Do Deepfakes Adequately Display Emotions? A Study on Deepfake Facial Emotion Expression

Recent technological advancements in Artificial Intelligence make it easy to create deepfakes and hyper-realistic videos, in which images and video clips are processed to create fake videos that appear authentic. Many of them are based on swapping faces without the consent of the person whose appearance and voice are used. As emotions are inherent in human communication, studying how deepfakes transfer emotional expressions from original to fakes is relevant. In this work, we conduct an in-depth study on facial emotional expression in deepfakes using a well-known face swap-based deepfake database. Firstly, we extracted the photograms from their videos. Then, we analyzed the emotional expression in the original and faked versions of video recordings for all performers in the database. Results show that emotional expressions are not adequately transferred between original recordings and the deepfakes created from them. High variability in emotions and performers detected between original and fake recordings indicates that performer emotion expressiveness should be considered for better deepfake generation or detection.


Introduction
Recent technological improvements have made it simple to construct "deepfakes" and hyper-realistic videos that use face swaps and leave little evidence of alteration [1]. Artificial intelligence (AI) apps blend, replace, and superimpose photos and video clips to generate phony videos that appear legitimate [2]. Without the approval of the individual whose image and voice are involved, deepfake technology can manufacture any kind of material [3]. e extent, scale, and sophistication of the technology involved in deepfakes are game-changing, as essentially anyone with a computer can create fake movies that are virtually indistinguishable from legitimate media [4]. Most early examples of deepfakes focused on joke videos about well-known individuals. However, risks are emerging as they are being used for revenge porn, bullying, fake video evidence in courts, political sabotage, terrorist propaganda, blackmail, market manipulation, and fake news [2]. e human face can express emotions faster than people can explain or even comprehend their sentiments, making facial expression one of the most direct ways humans transmit their emotions [5]. Automatic facial expression recognition (FER) is becoming a hot topic in academia.
Although most of these systems aim to recognize only a small number of prototypical emotional expressions, much progress has been made in developing computer systems that analyze this sort of human communication [6][7][8]. Many recent efforts on facial emotion recognition based on facial expressions have used deep learning to solve the problem, whether in static photos or dynamic video recordings [9][10][11][12][13]. Emotions in FER are becoming growingly important in many fields, such as health [14].
Literature on deepfakes and emotions is scarce. To the authors' knowledge, the only works that specifically tackle audio-visual deepfake detection using affective cues are [15,16], and [17]. Moreover, we could not find specific studies on how well deepfakes express emotions.
How well emotions are displayed in deepfakes when faces are swapped between different performers is an interesting research question. To that end, an in-depth study on facial emotional expression in deepfakes is carried out in this work. With that goal, a well-known face swap-based deepfake database is used. e photograms from their videos were extracted first, and the emotional expression was analyzed in the original and faked versions of the video recordings for all performers in the database. Results show that emotional expressions are not adequately transferred between original recordings and the deepfakes created from them. e high variability in emotions and performers detected between original and fake recordings indicates that performer emotion expressiveness should be considered for better deepfake creation. e structure of this document is as follows: the related work section presents studies found in the literature on deepfakes, FER, and emotion recognition. e methods section describes used databases and introduces the method used to categorize emotional photograms from video recordings.
e Results section shows obtained results in detail. Results are interpreted, and the discovered knowledge is displayed in the Discussion section, while the Conclusion section presents the conclusions.

