1/78
Looks like no tags are added yet.
Dataset Bias
Imbalanced datasets (women and people of color are underrepresented); Balancing gender in datasets books diagnostic performance for women and men
Physical Bias (Medical Device Bias)
When the medical device physically produces undesirable performance variation across demographic groups (i.e. hip joint replacements not being sized for women’s frames)
Computational Bias (Medical Device Bias)
When the device calculates results that create undesirable and uneven results from demographic groups. Due to dataset bias (imbalance and underrepresentation in datasets) and algorithm bias (the mathematics of data processing disadvantages certain groups).
Interpretation Bias (Medical Device Bias)
A medical device is subject to biased inference of readings based on previous research. Previous biases in medical system affecting ways the computer was told to interpret. (i.e. interpretation of spirometry data based on assumptions that Black & Asian people have lower lung capacity than White people)
Algorithmic Bias Types
Representation bias, omitted variable bias, aggregation bias, linking bias, label bias
Representation bias (Algorithmic Bias)
Data underrepresents or misrepresents subsets of the population measurement bias - data inaccurately represents the variables
Omitted variable bias (Algorithmic Bias)
Omitted data. For example, the “black box” nature of many algorithms makes it difficult to determine which variables were used and which were omitted.
Aggregation Bias (Algorithmic Bias)
Conclusions are drawn about individuals based on observations about larger groups
Linking bias (Algorithmic Bias)
Correlations that AI draws about particular users based on the characteristics of other users that may not be accurate in a heterogeneous population
Label Bias (Algorithmic Bias)
Biases that arise when the same object is labeled differently based on differing perceptions by the annotator
Stratified Approach
Medical devices designed and tailored to fit different groups (however, categorizing people into different groups creates other moral problems)
One-size-fits-all approach
Allows for a single off-the-shelf available product; assumes a certain end-user and caters to that typical user in the design and testing process (normalization)
Morally Significant Difference (bioethics)
A distinction between two actions, cases, or policies that alters their moral permissibility or weight
Data Transparency Issues in FDA AI/ML approved documents
Insufficient demographic data, age discrimination, sex discrimination, class discrimination, poor reporting on pre and post approval device performance
Formalist legal narrative of AI bias
Assumes discrimination can be resolved through technical corrections rather than structural reform.
EU AI Act
The European Union's law regulates artificial intelligence according to the risks it poses to individuals and society. It establishes requirements for transparency, human oversight, accountability, and the protection of fundamental rights.
3 intersectional dimensions of EU AI Act
Race and ethnicity: AI border technologies can reproduce racial hierarchies by treating people from certain racial or ethnic backgrounds as suspicious, dangerous, or higher-risk.
Gender and sexuality: Systems can discriminate against women and LGBTQ+ migrants when they rely on binary gender categories or heterosexual assumptions. For example, an asylum system may fail to understand someone's sexual orientation or gender identity.
Class and education: Differences in socioeconomic status and education can shape people's access to technology and their ability to navigate digital screening, documentation, and asylum procedures.
Examples of Androcentric bias in medical AI
Medical research: Historically, medical trials have often relied disproportionately on male participants, leaving gaps in knowledge about women's bodies.
Heart disease: AI trained on predominantly male medical data may be less effective at identifying symptoms and patterns of heart disease in women.
Pain assessment: Medical assumptions that women exaggerate or experience pain differently can be incorporated into the data used to train AI, potentially leading to undertreatment.
Drug dosages and treatment: AI models based on male-centered data may fail to account for sex-related differences in how patients respond to medication.
Race and gender together: An algorithm may disadvantage Black women because both racial and gender biases are embedded in its training data.
UChicago learning models of extensive cancer genetics tissue histology data
4 key gender inequality principles
User under-representation (in research), User vulnerability to security breaches, App gender biases (apps can reinforce gender bias or sexism), Absence of Leadership
Fairness agnostic/marginal fairness approaches
Fairness-agnostic approach: An approach that does not explicitly account for demographic fairness when developing or evaluating an algorithm. It may prioritize overall accuracy or efficiency without examining whether certain groups experience worse outcomes.
