Demographic Inaccuracies and Biases in the Depiction of Patients by Artificial Intelligence Text-to-Image Generators

doi:10.21203/rs.3.rs-4523331/v1

Download PDF

Article

Demographic Inaccuracies and Biases in the Depiction of Patients by Artificial Intelligence Text-to-Image Generators

https://doi.org/10.21203/rs.3.rs-4523331/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

The wide usage of artificial intelligence (AI) text-to-image generators raises concerns about the role of AI in amplifying misconceptions in healthcare. This study therefore evaluated the demographic accuracy and potential biases in the depiction of patients by two commonly used text-to-image generators. A total of 4,580 images of patients with 29 different diseases was generated using the Bing Image Generator and Meta Imagine. Eight independent raters determined the sex, age, weight group, and race and ethnicity of the patients depicted. Comparison to the real-world epidemiology showed that the generated images failed to depict demographical characteristics such as sex, age, and race and ethnicity accurately. In addition, we observed an over-representation of White as well as normal weight individuals. Inaccuracies and biases may stem from non-representative and non-specific training data as well as insufficient or misdirected bias mitigation strategies. In consequence, new strategies to counteract such inaccuracies and biases are needed.

Health sciences/Medical research/Epidemiology

Health sciences/Health care/Medical ethics

Generative AI

deep learning

transformers

generative adversarial networks

diffusion models

Generative artificial intelligence (AI) aims to create synthetic data that is indistinguishable from real data¹. The field is experiencing rapid advancements due to the development of new powerful algorithms, increasing computational power and availability of training data. One example of generative AI algorithms are text-to-image generators that can create synthetic images based on human text commands². These generators use algorithms from the field of natural language processing such as transformers to understand the text command. Next, they apply algorithms from the field of computer vision such as generative adversarial networks or diffusion models to create images based on the command.

Today, the images generated by publicly available text-to-image generators are often photorealistic and hardly distinguishable from real images. In addition, most generators are free to use, the commands to generate the images can be adapted to fit any need of the user, the copyright typically belongs to the individual creating the images, and no consent of the person depicted is required. The visual quality of the images together with the convenience of the image generation have led to increased popularity of these algorithms with millions of images being created every day. In the medical context, artificially created images of patients are being used in scientific and non-scientific publications and for teaching purposes (e.g., in presentation slides, reading material)^3–7. Furthermore, as medical data is scarce, the images are being used to augment data sets to train other AI algorithms^6,8,9. For example, AI is being used to obtain diagnoses from photos of patient faces¹⁰.

This increasing number of users and use cases bring risks and challenges. Especially when used in the medical context, it is not sufficient that images are photorealistic, they also need to be accurate. For example, when depicting patients with certain diseases, the pictures should accurately present important features of the disease including fundamental epidemiological characteristics. Although some diseases display typical phenotypes that could be reflected in facial images (e.g., a flattened nose and epicanthus in Down’s syndrome), disease-specific facial features are lacking for numerous conditions. On the other hand, most diseases predominantly occur in specific age ranges, and/or a specific sex, and/or specific races and ethnicities. Accordingly, a first step for generative AI should be the accurate representation of the epidemiological characteristics of diseases in generated images. Moreover there is vast literature on the impact of unconscious biases related to epidemiological characteristics such as age, sex, race/ethnicity, or body weight on medical decision making¹¹. AI should help reduce these biases rather than potentially amplifying them.

Yet, recent research has shown that also AI algorithms are susceptible to biases based on sex and gender as well as race and ethnicity^12–14. Especially females and people of color were shown to be under-represented in established training datasets¹². In consequence AI algorithms tend to misclassify and misinterpret such images^12–14. In addition, some diseases such as infectious^15,16, psychiatric^17,18, or internal medicine diseases¹⁹ carry disease-related stigmas with profound effects on patients. Stigmas reduce the likelihood of individuals seeking help and adhering to therapy regimens, they reduce treatment quality by medical personnel, and increase societal risk factors of patients^17,18. Thus, replication or amplification of biases and stigmas by generative AI models may have particularly adverse effects.

