Datasets for Automated Affect and Emotion Recognition from Cardiovascular Signals Using Artificial Intelligence— A Systematic Review

Simple Summary

We reviewed the literature on the publicly available datasets used to automatically recognise emotion and affect using artificial intelligence (AI) techniques. We were particularly interested in databases with cardiovascular (CV) data. Additionally, we assessed the quality of the included papers. We searched the sources until 31 August 2020. Each step of identification was carried out independently by two reviewers to maintain the credibility of our review. In case of disagreement, we discussed them. Each action was first planned and described in a protocol that we posted on the Open Science Framework (OSF) platform. We selected 18 works focused on providing datasets of CV signals for automated affect and emotion recognition. In total, data for 812 participants aged 17 to 47 were analysed. The most frequently recorded signal was electrocardiography. The authors most often used video stimulation. Noticeably, we did not find much necessary information in many of the works, resulting in mainly low quality among included papers. Researchers in this field should focus more on how they carry out experiments.

Abstract

Our review aimed to assess the current state and quality of publicly available datasets used for automated affect and emotion recognition (AAER) with artificial intelligence (AI), and emphasising cardiovascular (CV) signals. The quality of such datasets is essential to create replicable systems for future work to grow. We investigated nine sources up to 31 August 2020, using a developed search strategy, including studies considering the use of AI in AAER based on CV signals. Two independent reviewers performed the screening of identified records, full-text assessment, data extraction, and credibility. All discrepancies were resolved by discussion. We descriptively synthesised the results and assessed their credibility. The protocol was registered on the Open Science Framework (OSF) platform. Eighteen records out of 195 were selected from 4649 records, focusing on datasets containing CV signals for AAER. Included papers analysed and shared data of 812 participants aged 17 to 47. Electrocardiography was the most explored signal (83.33% of datasets). Authors utilised video stimulation most frequently (52.38% of experiments). Despite these results, much information was not reported by researchers. The quality of the analysed papers was mainly low. Researchers in the field should concentrate more on methodology.

1. Introduction

Facilitating access to databases seems to be an essential matter in the field of machine learning (ML). Publicly available, reliable datasets could drive research forward, making it unnecessary to re-run similar yet complicated experiments in order to obtain sufficient data. Credible work relies on proper arrangement, validation, adjustment, and fairness in artificial intelligence (AI) [1,2].

Moreover, sufficient descriptions of the scientific methods in AI are a constant challenge. It seems to be particularly valid in automated affect and emotion recognition (AAER) studies, which fall under the field of human–computer interaction (HCI), linking psychology, computer science. and biomedical engineering. As human emotions affect multiple channels, research on this topic is being conducted based on speech, facial expressions, gestures or physiological signals, which became exceptionally popular in the last decade [3].

Increasing interest in the field, among others, comes from broad application prospects. Recent studies point out the potential usage of emotion recognition techniques in medical fields, public security, traffic safety, housekeeping, and related service fields [4].

The topic is extensive, as it covers both data acquisition and computation. A typical experiment in AAER involves several steps [5]. Firstly, the researchers need to adopt a specific perspective on the field, as many exist that consider the universality [6,7] of emotions or their structure [8]. The theoretical approach imposes an understanding of emotions, selections of material used for stimulation, and interpretations. However, the general structure of elicitation experiments that are carried out to gather the data from human participants remains stable [9]. To evoke emotions, passive (e.g., video, music, or pictures presentation) or active stimulation (e.g., game playing, interaction with virtual reality, or conversation) is used [5]. Eliciting material may have different lengths, types, and quantities. After the stimulation phase, the subjects are asked how they felt. Several validated instruments enable it, e.g., Self-Assessment Manikin (SAM) [10].

During the stimulation phase, subjects are connected to measuring devices. Researchers use dedicated hardware [11,12,13] and experiment with smartphone [14] or wearable [15,16] technologies, especially with CV signals [17,18,19]. Among others, gathered data include physiological signals [20] (e.g., heart, skin, brain, respiratory system, and eye work), facial expressions [21], and speech [22,23]. Typically, several signals are collected in order to improve the accuracy of the AI system used for AAER [24].

