Age classification based on subject's physiological responses

Abstract

In this work we rely on physiological signals as honest indicator of people's instinctive behavior or emotions. Benefiting from the fact that these signals can be easily acquired from wearable devices, we here analyze the ability of these data to classify not only diferent human activities but also individuals' age. We consider Photoplethysmography (PPG) and Galvanic Skin Response (GSR) belonging to a dataset collected in a real laboratory environment at the University of Tokyo. In the experiment a population of Japanese young adults and a population of Japanese elderly people were involved and performed four diferent tasks: Reading, Comprehension, Audio Listening and Math Calculation. Four binary classifiers have been considered, one for each of the experiment tasks, to classify the population age. Diferent classification models have been tested (SVM, CART and XgBoost) with a LOSO validation strategy, obtaining classification accuracy between 69% and 78%. The task in which the two groups are most easily distinguishable is that of mathematical calculation. Finally, we also perform a multi-class classification considering age and tasks, for a total of six classes: Math Calculation, Reading and Audio Listening for each subject group (young adult and elderly), obtaining an overall good performance.

1. Introduction

In daily life, the concept of the Internet of Things plays an increasingly important role [ 1 ]. The advance in communication and computing technologies, as well as the reduction in sensor and electronic component cost, led to the creation of interconnected systems where also everyday objects, such as mobile phones, actuators, appliances or smart devices, are capable of sending and receiving data over the network [ 2 ]. In this Smart Environment, the connected devices must be able to acquire knowledge, understand how to apply it and working collaboratively to make human life more comfortable [ 3 ]. In future scenarios, urban and home environments are destined to become increasingly linked to technological aspects, for example with the introduction of self-driving vehicles. In this context, thus, the analysis of the behaviors and emotions of people during their daily activities, interacting with the environment, may bring to the definition of systems able to receive feedback from users and consequently adapt. In particular, these systems may help to define environments more friendly to vulnerable citizens, such as the elderly and people with disabilities [ 4 ]. In urban environments, for example, technologies capable of receiving feedback from users could be involved in the definition of self-driving cars able to adapt their dynamic behaviours to the safety perception of diferent categories of pedestrians [ 5 ] or in the realization of trafic lights able to adapt their waiting time according to the presence or absence of people with impaired mobility.

To implement such systems, thus, an essential aspect consists in profiling and recognizing the categories of individuals with which these systems could interact. In this respect, the aging of the population is a relevant factor that should be taken into account. It has been observed that subjects of diferent ages react diferently to particular stimuli both from emotional [ 6 ] and behavioral point of view [ 7 ] [ 8 ]. For instance, the elderly appear less reactive than young adults in response to audio and visual stimuli [ 9 ], as well as they appear slower in carrying out cognitive tasks as mouse pointing [10] or in driving ability [11]. In addition, also concerning walkability, the two populations appear diferent. In particular, in [ 7 ] it is reported how elderly tend to behave more cautiously when they have to face an obstacle, passing it only when they feel save.

For this reason, in a world where the average age of the population is destined to increase over the years [12, 13], the definition of systems capable of automatically recognizing the age of the users and behave consequently is becoming a primary issue. In this context, a fruitful research area concerns the use of person’s physiological responses such as heartbeat (Photopletismography or Electrocardiography), sweat glands activity (Electrodermal Activity) or electrical activity on the scalp (Electroencephalography). These signals are not voluntary and uncontrolled responses of our Autonomous Nervous System (ANS) and can thus be considered as reliable and honest indicators of the subject’s instinctive behavior or emotions. Besides, several studies underlined how these signals change with the increase of age. For example, in [14] it is shown how the shape of the Photoplethysmography (PPG) is afected by the subject’s age. In particular, the PPG signals of elderly appear as more rounded and characterized by the disappearance of the dicrotic notch and the inflection point. Even concerning Electrodermal Activity (EDA) and Electroencephalography (EEG), age-related diferences are reported. For example, from the analysis of EDA signal it emerges how, in general, some signal characteristics appear afected by the subject age with a lesser skin response in elderly with respect of young adults and middle-aged adults [15].