In this study, we evaluate the accuracy of the representation of disease-specific demographic characteristics (not key phenotypic features) of patient populations depicted by two commonly used text-to-image generators. More specifically, we use the Bing Image Generator from Microsoft and Imagine from Meta to create images of patients with 29 diseases. Among these are 14 diseases with distinct epidemiology (e.g., occurrence in only a specific age or sex), and 15 stigmatized diseases. Further, we analyze potential biases, with a focus on sex, and race and ethnicity of the depicted individuals.

A total of 4,580 images were generated: 2,320 with each of the two text-to-image generators, and 160 for each of the 29 diseases (Fig. 1). The sole exception was substance use disorder, for which only 20 images could be generated in Bing likely due to a software update. For exemplary AI-generated images and ratings see Fig. 2.

Ratings and Inter-Rater Reliability

For each image, two raters determined the following demographic characteristics:

Sex [female (F), male (M)];
Age [child (0–19 years), adult (20–60 years), elderly (> 60 years)] as suggested by the World Health Organization (WHO)²⁰ and the United Nations (UN)²¹;
A combined rating of “race/ethnicity” [Asian, Black or African American (BAA), Hispanic or Latino (HL), Native Hawaiian or Other Pacific Islander (NHPI), American Indian or Alaska Native (AIAN), White] as suggested by the United States Census Bureau²² and National Institutes of Health (NIH)²³;
Weight [underweight, normal weight, overweight].

The inter-rater reliability (IRR) for sex was κ = 0.965 (95%-CI = 0.958–0.973); for age κ = 0.910 (95%-CI = 0.896–0.924); for race/ethnicity κ = 0.938 (95%-CI = 0.930–0.947); and for weight κ = 0.836 (95%-CI = 0.808–0.864).

Accuracy of the Representation of Disease-Specific Demographic Characteristics

Across the 29 diseases, the representation of disease-specific demographic characteristics in the patient images was often inaccurate for both Bing and Meta (Fig. 1, Fig. 3). Accurate representation of age, sex and race/ethnicity was achieved only once, in images of patients with multiple sclerosis by Meta. In four diseases both Bing and Meta accurately depicted the fundamental demographic characteristics as defined in Fig. 1. These were medulloblastoma, where predominantly children were shown, Alzheimer’s disease, where predominantly elderly individuals were shown, as well as melanoma and multiple sclerosis, where race/ethnicity were accurate. For other diseases, demographic characteristics were depicted markedly wrong, e.g., for prostate cancer (only affecting males) or eclampsia (only affecting females), for which Meta depicted both female and male patients. On the other hand, Meta showed better accuracy in representing the race/ethnicity (Fig. 3). The incidence of the disease did not have a clear effect on the accuracy, e.g., images of patients with the more common disease pyloric stenosis did not show better accuracy than images of the rare disease medulloblastoma.

General Biases

Among all diseases, there was an over-representation of White individuals (Bing: 68%, Meta: 28%, pooled real-world patient data: 20%; Fig. 4A). There was no substantial over-representation of Asian, BAA, HL, NHPI, or AIAN in the 15 stigmatized diseases (Fig. 4B). Moreover, among all diseases, there was an over-representation of normal weight individuals (Bing: 88%, Meta: 93%, general population²⁴: 63%). Conversely, there was an under-representation especially of overweight individuals (Bing: 5%, Meta: 4%, general population²⁴: 32%). We did not observe a substantial over-representation of male sex (Bing: 55%, Meta: 48%, pooled real-world patient data: 52%).

Sex Differences in Stigmatized Diseases

Among the 15 stigmatized diseases, there were sex differences in age and weight. In both text-to-image generators, females were younger than males. More precisely, in images by Bing, females were more often children and less often elderly (F(1,1136) = 23.878, p < 0.001). In images by Meta, females were more often adults and less often elderly (F(1,1195) = 19.872, p < 0.001).

In both generators, depicted females were rated as having higher weight than males. More precisely, in images by Bing, females were less often underweight and more often normal weight (F(1,1135) = 4.949, p = 0.026). In images by Meta, females were less often normal weight and more often overweight (F(1,1194) = 8.082, p = 0.005). Notably, when analyzing all 29 diseases together, there were no sex differences in age or weight in Bing (for details see Supplementary Materials).