Next, the recognition phase begins. It involves data preprocessing, classification or regression, and finally, validation [5]. Due to its flexibility resulting in, e.g., reduced data preprocessing time [5], deep learning (DL) techniques are widely adopted [25,26,27], along with classical approaches in AAER [28,29,30,31].

As the data collection process in experiments within this field is complex and multi-stage, the problems may occur on many levels. It is thus crucial to plan the experiment and report upon it in adequate detail [32].

The replicability crisis in both psychology and computer science also affects studies on AAER [5,33,34]. Poor methodological conduct often makes it impossible for existing research to be replicated or reproduced. Even in renowned and well-established research that dictates the social order, the phenomenon is widely present [35,36].

Datasets collected inadequately might contribute to lowering the credibility of emerging research (influencing model development by introducing undesirable biases) and waste of time and resources. This issue has been widely discussed before and is known as the garbage in, garbage out problem [37,38]. Avoiding bias and proper validation of experiments are crucial to eliminating it [32].

Promisingly, publishing source codes and data is becoming a desirable standard in computer science [39,40,41,42]. Journal initiatives [43,44] on the topic emphasise the importance of computational research reproducibility and promote open research. In turn, preregistration of the research plan, taking into account the hypotheses and defining step-by-step the methodology allows for improving the quality of the research and its reproducibility from a psychological perspective [45,46].

To create a reliable model presenting a high degree of emotion or affect recognition precision, it is relevant to limit external and internal factors potentially confounding the collected measurements [47,48].

The confounding effect of incomplete control may arise from any stage of the study. For instance, subjects with somatic disorders might affect measures of features, mood disorders, or alexithymia, which is estimated to affect 13% of the population [49].

Each stage of an experiment leading to AAER should be repeatable and standardised among subjects. AAER concerns stimuli presentation, assessment of elicited emotions by the subject, collection of physiological parameters, and laboratory environment, including the presence of experimenter and individual factors [5].

While measuring emotional and affective responses in the laboratory environment using objective methods reduces the risk of self-reported bias, the risk of contextual non-integrity remains. This creates the need to document all the contextual environmental aspects that could influence the measurement [50].

Along with the pervasiveness of wearable devices available to register user psychological parameters during daily activities, AAER is reached [51,52]. Wearable devices are proven to measure efficiently CV signals while being offered at low prices [53,54]. However, the challenge remains to design credible ML models able to deal with the broad spectrum of possible emotions and lack of universality in this category among cultures [7].

Studies on ubiquitous computing are growing in number [55,56,57]. Due to the constraints of time and human resources, all these results could not be read. Therefore, creating summaries along with the analysis of evidence is now necessary [58]. Describing the data together with a critical appraisal helps to determine, for example, the actual accuracy of the methods and to highlight those articles whose results are derived from a high-quality methodological process. The selection of studies answering a similar research question may be chaotic, purposeful, or systematic [59]. The latter method reduces the risk of researchers steering conclusions, as it follows restrictive, transparent criteria [32,60,61].

Because of the above and since previous similar studies on AAER were of weak reliability [5], we decided to present a systematic review on the topic, corresponding to approved standards, to limit the risk of bias (RoB). We review public datasets available for AAER with the use of AI, utilising physiological modalities as an input with the focus on CV signals.

This paper is a part of the project on a systematic review of studies focused on AAER from CV signals with AI methods. For more details, see the protocol [62] and our previous conference paper [63].

Research Questions

• What are the datasets used for AAER from CV signals with AI techniques?
• What are the CV signals most often gathered in datasets for AAER?
• What were other signals are collected in analysed papers?
• What are the characteristics of the population in included studies?
• What instruments were used to assess emotion and affect in included papers?
• What confounders were taken into account in analysed papers?
• What devices were used to collect the signals in included studies?
• What stimuli are most often used for preparing datasets for AAER from CV signals?
• What are the characteristics of investigated stimuli?
• What is the credibility of included studies?