Finally, people’s physiological signals are also involved in analysis related to compare young and elderly during specific tasks. In this context, the diferent behavior of the young and elderly during driving tasks has been studied through EEG signals [ 8 ]. This research shows how high drive performance is associated, in the elderly, with increased mental efort and fatigue while, in the young, with a higher focus on the task and lesser attention to distractors. In [16], instead, an emotion classification task involving two populations of diferent ages is reported. In this analysis, two physiological signals are taken into account: Galvanic Skin Response (GSR) and Heart-Beat Variability (HRV). In most of the studies presented, however, the physiological signals have been used to analyze the general individual’s behaviour in a specific task (as driving) or to distinguish the emotion of diferent ageing people. Besides, rarely the physiological signals have been used in classification tasks to recognize people of diferent ages. Our study is developed in this context to automatically discriminate between young adults and the elderly while performing diferent tasks. To this end, starting from the dataset described in section 2, two types of analysis have been performed. In the first part of our study, we tested several binary classifiers in order to recognize the two populations while performing a specific cognitive activity. In the second part, instead, a multi-classification task has been carried out to define a classifier able to recognize both the age of the subject involved and the activity performed. In this work, diferent classifiers and feature sets have been tested. In particular, we focus our attention on the analysis of PPG and GSR signals. The preprocessing of these signals is reported in section 3 while the features extracted are described in section 4. In section 5, the classification settings and the results of the two analyses are presented and discussed. Finally, conclusions are drawn in section 6.

2. Experimental Protocol and Data Collection

All the analysis were performed on a dataset collected in a real laboratory environment at the University of Tokyo and already partially described in [17]. In the experiment, two diferent groups of subjects were involved: a population of young adults, composed of 16 Japanese master and PhD students, (average age = 24.7 years, standard deviation = 3.3, 4 women), and a population of Japanese elderly people (retired), 20 subjects, (average age of 65.15, standard deviation = 2.7, 10 women). All the subjects were healthy and no mental or heart diseases were reported.

All the participants performed the same tasks defined by the same experimental protocol, lasting about 30 minutes, and described below: • 3 minutes of Subject’s emotional profiling carried out filling the STAI questionnaires [18]. • 6 minutes of Reading and Comprehension tasks. Two diferent texts were proposed: a Fairy-tale (“The Wolf and the Seven Little Kids”) and a philosophy text (“Kant’s Critique of Pure Reason”). The subjects had 2 minutes to read each text and 1 minute to answer self assessment and Reading Comprehension questions. • 15 minutes of Audio Listening and Math Calculation tasks composed of six repetitions of a two steps sequence consisting of: 1. 2 minutes of audio listening in which the relaxation were induced by natural and real life sounds (Figure 1 right). 2. 30 seconds of mental arithmetic calculations like sums, subtractions and multiplications (Figure 1 left).

The audio tracks and the calculations proposed were the same for each subject but changed according to the iteration.

Between each couple of tasks, a period of resting time (Baseline acquisition) of about 1 minute was acquired.

During the whole experiment, the PPG and the GSR data of each participant were collected using the Shimmer3 GSR+ Unit [19], with a sampling frequency of 128 Hz. An example of the adopted sensors are showed in Figure 2.

3. Signal pre-processing

With the aim of remove acquisition artifacts and noise, both PPG and GSR signals have been pre-processed using a wavelet multiresolution denoising method similar to the one described in [20]. In particular, the PPG raw signal of each subject has been divided into frequency sub-bands using a Stationary Wavelet Transform (SWT) [21] with mother wavelet Fejer-Korovkin22 [22] and four levels of decomposition. A Soft Thresholding has been then applied to the detail coeficients of each sub-band. The threshold adopted for this purpose was the Universal Threshold calculated by the formula = √2( )), where is the length of the jth wavelet coeficient [23].

Likewise, in literature, the use of wavelet-based denoising strategies, in particular those based on SWT, proved to be very eficient in removing noise from the GSR signal [ 24]. Therefore a multiresolution denoising strategy based on SWT has been also used to pre-process the GSR signals. In this second case, however, a Coiflet3 mother wavelet with 7 levels of decomposition has been employed for the decomposition. Besides, the threshold used in the Soft Thresholding was fixed and determined trying to yield minimum of the maximum mean square error over a given set of functions (Minimax thresholding method [25]).