Related Work.
Although there have been numerous advanced algorithms for making realistic synthetic face films in recent years [18], most of them have not yet become widely available as open-source software tools that anybody can use. On the other hand, a much simpler method based on the work of neural image style transfer [19] has become the tool of choice beneath many deepfake videos or existing deepfake datasets. Moreover, there are several open-source implementations, such as Face2face [20] and DeepFaceLab [21]. [22] provides a comprehensive review of deepfake technologies. e encoder and decoder are usually two convolutional neural networks that make the autoencoder. e encoder turns the face of the input target into a code vector. ere is only one encoder, regardless of the individuals' identities, to ensure that identity-independent features are captured, such as facial expressions. Each identity, on the other hand, has its decoder, which uses the code to generate the face of the matching subject. In an unsupervised way, the encoder and decoder are trained in tandem using noncorresponding face sets of various participants.
Deepfakes are difficult to be identified effectively with current detection methods. Shad et al. implemented several methods to detect deepfake images and made a comparative analysis [23]. In detection models, to identify video tampering, stacked autoencoders, CNNs [24], long-short term memory (LSTM) networks [25], Siamese networks [26], or GANS have been investigated. e algorithm in [27] recognized false videos by detecting key video frames, which reduced calculation time and increased accuracy when the video featured more than one keyframes. EfficientNet-V2 has also been used with good results to detect deepfakes in large-scale fake face video datasets [28].
Multiple attempts have been made to release benchmark datasets because of the rise in AI-generated deepfake material. e amount and quality of early deepfake detection datasets, such as the UADFV dataset [29] and the Deep-fakeTIMIT dataset [30], are limited. Many of these issues are addressed by the more recent FaceForensics++ [31] and the DeepFake Detection Challenge dataset [32], as well as other viable deepfake detection databases. e Google DeepFake Detection dataset [33] contains 3,068 deepfake movies created from 363 original footage of 28 agreed individuals in 16 situations. e source video recordings of 100 hired actors are included in DeeperForensics [34], while 1,000 target videos are taken from FF++. Each source identity is swapped onto 10 target videos to create 1,000 phony videos. 590 real videos and 5,639 false videos make up Celeb-DF [35]. e original videos were obtained from YouTube, and the content consisted of 59 celebrity interviews. An enhanced face swap approach is used to create the faked videos.
In terms of facial expression emotion recognition, Ekman and Friesen created a facial action coding system (FACS) to represent facial expressions using action units (AUs) [36]. ey identified 30 FACS AUs, out of the 44 FACS AUs they described, that were anatomically associated with the contractions of facial muscles. Of these 30 FACS AUs, 12 correspond to the upper face and 18 to the lower face. AU scans can be done individually or in groups [6]. Human encoders can use FACS to encode all facial expressions using these 30 AUs manually. e emotional labels may be described by the AU combinations defined in the FACS. FACS has become a face behavior measurement criterion in various disciplines, including computer-based vision [37], because of its descriptive ability. Based on FACS, the emotion facial action coding system (EMFACS-7) was later suggested to determine whether basic emotions had prototypical facial expressions [38]. On the other hand, such archetypal utterances are uncommon in ordinary life. Instead, small changes in one or a few identifiable facial features, such as tightening of the lips in anger or obliquely dropping the lip corners in melancholy, are more typically used to transmit emotion [39].
Other works, such as [40,41], have provided FER automatically based on FACS. Face and eye detection, which includes facial landmarks, head attitude, and eye gaze, is the first step in automatic recognition. en, using a classifier, AU estimation is performed.
To the authors' knowledge, [15,16], and [17] are the only studies that exclusively address audio-visual deepfake detection using emotive cues. To detect falsifications or manipulations in the input video, the approach provided in [15] concurrently utilizes the audio (speech) and video (facial) modalities, as well as perceived emotion components collected from both modalities. ey used both modalities to detect similarity (or dissimilarity) between modality signals, and they discovered that perceived emotional information aids in detecting deepfake content. [16,17] offered a technique for detecting deepfakes based on semantic consistency in emotion, which was based on previous emotion identification work that extracted emotions over time from a 2 Computational Intelligence and Neuroscience subject's speech and faces separately. Synthesized voices or faces are then detected by analyzing these emotional signals.

Methods
e materials utilized in this study, including how they were created, are described in this section. Considered facial action units for emotion recognition are then specified. Finally, the procedure and design are described, including the models used to analyze the materials, how they are evaluated, and which metrics are employed.

Materials.
e Celeb-DF database includes 590 genuine videos and 5639 DeepFake videos in the Celeb-DF, encompassing over two million photograms. With a normal frame rate of 30 frames per second, the average length of all videos is around 13 seconds. e real videos were selected from publicly available YouTube videos showing interviews with 59 celebrities of various genders, ages, and ethnic backgrounds. Furthermore, the real films show a wide range of changes in factors, such as the size of the subjects' faces (in pixels), orientations, lighting conditions, and backgrounds. DeepFake films are made by switching the faces of each of the 59 subjects. e finished videos are in MPEG4.0 format [35].
Celeb-DF's videos were created with a DeepFake synthesis algorithm based on the original DeepFake maker framework [19].
is framework is based on variational autoencoders (VAEs) and generative adversarial networks (GANs). It consists of 6 subnetworks, including two domain image encoders, two domain image generators, and two domain adversarial discriminators. e algorithm learns translation in both directions in one shot, and it was improved in different ways to address the following unique visual artifacts that have been detected in prior datasets: (i) Enhancement of the simulated face to 256 × 256 pixels. (ii) e color disparity between the synthesized donor's face and the original target's face has been significantly reduced. (iii) Improved face mask creation phase by synthesizing a face with more surrounding context to cover the original facial parts after warping completely. e result is a smoother mask based on landmarks on the eyebrow and interpolated points on the cheeks and between the lower lip and the chin. (iv) Reduced temporal flickering by integrating temporal correlations among the observed facial landmarks.