Racial/Ethnical Differences in Stigmatized Diseases

Among the 15 commonly stigmatized diseases, there were racial/ethnical differences in age and weight. In images by Bing, White individuals were less often children or elderly and more often adults (F(1,1136) = 4,810, p = 0.029). In images by Meta, White individuals were less often adults and more often elderly (F(1,1195) = 50.692, p < 0.001).

Further, in images by both Bing and Meta, Asian, BAA, HL, NHPI, and AIAN individuals combined were rated as having more weight than White individuals. More precisely, Asian, BAA, HL, NHPI, and AIAN individuals were depicted as less underweight, less normal weight, and more overweight (Bing: F(1,1135) = 27.083, p < 0.001; Meta: F(1,1194) = 4.646, p = 0.031).

In summary, we found that the images of patients created by the Bing Image Generator and Meta Imagine often did not accurately represent the disease-specific demographic characteristics. In addition, we observed an over-representation of White as well as normal weight individuals across all analyzed diseases. Female individuals as well as Asian, BAA, HL, NHPI, and AIAN individuals were depicted as being more overweight compared to male and White individuals, respectively. Such inaccuracies raise concern about the role of AI in amplifying misconceptions in healthcare¹⁴ given its large numbers of users and use cases^3–9.

We found that images by both Bing and Meta displayed a broad range of demographic inaccuracies. This was most striking for Meta’s depictions of patients with prostate cancer, hemophilia B, premenstrual syndrome, and eclampsia, for which both female and male individuals were shown. Likewise, images by Bing often displayed substantial inaccuracies, especially regarding the races/ethnicities that were accurately depicted for only two diseases.

Presumably, these inaccuracies are caused by the composition of the training data of the generative AI models. They are typically trained on large non-medical datasets consisting of publicly available images from the internet, databases such as ImageNet, Common Objects in Context, and other sources⁶. Such large datasets are necessary to produce photorealistic images. However, as they do not contain large numbers of images of actual patients, information on disease-specific demographic characteristics as well as important risk factors are missing. Thus, their ability to generate accurate images of these patients and their diseases is limited.

Another factor influencing the quality of the output are bias mitigation strategies in the code of the algorithms that, for example, aim to counteract known biases in the training data. These bias mitigation strategies can result in an over-correction of biases as was shown previously²⁵. Thus, it can also be speculated that the depictions of both female and male patients in the images of sex-specific diseases by Meta were influenced by such code-based adaptions. In fact, there was no over-representation of any sex in the images of the two text-to-image generators. This may be interpreted as a positive sign as sex/gender biases have been a common phenomenon in generative AI algorithms^12,14. On the other hand, achieving accurate demographic representation by applying bias mitigation strategies seems challenging and representative training data may be necessary.

Moreover, we also found examples of insufficient representation in our data. We detected a bias toward White individuals in both generators. Interestingly, this bias was stronger in Bing than in Meta, which may be another sign of stricter bias mitigation in Meta compared to Bing. A similar over-representation of White individuals has previously been reported in a study on AI-generated images of healthcare professionals¹⁴. In addition, we detected a bias towards normal weight in both generators. Conversely, we found that especially individuals with overweight were under-represented which may be caused by a similar under-representation in the training data. However, these results need to be interpreted cautiously as the images did not depict the entire body.

In both text-to-image generators, the depicted females were younger than the males, which may represent a bias of the two algorithms, potentially veering towards gender stereotypes. However, the real-world epidemiology is complex. While females generally have a higher life expectancy than males²⁶, there are studies suggesting an earlier onset in females in some of the diseases included in our analyses, e.g., depression²⁷ or COVID-19²⁸. However, there are also studies suggesting an earlier symptom onset in males in other diseases, e.g., schizophrenia²⁹ or diabetes type 2³⁰; or no known sex differences, e.g., in multiple sclerosis³¹ or malaria³². There is additional research on sex differences in the age of diagnosis in contrast to the age of disease onset²⁹. Taken together, no conclusive interpretation of these findings is possible.