Since in our study the SWT is implemented with the algorithm a-trous [26], a preliminary operation of replicate padding has been applied to both the analyzed signals in order to obtain a length divisible by 2 [21]. In our study the value of “level” is diferent according to the signal considered: 4 in the case of PPG and 7 for GSR. In order to reduce both inter and intra subjects variability, the denoising task has been followed by a normalization phase. In case of PPG, a two-step normalization has been applied. Firstly, the amplitude of each signal has been normalized applying z-score operation. Then a subject’s heartbeat normalization method is applied, considering the heart beat rate of the baseline. In the case of GSR signals, instead, only an amplitude normalization has been performed. In particular, a z-score normalization has been applied to the whole signal before splitting it into diferent experimental trails.

Once segmented, the number of instances for each task appeared unbalanced as shown in Table 1. In particular, there are less instances in the case of reading and comprehension tasks. Therefore, a data augmentation strategy has been applied in order to create more balanced groups. The signals related to reading task were divided in non-overlapping segments of 40 seconds while the Reading Comprehension signals were segmented into two parts of equal length. The last two columns of the Table 1 show the new cardinality after the data augmentation procedure.

4. Features Extracted

After the pre-processing, each of the resulting segments was analyzed in order to identify characteristics that can be significant in discriminating young subjects from elderly ones. For this purpose, seven time-domain features have been extracted from the PPG signal: • Four statistical features (Minimum, Maximum, Mean and Standard Deviation of the signal); • Peak Rate, representing the mean number of peaks per second; • Inter Beat Interval (IBI), representing the mean distance between two peaks in a row; • Root Mean Square of Successive Distance (RMSSD) representing the variance of the distance between two peaks [27].

In the GSR, two types of signal components are usually taken into account during the feature extraction procedure: the Phasic component, related to rapid changes in skin conduction (Skin Conductance Responses) due to external stimuli or spontaneous responses, and the Tonic component related, instead, to the slow change in the signal and representative of the general arousal or stress level. In our analysis, features from both phasic and tonic components were considered. To this aim, the GSR signals were first decomposed into these two components applying the Cvx algorithm [28]. Diferent time domain features have been extracted, according to the signal component, as listed below: • Signal not decomposed: four statistical features (Maximum, Minimum, Mean and Standard Deviation of the signal) • Phasic Component: seven statistical and peak related features: – Maximum and Minimum – Peak Rate, representing the mean number of peaks per second – Peak Area and Peak Area per Second, representing respectively the mean area under the peaks and the mean area under peaks evaluated per second. – Peak Height representing the mean height of the peak detected on the phasic component. – Rise Time (or also Onset-to-Peak Time) defined as the mean number of samples from the onset of the skin conductance response to the top of the peak.[29, 30] • Tonic Component: the Regression Coeficient has been considered as representative of the signal slope.

Finally, the features have been normalized by z-scoring before using them as input to classiifers.

5. Results and discussion

5.1. Classification Setting

In order to determine if it is possible to recognize signals acquired from young adults with respect to signals acquired from elderly, four binary classification tasks have been performed, one for each activity in the dataset (reading, comprehension, math calculations and audio listening). For each task, the signals collected on young adults define the instances of the first class while the signals collected on elderly subjects characterize the elements of the second class. In addition, in all the experiments performed, three well-known classifiers have been involved: Classification and Regression Tree (Cart) [31], Support Vector Machine (SVM)[32] and gradient boosted decision trees algorithm implemented as XgBoost[33]. In the case of SVM, three diferent kernel have been tested: linear ( SVM-Linear ), gaussian (SVM-Gauss) and polynomial cubic (SVM-Cubic) kernel. It is important to underline that all the selected classifiers perform in general well even in the case of moderately unbalanced classes[34, 35]. This makes their use suitable in our datasets.

Finally, a Leave-One-Subject-Out procedure has been applied to evaluate the performance of the trained classifiers. In this method, during each iteration, all the signals of one subject were used as test set while the signals of the remaining subjects were used to train the model. Several evaluation metrics including accuracy, F1-score and the weighted F1-score [36] have been computed to evaluate the performance of the diferent classification tasks. In particular, the weighted F1-score (W-F1) is computed as the weighted mean of per-class F1 scores on the base of the following formula:

W-F1 = ∑ =1 ∗ 1 (1) where m is the number of classes considered (here 2), is the number of elements in class “c”, is the total number of elements analyzed and 1 is the F1-score for the ℎ -class.

