This study delves into the exploration of pupillometry as a modality for afect state recognition. It examines the propensity for bias in both feature-based and learning-based machine learning models that interpret afect through pupil responses. Our research lies at the intersection of afective computing and mental health, recognizing the paramount importance of accurately identifying afect states for efective mental health interventions. We rigorously evaluate the performance of these pupillometry-based models across diverse demographic groups, including variables such as ethnicity, gender, age, vision problems, and iris color. Our findings reveal notable disparities, particularly in gender and ethnicity. Bias levels are pronounced in both feature-based and learning-based models, with F1 score diferentials reaching up to 36.28%. Additionally, our analysis uncovers a slight bias related to iris color, significantly impacting the eficacy of afect state recognition models that rely on pupil responses. This underscores the critical need for fairness and accuracy in developing machine learning models within afective computing. By highlighting these areas of potential bias, our study contributes to the broader discourse on creating equitable AI systems and advancing mental health care, education, and social robotics. It emphasizes the ethical imperative of developing unbiased, inclusive technologies in healthcare systems.
CEUR
ceur-ws.org Biometrics⋆
tal health, the intricate relationship between afect state
recognition and cognitive and mental health outcomes
Afect state recognition, central to understanding and
managing various mental health disorders, encompasses
the complex process of identifying and interpreting
emotional states [
ies within psychiatry and psychology, particularly in
assessing cognitive load for memory-based tasks [
The advent of technological solutions, such as recog- lometry in diagnosing nonconvulsive status epilepticus
Machine Learning for Cognitive and Mental Health Workshop ical responses such as pupillometry are generally
conmore, the predictive nature of afect state recognition in
mental health conditions paves the way for early and
more efective intervention strategies [
employed pupillometry to assess atypical pupillary light
reflexes and the LC-NE system in Autism Spectrum
Disorder (ASD)[
This work aims to investigate the bias that exists in Several features can be extracted from the pupil reafect state recognition models based on physiological sponse, including mean and variance of the pupil resignals, specifically pupillometry, which plays a signifi- sponse, maximum dilation, minimum contraction, dilacant role in understanding cognitive and mental health tion speed, dilation duration, contraction duration, and applications. The main goal of this study is to shed light the diference between dilation and contraction. In total, on the potential bias that may exist in common learn- 30 features were manually extracted and used to train ing methods. The structure of the paper is as follows: a kernel SVM classifier. Diferent kernels were tested, ifrst, the methodology, which includes preprocessing and and the Gaussian Kernel showed the best performance learning models, is explained. Then, the experiments in general. and results are presented and validated using a dataset collected for this work. Finally, we discuss the findings 2.3. Learned-Based Model and conclude at the end.
The framework focuses on the use of pupillary responses and approaches the task of afect state recognition as a binary classification problem based on the targeted group. The first step is preprocessing the pupillometry data to mitigate the efect of noisy samples. Then, the data is used to develop the classification model either from handcrafted features specific to the pupillometry data or using a learned model. Finally, model training and testing are described at the end to investigate the diferent cases.
The initial processing of the pupillometry data is
paramount to remove any irrelevant and noisy samples
that may impact pupil size analysis. The raw data can
be contaminated with various outliers like system errors,
blinks, eye-tracker glitches, and eyelid occlusion, which
can be identified and eliminated during this stage.
Previous studies [
The feature-based method is a common approach in machine learning where specific features are extracted from the data and used to train a the algorithm. In this study, the pupil responses for each participant were divided into 150 sets of sequences, with each sequence corresponding to the pupil response for each image. Each sequence has a length of 300 samples, which were used to extract the features.
The long short-term memory (LSTM) [
The use of deep learning methods such as LSTM for feature learning and afect state recognition is efective in various machine learning tasks. This approach can improve the performance of the model, as it can capture temporal dependencies and relationships in the data that might be missed by manual feature extraction. In both approaches, feature-based and learned-based, we divided the data into training and testing datasets, allocating 80% to training and 20% to testing, respectively. While constructing the model, we utilized data from all demographic groups with the intention of creating a model that captures feature representations from all these groups. To assess the model’s fairness and prevent bias towards any particular group, we further divided the testing data into subgroups during the evaluation phase and assessed the model’s performance for each subgroup.