Facial Action Units in Emotion.
Face feature extraction is a critical step in recognizing accurate facial expressions. e FACS approach introduced by Ekman and Friesen [36] has been the beginning point for encoding facial traits. e facial action coding system (FACS) [36] explains facial expressions through action units (AUs), which are physically tied to the contractions of specific facial muscles and can occur singly or in combination. EMFACS [38], based on FACS, was designed with a subset of AUs associated with emotions in mind, as described in the previous section. It is possible to map AUs onto the fundamental emotion categories using a finite number of criteria, as described in the FACS Investigators' Guide [42]. e action units (AUs) and categorical emotions associated with them are shown in Table 1. Happiness (Ha), Sadness (Sa), Surprise (Su), Fear (Fe), Anger (An), and Disgust (Di) are the categorical emotions.
Faces that lacked any emotion-related AUs were classified as Neutral.

Procedure and Design.
e convolutional experts network model [43] is used in the OpenFace facial analysis tools [38]. An input image is given, and a region of interest of size n × n is retrieved from it based on the estimated landmark position.
is small region is passed via a contrast normalizing convolutional layer with a kernel shape 500 × 11 × 11 that performs Z-score normalization before the correlation operation, resulting in a 500 × n × n with n � n−10. e response maps are then fed into a 200 × 1 × 1 convolutional layer containing ReLU units. e mixture of expert layer employs a convolutional layer of 100 × 1 × 1 sigmoid probability decision kernels to develop an ensemble to capture ROI fluctuations. e output response map is a nonnegative and nonlinear combination of neurons in the mixture of expert layer using a sigmoid activation. e convolutional experts network model was used to extract a set of facial features for each frame of each video in the Celeb-DF database. ese features include face location and rotation, gaze direction, the location of face parts in 2D and 3D, the presence or absence of AUs, and the intensity of AUs if they are present.
Two possible AU prediction models exist, depending on dynamism. e static one uses a single image to estimate the existence or intensity of AUs (henceforth, the static model). In contrast, the dynamic model is calibrated to a person by performing person normalization in the video and seeking to adjust for over and under prediction of AUs (subsequently, the dynamic model). By default, static models are used for photos and dynamic models for image sequences and videos. However, some video sequences have limited dynamic range (the same expression is maintained throughout the clip), making postcalibration ineffective and possibly damaging [41]. We applied both prediction models to analyze the input videos and compared the results. e total number of photograms with faces analyzed was 225,390 for original videos and 2,116,768 for fake videos in the Celeb-DF database. Obtained data was saved to files in CSV format, including original and fake videos in the static and dynamic models, adding up to 23,4 GB of data to be analyzed.
After all videos from the Celeb-DF database were analyzed with both static and dynamic models, we developed a script to select the set of emotionally relevant photograms for each model. is selection was based on the mapping of AUs onto the basic emotion categories described by Ekman Computational Intelligence and Neuroscience using the rules displayed in Table 1. e selected set added up to 7,6 GB of data.
Subsequently, the selected set was analyzed using the following metrics: (i) e percentage of fake recordings with emotional photograms per performer and emotion (ii) For each emotion, the number of emotional photograms that are only in the original, only in the corresponding fake, and those in both (iii) Sample data and detected overall emotion (amount and percentage) per performer for dynamic and static models, considering emotions just in the original, the fake, and those in common Table 2 describes the detected emotional recordings in the fakes in the sample, as detected emotions vary considerably between performers in the database. It shows the percentage of recordings with emotional photograms in the Celeb-DF database for each emotion and performer, both in dynamic and static models. e recordings with emotional photograms are displayed by their percentage normalized to the number of fake recordings per performer. For example, the 14.04% in Sadness for performed with id0 means that the 14.04% of fake recordings of the performer with id0 included at least one sad photogram with the dynamic model. Most relevant values in each column (the highest values) are displayed in bold. Table 3 shows the results for recognized emotional photograms in the Celeb-DF database. e outcomes for each emotion are shown by the number of emotional photograms in common (EPC), emotional photograms in original alone (EPOA), and emotional photograms in fakes alone (EPFA). Photograms in common are the ones that share a given emotion by the original recording and the corresponding photogram of a fake recording made on the original one. In contrast, original alone photograms are the ones that showed emotion in the original but not in the fake, and fake alone are the ones that showed emotion in the fake but not in the original one. e percentages are also shown by the percentage of common emotional photograms (%C), emotional photograms in original alone (%O), and emotional photograms in fakes alone (%F). To adequately display the data, emotional photograms detected in original recordings were split between the ones in common with the corresponding fakes and those just in the original. erefore, values in columns of common and original alone for each emotion sum 100%. e percentage of fakes is calculated compared to the sum of common and original alone photograms. As both dynamic and static models were used in the study, both are included. e results shown are global, including all recordings in the database, and they intend to show similarity in emotivity between original and fake recordings. Most relevant values in each column (the highest values) are displayed in bold. Figure 1 displays the percentages of common, original alone, and fake alone photograms per emotion, both in dynamic and static models. 100% is depicted as the sum of common, original alone, and fake alone. e aim is to properly portray the similarity between emotional photograms in original and fake recordings for each emotion.

