated through integration of multiple neural signaling at the level of the sinoatrial node^{19}.
Major shortcomings of current signal processing methodologies used to objectively assess affective haptic stimuli are related to the stimulus duration. Standard linear and nonlinear HRV measures, in fact, require relatively longterm recordings to accurately characterize the emotional state of a subject. For example, observations from 30 seconds to 5 minutes are the least needed for simple HRV spectral analysis^{16}. Nevertheless, actual affective haptic stimuli, such as caresses, usually last for few seconds. Therefore, standard signal processing methodologies are not suitable to perform such an assessment. Even considering affective haptic stimuli lasting for more than 30 seconds, the actual emotional perception will be compromised by the saturation of the CT fibers activity^{3,5}. Furthermore, from a methodological point of view, standard methods are generally not suitable to provide accurate nonlinear fitting in the absence of information regarding the system phase space, and their estimates are biased by noise statistics (e.g., white or 1/f noise) underlying physiological dynamics, and interpolation techniques^{20,21}.
To overcome these limitations, we here propose the application of inhomogeneous pointprocess models of heartbeat dynamics to assess shortterm affective haptic perception in an instantaneous fashion, such that caresslike stimuli lasting for 4.3–25 seconds can be therefore automatically recognized. As an input, the proposed methodology takes unevenly sampled heartbeat interval information exclusively, derived from the electrocardiogram. To our knowledge, the use instantaneous measures of linear and nonlinear autonomic control derived from cardiovascular dynamics in order to characterize and automatically discern physical properties of affective haptic perception is novel in the current literature. Motivations of this study are related to scientific knowledge, and possible potential application in the field of experimental psychology and assessment of patients with mental disorders.
The model underlying these estimates comprises probability density functions (pdf) characterizing and predicting the time until the next event occurs, as a function of the past history. Of note, the derived instantaneous measures can be estimated without applying any interpolation techniques to the original series, and are associated to effective goodnessoffit measures^{20,21,22,23}. The framework relies on inverseGaussian pointprocesses with Laguerre expansion of the nonlinear WienerVolterra kernels, accounting for the longterm information given by the past heartbeat events^{20,21,23}. This methodology has been successfully applied to the study of brainheart interaction^{24}, assessment of Parkinson’s disease^{25} and depression in bipolar disorder^{26,27}.
In this study, the use of instantaneous nonlinear estimates is emphasized. We consider up to cubicorder nonlinearities, allowing for a comprehensive assessment through the dynamic spectrum and bispectrum of the considered cardiovascular dynamics^{23}, as well as for instantaneous measures of complexity, through Lyapunov exponents and entropy^{20,21}. As mentioned above, although the underlying physiological dynamics linked to these measures is still unknown, it has been widely accepted that the quantification of ANS nonlinear and complex dynamics provides meaningful information on psychophysiological and pathological states^{28,29,30,31,32}. Furthermore, measures of timevarying complexity have enhanced discriminant power with respect to standard complexity measures^{20,21}.
Shortterm caresslike stimuli were administered on the forearms of 32 healthy volunteers (16 females) through an adhoc wearable haptic device able to mimic caresses^{33}. In order to develop a processing chain able to actually discern affective perception using electrocardiogramderived information exclusively, we moved from looking for simple statistical differences between stimulus physical characteristics to building an automatic classification algorithm.
Next, we describe the mathematical bases of the inhomogeneous pointprocess modeling, extending the previously proposed approach reported in^{23}. Then, a brief description of related instantaneous estimates is reported, allowing with experimental setup and results which are expressed in terms of confusion matrix of classification and complementary statistics.
Materials and Methods
PointProcess Model of the Heartbeat
A random point process is a stochastic process characterizing the occurrence in time of discrete events. Here, this methodology aims to deriving estimates from the intrinsically discrete, unevenly sampled heartbeat events at each moment in time. To this extent, starting from the electrocardiogram (ECG), it is possible to define the pdf predicting the next ventricular contraction (Rpeaks), through a parametric formulation of the past heartbeat events. As such a pdf is a continuous function defined in the time domain, it is possible to obtain instantaneous cardiovascular estimates at any desired time resolution.
