Abstract
Human body expressions convey emotional shifts and intentions of action and, in some cases, are even more effective than other emotion models. Despite many datasets of body expressions incorporating motion capture available, there is a lack of more widely distributed datasets regarding naturalized body expressions based on the 2D video. In this paper, therefore, we report the multi-view emotional expressions dataset (MEED) using 2D pose estimation. Twenty-two actors presented six emotional (anger, disgust, fear, happiness, sadness, surprise) and neutral body movements from three viewpoints (left, front, right). A total of 4102 videos were captured. The MEED consists of the corresponding pose estimation results (i.e., 397,809 PNG files and 397,809 JSON files). The size of MEED exceeds 150 GB. We believe this dataset will benefit the research in various fields, including affective computing, human-computer interaction, social neuroscience, and psychiatry.
Similar content being viewed by others
Background & Summary
It is widely accepted that emotion is communicated via multiple models involving both verbal and non-verbal aspects, such as tone, eye movement, facial expression, and body language. Recent studies have demonstrated that body movements can effectively reflect changes in affective state1, even among primates2. People pay more attention to body expressions than facial expressions or voices when dealing with affective states such as information in high intensity3, perceptual ambiguity conditions4, or when information from these channels is incongruent5,6. As increasing psychological studies indicated the significant role of body movement in transmitting information and emotional states7,8,9, artificial intelligence for emotion recognition is changing from facial expression system10 or body expression system11,12 to a multi-channel information combination13.
Various domains of studies on body parts movement cover gait analysis14, body posture analysis, and gesture analysis. One focus of body movement is kinematic information of body movement such as velocity, acceleration, trajectory, and postures, which cannot be accurately and effectively represented by static pictures or verbal descriptions. In recent decades, motion capture technology has made it possible to precisely capture and analyze the kinematic data of each joint15,16,17,18. A variety types of stimulus sets have emerged, including point-light displays19,20, video clips21, images22, or virtual agents14,23,24. The study of body movement has gradually shifted from concepted research to data-based quantitative research.
However, kinematic information from 2D video is also essential for studying emotional body movements. It is not customary for individuals to equip themselves with sensors, as is commonly done in laboratory settings. Fortunately, many pose estimation projects, such as AlphaPose25, Pose Tensorflow26,27, OpenPose28, and Deeplabcut29,30, use machine learning to estimate the posture of persons or animals in videos or pictures and obtain various data, such as the coordinates of joints. They have been applied in some studies in the field of social neuroscience31,32,33,34,35. For example, de Gelder and Poyo Solanas proposed the radically distributed model36, which suggests an additional mid-level feature analysis between low-level feature and high-level concept analyses. The mid-level features – kinematic features (e.g., velocity, acceleration, vertical movement) and postural features (e.g., limb angle, limb contraction, symmetry, surface, shoulder ratio) – have a specific mapping with the brain. Poyo Solanas, Vaessen, and de Gelder found that the extra-striate body area and fusiform body area exhibit more sensitivity towards postural features than kinematic features37.
Therefore, we report a larger and standardized dataset with various emotions: the multi-view emotional expressions dataset (MEED). MEED contains 4102 recordings of six emotional (anger, disgust, fear, happiness, sadness, surprise) and neutral body movements from three views (left, front, right). Each recording consists of the frames extracted by OpenPose and the coordinates of pixel space for 25 body joints in each frame. MEED is freely available. We expect to encourage researchers in multiple fields (e.g., affective computing, human-computer interaction, artificial intelligence, social security, and social neuroscience) to fully explore the various features of emotional body movements in daily life. Interdisciplinary research in these fields should also be promoted.
Methods
Preparation phase
Twenty-four college students with acting experience from Dalian University of Technology were recruited with appropriate payment. All participants signed an informed consent, knowing that the recordings they performed would be shared publicly. Two actors dropped out, leaving 22 actors (19–24 years old, mean = 20.6 years) included in the MEED. This study was approved by the Human Research Institutional Review Board of Liaoning Normal University and followed the Declaration of Helsinki (1991).