Results
Performers' identifiers range from 0 to 61; however, there were no performers with ids 14, 15, and 18. Hence, there are no rows for them. As for the performer with id 36, he is in the original recordings of the Celeb-DF database; however, no fakes were performed because he was wearing glasses. Table 4 shows information about the sample per performer. e first four columns display information about performer id (Id), number of original recordings by the performer (OR), number of performers of deepfakes on the original recordings of the current performer (PF), and number of fake recordings made on recordings by the performer (FRP). e remaining columns show the number of photograms with emotions in common with the original recording and the corresponding photogram of a fake recording (CO), the number of photograms that showed emotion in the original recording but not in the corresponding fakes (OR), and the number of emotional photograms that appear in fakes but not in corresponding photograms on original recordings alone (FA). Next to them, their corresponding percentages are displayed. Values in columns of common (%CO) and original alone (%OR) sum 100%, and the percentage of fakes (%FA) is calculated compared to the sum of common and original alone photograms, as in Table 3. Some results over 100% appear in fake percentages, as the number of fake recordings is larger than the original ones, and the percentage of fakes is calculated compared to the sum of common and original alone photograms. As dynamic and static models were used in the study, both outcomes are displayed. Most relevant values in each column (the highest values) are displayed in bold.
As the amount of data per emotion was too big to display adequately using tables, we have used figures to represent global data and the most representative two emotions graphically.   Computational Intelligence and Neuroscience photograms. erefore, a high percentage in X represents a high percentage of emotional photograms in original recordings alone, while high percentage in Y represents a high percentage of emotional photograms in fake recordings alone. e circles with big diameters represent high common percentages between original and fakes. e id of the corresponding performer is placed next to its corresponding circle. At the bottom right part of each image, a scale details the percentages corresponding to the different diameters, which serve as a visual guide. e caption of each figure shows the emotion or emotions it represents. In some cases, fake recordings showed a specific emotion, and their corresponding originals did not. When that happened for all original recordings of a given performer, as the percentage increase should be performed over 0%, we did not include these recordings in the previous figures. e number of performers in which that happened, for each emotion and model, were as follows: Finally, there were no performers without neutral emotion in both originals and fakes using the dynamic model. However, there were 16 without neutral emotion detections using the static model.