2. Methods

2.1. Eligibility Criteria, Protocol

Papers in which more than half of the sample constitutes a specific population (e.g., children or people with illness) were excluded. All experiments needed to be carried out in laboratory settings. We considered any type of publication to be eligible in which CV signals and AI methods were used for AAER. The primary focus of our whole project [63] was the performance of these computer programs (e.g., specificity, sensitivity, accuracy). For this focused systematic review, we imposed additional inclusion criteria, namely public availability of the data.

Due to double referencing, some of the references were overlapping. These were post-conference books and full proceedings. We excluded them as they contained little information about specific chapters. Nevertheless, we did not reject these particular sections. We excluded introductions to Special Issues in a journal or section, letters to editors, reviews, post-conference books, full proceedings (but not qualified papers), and case studies.

The review protocol was published on the Open Science Framework (OSF) [64] and then registered there [62] on 18 March 2021. All additional information about methods can be found in the protocol.

2.2. Search Methods

We searched article databases (MEDLINE, Web of Science, dblp, EMBASE, Scopus, IEEE, Cochrane Library) and preprint databases (medRxiv, arXiv). The complete search was done on 31 August 2020.

To develop the MEDLINE strategy (see protocol on OSF [62]), we combined MeSH (controlled vocabulary) and free-text words related to AAER, CV signals, and AI. Then, these strings were translated for other sources utilised in the search. We adopted no date or language restrictions.

Additionally, we screened full texts of included papers for otherwise not identified studies. We included them in further steps of identification.

2.3. Definitions

We used the following definitions. AAER [65,66] refers to finding patterns with specific signals (e.g., behavioural, physiological) consistent with detected states. AI refers to software able to perform tasks as accurately as intelligent beings (e.g., humans) [67]. DL refers to the architecture of neural networks comprising at least two hidden layers [68]. Performance metrics, which refer to a mathematical evaluation of model predictions with ground truth [69]. CV signals refer to an electrocardiogram (ECG), pulse oximetry (POX), heart rate (HR), intracranial pressure (ICP), pulse pressure variation (PPV), heart rate variability (HRV), photoplethysmogram (PPG), blood volume pressure (BVP), and arterial blood pressure (ABP) [53,70].

2.4. Data Collection

EndNote (Claritive Analytics^®) and Rayyan [71] were utilised for deduplication of identified references. P.J., D.S., M.S., and M.M. used the Rayyan [71] application to screen the remaining references independently. Subsequently, full texts were assessed separately by P.J., D.S., M.S., and M.M. for meeting inclusion criteria.

P.J., D.S., M.S., M.M., W.Ż., and M.W.G. collected all necessary data independently using a pre-specified extraction form. We gathered bibliographic data (e.g., year, journal name) and information about authors, funding, and conflicts of interest. We also focused on population, models, and outcomes—AI methods and additional analyses, e.g., interpretability, as specified in the protocol (see OSF [62]).

Pilot exercises were conducted before each phase, namely screening of abstracts and titles, full text evaluation, and extraction of the data. By doing so, we aimed at improving the sense of understanding among the reviewers. When discrepancies occurred (at each step of data identification), they were resolved via discussion.

2.5. Quality Assessment

The methodological credibility of included studies was assessed using a tool developed by our team (see Appendix C). The method was based on well-grounded techniques, namely Quality Assessment of Diagnostic Accuracy Studies (QUADAS) [72], Prediction model Risk Of Bias ASsessment Tool (PROBAST) [73], and an instrument provided by Benton et al. [74] as it was dedicated to the same study design as included by us. The process of evaluation was preceded by pilot exercises. We rated RoB independently in pairs (P.J., D.S., M.S., M.M., W.Ż., and M.W.G.). Discussion resolved all discrepancies.

The utilised tool constituted of eight questions (items):

• Was the sample size pre-specified?
• Were eligibility criteria for the experiment provided?
• Were all inclusions and exclusions of the study participants appropriate?
• Was the measurement of the exposition clearly stated?
• Was the measurement of the outcome clearly stated?
• Did all participants receive a reference standard?
• Did participants receive the same reference standard?
• Were the confounders measured?