In addition, females were depicted as having higher weight than males. Among the general population, females tend to have a slightly higher BMI³³ and body-fat percentage³⁴ than males. If this is also the case for the diseases included in this article has yet to be investigated. Further, Asian, BAA, HL, NHPI, and AIAN individuals combined showed more weight than White individuals, which is inaccurate despite global shifts in the distribution of under- and overweight^33,35. Thus, not only did our data reveal an under-representation of Asian, BAA, HL, NHPI, and AIAN individuals combined but also a tendency to portray them as more overweight compared to White individuals.

AI-generated images are not yet ready to be used in the medical context without caution. Instead, they should be carefully evaluated for accuracy and potential biases. Such biases include but are not limited to over-representation of White individuals, male sex (although not observed in our sample), and normal weight. Importantly, inaccuracies and biases can be manually corrected, e.g., by selecting specific images so that the final sample has a fitting distribution of the most important demographic characteristics. However, this requires careful research on these characteristics in the real world. Given our findings, scientific and non-scientific publications should highlight if patient images are real or AI-generated.

There are limitations to our study. Firstly, the rating process is inherently limited. One can only approximate demographic characteristics from images. For example, race and ethnicity are aspects of a person’s identity that we could only estimate based on features such as skin color and facial characteristics. Moreover, all raters identify as White which may have led to biased ratings despite all measures to ensure accurate ratings. It is also difficult to estimate the weight category just from pictures of faces. Biological sex can only be determined by chromosomal analysis and ratings do not reflect gender identity. Secondly, our comparisons to real-world epidemiological data are limited by the availability and quality of the real-world epidemiological data itself. Thirdly, the field of generative AI is rapidly evolving. Thus, our findings are only a snapshot of the features and capabilities of these algorithms in February 2024. Fourthly, there are differences between the outputs of the Bing and Meta algorithms. Thus, our findings may not extend to other text-to-image generators.

Taken together, images of patients generated by two commonly used text-to-image generators did not accurately display fundamental demographic characteristics such as sex, age, and race/ethnicity. In addition, we observed an over-representation of White as well as normal weight individuals. In consequence, the use of AI-generated patient images requires caution and future software models should focus on ensuring adequate demographic representation of patient groups across the world.

Text-to-Image Generators

We used the latest versions of the two common text-to-image generators Bing Image Generator from Microsoft (bing.com/images/create) which is based on the DALL-E3 algorithm by OpenAI, and Imagine from Meta (imagine.meta.com). The following text prompt was used to generate images of patients: “Photo of the face of a patient with [disease]”. The blank was filled with the name of the specific diseases (Fig. 1). We chose 14 diseases with distinct epidemiological characteristics such as occurrence in a specific age range, sex, and race or ethnicity. For example, we chose the diseases pyloric stenosis and medulloblastoma as they predominantly occur during childhood^36,37. These were used to analyze the accuracy of the depiction of the demographical characteristics specific to each of the 14 diseases. In addition, we created images of patients with five infectious, five psychiatric, and five internal medicine diseases that are stigmatized^15–19 to investigate biases among these.

All images were created in February 2024. We used eight computers, four internet browsers, and 16 accounts to minimize the influence of user data on the image generation. We generated 80 images for each of the 29 diseases in each Bing and Meta. Importantly, we were only able to generate 20 images of patients with substance use disorder in Bing. This was likely due to a sudden software update prohibiting the generation of additional images of individuals with substance use disorders.

Ratings

Determining demographic characteristics from images of faces is challenging. We thus took several measures to standardize ratings and to reduce subjectivity: First, the ratings were performed by eight M.D. Ph.D. researchers (T.L.T.W., L.S.S., P.Z.M., J.F.R., L.I.V., J.A.G., L.K., and L.B.J; 3 female, 5 male). Second, the raters adhered to the established multiracial Chicago face dataset which includes standardized images and descriptions of faces³⁸. Third, a separate practice data set with images from both text-to-image generators was created and all ratings were performed and discussed in the entire group of raters in accordance with the Chicago face dataset. Fourth, each of the images was rated by two raters, independently. In case of disagreement between the ratings, a third rater was included, and the final rating achieved by discussion and majority voting.