Discussion
ere are no known studies about emotions in deepfake databases. We address this shortcoming by analyzing a database that includes recordings with emotional content and is based on the most used deepfake strategy for deepfakes and face-swapping. Our study focuses on how well emotions are displayed in these databases when faces are swapped between different performers. e selected deepfakes database, Celeb-DF [35], is a wellknown deepfake database whose original videos were chosen from publicly available YouTube videos corresponding to interviews of celebrities. It includes a wide range of facial expressions by the performers. Although it is a fairly new database, it has already been cited many times, over 170 in Scopus and 300 in Google Scholar as of May 2022.
As recordings in the Celeb-DF database were not selected with emotions in mind, we first determined how emotional the fakes used the OpenFace emotion detection tool. Table 2 shows that fake recordings included emotions, regardless of the type of model used to analyze them. e use of two different models was decided because of the difficulty to capture spatial-temporal information of expressions with a slight motion in facial emotion databases [10]. Still, the percentage of fake recordings that included emotions varied notably, considering each specific performer, emotion, and model. Overall, the static model detected considerably more emotional recordings than the dynamic one. Besides, there are significant differences among different emotions. Almost no performer expressed fear in any recording, which is not surprising as it is the most difficult to display, considering it involves more action units (AUs) [36] than other emotions. On the other end, happiness was portrayed by almost all performers.
To properly compare emotions detected in fakes with the originals, we split the percentage of emotional photograms detected in original recordings between those just in the original and those in common with the corresponding fakes. Results in Table 3 show that the metric that indicates the "goodness" of emotions in fakes, the percentage of common emotional photograms between original and fake recordings, is higher than 50% only for Happiness, Disgust, and Sadness just in the case of the static model. e percentage varied widely, between 9.08% and 66.77%, depending on the emotion and the model. e difference in categorical emotion display is consistent with the literature. e difference in emotion recognition is usually shown in the confusion matrices of the results obtained by machine learning systems. Happiness is usually the emotion with better recognition rates, and Fear is the one with the worst in posed expressions [44]. Figure 1 shows that normalizing all detected emotional photograms to 100% per emotion, and these differences are also evident when considering global data. As for the used models, the proportions of emotions are similar in both cases. e only exceptions are for the static model, in which Sadness is noticeably higher and Neutral drops considerably.
In this regard, the results for Neutral photograms stand out. e percentages in common with the corresponding fakes are medium compared to the other analyzed emotions in the dynamic model and the lowest in the static. ere were 16 out of 59 performers for which no neutral photograms were detected using the dynamic model, neither for original nor for fake recordings. Besides, data from 23 performers showed no neutral photograms in fakes when none was found in original recordings. Although the number of  Figure 3), the percentage of common neutral photograms is the lowest of all. In contrast, photograms in common are similar or better for all emotions compared to the dynamic model. It indicates that person normalization-based calibration and correction of AUs performed in the static model works better for emotion recognition. However, it does not perform well for neutral emotion.
Most human facial emotion works focus on specific emotion category recognition, and few references are found on the neutral recognition rate. Most times, such information is shown in confusion matrices. When such information is available, results in neutral pictures are generally good, comparable to the best emotions identified in the same works [45].
Regarding facial emotion recognition in video recordings or dynamic pictures, many recent works tackle it using deep learning. Video recordings in selected emotional video recording databases are preprocessed, and prior knowledge about emotional transitions is used when developing deep learning models. Evaluation protocols select specific frames in the video recordings for model training, typically only the video recordings that have one of the labeled emotions and a neutral frame at the beginning [9]. Or neutral expression frames at the beginning and then a different number of frames near the emotional peak of the sequence [10][11][12]. As it happens with images, in most cases, neutral is not included when detecting emotions in video recordings [13,46]. Different problems arise when working with facial emotion databases. Usually, all recordings in benchmark databases reflect the same temporal activation patterns. Taking two of the most referenced facial emotion databases as a guide, recordings in the CK + database [37] include transitions from neutral to the peak of emotion, while recordings in the MMI database [47] follow onset-apex-offset temporal segments. While the CK + database is widely used without much trouble or need for preprocessing, the MMI database is more challenging for a series of reasons. First, the subjects perform emotional expressions in a non-uniform way, as different people perform the same expression in different ways. Second, some subjects have a mustache or wear accessories, such as glasses or headcloths. ird, in some recordings, the apex frames are not with high expression intensity [48]. Finally, some recordings are more complicated as they include several different emotional expressions in the same sequence [49].