Items were assessed using a three-point scale with the following answers: yes/partial yes, no/partial no, and not reported resulting in high, low, or unclear RoB. For more details, see Appendix C.

2.6. Analyses

We concentrate on descriptive synthesis regarding characteristics of populations and collected datasets, i.e., stimuli, signals, devices, emotions, and affect. We also present results regarding credibility.

The quantitative summary with sensitivity, heterogeneity, and subgroup analysis of all papers is not the purpose of this focused review. For more details, please refer to the protocol [62] and other papers from the project [63].

3. Results

From 4649 records, we identified 195 studies that met our eligibility criteria. Then, we selected a sub-sample of 18 papers. Each paper provides one validated, publicly available dataset, including CV signals with labels regarding emotions or affect.

Names of datasets described in included papers are as follows: Database for Emotion Analysis using Physiological signals (DEAP) [75]; Multimodal Analysis of Human NOnverbal Behaviour in real-world settings–Human-Computer Interaction (MAHNOB-HCI) [76]; MEG-based multimodal database for DECoding AFfective physiological responses (DECAF) [77]; a dataset for Affect, personality and Mood research on Individuals and GrOupS (AMIGOS) [78]; a multimodal databASe for impliCit pERsonaliTy and Affect recognitIoN using commercial physiological sensors (ASCERTAIN) [79]; AUgsburg Database of Biosignal 4 (AuDB-4) [80]; Emotion Recognition Smartwatch (ERS) [81]; IT Multimodal Dataset for Emotion Recognition (ITMDER) [82]; Database for Affective Gaming (DAG) [83]; Quality Adaptive Multimodal AFfect recognition system for user-centric multimedia indexing (QAMAF) [84]; Virtual Reality Affective Dataset (VRAD) [85]; NEME [86]; WEarable Stress and Affect Detection (WESAD) [87]; a Multi-modal Physiological Emotion Database for discrete emotion recognition (MPED) [88]; database of multimodal (Face, Body gesture, Voice and Physiological signals) recordings (emoFBVP) [89]; a database for emotion recognition through EEG and ECG signals from wireless low-cost off-the-shelf devices DREAMER [90]; Multi-subject Affective Physiological Database (MAPD) [91]; and Mazeball Dataset (MD) [92].

Supplementary File S1 (OSF [64]) and Appendix A and Appendix B contain the list of all included studies, the subgroup of datasets analysed in this review, and the excluded studies with reasons, respectively. The remaining included studies are considered in other articles from the project [63]. The flow of our study is presented in Figure 1. Our reporting is consistent with Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines with diagnostic test accuracy (DTA) extension [93].

3.1. Included Studies

Included studies were published mainly in scientific journals (66.67% of papers, mean Impact Factor = 7.01) [75,76,77,78,79,80,81,85,86,88,90,92]. The most popular was IEEE Transactions on Affective Computing. Out of all the authors, 27.78% did not report on funding [77,82,87,89,91], while 88.89% did not inform about competing interests [75,76,77,78,79,80,82,83,84,86,87,88,89,90,91,92].

None of the studies provided source code of executed analyses, while only one study (5.56%) reported registering protocol [92].

3.2. Experiments

The total number of elicitation experiments was 21, presented in 18 papers. Three of the studies (16.67%) carried out two trials each [77,78,81]. It was found that 76.19% of experiments were conducted using passive stimulation solely [75,76,77,78,79,80,81,82,84,86,88,90,91], while 19.05% used only active elicitation (e.g., video games) [83,85,89,92]. One experiment (described in Schmidt et al.’s paper [87]) used both passive (video) and active stimulation (meditation and Trier Social Stress Test (TSST). The essential characteristics of experiments regarding stimuli are presented in

Table 1. Characteristics of stimuli in 18 included studies (21 experiments).

Variable (No. of Experiments Available for Calculations)		No (%) Mean (Range)
Type of stimuli (21)
	Video (music, movie, ads)	11 (52.38)
	Audio (music excerpts)	4 (19.05)
	Game (FIFA 2016, Maze-Ball)	2 (9.52)
	Virtual Reality (videos, scenes)	2 (9.52)
	Self elicitation (actors)	1 (4.76)
	Mixed (TSST 1, video and meditation in one experiment)	1 (4.76)
Length of stimuli [seconds] (17)		304.60 (32–1200)
No. of stimuli in dataset (20)		27.70 (4–144)
No. of elicited emotions [classes] (18)		6.06 (3–23)

¹ Trier Social Stress Test.