Wherever possible, real-world epidemiological data was obtained from official sources such as the WHO or large-scale epidemiological reviews such as global burden of disease studies. If such sources were not available, other epidemiological publications were used (see references in Fig. 1).

Statistical Analyses

Firstly, we calculated the IRR for each variable based on the ratings by raters 1 and 2 using Cohen’s κ.

Secondly, we analyzed the accuracy of the representation of disease-specific demographic characteristics in the generated patient images. Here, for each disease and text-to-image generator we compared the age, sex, and race/ethnicity combined to the real-world epidemiology. We evaluated whether the patients’ age, sex, and race/ethnicity as depicted in the images were “accurate” in comparison to the real-world epidemiological data (green background in Fig. 1), “imprecise” (yellow background), or “wrong” (red background).

Age was rated as “accurate” if both the most common age group as well as the age distribution matched the real-world data, as “imprecise” if only one of the two matched, and as “wrong” if none of the two matched.

Sex was rated as “accurate” if the F:M ratio was less than factor 1.50 different to the real world, as “imprecise” if there was a difference of factor 1.50-3.00, or “wrong” if the difference was larger than factor 3.00. For example, a ratio of 60 F:40 M in the generated images and 45 F:55 M in the real world would correspond to a factor of (60/40)/(45/55) = 1.83 (“imprecise”).

The race/ethnicity was rated as “accurate” if both the most common race/ethnicity as well as the distribution matched the real-world data, as “imprecise” if only one of the two matched, and as “wrong” if none of the two matched (for details see Supplementary Materials).

Thirdly, we compared the real-world data and the image ratings among all diseases combined to analyze more general biases such as an over-representation of male and White individuals as reported previously¹⁴.

Fourthly, we investigated biases regarding sex and race/ethnicity in the images of patients with stigmatized diseases. We used analyses of covariance (ANCOVA) to identify sex differences as well as racial/ethnical differences in weight and age. Based on the literature, we expected biases especially in the depiction of White individuals in comparison to people of color¹⁴. Thus, we dichotomized the race/ethnicity variable into White vs. Asian or BAA or HL or NHPI or AIAN. Analyses on sex differences were controlled for the effects of the disease depicted, race/ethnicity, and age (not in analyses on sex differences in age). Analyses on racial/ethnical differences were controlled for the effects of the disease depicted, sex, and age (not in analyses on racial/ethnical differences in age). The analyses were performed in IBM SPSS Statistics version 29.0.2.0. P-levels < 0.05 were considered statistically significant.

DATA AVAILABILITY

The 4,580 generated patient images are available upon request directed at the corresponding author.

CODE AVAILABILITY

The text-to-image generators are available online: Bing Image Generator from Microsoft (bing.com/images/create) and Imagine from Meta (imagine.meta.com).

CONFLICT OF INTEREST

T.L.T.W. and L.I.V. receive royalties for books published by ELSEVIER. I.K.K. receives funding for a collaborative project from Abbott Inc. She receives royalties for book chapters. Her spouse is an employee at Siemens AG and stockholder of Siemens AG and Siemens Healthineers. The remaining authors declare no competing interests.

Funding:

None.

AUTHOR CONTRIBUTIONS

Idea and study design: T.L.T.W., and L.B.J.