Mathematically, motivated by both goodnessoffit and physiological reasons, the pdf characterizing cardiovascular control dynamics follows an inverseGaussian model^{20,21,23}:
with:
• t ∈ (0, T], the observation interval;
• the Rwave events, in this study, detected from the ECG, and the index of the previous Rwave event before time t
• the times of the events;
• RR_{j} = u_{j} − u_{j−1} > 0 the j^{th} RR interval;
• N(t) = max{k : u_{k} ≤ t} the sample path of the counting process of the RR interval seris;
•
• ;
• ξ(t) the vector of the model timevarying parameters;
• the firstmoment statistic (mean) of the distribution;
• ξ_{0}(t) > 0 the shape parameter of the inverse Gaussian distribution;
Accounting for history dependence, is thus able to predict the next heartbeat event, being parametrized in its firstorder moment as a Nonlinear Autoregressive model with Laguerre expansions (NARL) of the Volterra terms:
where
is the output of the Laguerre filters just before time t, and
is the i^{th}order discrete time orthonormal Laguerre function, with n ≥ 0 and α is the discretetime Laguerre parameter (0 < α < 1) which determines the rate of exponential asymptotic decline of these functions. The coefficients g_{0}, {g_{1}(i)}, and {g_{2}(i, j)} correspond to the timevarying zero, first, secondorder NARL coefficients, respectively. Considering the derivative RR series improves the achievement of stationarity within the sliding time window W (in this work we have chosen W = 90 sec.)^{23}.
Of note, the corresponding nonlinear autoregressive (NAR) WienerVolterra model with degree of nonlinearity 2 and longterm memory^{34} becomes:
As mentioned above, as is defined in continuous time, it is possible to obtain an instantaneous R–R mean estimate at a very fine timescale, which requires no interpolation between the arrival times of two beats. In this work, we consider nonlinearities associated to eq. 5 up to the thirdorder. Cubic terms, in fact, allow for the estimation of the dominant Lyapunov exponent, whereas quadratic terms account for the highorder spectral estimation (see sections below).
Parameter Estimation, Model Selection, GoodnessofFit
The optimal timevarying parameter vector ξ(t) is defined as the set maximizing the following local loglikelihood^{23}:
where:
• (t − l, t] is a local observation interval of duration l;
• is a subset of Rwave events;
• m_{1} = N(t − l) + 1;
• m_{2} = N(t);
• i_{max} = max{i_{1}, i_{2}, …, i_{n}};
• w(τ) = e^{−ϖτ} is an exponential weighting function.
We use a NewtonRaphson procedure to maximize the local loglikelihood in eq. 6. This formulation is also used to preprocess all the actual heartbeat data with a previously developed algorithm^{35}, performing a realtime RR interval error (e.g., peak detection errors and ectopic beats) detection and correction.
Because there is significant overlap between adjacent local likelihood intervals, we start the NewtonRaphson procedure at t with the previous local maximumlikelihood estimate at time t − Δ, where Δ defines the time interval shift to compute the next parameter update.
We determine the optimal model orders based on the KolmogorovSmirnov (KS) test and associated KS statistics^{23}. Autocorrelation plots are also considered to test the independence of the modeltransformed intervals^{23}.
Instantaneous Time, Frequency, and HigherOrder Spectral Analysis
Estimation of the InputOutput Volterra Kernels
In order to provide quantitative tools related to representations defined in the time, frequency, and higher order spectral domains, considering quadratic nonlinearities (with n = 2), estimated parameters of the fully autoregressive eq. 5 have to be linked to classical inputoutput WienerVolterra model^{23}.
This transformation can be performed in the frequency domain by using the following relationships^{23}
where
• H_{p}(f_{1}, …, f_{n}) is the Fourier transforms of the WienerVolterra kernels of order p
• and are the Fourier transforms of the extended terms of γ_{1}(i) and γ_{2}(i, j)^{23}, respectively
• M is a given integer representing the kernel order
•
• r = 2p − M and σ_{M} is the permutation set of
Since the Volterra kernels induced by the NAR model are nested, the M^{th}order kernel can be deduced recursively^{23}.
Instantaneous Analysis in the Time and Frequency Domain
The timedomain characterization is based on the first and the second order moments of the underlying probability structure^{22}. Namely, given the timevarying parameter set ξ(t), the instantaneous estimates of mean , RR interval standard deviation , mean heart rate , and heart rate standard deviation can be derived at each moment in time as follows^{22}:
The linear power spectrum can be estimated as:
where .