.

Most of the experiments did not use stimuli from validated databases (71.43%, e.g., FIFA 2016, YouTube videos) [75,76,78,80,81,83,84,85,86,87,88,89,91,92], whereas public sources (e.g., DEAP, DECAF) accounted for 23.81% [77,79,81,90]. Pinto [82] used both forms. In 47.62% of experiments, the justification for the choice of the database was not reported [77,79,81,83,84,86,87,89,91,92]. Pinto [82] partially reported on it. In 42.85% of experiments, validation was provided by conducting a pilot study or preliminary classification by researchers [75,76,77,78,81,85,88,90]. Stimuli were described by authors most frequently in terms of valence (52.38% of experiments) [75,76,77,78,80,82,85,86,90,92], arousal (52.38%) [75,76,77,78,80,82,85,86,90,92], and discrete emotional tags (38.10%) [76,77,81,88,89,91]. Four experiments [77,79,83,84] did not report on it at all.

The presence of diseases or disorders was the most often controlled factor in participants (61.90% of experiments) [75,76,78,81,82,85,86,87,88,90,91]. The mood was controlled using the Positive and Negative Affect Schedule (PANAS) tool in two experiments (9.52%) [78]. In only one experiment, the authors checked if the participants were able to recognise emotions or affective states correctly [85]. In examining factors controlled in the laboratory, the most frequent was found to be brightness (33.33% of experiments) [75,77,78,90,91], followed by volume (28.57%) [78,80,82,84,91], presentation of stimuli (14.29%) [75,80,85], the comfort of participants (9.52%) [77,86], and time of the day (9.52%) [81]. In 23.81% of experiments, it was not reported which factors were controlled [77,80,83,84,92]. Additionally, in four experiments (19.05%), personality was measured in participants using the Big Five Personality Test [78,79] or the Eysenck Personality Questionnaire (EPQ) [91].

In the assessment of emotions and affect by participants, the most prevalent instruments used were: SAM (38.10% of experiments) [75,78,82,85,87,88,90] for valence, arousal and dominance, selecting a discrete emotion from the provided list (23.81%) [76,78,83,89], and PANAS (19.08%) [81,87,91].

3.3. Signals and Devices

Table 2. Characteristics of devices and signals in 18 included studies (21 experiments).

Variable (No. of Datasets Available for Calculations) 1		No (%) Mean (Range)
Used devices (16)
	Shimmer 2R	3 (18.75)
	BIOPAC MP150	3 (18.75)
	Biosemi ActiveTwo	2 (12.50)
	NeXus-10	1 (6.25)
	ProComp Infiniti	1 (6.25)
	BIOPAC BioNomadix	1 (6.25)
	BItalino	1 (6.25)
	RespiBAN Professional	1 (6.25)
	B-Alert x10	1 (6.25)
	Empatica E4	1 (6.25)
	Polar H7	1 (6.25)
	Zephyr BioHarness	1 (6.25)
	IOM Biofeedback	1 (6.25)
CV 2 signals recorded (18)
	ECG 3	15 (83.33)
	HR 4	3 (16.67)
	BVP 5	3 (16.67)
	PPG 6	1 (5.56)
Sampling frequency [Hz] (12)		543.31 (32–2048)
Length of baseline recording [seconds] (7)		292.14 (5–1200)

¹ some studies used more than one device or cardiovascular signal; ² cardiovascular; ³ electrocardiogram; ⁴ heart rate; ⁵ blood volume pressure; ⁶ photoplethysmogram.

summarises applied devices and recorded CV signals. Apart from CV signals, electrodermal activity (EDA) [75,76,78,79,80,82,83,84,87,88,91,92] is available in 66.67% of datasets. Thus, these are the most prevalent data. The next most common are face video [75,76,77,78,79,83,84,89,91], electroencephalography (EEG) [75,76,78,79,84,85,88,90], respiration [75,76,80,82,83,88], and electromyography (EMG) [75,77,80,83,87] in 50%, 44.44%, 33.33%, and 27.78% of datasets, respectively. The remaining signals include, e.g., magnetoencephalography (MEG), gyroscope, accelerometer, and audio. There is only one dataset that focuses solely on CV signals [86]. The authors recognise devices used for recording CV signals as wearable in 38.89% of studies [78,79,84,85,87,89,90].