Thirty-five standardized daily event scenarios (five for each emotion and neutral) with high recognition accuracy (82.9% - 100%, mean = 93.4%) were created to guide the actors in the recording phase. The specific content and validation of these scenarios and performances were introduced in our previous work20,24.
Three Microsoft Kinect 2.0 cameras, with a resolution of 15 fps, were placed respectively at the front, left, and right of a 1 m × 1 m sized stage, 1.05 m high from the floor, 2.5 m from the center of the stage, and were controlled by a laptop computer (Microsoft Surface Pro 4). More details can be found in our previous work38.
Recording phase
Actors, wearing in black tights, performed six seconds according to the randomly presented scenario, and several performances were selectively repeated to guarantee robustness. Actors were asked to face the center camera, standing naturally with arms hanging down. All three cameras started recording simultaneously after the actor indicated he/she was ready. The recording phase took approximately two hours, during which the actors may rest at any time.
Pose estimation
OpenPose (v1.7.0), an advanced, reliable bone-extraction library28, uses a convolutional neural network to estimate skeletal joints and coordinates (x, y) of actors’ physical joint points. This dataset is based on 25 points model (i.e., nose, neck, right shoulder, right elbow, right wrist, left shoulder, left elbow, left wrist, mid hip, right hip, right knee, right ankle, left hip, left knee, left ankle, right eye, left eye, right ear, left ear, left big-toe, left small-toe, left heel, right big-toe, right small-toe, and right heel; see Fig. 1).
Each video has 97 image frames (see Fig. 2), except part of which are slightly fewer. The horizontal and vertical coordinates (x, y) of 25 keypoints in the pixel space of each frame for each video, as well as the confidence level for determining joint position, were available through pose estimation. Results from pose estimation have two forms: images and data files of joints position. All image files were composed of image frames, skeletal joints, and 25 keypoints (see Fig. 1). For individual recordings, the information in image files were digitized to the datafile of each frame.
Data Records
Due to a malfunction in the equipment, there were no frontal view videos recorded of the actor M01. Eventually, 4,162 videos were collected, and the following files were excluded from analysis: one file (left_M04H0V2) was corrupted, two actors (F04 and F13; 54 videos) dropped out, two dance videos (right_F06dance, right_M07dance) were test files, and three videos (front_M03H0V2, front_M06SA0V1, right_M01SA2V1) with severe limb obscuration failed to be estimated by OpenPose (v1.7.0). Therefore, MEED retains 4102 recordings (see Table 1). Among them, 4092 videos contain 97 frames each, while the remaining videos have frames of 96, 77, 95, 87, 75, 98, 98, 98, 68, and 93 respectively for left_F07N3V1, left_F11SA4V1, left_M09SU0V2, front_M09SU0V2, front_M10N4V1, right_F02N4V1, right_F07SA5V1, right_M06h5v1, right_M09N1V2, and right_M09SU0V2. MEED is freely available on Zenodo39.
All remaining recordings were systematically named as “<view> <actor_id> <emotion> <scenario_id> <version>”, where “view” refers to the point of view, “actor_id” refers to the actor ID, and “emotion” includes anger (A), disgust (D), fear (F), happiness (H), neutral (N), sadness (SA), and surprise (SU). “scenario_id” refers to scenario (1~5) performance and free performance (0), and “version” is the number of repetitions.
The main folder of MEED has 21 actor folders for front view, 22 actor folders for left view, and 22 actor folders for right view. Pose estimation results include PNG files of each frame in individual performance and JSON files about the coordinates of 25 keypoints, named by recording name and the frame number of each frame. MEED totally has 397,809 PNG files and 397,809 JSON files. Moreover, to facilitate the subsequent research, MAT files of coordinates for each recording are available in the corresponding recording folder, and all coordinate files for a single view are summarized in MEED. In the main folder, there is one quality .csv file and one quality .mat file to show the technical validation result of MEED (see Technical Validation section).
Technical Validation
Proportion of unrecognized keypoints
The effectiveness of OpenPose in extracting coordinates depends on various factors such as the velocity of the actor’s movement, fps, physical occlusion, etc. A high velocity may cause blurring in some frames and deviations in the position of keypoints.