By integrating the (11) in each frequency band, we can compute the index within the very low frequency (VLF = 0.01–0.05 Hz), low frequency (LF = 0.05–0.15 Hz), and high frequency (HF = 0.15–0.5 Hz) ranges^{22}.
Instantaneous Bispectral Analysis
The higherorder spectral representation allows for statistics beyond the second order, and phase relations between frequency components otherwise suppressed.
The analytical solution for the bispectrum of a nonlinear system response subject to stationary, zeromean Gaussian input can be found in^{36}.
Further details on the instantaneous bispectrum derivation from pointprocess nonlinear models can be found in^{23}.
Through this powerful computational tool, we here evaluate the instantaneous presence of nonlinearity in heartbeat series by calculating the nonlinerar sympathovagal interactions. Specifically, by integrating Bis(f_{1}, f_{2}, t) in the appropriate frequency bands, it is possible to obtain:
Instantaneous Measures of Complexity
Instantaneous Lyapunov Exponents
Considering cubic nonlinearities (with n = 3) in the fully autoregressive eq. 5, and using a Fast Orthogonal Search algorithm, it is possible to estimate the complete Lyapunov Exponents (LE) spectrum at each moment in time^{21}. In this study, we use the Instantaneous Dominant Lyapunov Exponent (IDLE, λ_{1}), which is the first exponent of the LE spectrum, along with λ_{2}:
where Δ is the sampling time step, N the length of data samples, and R_{(j)ii} is part of the QR decomposition of the Jacobian of the time series^{21}.
Instantaneous Entropy Measures
In this study, the estimated instantaneous entropy measures refer to the inhomogeneous pointprocess approximate and sample entropy, A_{I} and S_{I}, respectively^{20}. These measures have their foundation in the instantaneous phase space estimation, in which the distance between two points is calculated through KS distance (i.e. the maximum value of the absolute difference between two cumulative distribution functions) between the two pdfs associated to these points. The timevarying radius r(t) is instantaneously expressed as ^{20}.
Moreover, A_{I} and S_{I} dynamical values are not seriously affected by the kind of noise underlying the complex system, thus ensuring truly instantaneous tracking of the dynamic system complexity^{20}.
Experimental Setup
Thirtytwo participants aged 27 ± 2 (16 females) gave their informed consent to take part in the study. No participants reported physical limitations and any experience of mental or personality disorder in their life that would affect the experimental outcomes. All participants gave written informed consent to participating in the study, which was approved by the Ethical Committee of the University of PisaPisa University Hospital, Pisa (Italy). All experimental procedures and analyses were carried out in accordance with such approved guidelines. During the experimental protocol the participants were comfortably seated, the right forearm was horizontal and placed on the forearm support, hand palm down. For all trials, participants wore earplugs in order to prevent any auditory cues.
The adhoc robotic device uses a layer of elastic fabric to convey haptic likecaressing stimuli^{33} (see Fig. 1). The system is endowed with a load cell placed at the basis of the forearm support, prior to the experiment, the load cell was auto calibrated with respect to the forearm weight. When the device was active, two different operating phases were distinguishable: a calibration phase, when forces exerted by the fabric on the user were calibrated, and a movement phase. In this phase, the motors started to coherently rotate and the fabric moves forward and backward over the user forearm, simulating a caress. For further technical details on the device, the reader is invited to refer to ref. 33.
In this study, we considered 4 different stimuli among 2 levels of force (F1 = 2 N, F2 = 6 N) and 2 levels of velocity (V1 = 9.4 mm/s, V2 = 65 mm/s), obtained by feeding the motors with two different sinusoidal input trajectories, at the frequencies of 0.1 Hz and 0.4 Hz. These values were chosen according to previous studies^{4,33}. Exemplary timeline is shown in Fig. 1. Between two stimuli, the motors were stopped and the force was set to 0 N, in this case the fabric was only lightly in contact with the forearm. As the force increases, the fabric is more closely wrapped around the forearm and there is no more pure sliding (as with light forces) but also skin torsion. This behavior was coherent with the goal to reproduce as exhaustively as possible the behavior of the human caress. Throughout the experiment, there were two phases of resting sessions with a duration of two minutes: the first at the beginning of the protocol, the second at the end of the stimulation.