5.2. Classification Results

Four diferent binary classifications, one for each activity, have been performed to recognizing young adults’ signals from elderly ones. The classification performance obtained are summarized in Tables 2a (Reading), 2b (Comprehension), 2c (Math Calculation) and 2d (Audio Listening). The performance metrics (accuracy, F1-score and weighted F1-score) are reported, considering ifve classification models and varying the set of features used (PPG, GSR or joining PPG and GSR). To reduce bias in the results, a Leave One Subject Out (LOSO) strategy has been adopted for all the classifiers.

The best performance for each task is obtained using both GSR and PPG features. In particular, the best of all results has been observed in the Math Calculation task, where an accuracy of 78% has been reached using a SVM with a linear kernel. In this case, the two classes (young and elderly) are well discriminable with F1-score values greater than 70%. On the other hand, the two populations are less distinguishable in the case of audio listening. In this case, XgBoost with features concatenated from both signals allowed to reach an accuracy of 69%. Considering the classification performance that can be reached using only one of the two physiological signals, the PPG seems in general to be more useful in discriminating the two populations. In fact, in almost all the studies carried out, features related to the subjects heartbeat outperformed the results generated using features related to skin conductance. Moreover, we recall that all the PPG signals were normalized not only with respect to the amplitude but also with respect to subject’s heartbeat during baseline. This procedure reduces the inter-subject heterogeneity making the signals subject-independent. Finally, another general consideration regards the classifiers that allowed to reach the best performances. In all the conducted experiments, the highest accuracy has been achieved using XgBoost or SVM with Linear Kernel while the worst performances have been obtained using CART with accuracies around 55%. The results described so far have shown, in general, positive performance in recognizing young adults from elderly when a given task is analyzed. To discriminate not only the population group with respect to age but also the task performed, a multi-class classification analysis has been carried out. In this latter, six classes have been considered: three activities (Math Calculation, Reading and Audio Listening) for the two population groups. The Comprehension task has been excluded by the analysis due to its limited number of instances compared to the others. To train the diferent classifiers, the features extracted from both the types of physiological signals have been employed as suggested by the analyses of the binary classifiers.

In Table 3 the results of this multi-class analysis are summarized. As in binary classification, the best performance has been reached using the SVM with Linear kernel. This model allowed to reach an accuracy of 62%, outperforming the accuracy of the other classifiers. Furthermore, in order to better understand the misclassification errors, an in-depth analysis of the confusion matrix has been performed. From this matrix, shown in Table 4, it emerges that the classes that are better recognized are those related to Math Calculation, while on the opposite the lower performances are obtained in the Audio listening tasks. In case of misclassification, the algorithm tends to well classify the task performed but to misunderstand the population group. Performance of the binary classifiers in discriminating Young adults (Yng) from Elderly (Eld) in the diferent tasks analyzed, varying the feature set and adopting a LOSO validation strategy. Three performance metrics are reported: accuracy, F1-score (F1) and Weighted F1-score (W-F1). In each table, the accuracies in bold represent the best performances achieved for each feature set considered, while in red is highlighted the best accuracy at all.