Marginal fairness: An approach that evaluates fairness across one characteristic, or attribute, at a time.
Intersectional Debiasing
A study found that debiasing using the combination of ethnoracial and gender was associated with better predictive model performance than considering either attribute alone or additively.
Sufficiency vs separation fairness algorithms
Sufficiency: An algorithm satisfies sufficiency when its predictions have the same meaning and reliability across demographic groups.
Separation: An algorithm satisfies separation when its error rates are equal across demographic groups.
4 functional subtypes of AI (Bowleg)
1. Reactive machines | AI that responds to current information but does not retain memories of previous experiences. |
2. Limited memory | AI that uses past information or observations to inform current decisions. |
3. Theory of mind | A proposed type of AI that could understand human emotions, beliefs, intentions, and perspectives. |
4. Self-aware AI | A hypothetical type of AI that possesses consciousness and awareness of its own existence. |
Racial Capitalism
A socio-economic framework asserting that capitalism and racism are co-constitutive; the expansion of capitalist systems has relied on racial hierarchies, divisions, and exploitation.
Fiscal solvency
An individual or a business has enough total assets to cover all of its debts and long-term financial obligations
Outcast surplus communities and population
Human and nonhuman bodies/ecosystems rendered permanently superfluous, unpriced, and expendable relative to capitalist accumulation
Vicarious structural racism
Witnessing the effects of racist structural conditions or practices
Collective trauma
Trauma experienced directly by a population during & after a large-scale disaster where individuals who are not directly exposed to the disaster identify with individuals directly exposed based on common characteristics (e.g., religion, racial/ethnic group, gender, etc.)
Redlining
The systematic, discriminatory practice of categorically denying financial services, mortgages, insurance, healthcare, or other essential resources to residents of specific neighborhoods based on their race or ethnicity, rather than on individual creditworthiness
Environmental exposures that disproportionately impact LGBTQ+ populations
Social, institutional, and structural: discriminatory housing policies, employment, and healthcare access, delayed care due to anticipatory discrimination or inadequate funds, consistently higher unemployment rates (esp. for transgender individuals); disproportionately work in sectors with high levels of occupational hazards, proximity to secondhand smoke, underlying health conditions linked to identity/gender presentation management strategies, ambient air pollution, environmental disasters, secondhand smoke
Downstream vs. Upstream risk factors
Downstream (individual health behaviors) vs Upstream (community/structural)
Food Insecurity (Defined and measured)
Lack of access to nutritious substances due to financial or resource constraints, and measured through nutrition, worries about food running out, meal size, and money for food
Food Deserts
Areas devoid of food options (critique: inaccurately depicts the structurally racist practices that created the lack of healthy food options)
Food Swamps
Areas concentrated with energy-dense, low nutrient foods
Food Apartheids
The structurally racist structures and practices that lead to inequitable food environments
American Heart Association’s “Life’s Essential 8”
Smoking status, Body Mass Index (BMI), Physical activity, Diet, Total cholesterol, Blood pressure, Fasting glucose, Sleep duration (understudied/recently identified risk factor)
Ideal (and modified) cardiovascular health metrics (mICVH)
Smoking status (never smoked or quit smoking > 12 months ago), BMI (within “normal” range), Physical activity (≥ 150–300 min/week moderate or ≥ 75–150 min/week vigorous), Sleep duration (7 to 9 hours/night), Hypertension, Pre-diabetes/type 2 diabetes, Dyslipidemia (imbalance of fat in bloodstream; high LDL “bad cholesterol” or low HDL “good cholesterol”)
Structural heteropatriarchy
Mutually constitutive structures of sexism and homophobia that ensure cisgender males’ superordinate social positions under capitalism; the system of regulation that privileges cisgender men and heterosexuality, which serves to reinforce the dominance of traditional heterosexuality, including women’s subservient role in the household (a social determinant of health)
