Anthropologist's Digest

Tag: doctor

  • The AI Bias Problem: Ethnocultural Discrimination and Asthma

    The AI Bias Problem: Ethnocultural Discrimination and Asthma

    Hello all, and welcome back to another post! This is the third article in a series about folk illness and medical anthropology, but it will go more in-depth on the subject than the previous articles have. Shifting from a focus on Latin American cultures and popular folk illnesses in their cultures, this article will focus on a range of ethnic groups and how they describe asthma symptoms.

    Introduction

    You may be asking yourself, why asthma? Asthmatic symptoms can be induced by methacholine, a drug used in the methacholine challenge test. Working as a non-specific cholinergic agonist, the drug stimulates muscarinic receptors in the lungs to narrow, or “bronchoconstrict”, the airways in a way similar to asthma. In the classic methacholine challenge, a patient’s standard lung function is measured before they are administered increasing doses of methacholine through a nebulizer. The doses stop once the patient’s airways begin to narrow, and if one’s lung functionality drops 20% during the low-dosage phase, the individual likely suffers from asthma due to airway hyperresponsiveness. In a number of the studies I will reference, methacholine is used both in the methacholine challenge test to test for asthma and as a simulator of asthmatic symptoms. Due to its simulatability with this medication, asthma is the perfect sickness to measure cultural differences in illness description.

    Research on Patient Reporting Styles

    A multitude of studies looking into the phenomenon of different reporting styles have been conducted over the years. Firstly, a study conducted by Gemma et al. induced asthma symptoms in a population of White and African American participants in Northern California using the methacholine challenge to study the manner in which they would report symptoms. The results show that African Americans used upper-airway word descriptors such as “tight throat”, “scared-agitated”, “voice tight”, and “itchy throat”, among others. Conversely, White participants tended to use lower airway and chest-wall descriptors such as “deep breath”, “light-headed”, “out of air”, and “hurts to breathe” (Gemma et al.). Before interpreting these results, it is important to note that African Americans required a significantly lower dose of methacholine to achieve a 30% drop in cardiovascular function, meaning that their group could experience asthma with less provocation than other ethnic groups. The disparity in the levels at which the disease is experienced means that other differences, such as how the condition is experienced, could affect reporting styles. Another similar study conducted by the same team (Gemma et al.) examined the effects on a more ethnically diverse population, including Hispanic, Asian-Pacific, White, and African American individuals. The results of this study were similar, showing that African Americans and Asian-Pacific Islanders used upper-airway descriptors, Hispanic individuals used both upper and lower airway descriptors, and Whites used only lower airway descriptors (Gemma et al.). Overall, these studies demonstrate the diverse reporting styles across ethnic groups and the significant differences between these styles.

    Unintended Discrimination

    A quick search on the popular LLM-powered clinical support tool OpenEvidence tells physicians that the top symptoms of asthma are wheezing, cough, and shortness of breath. Problems arise when considering the differences between semantics of illness in the context of artificial intelligence (AI)-powered diagnostic tools. If you are not familiar with new developments in medical AI technology, you can read about them here. In brief, AI-powered medical care is rapidly advancing, but to ensure that new intelligent assistants can serve all patients equally, misdiagnoses must be prevented. I previously conducted a small-scale hypothesis-generating study that revealed semantic differences in illness descriptors across cultural groups, but it did not use an actual illness simulation to collect them.

    The reviewed studies reveal that semantic differences still exist when symptoms are simulated. Training of medical AI models must be thorough as a result; failure to train a model on responses typical to multiple different ethnic groups can result in structural violence in the form of misdiagnoses.

    If African American individuals consult medical AI programs that were trained on the data of white patients and describe their asthma in the typical style of the African American ethnic group, the AI could misinterpret the symptom reports as a different illness. Such an issue has been observed before in AI programs designed to expedite the process of hiring at large companies. A prime example of this is the company Amazon; its AI hiring assistant had to be shut down after it discriminated against women in the hiring process due to training cases being primarily those of males (Chen, np). With a multitude of factors that could each lead to the development of structural discrimination, companies developing AI-powered medical tools must take into deep consideration the culturally bound reporting styles of different ethnocultural groups.

    In conclusion, awareness about ethnocultural divides in patient-provider communications is key. A perfect solution to this heaping issue has not been found, but awareness is the first step to developing a system that treats every individual equally.

    References (MLA)

    Hardie, Gemma, et al. “Ethnic Differences in Methacholine Responsiveness and Word Descriptors in African Americans, Hispanic-Mexican Americans, Asian-Pacific Islanders, and Whites with Mild Asthma.” Journal of Asthma, vol. 47, no. 4, 2010, pp. 388–396. PubMed, https://pubmed.ncbi.nlm.nih.gov/20528591/. Accessed 22 May 2026.