(a) Binary classifiers performance for the

Reading Task PPG Features

GSR Features

PPG and GSR Features (b) Binary classifiers performance for the

Comprehension Task PPG Features

GSR Features

PPG and GSR Features (c) Binary classifiers performance for the

Math Calculation Task PPG Features

GSR Features

PPG and GSR Features

6. Conclusion

In their daily life, people are subjected to diferent stimuli that could afect their behavior and emotions. In particular, the age of a person seems a relevant factor in the definition of how an individual responds to specific stimuli. In this paper diferent binary and multi-class classification tasks have proved that physiological signals permit to well discriminate between young adults and elderly, while performing diferent actions. PPG seems in general to be more useful in all the classification tasks, however the best results are achieved considering both PPG and GSR. These results, together with the increasing availability and reliability of wearable devices, are promising in the perspective of the definition of systems that, interacting with subjects, can recognize their emotions and behaviors as well as their age group, and consequently adapt. Concerning this topic, several factors like diferent cultural aspects or daily habits could be taken into account in future analysis to create systems able to interact with the largest possible number of heterogeneous users. Furthermore, the age of the individuals could be also used as an additional input variable, together with other parameters like the subject’s health, lifestyle or nutritional habits, in the definition of accurate measures of “physiological age” that could be used by industrial designers and product developers to guide their work in the development of appropriate technology able to provide eficient and personalized assistance to individuals of diferent ages and needed.

Acknowledgments

This research is partially supported by the FONDAZIONE CARIPLO “LONGEVICITY-Social Inclusion for the Elderly through Walkability” (Ref. 2017-0938) and by the Japan Society for the Promotion of Science (Ref. L19513). We want to give our thanks to Prof. Katsuhiro Nishinari and his staf, in particular Kenichiro Shimura and Daichi Yanagisawa for their indispensable support during the experiment held at RCAST - The University of Tokyo. reaction time in young and elderly males, International Journal of Health Sciences and Research 10 (2020) 333–337. [10] M. Hertzum, K. Hornbaek, How age afects pointing with mouse and touchpad: A comparison of young, adult, and elderly users, Intl. Journal of Human–Computer Interaction 26 (2010) 703–734. [11] H.-k. Shin, H.-C. Lee, Characteristics of driving reaction time of elderly drivers in the brake pedal task, Journal of Physical Therapy Science 24 (2012) 567–570. [12] U. C. Bureau, 2017 national population projections datasets, 2017. [13] M.-A. Wren, C. Keegan, B. Walsh, A. Bergin, J. Eighan, A. Brick, S. Connolly, W. Dorothy, J. Banks, Projections of demand for healthcare in Ireland, 2015-2030: first report from the Hippocrates model. Esri research series number 67 October 2017, 2017. [14] Q. Yousef, M. Reaz, M. A. M. Ali, The analysis of ppg morphology: investigating the efects of aging on arterial compliance, Measurement Science Review 12 (2012) 266. [15] D. S. Bari, H. Y. Yacoob Aldosky, Ø. G. Martinsen, Simultaneous measurement of electrodermal activity components correlated with age-related diferences, Journal of Biological Physics 46 (2020) 177–188. [16] L. Fernández-Aguilar, A. Martínez-Rodrigo, J. Moncho-Bogani, A. Fernández-Caballero, J. M. Latorre, Emotion detection in aging adults through continuous monitoring of electrodermal activity and heart-rate variability, in: International Work-Conference on the Interplay Between Natural and Artificial Computation, Springer, 2019, pp. 252–261. [17] F. Gasparini, A. Grossi, S. Bandini, A deep learning approach to recognize cognitive load using ppg signals, in: The 14th PErvasive Technologies Related to Assistive Environments Conference, 2021, pp. 489–495. [18] C. D. Spielberger, State-trait anxiety inventory for adults (1983). [19] A. Burns, E. P. Doheny, B. R. Greene, T. Foran, D. Leahy, K. O’Donovan, M. J. McGrath, Shimmer™: an extensible platform for physiological signal capture, in: 2010 annual international conference of the IEEE engineering in medicine and biology, IEEE, 2010, pp. 3759–3762. [20] A. Biswas, M. S. Roy, R. Gupta, Motion artifact reduction from finger photoplethysmogram using discrete wavelet transform, in: Recent Trends in Signal and Image Processing, Springer, 2019, pp. 89–98. [21] G. P. Nason, B. W. Silverman, The stationary wavelet transform and some statistical applications, in: Wavelets and statistics, Springer, 1995, pp. 281–299. [22] W. Li, Wavelets for electrocardiogram: overview and taxonomy, IEEE Access 7 (2018) 25627–25649. [23] D. L. Donoho, J. M. Johnstone, Ideal spatial adaptation by wavelet shrinkage, biometrika 81 (1994) 425–455. [24] W. Chen, N. Jaques, S. Taylor, A. Sano, S. Fedor, R. W. Picard, Wavelet-based motion artifact removal for electrodermal activity, in: 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), IEEE, 2015, pp. 6223–6226. [25] C. M. Stein, Estimation of the mean of a multivariate normal distribution, The annals of