Racialized heteropatriarchy
The mutually constitutive and co-occurring structural relationship between race-, gender-, and sexuality-related discrimination that privileges White, cisgender, heterosexual men; unfairly allocates power and resources to White, cisgender, heterosexual men and oppresses women, LGBTQ+ people, and/or those racialized as non-White
Anti-LGB structural stigma
The co-occurrence of labeling, stereotyping, separation, status loss, and discrimination in a context in which power is exercised; a fundamental cause of sexual minority health inequities; operationalized through measures of government policies related to LGBT civil rights
Family values panic as a manifestation of structural heteropatriarchy
Example of distal stressor
Proximal vs. distal stressors (minority stress model of stress and health)
Proximal (internal psychological processes and high-level coping mechanisms associated with exposure to discrimination) vs. Distal (discrete external experiences that are directly and/or psychologically taxing; two main forms: interpersonal discrimination & violence victimization, and biopsychosocial processes)
Allostasis
Long-term overactivation elevates baseline levels of physiological homeostasis (Increases distress, reduces the body’s immune capacity, heightens systemic inflammation, linked to cardiometabolic disease risk)
Technosolutionism
The belief that technology can “solve” complex social and structural problems that ultimately prevent us from enacting necessary structural changes to how we deliver care
Race correction as a conduit for the innovation of inequity
The adoption of algorithms and practice guidelines that correct or adjust outputs of diagnostic technologies to presumably equitably distribute access to resources and interventions
Atrial fibrillation and the AF ethnic paradox
Black individuals have a lower incidence and prevalence of atrial fibrillation compared to White individuals, despite having a significantly higher burden of established modifiable cardiovascular risk factors such as hypertension, obesity, diabetes, and left ventricular hypertrophy
Implementation science
Structured guides used to study and support how research findings, clinical guidelines, and evidence-based practices are put into real-world use
Good vs. Bad Death
Death that occurs at the right time and place vs. death that is struggled and fought against
Choice
An act of autonomy and self-determination that is premised on the freedom of the subject
Suicide
An individual act to end one’s life; largely decriminalized in the U.S. in the 1970’s
Medically Assisted Death
A practice involving the assertion and exercise of individual freedom that necessarily involves numerous actors, notably medical practitioners. It implicates broader social relationships as a subject of bioethical debate, biomedical practice, and legal regulation.
Imperative to treat
Renders medical intervention natural and normal; doctors aim to keep people alive at all costs without regard to quality of life
Natural vs unnatural death
Natural death takes place at the end of a life cycle; associated with “good death”
as an organic culmination of life, unnatural death as premature
Enhanced Natural Death
An argument for using technology to approximate an idealized natural death that is painless, fast, with minimal suffering, limited technology, and removed from medical institutions
Census tracts
Small subdivisions of counties that better estimate socioeconomic and community-level factors
7 structural variables comprised in the Yost Index
Median household income, Median house value, Median rent, Percent below 150% of the poverty line, Education index, Percent working class, Percent employed
Types of breast cancer outcomes
Stage of presentation, Surgical treatment status, Survival rates
Breast cancer outcome disparity dimensions
Demographics, Disease presentation disparities, Treatment disparities, Survival Disparities
SOGI data collection
Sexual orientation and gender identity, underrepresented in data collection, lack of culturally competent providers, lack of appropriate screening guidelines, poor data collection standards, gaps in funding access
Known LGBTQ+ oncology diagnosis and outcome disparities