Due to the limited number of samples, we introduced augmentation to enhance the training data. This augmentation was applied later in the evaluation, allowing us to assess its impact on the results. The pupil data sequences were augmented using noise injection and time-shifting methods [22]. Specifically, we added white noise to the original pupil data and performed 50 sample shifts. Importantly, the augmentation was applied to samples from the non-dominant group to ensure that our findings were not influenced by this imbalance. • Iris Color : The eye color case is a unique factor relevant to models using pupillometry data for their applications. Eye color afects the precision of detecting pupils and measuring their dilation and contraction. Thus, we categorize the data into light (light brown, green, blue, hazel) versus dark (black, brown, dark brown) iris colors. • Vision: This case evaluates the model’s efectiveness in capturing emotional states in data from individuals wearing glasses versus those not wearing glasses.
The proposed system was evaluated using a dataset collected at the University of Toronto. In the experiment, participants viewed a series of visual stimuli intended to 3. Experiments and Results elicit emotions spanning diferent valence and arousal values. The visual stimuli were selected from the InternaBias can be seen as the disparity in performance mettional Afective Picture System (IAPS) dataset [ 23]. The rics across diferent groups for a given task. Assuming IAPS database provides normative ratings of emotional we have = { 1, 2, … , } be the set of groups for bias valence and arousal for a large set of images. The rating investigation. For each group , we compute the per- scales are based on the Self-Assessment Manikin (SAM), formance metrics of a recognition model ( ). Then, a 9-point rating scale where a score of 9 represents a tahbesobliuatse diifesriednecnetiifiendtfhoerirampeatirriocsf:groups ( , ) as the high rating (i.e., high pleasure, high arousal), a score of 5 indicates a neutral rating, and a rating of 1 represents a ( , ) = | ( ) − ( )| low rating (i.e., low pleasure, low arousal).
The selected visual stimuli elicit the emotions of
interest, which include the two quadrants of the VA
di3.1. Data mensional model (i.e., HA, LA, or HV, LV). Each of the
To conduct a thorough assessment of bias in pupillom- aforementioned emotional states is achieved by
displayetry afect state recognition, we collected a dataset that ing 30 images of the same emotional target for 5 seconds
encompasses pupillometry data in response to visual stim- each. The images were selected to statistically produce
uli, taking into account a diverse range of demographics. the same response for diferent groups of people. All
The study involved 35 university students aged between images were presented on a screen with a resolution of
18 and 40 years, with a mean age of 24.6 and a standard 1920 by 1080 pixels. Following the recommendations of
deviation of 5.17. Participants were required to have the device manufacturers, the Gazepoint eye-tracking
no history of vision disorders, and they were also asked system was placed approximately 45 cm in front of the
about any medications they might be taking that could af- participant at an angle of around 30 degrees. The total
fect their responses, such as depression medication. The number of participants was 35.
data collected from the participants is categorized into The data collection process was approved by the
rediferent cases based on various demographic factors: search ethics committee at the University of Toronto. All
participants signed a consent form that clearly explained
• Gender : This case examines the algorithm’s abil- the data collection procedure and the privacy of their data.
ity to fairly recognize emotional states in females Furthermore, all participants received compensation in
versus males. the form of a gift card.
• Ethnic Group: This case assesses the model’s
ability to impartially detect emotional states 3.3. Metrics:
based on participants’ ethnic groups, including
Asian (Chinese), White (North American or Eu- In the evaluation process, two common metrics were
ropean), Black (African American or Caribbean), employed: accuracy and F1 score. Accuracy gauges the
and South Asian (Pakistani or Indian) [
We employed a feature-based algorithm for emotion recognition and assessed the presence of bias among different demographic groups, focusing on valence-based and arousal-based classifications. Our evaluation yielded results presented in Tables 1 and 2, along with Figures 2 and 3.