Additionally, limb occlusion lowers the confidence level for confirming joint positions, and long-term physical occlusion may make subsequent joints unrecognized due to the lack of prior information. Consequently, the coordinates of the unrecognized keypoints in some frames would appear as (0, 0). We consider the proportion of the number of these unrecognized keypoints to the number of all keypoints in all frames of each recording as one of the quality metrics for the 2D pose estimation dataset, called proportion of unrecognized keypoints (PUK), which is defined as
where N(0,0) is the total number of unrecognized keypoints in all frames of each recording, and Nkeypoints and Nframe separately refers to 25 keypoints of body pose estimation and total number of frames of each recording.
The results showed that the PUK was lowest in the frontal view, with mean values ranging from 0.003 to 0.048 under all conditions (see Table 2 and Fig. 3).
Confidence level
OpenPose uses confidence maps to assess the predicted data, which is created by the annotated keypoints28. Every confidence map is a 2D indication of the possibilities that the body part appears at each pixel location. It will generate a possible area of Gaussian distribution, the center of which is the keypoints. The Gaussian center has a maximum confidence of 1. The further away from the center, the lower the confidence is. In other words, each pixel position in the confidence map has a corresponding confidence value. The number of confidence peaks equals the number of people in the picture being predicted. MEED contains only single-person situations, so there is only one peak per confidence map. The ground-truth confidence map generated by the network is to take the maximum confidence value through a non-maximum suppression algorithm.
This confidence peak is expressed in the pose estimation results as the confidence level (CL) attached to each keypoint estimation. Therefore, we regard the mean CL of 25 keypoints within each recording as the second quality metric for this dataset, which is defined as
where Nframe and Nkeypoint refer to the number of frames in each recording and 25, respectively. To compare the pose estimation in this dataset with the normal level of OpenPose28, we analyzed CL in all conditions. Results showed that the CL in the frontal view is the highest. The mean values of CL ranged from 0.748 to 0.840 under all conditions (see Table 3 and Fig. 4). The results of two quality metrics suggest that the pose estimation results are good enough for further analysis.
Performance reliability
To ensure that all 22 actors expressed the instructed emotions equally well, that is, the reliability of these performances, we examined the consistency of the objective movement value across all of them. The objective movement of the recordings of frontal view in MEED was quantified using a customized MATLAB code40,41, and prior research has been demonstrated that this movement positively correlates with the intensity of emotion and the motion that observers can perceive from human body20,42,43. Specifically, if a pixel in two consecutive frames had a luminance change of more than 10 units, it was considered a pixel motion. The objective movement values were depicted by computing the average number of pixel motions in each frame and video, which were then saved in the frontMovement.csv.
We then conducted a reliability analysis of the objective movement value for each emotional and neutral condition across all actors using SPSS 26.0 (https://www.ibm.com/products/spss-statistics). The result showed that the Cronbach alpha coefficient was high under all emotional and neutral conditions (anger = 0.900, disgust = 0.939, fear = 0.919, happiness = 0.875, sadness = 0.929, surprise = 0.927, and neutral = 0.974), suggesting a high reliability of these performances and all actors in MEED express these emotions and scenarios equally.
Usage Notes
MEED is an open-source library that stores the results of 2D pose estimation with six emotions and neutral expression as well as three views. JSON and MAT files can be easily used by data processing software such as MATLAB (https://ww2.mathworks.cn/en/products/matlab.html), R (https://www.r-project.org), and Python (https://www.python.org). For example, the coordination data can be analyzed using representational similarity analysis44 for the association between kinematic features and postural features of body expressions and decision tree classifier45 for the relative importance of these features and body parts46.
Moreover, the unrecognized coordinates must be fixed if users want to involve them in their analyses. We suggest that users perform interpolation correction, such as linear, polynomial interpolation, and spline interpolation, on the coordinates of individual keypoint in the videos on the time scale as data streams. Given that linear interpolation is limited to the case of non-continuous unrecognized keypoints, we suggest fitting curve instead, such as the Curve Fitting Toolbox in MATLAB (https://ww2.mathworks.cn/products/curvefitting.html) or its built-in functions (spline, makima, pchip). We also recommend using Photoshop (https://www.adobe.com/products/photoshop.html) for PNG correction when necessary.