Higher breast cancer risk among SMW (perhaps due to nulliparity), Higher rates of smoking & alcohol use (linked w/ greater lung, colorectal, & breast cancer), SMW have higher rates of obesity (linked w/ breast, colorectal, liver, ovary, & pancreatic cancer risk), Gay, Bi, & MSM report higher rates of prostate & colorectal cancer (less screening), Highest risk of anal cancer among MSM (linked to prevalent high-risk HPV; 40X more likely among HIV-positive MSM), HIV positive MSM highest risk for AIDS related cancers (Kaposi sarcoma & NHL), HIV linked to increased risk of cancer-specific mortality
Structural barriers to LGBTQ+ oncology care
Lower screening rates, More likely to live in poverty, Less likely to have health insurance, Clinicians unaware of socio-culturally specific needs (low cultural competence), SGM people delay seeking care due to anticipatory discrimination, stigma, & microaggressions
NIH SGMRO (central aim, year established, goals, disruptions)
Goal to improve understanding of SGM Health; Est. 2015; Advance rigorous research on SGM population health (extramural & intramural), Foster strategic partnerships & collaborations w/ internal & external stakeholders, Foster a highly skilled and diverse workforce in SGM health research, Encourage SOGI data collection in biomedical research & workforce; was shut down by Trump in January 2025
7 Key Principles of Data Feminism
Examine power, Challenge power, Elevate marginalized voices, Rethink binary systems, Embrace complexity, Make data work for justice, Recognize that data science is never neutral
Six recommendations for improving regulatory reporting of AI/ML devices
Explicit representation standards, Inclusive validation criteria, Transparency requirements, Post-market surveillance for equity, Interdisciplinary review panels, Incentives for diversity and equity
Political Intersectionality
A specific dimension of the broader intersectionality framework—coined by legal scholar Kimberlé Crenshaw—that examines how political movements, agendas, and anti-discrimination strategies often fail people with multiple intersecting identities
Necropolis
This disposability allows both capital and the state to pursue policies and practices that are catastrophic to the planet and its many life forms because much of the cost is borne by “surplus” people and places.
Biopolis
Places where the well-being and reproduction of subjects matter
Living Organ Donation
A healthy person who chooses to donate an organ or a part of an organ to someone in need while they are still alive. Can donate: kidney, liver, tissue, blood, plasma
Deceased Organ Donation
The process of giving organs, corneas, or tissues after a person has died so they can be transplanted into patients with end-stage organ failure. Can donate: kidney, lungs, liver, heart, pancreas, intestines, eyes, tissues
Barriers to donation willingness
Lack of transplantation awareness, religious beliefs and misperceptions, medical distrust, fear of premature death, and fear of racism, religion, concerns about body disfiguration, and income disparities
Associations between attitudes about the organ allocation system and willingness to donate
Greater knowledge about the organ donation process; believing that the allocation system is equitable, perceived trustworthiness of the organ recovery and allocation process all increase willingness to doonate
Groups previously prohibited from donating blood in the US (Temporary or permanent)
Individuals taking PrEP (prevent HIV) or exposed to hepatitis, those who recently gave birth, 3-month deferral after a tattoo or piercing, individuals with “uncontrolled” high blood pressure, and individuals in remission with many types of cancer
Groups currently subject to 3-month abstinence requirements for blood donation eligibility
Gay/MSM Men
Risk of contracting HIV vs HBV from a blood transfusion
Behavior vs. identity-based blood donation regulation policies
Gender diverse populations (TGD)
An umbrella term for the ever-evolving array of labels people may apply when their gender identity, expression, or even perception does not confirm to the norms and stereotypes that people expect, including transgender and nonbinary people
Structural barriers to healthcare access for TGD
Medical provider ignorance in TGD-specific care, non-inclusive organizational policies, static electronic health records, lack of consistent and comprehensive documentation standards, cis-normative hospital policies, financial barriers, and poor health insurance, Medicaid coverage for gender affirming hormone therapy, function test has not been evaluated in TGD populations, Gaps about influence of hormone therapy on transplantation process
Issues with kidney function measurement in TGD patients
Types of pre- and post-operative transplantation risk barriers for TGD patients