Algorithmic Fairness and Bias Mitigation for Clinical Machine Learning: Insights from Rapid COVID-19 Diagnosis by Adversarial Learning

Machine learning is becoming increasingly promi- nent in healthcare. Although its benefits are clear, growing attention is being given to how machine learning may exacerbate existing biases and disparities. In this study, we introduce an adversarial training framework that is capable of mitigating biases that may have been acquired through data collection or magnified during model development. For example, if one class is over-presented or errors/inconsistencies in practice are reflected in the training data, then a model can be biased by these. To evaluate our adversarial training framework, we used the statistical definition of equalized odds. We evaluated our model for the task of rapidly predicting COVID-19 for patients presenting to hospital emergency departments, and aimed to mitigate regional (hospital) and ethnic biases present. We trained our framework on a large, real-world COVID-19 dataset and demonstrated that adversarial training demonstrably improves outcome fairness (with respect to equalized odds), while still achieving clinically-effective screening performances (NPV>0.98). We compared our method to the benchmark set by related previous work, and performed prospective and external validation on four independent hospital cohorts. Our method can be generalized to any outcomes, models, and definitions of fairness.

The Intersection of Candidate Gender and Ethnicity: How Voters Respond to Campaign Messages from Latinas

Despite the recent surge of women of color in elected political office in the U.S., the representation of Latinas is strikingly low. Past research offers unclear conclusions as to whether Latina political candidates face biases due to the intersection of their identities as women and as ethnic minorities, and how Latinas can navigate such biases. In this study, we identify how Latinas draw on their intersectional identities as both women and ethnic minorities to develop strategic campaign messages and how voters respond to such messages. Through an analysis of campaign advertising data and an original survey experiment, we show that Latina candidates do not face an automatic disadvantage based in gender and ethnic biases, but they can benefit from the intersection of these two identities, especially among female minority voters. These results are consequential because they offer insights into how to improve the descriptive and substantive representation of marginalized groups in the U.S.

OTHR-01. SNO 2020 DIVERSITY SURVEY: DEFINING DEMOGRAPHICS, RACIAL BIASES, CAREER SUCCESS METRICS AND A PATH FORWARD FOR THE FIELD OF NEURO-ONCOLOGY

BACKGROUND Neuro-oncology has grown tremendously since 2010, marked by increasing society membership, specialized clinical expertise, and new journals. Yet, modest improvement in racial/ethnic diversity amongst clinical trial participants, researchers and clinicians led us to conduct a survey identifying opportunities to enhance diversity and inclusiveness amongst neuro-oncology professionals. METHODS Summer 2020, the Women and Diversity Committee of the Society for Neuro-Oncology (SNO) distributed an anonymous online survey to members and affiliates (European Association of Neuro-Oncology, Asian Society for Neuro-Oncology, Society for Neuro-Oncology Latin America and Society for Neuro-Oncology Sub-Saharan Africa). The survey captured personal and professional characteristics, biases, effective mentorship qualities, career success metrics, and suggested field/society changes. Results were analyzed by geography, profession, age, racial/ethnic and sexual identity. Standard descriptive statistics characterized the study population. RESULTS The 386 respondents were predominantly female (58%) with a median age range of 40-49 years (31%), White (65%), and SNO members (97%). Most worked in North America (77%) in a research profession (67%). A majority of White respondents reported never experiencing biases (64%), while the majority of non-White respondents reported unconscious biases/microaggressions, followed by a lack of/limited mentorship. Qualitative assessments showcased that personal/professional success metrics were linked to needed improvements in diversity and inclusion efforts within the neuro-oncology community. CONCLUSIONS: The prevalence of racial/ethnic biases and poor mentorship rates amongst underrepresented groups in neuro-oncology is high and potentially linked to the limited diverse representation amongst members and affiliates. These findings warrant a swift implementation of equity and inclusion practices within the neuro-oncology field.

Are tracking recommendations biased? A review of teachers’ role in the creation of inequalities in tracking decisions

Sorting students into hierarchically ordered tracks or streams on the basis of their academic performance (i.e., tracking) is ubiquitous in educational systems, and oftentimes based on teachers’ track recommendations. International surveys indicate that tracking is associated with educational inequalities. To determine if inequalities in tracking may be due to teacher recommendations being biased against students from disadvantaged socio-economic and/or ethnic backgrounds, we conducted a systematic review of 26 recent articles on tracking recommendations and students’ socio-economic or ethnic background. We find that teacher recommendations are biased against students from disadvantaged socio-economic backgrounds, yet evidence with respect to ethnic biases is more mixed. We also conducted an integrative review to examine which factors may account for social and ethnic inequalities in teacher tracking recommendations. We conclude that students’, parents’ and teachers’ attitudes and behaviours play a role in tracking recommendations but cannot fully account for the inequality in these recommendations. We discuss promising areas for future study, and argue that research may want to focus on finding institutional moderators in order to combat biases in educational institutions.

Candidate selection and ethnic minority aspirants: Exploring the effect of party selectors’ biases in a PR system

This study examines whether ethnic majority party selectors’ ethnic biases (e.g. beliefs about the political traits of ethnic minority aspirants) impede the selection of ethnic minority aspirants in a proportional representation system. To this end, a quasi-experiment was conducted among local party chairs in Flanders (Belgium). The participants were asked to evaluate both a hypothetical ethnic minority aspirant and an ethnic majority aspirant. When the participants were negatively biased against the ethnic minority aspirant, the selection chances of the ethnic minority aspirant vis-à-vis the ethnic majority aspirant declined. Additionally, the quasi-experimental data were linked to an analysis of the real-life candidate lists composed by the participants for the 2018 local elections in Flanders. The participants’ ethnic biases could not predict the percentage of ethnic minority candidates on the real-life candidate lists. Instead, the presence of ethnic minority voters, aspirants and co-selectors determined the diversity of the lists.