MEED is applicable in multiple fields, such as the affective computing of body expressions and corresponding brain mechanisms37,46 in social neuroscience. Researchers in human-computer interaction, machine learning, sports motion analysis, psychiatry, and social security will also be interested in this dataset. We hope that MEED will be of further assistance to them.
Code availability
The MATLAB code for parsing the JSON file and processing the coordinates can be found at https://doi.org/10.5281/zenodo.8185369.
References
de Gelder, B. Why bodies? Twelve reasons for including bodily expressions in affective neuroscience. Philos. Trans. R. Soc. Lond., B, Biol. Sci. 364, 3475–3484, https://doi.org/10.1098/rstb.2009.0190 (2009).
Taubert, J. et al. A broadly tuned network for affective body language in the macaque brain. Sci. Adv. 8, eadd6865, https://doi.org/10.1126/sciadv.add6865 (2022).
Atias, D. & Aviezer, H. Real-life and posed vocalizations to lottery wins differ fundamentally in their perceived valence. Emotion 22, 1394–1399, https://doi.org/10.1037/emo0000931 (2020).
Chen, Z. & Whitney, D. Tracking the affective state of unseen persons. Proc. Natl. Acad. Sci. USA 116, 7559–7564, https://doi.org/10.1073/pnas.1812250116 (2019).
Van den Stock, J., Righart, R. & de Gelder, B. Body expressions influence recognition of emotions in the face and voice. Emotion 7, 487–494, https://doi.org/10.1037/1528-3542.7.3.487 (2007).
Meeren, H. K., van Heijnsbergen, C. C. & de Gelder, B. Rapid perceptual integration of facial expression and emotional body language. Proc. Natl. Acad. Sci. USA 102, 16518–16523, https://doi.org/10.1073/pnas.0507650102 (2005).
Abramson, L., Petranker, R., Marom, I. & Aviezer, H. Social interaction context shapes emotion recognition through body language, not facial expressions. Emotion 21, 557–568, https://doi.org/10.1037/emo0000718 (2021).
de Gelder, B. et al. Standing up for the body. Recent progress in uncovering the networks involved in the perception of bodies and bodily expressions. Neurosci. Biobehav. Rev. 34, 513–527, https://doi.org/10.1016/j.neubiorev.2009.10.008 (2010).
Sowden, S., Schuster, B. A., Keating, C. T., Fraser, D. S. & Cook, J. L. The role of movement kinematics in facial emotion expression production and recognition. Emotion 21, 1041–1061, https://doi.org/10.1037/emo0000835 (2021).
Prince, E. B., Martin, K. B. & Messinger, D. S. Facial action coding system. in Environmental Psychology & Nonverbal Behavior. (2015).
Huis In ‘t Veld, E. M., van Boxtel, G. J. & de Gelder, B. The body action coding system II: Muscle activations during the perception and expression of emotion. Front. Behav. Neurosci. 8, 330, https://doi.org/10.3389/fnbeh.2014.00330 (2014).
Huis In ‘t Veld, E. M., Van Boxtel, G. J. & de Gelder, B. The body action coding system I: Muscle activations during the perception and expression of emotion. Soc.Neurosci. 9, 249–264, https://doi.org/10.1080/17470919.2014.890668 (2014).
Gunes, H. & Pantic, M. Automatic, dimensional and continuous emotion recognition. Int. J. Synth. Emot. 1, 68–99, https://doi.org/10.4018/jse.2010101605 (2010).
Randhavane, T. et al. in Proceedings of the 15th ACM SIGGRAPH Conference on Motion, Interaction and Games Article 5 (Association for Computing Machinery, Guanajuato, Mexico, 2022).
Rong, Y., Shiratori, T. & Joo, H. Frankmocap: A monocular 3D whole-body pose estimation system via regression and integration. in 2021 IEEE/CVF International Conference on Computer Vision Workshops(ICCVW). 1749–1759 (2021).