Notably, our findings reveal significant performance diferences between males and females in both arousal and valence. Specifically, our analysis indicated that males scored 20.28% higher in arousal and 17.46% higher in valence compared to females. The F1 score exhibited a similar gender-based pattern of diferences.
Further examination of the model based on ethnicity factors showed significant variations in accuracy and F1 scores across diferent groups. Notably, the Asian group, despite having the highest number of samples, displayed the lowest accuracy and F1 scores in terms of arousal classification. In contrast, the South Asian group, with the second-lowest number of samples, demonstrated the highest performance. The percentage diference between the highest-performing group (South Asian) and the lowest-performing group (Asian) was 28.93% in accuracy and 21% in F1 score for arousal classification. These ifndings suggest that obtaining accurate feature representations for the Asian group in terms of arousal classiifcation may be more challenging based on the provided stimuli.
Regarding valence classification, our analysis revealed similar performance among the Asian, White, and South Asian groups, while the Black group exhibited significantly lower accuracy and F1 scores. Specifically, the percentage diference between the Black group and the group with the highest performance was 26.99% in accuracy and 46.71% in F1 score, respectively.
We employed an LSTM-based approach to investigate bias across diferent demographic groups. The results of Ethnic Group
Asian
White South Asian
Black in arousal accuracy and a 10.15% bias toward males in valence accuracy. Concerning ethnic groups, accuracy exhibited substantial variations across diferent ethnicities, as depicted in Tables 3 and 4. In terms of arousal, the Asian group had the lowest performance, while the White group achieved the highest accuracy, resulting in a significant 20.90% advantage favoring the White group. The other ethnic groups showed similar performance. In terms of valence, the Black group displayed the lowest performance, while the South Asian group achieved the highest, with a diference of 36.28%. In the remaining cases, there were no significant diferences between individual groups, suggesting that these factors share common representations that can be captured by the algorithms.
Based on the results presented above, it is evident that both models exhibit significant diferences in accuracy and F1 scores concerning ethnic groups and gender. This indicates that these two factors play a pivotal role in the development of afect recognition from pupillometry data, as the models struggled to find efective representations for them. In contrast, the other four cases displayed minor diferences in terms of accuracy and F1 scores, suggesting that these factors share common representations across all groups and do not adversely afect the data’s quality. For example, iris color had a limited impact on recognition performance, albeit not as pronounced as with gender and ethnic groups.
Despite the dataset including diverse groups during model training, the quality of the representations failed to adequately capture the diverse group responses within the studied population. We acknowledge that the unbalanced number of samples in each group might contribute to the bias observed in the results. To address this potential issue, we implemented data augmentation techniques (see 2.4) for the non-dominant groups (groups with fewer samples) to increase their sample size. Subsequently, we followed the same procedure as in the original case. However, our results demonstrated that even with the implementation of data augmentation, the performance did not change significantly. The bias in performance persisted in both the ethnic groups and gender-based cases, while the remaining cases exhibited similar performance.
In this study, we investigated the performance of featurebased and learned-based afect recognition models across various group factors, including ethnicity, gender, vision, iris color, and age, focusing on pupillometry as a the modality. Our research, involving a dataset from 35 diverse participants, revealed significant gender and ethnic biases in standard afect recognition algorithms, impacting both arousal and valence-based classifications. We also identified minor biases related to other factors, such as iris color. These findings emphasize the potential bias in afect recognition systems, highlighting the need for more inclusive and representative training data, rigorous fairness evaluation, and enhanced transparency in model development. Our study not only sheds light on the inherent biases in afective computing but also underscores the importance of considering demographic factors in the development of more equitable and efective afect recognition technologies, particularly given their direct relation to cognitive and mental health. [22] T. Ko, V. Peddinti, D. Povey, S. Khudanpur, Audio augmentation for speech recognition, in: Sixteenth annual conference of the international speech communication association, 2015. [23] P. J. Lang, International Afective Picture System (IAPS): Afective ratings of pictures and instruction manual, Technical report (2005).