The four kinds of stimulation were suitably randomized, with a prestimulus and a poststimulus interval of 35 seconds each. During the elicitation, the ECG was continuously acquired, following the Einthoven triangle configuration, by means of a dedicate hardware module, i.e. the ECG100C Electrocardiogram Amplifier from BIOPAC inc. with a sampling rate of 500 Hz. To obtain the HRV series from the ECG, a QRS complex detection algorithm was used, i.e. the automatic algorithm developed by PanTompkins^{37}. Peak detection and consequent correction of ectopic beats was performed with an automatic technique previously described in ref. 35.
Statistical Analysis and Pattern Recognition
All features were instantaneously calculated with a Δ = 5 ms temporal resolution. We considered the timevarying dynamics of a given feature X estimated throughout the shorttime window when the affective stimulus occurs. Although the length of such a shorttime window depends on the caressing velocity ranging from 4.3 to 25 seconds, in order to avoid possible biases related to signal length, we considered feature values derived from the last second of each caressing stimulus. In order to reduce possible intra, and intersubject variability, this value was then normalized by the feature value gathered from the fist 10 s of the initial resting state. Quantification of the feature dynamical information was performed through the maximum value , the median and its respective absolute deviation , all calculated along the time of last second of caressing. Accordingly, since we derived 13 instantaneous features, the total number of parameter used for the force/velocity classification was 39. A summary of all considered features is reported in Table 1. The median is considered as a measure of central tendency, which can be related to classical estimation methods that provide one finite value when considering a given time window, whereas the median absolute deviation and the maximum value depend on the feature variability along the time. Then, to average among multiple subjects, we consider group values expressed as median and its respective absolute deviation (i.e., for a feature , where ).
As an exploratory preliminary step, for each feature, we evaluated the statistical differences between groups of caresses, identified by levels of force/velocity. The difference was expressed in terms of pvalues from a nonparametric Wilcoxon test for paired data, under the null hypothesis that the medians of the two sample groups are equal.
In order to actually discern between force and velocity of the caresses using heartbeat dynamics exclusively, an automatic classification algorithm is needed. To this extent, each feature constituted a single dimension of the feature space. A multidimensional point was considered an outlier if zscores associated to its dimensions were greater than 4. The obtained feature set is taken as an input of the LeaveOneOut (LOO) procedure applied on a Support Vector Machine (SVM)based pattern recognition^{38} (see Fig. 2). Within the LOO scheme, the training set was normalized by subtracting the median value and dividing by the MAD over each dimension. More specifically, we used a nuSVM (nu = 0.5) having a radial basis kernel function with γ = n^{−1}, with n = 39 equal to the number of features.
Additionally, in order to explore the relative importance of all features in the classification problem, we employed a support vector machine recursive feature elimination (SVMRFE) procedure in a wrapper approach (RFE was performed on the training set of each fold and we computed the median rank for each feature over all folds). We specifically chose a recently developed nonlinear SVMRFE which employes a radial basis function kernel and includes a correlation bias reduction strategy into the feature elimination procedure^{39}.
Classification results are here expressed in terms of recognition accuracy, and in form of confusion matrix. The generic element r_{ij} of the confusion matrix indicated a percentage as to how many times a pattern belonging to class i was classified as belonging to class j. A more diagonal confusion matrix corresponded to a higher degree of classification. All of the algorithms were implemented by using Matlab© v8.4 endowed with an additional toolbox for pattern recognition, i.e., LIBSVM^{40}.
In the next section, we report experimental results in classifying force and velocity level of the administered caresslike stimuli. Specifically, we show best classification results given the adhoc number of feature identified through the SVMRFE procedure.
Results
Instantaneous series from a representative subject are shown in Figs 3 and 4. The model order selection analysis revealed optimal NARL order as p = 3 and M_{1} = 1, M_{2} = 0. Over all the considered subjects, 31 out of a total of 32 recordings showed KS plots and more than 98% of the autocorrelation samples within the 95% of the confidence interval. Of note, KS distances were as low as 0.0313 ± 0.0062. Figures 5 and 6 show an exemplary KolmogorowSmirnov and autocorrelation plots demonstrating how our model well predicted all heartbeat events of a given RR interval series.
As a first preliminary investigation, we performed a statistical analysis of all the features between force and velocity levels. Significant results (p < 0.05) from such an analysis are shown in Table 2.