As a consequence of these challenges, in many cases, specific subsets of recordings are selected with the less troublesome recordings to achieve better recognition rates [12,49], including discarding the recordings that do not start with entirely neutral photograms [48]. Neutral emotion is not explicitly classified in many works, even though transitions are a crucial factor when identifying emotions, as most databases do not include specific neutral recordings and initial photograms are considered neutral [49]. Besides, the difficulty in capturing spatial-temporal information of expressions with slight motion is a common trait in all facial emotion databases [10].
Neutral is understood as the lack of a specific categorical emotion. In the field of facial AUs, it can be interpreted as the lack of any AUs or the lack of any emotion-related AU, which is the approach we used in this work. Bad results in common neutral percentages indicate that neutrality is not adequately transferred to fakes. It is important to note that emotional video affects recognition systems are mainly based on transitions from neutral to specific emotions. Consequently, the effect of bad neutral photograms in fakes, mainly in initial photograms, may suggest that emotion recognition is prone to work worse than in their corresponding original recordings. Works on how to include emotional cues in deepfake detection are starting to be   [15][16][17], which indicates that emotional cues in video recordings can be used effectively for deepfake detection.
Results show that the improved algorithm based on the basic DeepFake maker algorithm [19], used in the Celeb-DF database, does not adequately transfer emotions expressed with facial expressions to deepfakes created from original recordings with emotional photograms. Our study also shows the great extent of the differences between emotions and performers when adequately displaying emotions in performed fakes (see Table 4). We can draw two main conclusions regarding deepfake creation. First, emotions should be considered criteria for selecting original recordings to make fakes. Second, performer selection should also be tuned to performers with similar emotional expressiveness.
Of course, using one deepfake database, even though it is widely used and includes a reasonable amount of original and fake recordings, limits the generalizability of the findings. erefore, more databases should be considered, their emotional expressivity analyzed, and comparative analysis performed for originals and fakes. Considering static facial expressions has improved the percentages of common detections in original and fake recordings for some emotions, while it has worsened them for others. is difference indicates that mixed approaches should be considered when analyzing emotions in recordings. Moreover, the results suggest that the scope of neutrality, although not discussed enough in the related literature, can be considered a factor affecting achieved results. As for the dynamic and static models used to extract facial expressions, although they are widely used, they may not have adequately identified all AUs. In this regard, the OpenFace implementation was selected because it has been trained and tested against multiple facial expression databases, including databases with emotional recordings [41].
Another aspect to discuss is that presented study has been performed on a database with images with good quality and contrast. In case a database or sample images do not have adequate quality or contrast, the use of contrast enhancement techniques would be necessary for better affective image processing. In these scenarios, the scientific literature offers a wide range of fuzzy image preprocessors that could be used or adapted, such as color extraction methods [50] or fuzzy image preprocessors based on geometric computations in Euclidean spaces [51]. ese computations are characterized by a reduced computational load, which is particularly useful for any real-time applications. the original recording and the corresponding photogram of a fake recording (CO), the number of photograms that showed emotion in the original recording but not in the corresponding fakes (OR), and the number of emotional photograms that appear in fakes but not in corresponding photograms on original recordings alone (FA) are displayed followed by their corresponding percentages, for dynamic and static models.   Computational Intelligence and Neuroscience 9

Conclusion
As far as we know, the presented study is the first one that addresses explicitly facial emotion expression in deepfakes. According to the findings, the deepfakes that were produced utilizing a face swap algorithm from the original recordings are not sufficiently capable of recreating emotional expressions. Concretely, a much higher proportion of emotional photograms is found in fake recordings compared to the authentic ones. Additionally, we saw that emotions varied widely from one another. e percentage of times the same feeling was present when a photogram revealed emotion in the false recording and the corresponding original was high. ese incidences, however, were insignificant when compared to the percentages of emotional photograms that were only found in genuine or fraudulent recordings. Besides, a high variability has been observed between emotions and performers, while adjustments to face dynamism show a better common emotion recognition between originals and fakes, and worse for neutral. erefore, performer emotion expressiveness should be considered for better deepfake creation. e results of this study have ramifications for the development and detection of deep fakes. When constructing fakes from authentic recordings, deepfake algorithms should consider how well they convey emotions. Regarding deepfake identification, our results support the variation in emotional expressions between authentic and fake recordings as a viable data source for identifying fake recordings.

Data Availability
e tabular data used to support the findings of this study have been deposited in the CORA repository (https://doi. org/10.34810/data262).

Conflicts of Interest
e authors declare no conflicts of interest regarding the study.