3.4. Validation

As we included only validated datasets in this analysis, all of the papers explored AAER with AI. Out of all the papers, 55.56% [75,76,77,78,82,85,86,90,91,92] conducted experiments with only one type of ML algorithm, while the rest explored more methods. In total, the data were validated using AI methods 33 times. Support vector machine was used most frequently (33.33%) [76,77,79,82,83,84,85,86,89,90,91]. Naive Bayes (NB) [75,78,79,84], random forest (RF) [81,83,87,91], and DL [88,89] were the second most explored techniques (12.12% each).

The authors classified 61 different discrete states in total. The most commonly classified one was sadness [81,86,89,91], occurring in 22.22% of papers. The following states were examined in two papers each (11.11%): fear [89,91], anger [89,91], amusement [87,91], anxiety [89,92], boredom [89,92], happiness [81,89], and neutral state [88,89]. Additionally, the authors used affect space in 12 (66.67%) papers [75,76,77,78,79,80,82,83,84,85,86,90].

All of the datasets were validated in classification experiments. Authors of only two datasets (11.11%) [86,89] compared their results with other publicly available data.

3.5. Population

The total number of analysed people was 916, with a mean number of 43.62 participants and a range from 3 [80] to 250 [91]. However, due to, e.g., missing data, the datasets contain complete information for only 812 of them.

The remaining characteristics of the population are shown in

Table 3. Characteristics of population in 18 included studies (21 experiments).

Variable (No. of Experiments Available for Calculations)	No. (%) Mean (Range)
Participating people (21)	916 43.62 (3–250)
Eligible people (20)	812 40.60 (3–250)
Age (18)	23.8 (17–47)
Percentage of females (16)	45.13 (0–86)
Ethnicity (4)
Chinese	2 (9.52)
European	2 (9.52)

. five experiments (23.81%) were approved by the ethics committee [81,85,88,90]. Participant consent was obtained in 15 experiments (71.43%) [75,76,77,78,79,82,85,86,87,88,89,90,91]. Only one experiment ensured the privacy (by anonymisation) of participants [90].

3.6. Credibility

The general RoB was analysed in two scenarios—with or without the first item of the proposed tool (see Section 2.5). We excluded the first question in the second condition because none of the included studies reported on pre-specification of sample size.

Of all the studies, 77.78% were of low quality in both scenarios, whereas 22.22% [78,85,86,90] and 11.11% [86,90] were of unclear quality in the first and second conditions, respectively. Two studies [78,85] were of high quality according to the latter scenario. The RoB across all RoB items is presented in Figure 2. The reference standard was provided for all participants in the same way in 16 studies (88.89%) [75,76,77,78,79,81,82,83,84,85,86,87,88,89,90,91]. For participants from 14 studies (77.78%) [75,76,77,78,79,81,82,83,84,85,86,89,90,91], the same reference standard was given. Thirteen studies (72.22%) [75,76,78,79,81,82,85,86,87,88,89,90,92] did not show any flaws in terms of providing eligibility criteria for the experiment. These are the most satisfied questions.

However, 14 papers (77.78%) [75,76,77,79,81,82,83,84,86,87,89,90,91,92] did not control confounders or did not report it. Measurement of exposition [80,83,84,86,87,89,91] and outcomes [76,77,79,80,83,84,89] was flawed, or authors did not mention it, in seven studies (38.89%).

All ratings are presented in

Table 4. Risk of Bias (RoB) among 18 included studies.