    Hardie, Gemma E., et al. “Ethnic Differences: Word Descriptors Used by African-American and White Asthma Patients during Induced Bronchoconstriction.” Chest, vol. 117, no. 4, 2000, pp. 935–943. PubMed, https://pubmed.ncbi.nlm.nih.gov/10767221/. Accessed 22 May 2026

    Chen, Zhisheng. “Ethics and Discrimination in Artificial Intelligence-Enabled Recruitment Practices.” Humanities and Social Sciences Communications, vol. 10, 2023, article no. 567, Springer Nature, https://doi.org/10.1057/s41599-023-02079-x. Accessed 27 May 2026.

  • Research Review: Semantics of Illness in AI Healthcare

    Research Review: Semantics of Illness in AI Healthcare

    Hello everyone, this is my first post in a while as school got off to a busy start, but I will be posting more frequently in the future, so stay tuned for more anthropology articles.

    I recently participated in a guided research program where I wrote my first research paper. Choosing a topic was slightly hard, but I eventually settled on medical anthropology. I had the most difficult task to deal with next: finding a research question. After looking into many different areas in which medical anthropology could be applied, my mentor and I looked into the development of artificial intelligence (AI) healthcare models and applications of anthropology to them.

    The concept of healthcare AI models is a relatively new one, as the technology needed to produce something effective enough to be worth developing has only just recently become available. Though experimental research could only be performed on existing prototype models (the soonest beta releases are years away), I wanted to look into issues that could arise in a fully functional healthcare model, specifically in regards to cultural differences. These theoretical issues would occur (most likely) primarily by way of cross-cultural differences in semantics or communication style.

    In terms of methodology, I interviewed six participants: a man and a woman each from Walnut Creek, USA; Istanbul, Turkiye; and Develi, Turkiye. The interviews were conducted as series of questions over international encrypted messaging app WhatsApp.

    Image credit: National Cancer Institute

    Semantic Differences

    Fortunately, there exists some research that looks into cultural differences in semantics of illness. Acclaimed Harvard professor Byron J Good’s The Heart of What’s the Matter: Semantics of Illness in Iran examines the Persian folk illness “heart distress” and its perceived causes in Iranian culture. The paper draws on observed semantic networks to trace how people describe their illness, and the result was that Iranians understand their illnesses very differently than Americans understand theirs. While outdated (it was published in 1977), Good’s study provided a strong theoretical basis for me to do more research. In a theoretical healthcare AI model, misdiagnoses could stem from semantic differences, resulting in an unintentionally biased model. This potential harm, should it become an issue, would be classified as structural violence, a form of social harm that prevents groups of people from having their essential needs met (Read more in the paper linked below).

    The first two interview questions were oriented towards semantic differences. The first asked participants to describe symptoms of a common cold to a doctor, and the second asked them to describe symptoms to an AI model. In theory, these two questions would reveal differences across cultures in both how patients interacted with physicians and how they interacted with AI. Provided symptoms were compared in a table (included below), and the style of descriptions (symptom vs biomedical, list vs na

    Cultural Comfort Differences

    The other two questions were less important to the study; these dealt with contextual differences in communication across cultures. The participants were asked to honestly give any sensitive or embarrassing symptoms that they would not feel comfortable sharing with an in-person physician. They were then asked if they would be more comfortable sharing these symptoms with an AI doctor. The purpose of these questions was to evaluate two things. First, if patients felt more comfortable sharing symptoms with a doctor and, consequently, omitted certain symptoms from a report to an AI model, misdiagnosis could occur from a skewed description. Second, if patients felt more comfortable sharing details with an AI model, they would be more likely to share full details and receive an accurate diagnosis as a result.

    Results

    It was hypothesized based on existing theory and assumptions that there would be a good deal of cultural differences in semantics of illness and comfortability interacting with an AI model. The results, however, showed minimal differences across the three cultures (see tables in the paper). There were little to no differences within the country of Turkey, and Turks tended to describe their symptoms and comfortability with in relatively similar ways. The only major difference within Turkey was that rural Turks tended to describe their illness in a symptomatic hybrid difference, while urban Turks tended to describe their illnesses in a list fashion. American and Turkish responses displayed similar levels of variation, as the only major difference between the countries was the symptoms of the common flu that they provided (more detail in the paper).

    There are many potential explanations for the high level of similarity between cultures, but experimental research is needed to verify any one of these explanations. A possible and likely reason that the results of my study were so different from Dr. Good’s is that globalization has made the world a much smaller place since 1977. With the technological revolution putting internet access into the hands of almost all people, information can be spread much more easily, and as a result, cultural mannerisms and ways of interacting bleed into each other more than they used to.

    This study was non-experimental and hypothesis-generating, so no conclusions can be drawn from the results. It does, however, raise interesting questions about how important cultural differences really are to the development of healthcare AI models.

    I would like to thank the team at Lumiere education for providing excellent services and making the experience of writing a paper much easier than it would have been if I had attempted to do it myself. I highly recommend them to anyone who needs guidance in writing and publishing their first paper.

    Read the paper here