Reddy, S. Generative AI in healthcare: an implementation science informed translational path on application, integration and governance. Implementation Science 19, 27 (2024).
Ramzan, S., Iqbal, M.M. & Kalsum, T. Text-to-Image Generation Using Deep Learning. Engineering Proceedings 20, 16 (2022).
Noel, G. Evaluating AI-powered text-to-image generators for anatomical illustration: A comparative study. Anat Sci Educ (2023).
Kumar, A., Burr, P. & Young, T.M. Using AI Text-to-Image Generation to Create Novel Illustrations for Medical Education: Current Limitations as Illustrated by Hypothyroidism and Horner Syndrome. JMIR Med Educ 10, e52155 (2024).
Koljonen, V. What could we make of AI in plastic surgery education. J Plast Reconstr Aesthet Surg 81, 94–96 (2023).
Fan, B.E., Chow, M. & Winkler, S. Artificial Intelligence-Generated Facial Images for Medical Education. Med Sci Educ 34, 5–7 (2024).
Reed, J.M. Using Generative AI to Produce Images for Nursing Education. Nurse Educ 48, 246 (2023).
Koohi-Moghadam, M. & Bae, K.T. Generative AI in Medical Imaging: Applications, Challenges, and Ethics. J Med Syst 47, 94 (2023).
Adams, L.C., et al. What Does DALL-E 2 Know About Radiology? J Med Internet Res 25, e43110 (2023).
Rokhshad, R., Keyhan, S.O. & Yousefi, P. Artificial intelligence applications and ethical challenges in oral and maxillo-facial cosmetic surgery: a narrative review. Maxillofac Plast Reconstr Surg 45, 14 (2023).
Meidert, U., Dönnges, G., Bucher, T., Wieber, F. & Gerber-Grote, A. Unconscious Bias among Health Professionals: A Scoping Review. Int J Environ Res Public Health 20(2023).
Buolamwini, J. & Gebru, T. Gender shades: Intersectional accuracy disparities in commercial gender classification. in Conference on fairness, accountability and transparency 77–91 (PMLR, 2018).
Bianchi, F., et al. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. in Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency 1493–1504 (2023).
Ali, R., et al. Demographic Representation in 3 Leading Artificial Intelligence Text-to-Image Generators. JAMA Surg 159, 87–95 (2024).
Saeed, F., et al. A Narrative Review of Stigma Related to Infectious Disease Outbreaks: What Can Be Learned in the Face of the Covid-19 Pandemic? Front Psychiatry 11, 565919 (2020).
Mak, W.W., et al. Comparative stigma of HIV/AIDS, SARS, and tuberculosis in Hong Kong. Soc Sci Med 63, 1912–1922 (2006).
Committee on the Science of Changing Behavioral Health Social, N., et al. in Ending Discrimination Against People with Mental and Substance Use Disorders: The Evidence for Stigma Change (National Academies Press (US). Copyright 2016 by the National Academy of Sciences. All rights reserved., Washington (DC), 2016).
Wood, L., Birtel, M., Alsawy, S., Pyle, M. & Morrison, A. Public perceptions of stigma towards people with schizophrenia, depression, and anxiety. Psychiatry research 220, 604–608 (2014).
Wahlin, S. & Andersson, J. Liver health literacy and social stigma of liver disease: A general population e-survey. Clin Res Hepatol Gastroenterol 45, 101750 (2021).
WHO. Definition of Key Terms. in Consolidated Guidelines on the Use of Antiretroviral Drugs for Treating and Preventing HIV Infection: Recommendations for a Public Health Approach. 2nd edition (2016).
UN. World Population Ageing 2019, (2019).
Jensen, E., et al. Measuring Racial and Ethnic Diversity for the 2020 Census. (United States Census Bureau, 2021).
Lewis, C., Cohen, P.R., Bahl, D., Levine, E.M. & Khaliq, W. Race and Ethnic Categories: A Brief Review of Global Terms and Nomenclature. Cureus 15, e41253 (2023).
WHO. Malnutrition. (2024).
Alba, D., Love, J., Ghaffary, S. & Metz, R. Google Left in ‘Terrible Bind’ by Pulling AI Feature After Right-Wing Backlash. (TIME, 2024).
Global age-sex-specific fertility, mortality, healthy life expectancy (HALE), and population estimates in 204 countries and territories, 1950–2019: a comprehensive demographic analysis for the Global Burden of Disease Study 2019. Lancet 396, 1160–1203 (2020).
Smith, D.J., et al. Differences in depressive symptom profile between males and females. Journal of Affective Disorders 108, 279–284 (2008).