Limits of a Second Language: Native and Second Languages in Management Team Communication

Cultural differences in speech acts are common challenges in management involving Chinese and Western managers. Comparing four groups – Native-speaking Chinese, English-speaking Chinese, Chinese-speaking Westerners, and non-Chinese- speaking Westerners, we assessed the effects of language and ethnicity on the ability to predict communication obstacles in a management team scenario. Bilingual subjects were less likely to be influenced by ethnic biases. Still, bilinguals were not more likely to adjust their metacognitions about communication toward those of the native speakers. The study creates a link between management, cognition and linguistics, as well as having consequences for the study of metacognition in cross-cultural management.

Breastfeeding Perspectives: Reactions to Breastfeeding Imagery and Social Norms

Objectives Normative behaviors surrounding breastfeeding in public and the community can influence breastfeeding behaviors and inhibit sustained breastfeeding. The goal of this research is to better understand how a diverse group of women perceive images of racially/ethnically similar and different women breastfeeding. The aim is to uncover biases that may impede breastfeeding goals and promotion. Methods An online mixed-methods survey was completed by a sample of 144 mothers with children 2 years of age or younger. The survey included images of women of varying racial/ethnic groups breastfeeding in various settings. Respondents generated word associations for the images and indicated the appropriateness and ease of breastfeeding for each image. Responses were timed to ascertain cognitive dissonance. Supplemental questions examined descriptive, injunctive and group norms, in line with the Theory of Normative Social Behavior. Results A sample that included 42% non-White individuals and 31% Hispanic/Latino individuals, 73% of whom breastfed their most recent child, responded to the survey. On average, respondents took longer to respond to questions about the appropriateness of images of Black and Hispanic women breastfeeding than images of White or Asian women, even though their responses of appropriateness were roughly equal, demonstrating cognitive dissonance in norms and perceptions. Further, with direct comparisons, women were twice as likely to say breastfeeding was easier for an image of a White mother breastfeeding than an image of a Black or Hispanic mother breastfeeding, regardless of respondent racial/ethnic identity. Finally, differences in image appropriateness differed most significantly by the setting (i.e., public vs. private). Conclusions Women are influenced by racial/ethnic biases about breastfeeding ease and appropriateness as well as norms surrounding behaviors of breastfeeding in public settings or in front of others. Future work is needed to dismantle these biases that may influence breastfeeding behaviors and inequitably impact the feeding choice and health of mothers and infants. Funding Sources This research was funded by an internal research grant from Fairfield University.

Building a future through multicultural early childhood music classes in a conflict-affected city

The aim of this exploratory study was to understand the experience of several Palestinian and Jewish parents participating in a culturally mixed early childhood music class, held in a conflict-affected city. Collective music-making is considered to be a powerful way for people to connect and socialize, regardless of their backgrounds, and for them to get to know each other through music. It also has the potential of reducing stereotypes, which is especially needed in a conflict-affected area. Data sources included records of the classes (videos and a field diary), parents’ sharing of audio and video recordings of musical episodes at home through the music class WhatsApp group and semi-structured interviews with three parents. Data were analysed inductively through open coding and three themes emerged: why music and music classes?; the music class experience; and music at home. These themes reflect the reasons for participating in the programme and the contexts in which the children and their parents experience music. Findings show that those families who chose to participate in this programme hold positive attitudes towards the Other and reported neither observable nor vocally expressed prejudice of their children against the Other. The toddlers who participated in the classes have the potential to not develop negative stereotypes towards the Other (Arab or Jewish), since their families do not expose them to such stereotypes. Nevertheless, they are members of societies (East and West Jerusalem) holding systemic ethnic biases, and therefore joint music classes may strengthen positive attitudes.

Social Services in Ethnically Mixed Cities: Street-Level Bureaucracy at the Crossroads of Ethno-National Conflict

There is a paucity of research examining street-level bureaucracy in cities affected by ongoing ethnopolitical conflict. This study addresses this limitation by exploring the work of social workers in the public services of mixed cities in Israel. It shows the interconnection between ambiguous institutional policies, varying workers’ views of the role of social services, and changing discretion patterns. Findings also suggest that episodes of conflict escalation intensify staff ethnic sectarianism, as well as increase workers’ own ethnic biases, which affect the ways in which they act as a liaison between the welfare system and citizens through their use of discretion.

Do Sociotropic Concerns Mask Prejudice? Experimental Evidence on the Sources of Public Opposition to Immigration

Does opposition to immigration mostly stem from prejudice or from sociotropic concerns about broad economic and cultural implications on the nation as a whole? Previous work on immigration preferences cannot answer this question because the two explanations are observationally equivalent when focusing on the attitudes of natives. I analyze a unique survey experiment that asks both natives and immigrants of various origins to evaluate different profiles of visa applicants to the Netherlands. The experiment also assigns an “ingroup treatment”—applications by individuals of the same ethnocultural background as the respondent. Using this rich data, I show that sociotropic concerns are the major source of immigration preferences, while ethnic biases play a moderate role. Remarkably, the ingroup treatment has limited effects on admission. However, bias against specific immigrant groups is detected in preferences of immigrant respondents and of those who sympathize with the far-right Freedom Party.