Study ID	RoB Item 1								Overall Quality 2
	1	2	3	4	5	6	7	8	Scenario 1	Scenario 2
[75]	NR	PY	PY	Y	Y	Y	Y	PN	Low	Low
[76]	NR	PY	PY	Y	PN	Y	Y	NR	Low	Low
[77]	NR	NR	NR	Y	PN	Y	Y	PN	Low	Low
[78]	NR	Y	PY	Y	Y	Y	Y	PY	Unclear	High
[79]	NR	PY	PN	PY	PN	Y	Y	N	Low	Low
[80]	NR	PN	N	PN	PN	N	N	PY	Low	Low
[81]	NR	Y	PN	PY	Y	Y	Y	PN	Low	Low
[82]	NR	Y	Y	PY	Y	Y	Y	PN	Low	Low
[83]	NR	NR	NR	PN	PN	Y	Y	NR	Low	Low
[84]	NR	NR	NR	NR	PN	Y	Y	N	Low	Low
[85]	NR	Y	Y	Y	Y	Y	Y	Y	Unclear	High
[86]	NR	PY	PY	NR	Y	Y	Y	NR	Unclear	Unclear
[87]	NR	Y	Y	NR	Y	Y	N	NR	Low	Low
[88]	NR	Y	Y	Y	Y	Y	PN	Y	Low	Low
[89]	NR	Y	PN	NR	PN	Y	Y	NR	Low	Low
[90]	NR	PY	PY	PY	Y	Y	Y	NR	Unclear	Unclear
[91]	NR	PN	PY	NR	Y	Y	Y	NR	Low	Low
[92]	NR	Y	NR	Y	Y	NR	N	NR	Low	Low

¹ Y, PY, NR, PN, N stands for yes, partial yes, not reported, partial no, no; ² for more details, see Appendix C.

. Among them, the most frequent was yes, marked in 43.06% of cases. However, the second most prevalent was not reported, which was assessed 27.08% times.

3.7. Additional Analyses

Please refer to the protocol [62] and our other papers [63] from the project on AAER from CV signals with AI methods for additional analyses.

4. Discussion

The paper search conducted in this study revealed that there are 18 publicly available validated datasets for AAER from CV signals. The methodological credibility assessment showed that only two studies are of high quality, suggesting a significant need for developing good scientific practices.

Furthermore, none of the studies provided a source code used for the validation experiments. It opens a discussion on replicability, which we are witnessing in science nowadays [5]. Experiments in included papers were conducted on small samples. The number of participants exceeded one hundred only in one study.

What is more, the subjects’ background information was poorly described. Only four studies established that the participants were either Chinese or European. According to Wierzbicka [94,95], the history behind a person (and the language he or she speaks) may play a crucial role in the emotional states they experience and thus should be controlled. Feldman later disseminated this belief in her approach [7].

Another bothering aspect of the analysis is that an ethical commission approved experiments described by only four papers, and only one study mentioned ensuring the privacy of participants. It lights up red flags in terms of maintaining ethical standards or suggests negligence of reporting crucial information. Authors of experimental studies should more carefully examine this aspect.

Additionally, the authors either selectively controlled the influence of potential confounders or did not do so at all. Various CV diseases, mental disorders [49], and participants’ moods and personalities may affect AAER from physiological signals [78]. Therefore, we believe authors should include such information.

The problem in assessing quality in systematic reviews is about distinguishing how much the authors did not take care of the methodological regime and how much they did not report the details of the research process [60]. Therefore, it is recommended that when submitting an article to the journals’ editorial office, the authors fill in a checklist and mark the exact place where they have included the minimum necessary descriptions of the operation process [32,96].

On the other hand, we observed great diversity in the choice of physiological signals, stimuli type and length. What is more, 38.89% of the studies used wearable devices to perform measurements. Considering the increasing popularity and facility of these instruments [78], it gives the excellent potential for future adoption of proposed methods in real-life scenarios. Thanks to recent advances in the field of sensors technology, such devices are well-suited for daily usage. They do not require complicated instalments, are comfortable to wear, and are easy to use [97]. However, one should remember that there are still many limitations standing in the wy of the wider use of wearable devices in AAER. First of all, the quality of physiological signals is still noticeably lower than that of medical-level equipment [98]. What is more, the data gathered by such instruments in non-laboratory settings are often flawed, with noise coming from motion or misplacement [99].