Statistics (1981) 1135–1151. [26] M. Holschneider, R. Kronland-Martinet, J. Morlet, P. Tchamitchian, A real-time algorithm for signal analysis with the help of the wavelet transform, in: Wavelets, Springer, 1990, pp. 286–297. [27] P. K. Stein, M. S. Bosner, R. E. Kleiger, B. M. Conger, Heart rate variability: a measure of cardiac autonomic tone, American heart journal 127 (1994) 1376–1381. [28] A. Greco, G. Valenza, A. Lanata, E. P. Scilingo, L. Citi, cvxeda: A convex optimization approach to electrodermal activity processing, IEEE Transactions on Biomedical Engineering 63 (2015) 797–804. [29] J. J. Braithwaite, D. G. Watson, R. Jones, M. Rowe, A guide for analysing electrodermal activity (eda) & skin conductance responses (scrs) for psychological experiments, Psychophysiology 49 (2013) 1017–1034. [30] W. Boucsein, Principles of electrodermal phenomena, in: Electrodermal activity, Springer, 2012, pp. 1–86. [31] L. Breiman, J. Friedman, R. A. Olshen, C. J. Stone, Classification and regression trees chapman & hall, New York (1984). [32] C. Cortes, V. Vapnik, Support-vector networks, Machine learning 20 (1995) 273–297. [33] T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 2016, pp. 785–794. [34] R. D. Abdu-Aljabar, O. A. Awad, A comparative analysis study of lung cancer detection and relapse prediction using xgboost classifier, in: IOP Conference Series: Materials Science and Engineering, volume 1076, IOP Publishing, 2021, p. 012048. [35] R. Akbani, S. Kwek, N. Japkowicz, Applying support vector machines to imbalanced datasets, in: European conference on machine learning, Springer, 2004, pp. 39–50. [36] N. Y. Hammerla, S. Halloran, T. Plötz, Deep, convolutional, and recurrent models for human activity recognition using wearables, arXiv preprint arXiv:1604.08880 (2016).

References

[1] L. Atzori , A. Iera , G. Morabito, The internet of things: A survey , Computer networks 54 ( 2010 ) 2787 - 2805 .

[2] S. Majumder , E. Aghayi, M. Noferesti , H. Memarzadeh-Tehran, T. Mondal , Z. Pang , M. J. Deen , Smart homes for elderly healthcare-recent advances and research challenges , Sensors 17 ( 2017 ) 2496 .

[3] E. Ahmed , I. Yaqoob , A. Gani , M. Imran , M. Guizani , Internet-of-things-based smart environments: state of the art, taxonomy, and open research challenges , IEEE Wireless Communications 23 ( 2016 ) 10 - 16 .

[4] A. Rasouli , J. K. Tsotsos , Autonomous vehicles that interact with pedestrians: A survey of theory and practice , IEEE transactions on intelligent transportation systems 21 ( 2019 ) 900 - 918 .

[5] S. Bandini , S. Fontana , F. Gasparini , D. Sorrenti , Interaction autonomous vehiclepedestrian: Dynamic vehicle behaviour as a function of subjective safety perception , in: IEEE International Conference on Robot & Human Interactive Communication ROMAN2020 , 2020 .

[6] L. Fernández-Aguilar , J. Ricarte , L. Ros , J. M. Latorre , Emotional diferences in young and older adults: Films as mood induction procedure , Frontiers in psychology 9 ( 2018 ) 1110 .

[7] F. Gasparini , A. Grossi , K. Nishinari , S. Bandini , Age-related walkability assessment: A preliminary study based on the emg , in: International Conference of the Italian Association for Artificial Intelligence , Springer, 2020 , pp. 423 - 438 .

[8] S. Getzmann , S. Arnau , M. Karthaus , J. E. Reiser , E. Wascher, Age-related diferences in pro-active driving behavior revealed by eeg measures, Frontiers in human neuroscience 12 ( 2018 ) 321 .

[9] S. Doiphode Rupali , S. Vinchurkar Aruna , A comparative study of auditory and visual