Qiu, S. et al. Sensor network oriented human motion capture via wearable intelligent system. Int. J. Intell. Syst. 37, 1646–1673, https://doi.org/10.1002/int.22689 (2022).
Mahmood, N., Ghorbani, N., Troje, N. F., Pons-Moll, G. & Black, M. AMASS: Archive of motion capture as surface shapes. in 2019 IEEE/CVF International Conference on Computer Vision (ICCV). 5441–5450 (2019).
Liu, S., Zhang, J., Zhang, Y. & Zhu, R. A wearable motion capture device able to detect dynamic motion of human limbs. Nat. Commun. 11, 5615, https://doi.org/10.1038/s41467-020-19424-2 (2020).
Atkinson, A. P., Dittrich, W. H., Gemmell, A. J. & Young, A. W. Emotion perception from dynamic and static body expressions in point-light and full-light displays. Perception 33, 717–746, https://doi.org/10.1068/p5096 (2004).
Zhang, M. et al. Construction and validation of the Dalian emotional movement open-source set (DEMOS). Behav. Res. Methods 55, 2353–2366, https://doi.org/10.3758/s13428-022-01887-4 (2023).
Kret, M. E., Denollet, J., Grèzes, J. & de Gelder, B. The role of negative affectivity and social inhibition in perceiving social threat: An fMRI study. Neuropsychologia 49, 1187–1193, https://doi.org/10.1016/j.neuropsychologia.2011.02.007 (2011).
de Gelder, B. & den Stock, V. J. The bodily expressive action stimulus test (BEAST). Construction and validation of a stimulus basis for measuring perception of whole body expression of emotions. Front. Behav. Neurosci. 2, 181, https://doi.org/10.3389/fpsyg.2011.00181 (2011).
Fourati, N. & Pelachaud, C. Emilya: Emotional body expression in daily actions database. in Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14). 3486–3493 (European Language Resources Association (ELRA), 2014).
Zhang, M. et al. Kinematic dataset of actors expressing emotions. Sci. Data 7, 292, https://doi.org/10.1038/s41597-020-00635-7 (2020).
An, W. et al. Performance evaluation of model-based gait on multi-view very large population database with pose sequences. IEEE Trans. Biom. Behav. Identity Sci. 2, 421–430, https://doi.org/10.1109/TBIOM.2020.3008862 (2020).
Cao, Z., Simon, T., Wei, S. E. & Sheikh, Y. Realtime multi-person 2D pose estimation using part affinity fields. in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 1302–1310 (2017).
Abadi, M. et al. in Proceedings of the 12th USENIX conference on Operating Systems Design and Implementation Vol. abs/1605.08695 265–283 (USENIX Association, Savannah, GA, USA, 2016).
Cao, Z., Hidalgo, G., Simon, T., Wei, S. E. & Sheikh, Y. Openpose: Realtime multi-person 2D pose estimation using part affinity fields. IEEE Trans. Pattern Anal. Mach. Intell. 43, 172–186, https://doi.org/10.1109/TPAMI.2019.2929257 (2021).
Mathis, A. et al. Deeplabcut: Markerless pose estimation of user-defined body parts with deep learning. Nat. Neurosci. 21, 1281–1289, https://doi.org/10.1038/s41593-018-0209-y (2018).
Nath, T. et al. Using deeplabcut for 3D markerless pose estimation across species and behaviors. Nat. Protoc. 14, 2152–2176, https://doi.org/10.1038/s41596-019-0176-0 (2019).
Mroz, S. et al. Comparing the quality of human pose estimation with blazepose or openpose. in 2021 4th International Conference on Bio-Engineering for Smart Technologies (BioSMART). 1–4 (2021).
Cataldi, S. et al. Decreased dorsomedial striatum direct pathway neuronal activity is required for learned motor coordination. eNeuro 9, ENEURO.0169-0122.2022, https://doi.org/10.1523/ENEURO.0169-22.2022 (2022).
Sabo, A., Mehdizadeh, S., Iaboni, A. & Taati, B. Estimating parkinsonism severity in natural gait videos of older adults with dementia. IEEE J Biomed Health Inform 26, 2288–2298, https://doi.org/10.1109/jbhi.2022.3144917 (2022).