Significant differences were found in discerning caresses performed at different velocities only. No relevant changes, in fact, were found in instantaneous heartbeat linear and nonlinear dynamics between caresses performed at different force levels. On the other hand, caresses performed at higher velocity affected the variability of all of the instantaneous measures, as well as median value of four features, one of which was related to heartbeat complex dynamics, and maximum value of 8 features, six of which were related to heartbeat complex/nonlinear dynamics (see Table 2).
Caressing Force Classification
Table 3 shows the confusion matrix as well as the total average accuracy whilst discerning caresseslike stimuli performed at different force levels, considering first 26 features selected by the nonlinear SVMRFE algorithm. This specific feature set, giving the best total average accuracy, which was 69.79%, is partially listed in Table 4 ordered by median rank over every fold computed during the LOO procedure.
Figure 7 shows the recognition accuracy while discerning caressing force levels, as a function of the feature rank estimated through the SVMRFE procedure.
Furthermore, we performed the classification of caressing at low force vs. high force levels using features only related to the central tendency (median) of the timevarying estimates, such as , , , , , and , , , , , , , , avoiding any measure of variability (i.e., MAD and maximum values). Results of this classification are shown in Table 5. In this case, the total average accuracy was 57.66%.
Caressing Velocity Classification
Table 6 shows the confusion matrix as well as the total average accuracy whilst discerning caresses performed at different velocity levels, considering first 35 features selected by the nonlinear SVMRFE algorithm. This specific feature set, giving the best total average accuracy, which was 81.25%, is partially listed in Table 7 ordered by median rank over every fold computed during the LOO procedure.
Figure 8 shows the recognition accuracy, while discerning caressing velocity levels, as a function of the feature rank estimated through the SVMRFE procedure.
Likewise performed for the caressing force classification, we performed the classification of caressing at low velocity vs. high velocity levels using features only related to the central tendency (median) of the timevarying estimates, avoiding any measure of variability. Results of this classification are shown in Table 8. In this case, the total average accuracy was 57.14%.
Discussion and Conclusion
In conclusion, we proposed a comprehensive signal processing framework to assess shortterm affective haptic stimuli through the analysis of heartbeat dynamics exclusively. We built on a previously proposed set of equations defining inhomogeneous pointprocess nonlinear models^{23} to obtain instantaneous linear, nonlinear, and complex estimates of cardiovascular dynamics. Key concepts relay on the definition of a continuous pdf predicting the next heartbeat event, identified through Rwaves from the ECG, parametrized using a nonlinear autoregressive model with Laguerre expansion of the WienerVolterra terms^{23}. Such a parametrization is defined up to the cubic terms, allowing for the instantaneous estimation of Lyapunov exponents^{21} and approximate and sample entropy^{20}, as well as instantaneous nonlinear bispectral measures, and linear estimates defined in the time and frequency domain^{23}.
To our knowledge, this is the only signal processing method proposed so far able to objectively assess shortterm (4.3 to 25 seconds) affective haptic stimuli such as caresses. Standard linear and nonlinear HRV measures^{16}, in fact, would be unable to perform such an assessment as they require relatively longterm recordings to accurately characterize the emotional state of a subject. On the other hand, recognition of longterm caressing (i.e., >30 seconds) would lead to the saturation of the CT fibers activity^{3,5}, strongly biasing the affective part of the emotional perception. Of note, motivation of this study are strictly related to objective assessments performed in experimental psychology and, more in general, in the evaluation of patients with mental disorders.