Similarly to our study, the CV databases were also explored by Merone et al. [100]. The authors investigated 12 datasets with the inclusion criteria of having an ECG signal. In addition, they analysed included sets in terms of many parameters, e.g., the number of ECG channels and electrodes type. However, they did not primarily focus on emotions or affect. They included only one paper [101] covering this scope, which we did not consider eligible for inclusion as it did not meet the criteria. Since datasets including CV signals are still unexplored, we cannot discuss our results with other authors. Furthermore, Hong et al. [102] analysed ECG data systematically using DL. Still, they identified only one study about AAER [103], but it was not in their primary interest, so they did not describe it in detail.

However, in the literature, there are plenty of reviews (systematic and not) focusing on AAER from multiple signals or focusing on specific ones, e.g., EEG [104,105,106], or covering multiple modalities [53,107,108,109,110,111]. Still, their quality has been thoroughly criticised in our recently published umbrella review [5].

In line with these results, in the current literature, we found a shortage of highly credible and methodologically reliable publications and thus datasets that could form the basis of further AI research. This review shows a need to create guideline-compliant datasets with a transparent, fully reported methodology and limited RoB.

Models able to accurately recognise emotions using physiological parameters can contribute to the development of many disciplines. They create the possibility of reaching more advanced levels of HCI, where a computer (or system, in general) can modify its behaviour depending on the identified interlocutor’s state and choose the reaction closest to natural social schemes [112].

While using wearable devices, users might be supported in maintaining a psychological and healthy life balance, e.g., by identifying sources of stress, anxiety, or tension during their everyday activities and receiving feedback about their organisms reactions and resources [113]. Furthermore, assessments made on the basis of their CV signals can be used to investigate the impact of different emotional and affective states on the risk of developing CV diseases [114].

Well-validated AI models can significantly support research in the field of health and medical sciences and emotion theory by facilitating the simple, quick, and more matter-of-fact evaluation of emotions and other states and, therefore, reducing the RoB resulting from participants’ incorrect reporting.

Among the implications of our study, we should first include the recommendation to incorporate current, reliable guidelines and standards in the methodology development process and use quality assessment and reporting tools, as this translates into more reliable data, which may result in developing better recognition models [32]. For primary studies, we suggest following the proposed checklist for RoB (see Appendix C) or other available tools, e.g., [32].

Strengths and Limitations

The performed review has high standards [32,60,61,115]. The research question was precisely defined. We utilised multiple resources for collecting studies mentioned in Section 2. Inclusion and exclusion criteria were firstly discussed and recorded. Researchers who participated in this review have knowledge in multiple disciplines: computer science, psychology, HCI, medicine, and methodology. To ensure transparency, we provide all necessary information in the Appendices and Supplements with a permanent DOI [64].

On the other hand, we did not search any Chinese databases. Considering the growing amount of evidence in this language, we might not have considered a large amount of evidence and thus weakened our conclusions. Moreover, the use of the search strategy itself and the stages of identifying articles based on titles and abstracts may be a limitation. Due to such action, we may miss an extraordinary piece of work that did not meet our criteria due to its original form.

5. Conclusions

This paper systematically reviewed the datasets that include CV signals for AAER with AI methods and assessed their quality.

Due to poor reporting and not following methodological guidelines, the evidence, however, is limited. Nevertheless, according to our review, the most up-to-standards research was proposed by Correa et al. [78] and Marin et al. [85].

In the future, more attention should be put into controlling bias in research to ensure incremental knowledge gain. The quality of papers and reporting needs to be improved in order to propose and develop models that do not introduce biases. Preferably, authors should focus more on methodology and describe procedures thoroughly. We recommend following standardised guidelines of reporting [116].

Our next steps include the synthesis of gathered evidence with other physiological signals. Furthermore, we want to propose our own unbiased dataset for AAER for public use. Based on these data, we plan to improve our affective games [117,118,119].