Park, K. W. et al. Machine learning–based automatic rating for cardinal symptoms of parkinson disease. Neurology 96, e1761, https://doi.org/10.1212/WNL.0000000000011654 (2021).
Sabo, A., Mehdizadeh, S., Iaboni, A. & Taati, B. Prediction of parkinsonian gait in older adults with dementia using joint trajectories and gait features from 2D video. in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). 5700–5703 (IEEE, 2021).
de Gelder, B. & Poyo Solanas, M. A computational neuroethology perspective on body and expression perception. Trends Cogn. Sci. 25, 744–756, https://doi.org/10.1016/j.tics.2021.05.010 (2021).
Poyo Solanas, M., Vaessen, M. & de Gelder, B. Computation-based feature representation of body expressions in the human brain. Cereb. Cortex 30, 6376–6390, https://doi.org/10.1093/cercor/bhaa196 (2020).
Liu, S. et al. Multi-view laplacian eigenmaps based on bag-of-neighbors for RGB-D human emotion recognition. Inf.Sci. 509, 243–256, https://doi.org/10.1016/j.ins.2019.08.035 (2020).
Zhang, M. et al. Multi-view emotional expressions dataset. Zenodo https://doi.org/10.5281/zenodo.8185369 (2023).
Cross, E. S. et al. Robotic movement preferentially engages the action observation network. Hum Brain Mapp 33, 2238–2254, https://doi.org/10.1002/hbm.21361 (2012).
Williams, E. H., Bilbao-Broch, L., Downing, P. E. & Cross, E. S. Examining the value of body gestures in social reward contexts. NeuroImage 222, 117276, https://doi.org/10.1016/j.neuroimage.2020.117276 (2020).
Ross, P., de Gelder, B., Crabbe, F. & Grosbras, M. H. Emotion modulation of the body-selective areas in the developing brain. Dev Cogn Neurosci 38, 100660, https://doi.org/10.1016/j.dcn.2019.100660 (2019).
Ross, P., de Gelder, B., Crabbe, F. & Grosbras, M. H. A dynamic body-selective area localizer for use in fMRI. MethodsX 7, 100801, https://doi.org/10.1016/j.mex.2020.100801 (2020).
Kriegeskorte, N., Mur, M. & Bandettini, P. Representational similarity analysis - connecting the branches of systems neuroscience. Front. Syst. Neurosci. 2, 4, https://doi.org/10.3389/neuro.06.004.2008 (2008).
Loh, W.-Y. Classification and regression trees. Wiley Interdiscip Rev. Data Min. Knowl. Discov. 1, 14–23, https://doi.org/10.1002/widm.8 (2011).
Poyo Solanas, M., Vaessen, M. J. & de Gelder, B. The role of computational and subjective features in emotional body expressions. Sci. Rep. 10, 6202, https://doi.org/10.1038/s41598-020-63125-1 (2020).
Acknowledgements
We also thank K. Zhang, B. Du, S. Chen, B. Zhan, S. Guo, X. Jiang, Y. Wang, and B. Wang for contribution to the data collection. This work was supported by the National Natural Science Foundation of China (32020103008), Dalian Outstanding Young Scientific and Technological Talents Project (2022RY20), and Liaoning Normal University High-level Scientific Research Achievements Cultivation Project (23GDL008). We also thank Dr. Emily S. Cross and Dr. Elin H. Williams for sharing the code to compute objective movement value in the present study.
Author information
Authors and Affiliations
Contributions
M.Z., S.L. and W.L. designed this study. M.Z., Y.Z., X.X. and Z.R. collected the data. M.Z., Y.Z., X.X. and Z.R. organized and analyzed the data. M.Z., Y.Z., X.X., Y.Z., S.L. and W.L. wrote the manuscript. All authors approved the final version of the manuscript for submission.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Zhang, M., Zhou, Y., Xu, X. et al. Multi-view emotional expressions dataset using 2D pose estimation. Sci Data 10, 649 (2023). https://doi.org/10.1038/s41597-023-02551-y
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41597-023-02551-y