Our analysis revealed that instantaneous cardiovascular dynamics is significantly affected by the velocity of the affective haptic stimuli (see Table 2). As a matter of fact, no relevant changes were found between caresses performed at different force levels. Of note, 14 out of the 25 features which were sensitive to caressing velocity level are derived in a nonlinear fashion, revealing a decrease of complexity variability and maximum values during the haptic stimulus. This result further justify the use of realtime estimates of heartbeat dynamics through pointprocess modeling, also highlighting the great role of nonlinear dynamics in cardiovascular control, fervently pointed out in the current literature^{16}. In order to develop a processing chain able to actually discern affective perception, we moved from looking for simple statistical differences between stimulus physical characteristics to building an automatic classification algorithm. The automatic classification, embedded in the online processing chain, was performed by means of nuSVM, and following a LOO and nonlinear SVMRFE procedures for crossvalidation and feature selection, respectively. Recognition accuracies were very satisfactory, being of 69.79% along the force dimension, and 81.25% along the velocity dimension. This suggest that, in line with results from the statistical analysis, cardiovascular dynamics is strongly affected by the caressing velocity, rather than the caressing force. In order to further emphasize the role of our processing framework, which allows for the calculation of features’ variability measures through the definition of instantaneous estimates, we performed the caressing force and velocity classification using features only related to the central tendency. In these cases, recognition accuracy were as low as <58%. This result demonstrates that information on central tendency only are not sufficient to achieve an automatic discrimination of affective haptic perception using heartbeat dynamics exclusively. This is also demonstrated by the results of the SVMRFE feature selection procedure, pointing out how complexity variability measures are the most needed for the classification (see Tables 4 and 7). Indeed, as shown in Fig. 3, the timevarying dynamics of the heartbeat features is highly nonstationary, with most significant changes in nonlinear and complex dynamics (i.e., LL, LH, HH, λ_{1}, λ_{2}, A_{I}, S_{I}) during caressing at slower velocity. Importantly, our results are not biased by possible stochastic components underlying physiological dynamics, as our instantaneous complexity measures are not significantly affected by the statistical properties of the physiological noise behind the observed dynamics^{20,21}.
Finally, we remark that in this study caressinglike stimuli were administered through an adhoc robotic device^{29}. Further research is needed to investigate possible differences/similarities with actual caresses by a human hand, possibly investigating more than two levels of caressing force and velocity.
Additional Information
How to cite this article: Valenza, G. et al. Inhomogeneous PointProcesses to Instantaneously Assess Affective Haptic Perception through Heartbeat Dynamics Information. Sci. Rep. 6, 28567; doi: 10.1038/srep28567 (2016).
References
 1.
Liljencrantz, J. & Olausson, H. Tactile c fibers and their contributions to pleasant sensations and to tactile allodynia. Frontiers in behavioral neuroscience 8, 1–6 (2014).
 2.
Macefield, V. G. Tactile c fibers. In Encyclopedia of Neuroscience, Springer, 3995–3998 (2009).
 3.
Klöcker, A., Oddo, C. M., Camboni, D., Penta, M. & Thonnard, J.L. Physical factors influencing pleasant touch during passive fingertip stimulation. Plos One 9, e101361 (2014).
 4.
Löken, L. S., Evert, M. & Wessberg, J. Pleasantness of touch in human glabrous and hairy skin: order effects on affective ratings. Brain research 1417, 9–15 (2011).
 5.
McGlone, F., Vallbo, A. B., Olausson, H., Loken, L. & Wessberg, J. Discriminative touch and emotional touch. Canadian Journal of Experimental Psychology/Revue canadienne de psychologie expérimentale 61, 173–183 (2007).
 6.
Ackerley, R., Carlsson, I., Wester, H., Olausson, H. & Wasling, H. B. Touch perceptions across skin sites: differences between sensitivity, direction discrimination and pleasantness. Frontiers in behavioral neuroscience 8, 1–10 (2014).
 7.
Agrafioti, F., Hatzinakos, D. & Anderson, A. K. ECG pattern analysis for emotion detection. Affective Computing, IEEE Transactions on 3, 102–115 (2012).
 8.
Calvo, R. & D’Mello, S. Affect detection: An interdisciplinary review of models, methods, and their applications. Affective Computing, IEEE Transactions on 1, 18–37 (2010).
 9.
Eyben, F., Wollmer, M. & Schuller, B. OpenEARintroducing the Munich opensource emotion and affect recognition toolkit. InAffective Computing and Intelligent Interaction and Workshops, ACII 3rd International Conference on, 1–6 (2009).
 10.
Haag, A., Goronzy, S., Schaich, P. & Williams, J. Emotion recognition using biosensors: First steps towards an automatic system. InADS, 36–48 (2004).
 11.
Hatayama, T., Kitamura, S., Tamura, C., Nagano, M. & Ohnuki, K. The facial massage reduced anxiety and negative mood status, and increased sympathetic nervous activity. Biomedical research (Tokyo, Japan) 29, 317–320 (2008).
 12.
Kim, K. H., Bang, S. W. & Kim, S. R. Emotion recognition system using shortterm monitoring of physiological signals. Medical and biological engineering and computing. 42, 419–427 (2004).
 13.
Lindgren, L., Rundgren, S., Winsö, O., Lehtipalo, S. & Wiklund, U. et al. Physiological responses to touch massage in healthy volunteers. Autonomic Neuroscience 158, 105–110 (2010).
 14.
Olausson, H., Cole, J., Rylander, K., McGlone, F. & Lamarre, Y. et al. Functional role of unmyelinated tactile afferents in human hairy skin: sympathetic response and perceptual localization. Experimental brain research 184, 135–140 (2008).
 15.
Wiens, S., Mezzacappa, E. S. & Katkin, E. S. Heartbeat detection and the experience of emotions. Cognition & Emotion 14, 417–427 (2000).
 16.
Acharya, U. R., Paul Joseph, K., Kannathal, N., Lim, C. & Suri, J. Heart rate variability: a review. Medical and Biological Engineering and Computing 44, 1031–1051 (2006).
 17.
Delaney, J., Leong, K. S., Watkins, A. & Brodie, D. The shortterm effects of myofascial trigger point massage therapy on cardiac autonomic tone in healthy subjects. Journal of advanced nursing 37, 364–371 (2002).
 18.
Lindgren, L., Lehtipalo, S., Winsö, O., Karlsson, M. & Wiklund, U. et al. Touch massage: a pilot study of a complex intervention. Nursing in critical care 18 269–277 (2013).
 19.
Sunagawa, K., Kawada, T. & Nakahara, T. Dynamic nonlinear vagosympathetic interaction in regulating heart rate. Heart and Vessels 13, 157–174 (1998).
 20.
Valenza, G., Citi, L., Scilingo, E. P. & Barbieri, R. Inhomogeneous pointprocess entropy: An instantaneous measure of complexity in discrete systems. Physical Review E 89, 052803 (2014).
 21.
Valenza, G., Citi, L. & Barbieri, R. Estimation of instantaneous complex dynamics through Lyapunov exponents: a study on heartbeat dynamics. PloS one 9, e105622 (2014).
 22.
Barbieri, R., Matten, E., Alabi, A. & Brown, E. A pointprocess model of human heartbeat intervals: new definitions of heart rate and heart rate variability. American Journal of PhysiologyHeart and Circulatory Physiology 288, H424 (2005).
 23.
Valenza, G., Citi, L., Scilingo, E. P. & Barbieri, R. Pointprocess nonlinear models with Laguerre and Volterra expansions: Instantaneous assessment of heartbeat dynamics. Signal Processing, IEEE Transactions On 61, 2914–2926 (2013).
 24.
Valenza, G., Greco, A., Gentili, C., Lanata, A., Sebastiani, L., Menicucci, D., Gemignani, A. & Scilingo, E. P. Combining electroencephalographic activity and instantaneous heart rate for assessing brain–heart dynamics during visual emotional elicitation in healthy subjects. Phil. Trans. R. Soc. A 374(2067), 20150176 (2016).
 25.
Valenza, G., Orsolini, S., Diciotti, S., Citi, L., Scilingo, E. P., Guerrisi, M., Danti, S., Lucetti, C., Tessa, C., Barbieri, R. & Toschi, N. Assessment of spontaneous cardiovascular oscillations in Parkinson’s disease. Biomedical Signal Processing and Control 26, 80–89 (2016).
 26.
Valenza, G., Citi, L., Gentili, C., Lanatá, A., Scilingo, E. P. & Barbieri, R. Pointprocess nonlinear autonomic assessment of depressive states in bipolar patients. Methods of information in medicine, 53(4), 296–302 (2013).
 27.
Valenza, G., Citi, L., Gentili, C., Lanata, A., Scilingo, E. P. & Barbieri, R. Characterization of depressive states in bipolar patients using wearable textile technology and instantaneous heart rate variability assessment. Biomedical and Health Informatics, IEEE Journal of 19(1), 263–274 (2015).
 28.
Chang, H. A., Chang, C.C., Kuo, T. B. & Huang, S.Y. Distinguishing bipolar ii depression from unipolar major depressive disorder: Differences in heart rate variability. The World Journal of Biological Psychiatry 0, 1–10 (2015).
 29.
Costa, M. D. & Goldberger, A. L. Generalized multiscale entropy analysis: Application to quantifying the complex volatility of human heartbeat time series. Entropy 17, 1197–1203 (2015).
 30.
Glass, L. Dynamical disease: Challenges for nonlinear dynamics and medicine. Chaos: An Interdisciplinary Journal of Nonlinear Science 25, 097603 (2015).
 31.
Lo, M. T., Chang, Y.C., Lin, C., Young, H.W. V. & Lin, Y.H. et al. Outlierresilient complexity analysis of heartbeat dynamics. Scientific reports 5, 1–7 (2015).
 32.
Valenza, G., Nardelli, M., Bertschy, G., Lanata, A. & Scilingo, E. Mood states modulate complexity in heartbeat dynamics: A multiscale entropy analysis. EPL (Europhysics Letters) 107, 18003 (2014).
 33.
Bianchi, M., Valenza, G., Serio, A., Lanata, A. & Greco, A. et al. Design and preliminary affective characterization of a novel fabricbased tactile display. In Haptics Symposium (HAPTICS), IEEE, 591–596 (2014).
 34.
Marmarelis, V. Identification of nonlinear biological system using Laguerre expansions of kernels. Ann. Biomed. Eng. 21, 573–589 (1993).
 35.
Citi, L., Brown, E. N. & Barbieri, R. A realtime automated pointprocess method for the detection and correction of erroneous and ectopic heartbeats. Biomedical Engineering, IEEE Transactions on 59, 2828–2837 (2012).
 36.
Nichols, J., Olson, C., Michalowicz, J. & Bucholtz, F. The bispectrum and bicoherence for quadratically nonlinear systems subject to nonGaussian inputs. Signal Processing, IEEE Transactions on 57, 3879–3890 (2009).
 37.
Pan, J. & Tompkins, W. J. A realtime qrs detection algorithm. Biomedical Engineering, IEEE Transactions on 32, 230–236 (1985).
 38.
Schölkopf, B., Smola, A. J., Williamson, R. C. & Bartlett, P. L. New support vector algorithms. Neural computation 12, 1207–1245 (2000).
 39.
Yan, K. & Zhang, D. Feature selection and analysis on correlated gas sensor data with recursive feature elimination. Sensors and Actuators B: Chemical 212, 353–363 (2015).
 40.
Chang, C. C. & Lin, C.J. LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology 2, 1–27 (2011).
Acknowledgements
The research leading to these results has received partial funding from the Department of Anesthesia, Critical Care & Pain Medicine, Massachusetts General Hospital, and Harvard Medical School, Boston, MA, USA, and European Union Seventh Framework Programme FP7/2007–2013 under Grant No. 601165 of the project “WEARHAP”, and, ERC Advanced Grant No. 291166 “Soft Hands: A Theory of Soft Synergies for a New Generation of Artificial Hands”. European H2020 project “SoftPro: Synergybased Opensource Foundations and Technologies for Prosthetics and RehabilitatiOn” (No. 688857). European Commission—Horizon 2020 Program under Grant 689691 “NEVERMIND” (NEurobehavioural predictiVE and peRsonalised Modelling of depressIve symptoms duriNg primary somatic Diseases with ICTenabled selfmanagement procedures).
Author information
Affiliations
Neuroscience Statistics Research Laboratory, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA
 G. Valenza
 & R. Barbieri
Department of Information Engineering and Research Centre E Piaggio, University of Pisa, Pisa, Italy
 G. Valenza
 , A. Greco
 , M. Bianchi
 & E. P. Scilingo
School of Computer Science and Electronic Engineering, University of Essex, Colchester, CO43SQ, UK
 L. Citi
Department of Advanced Robotics, Istituto Italiano di Tecnologia, via Morego 30, 16163 Genova, Italy
 M. Bianchi
Authors
Search for G. Valenza in:
Search for A. Greco in:
Search for L. Citi in:
Search for M. Bianchi in:
Search for R. Barbieri in:
Search for E. P. Scilingo in:
Contributions
G.V., A.G., M.B. and E.P.S. conceived the experiment; M.B. realized the haptic device; G.V., A.G., L.C. and R.B. analyzed data; All authors discussed the experimental results; All authors wrote the manuscript.
Competing interests
The authors declare no competing financial interests.
Corresponding author
Correspondence to G. Valenza.
Rights and permissions
This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/
About this article
Further reading

Genetic algorithm for the optimization of features and neural networks in ECG signals classification
Scientific Reports (2017)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.