Kharroubi, S.A. & Diab-El-Harake, M. Sex-differences in COVID-19 diagnosis, risk factors and disease comorbidities: A large US-based cohort study. Front Public Health 10, 1029190 (2022).
Solmi, M., et al. Age at onset of mental disorders worldwide: large-scale meta-analysis of 192 epidemiological studies. Mol Psychiatry 27, 281–295 (2022).
Wright, A.K., et al. Age-, sex- and ethnicity-related differences in body weight, blood pressure, HbA(1c) and lipid levels at the diagnosis of type 2 diabetes relative to people without diabetes. Diabetologia 63, 1542–1553 (2020).
Romero-Pinel, L., et al. The age at onset of relapsing-remitting multiple sclerosis has increased over the last five decades. Mult Scler Relat Disord 68, 104103 (2022).
Paintsil, E.K., Omari-Sasu, A.Y., Addo, M.G. & Boateng, M.A. Analysis of Haematological Parameters as Predictors of Malaria Infection Using a Logistic Regression Model: A Case Study of a Hospital in the Ashanti Region of Ghana. Malar Res Treat 2019, 1486370 (2019).
Trends in adult body-mass index in 200 countries from 1975 to 2014: a pooled analysis of 1698 population-based measurement studies with 19·2 million participants. Lancet 387, 1377–1396 (2016).
Karastergiou, K., Smith, S.R., Greenberg, A.S. & Fried, S.K. Sex differences in human adipose tissues - the biology of pear shape. Biol Sex Differ 3, 13 (2012).
Haslam, D.W. & James, W.P.T. Obesity. The Lancet 366, 1197–1209 (2005).
Ma, D.S., Correll, J. & Wittenbrink, B. The Chicago face database: A free stimulus set of faces and norming data. Behav Res Methods 47, 1122–1135 (2015).
Li, Z.Z., et al. Global, regional, and national burden of gallbladder and biliary diseases from 1990 to 2019. World J Gastrointest Surg 15, 2564–2578 (2023).
Banerjee, P., Jain, A., Kumar, U. & Senapati, S. Epidemiology and genetics of granulomatosis with polyangiitis. Rheumatol Int 41, 2069–2089 (2021).
Mayeux, R. & Stern, Y. Epidemiology of Alzheimer disease. Cold Spring Harb Perspect Med 2(2012).
Schmidt, R., et al. [Sex differences in Alzheimer's disease]. Neuropsychiatr 22, 1–15 (2008).
Li, X., et al. Global, regional, and national burden of Alzheimer's disease and other dementias, 1990–2019. Front Aging Neurosci 14, 937486 (2022).
Padala, S.A., et al. Epidemiology, Staging, and Management of Multiple Myeloma. Med Sci (Basel) 9(2021).
Rawla, P. Epidemiology of Prostate Cancer. World J Oncol 10, 63–89 (2019).
Global, regional, and national incidence, prevalence, and years lived with disability for 354 diseases and injuries for 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet 392, 1789–1858 (2018).
McHugh, J., et al. Prostate cancer risk in men of differing genetic ancestry and approaches to disease screening and management in these groups. Br J Cancer 126, 1366–1373 (2022).
Hantsoo, L., et al. Premenstrual symptoms across the lifespan in an international sample: data from a mobile application. Arch Womens Ment Health 25, 903–910 (2022).
Zhu, L., et al. Global burden and trends in female premenstrual syndrome study during 1990–2019. Arch Womens Ment Health (2024).
Abalos, E., Cuesta, C., Grosso, A.L., Chou, D. & Say, L. Global and regional estimates of preeclampsia and eclampsia: a systematic review. Eur J Obstet Gynecol Reprod Biol 170, 1–7 (2013).
Saginala, K., Barsouk, A., Aluru, J.S., Rawla, P. & Barsouk, A. Epidemiology of Melanoma. Med Sci (Basel) 9(2021).
Morgese, F., et al. Gender Differences and Outcomes in Melanoma Patients. Oncol Ther 8, 103–114 (2020).
Walton, C., et al. Rising prevalence of multiple sclerosis worldwide: Insights from the Atlas of MS, third edition. Mult Scler 26, 1816–1821 (2020).
Global, regional, and national burden of multiple sclerosis 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol 18, 269–285 (2019).
Shi, D., et al. Trends of the Global, Regional and National Incidence, Mortality, and Disability-Adjusted Life Years of Malaria, 1990–2019: An Analysis of the Global Burden of Disease Study 2019. Risk Manag Healthc Policy 16, 1187–1201 (2023).
Global, regional, and national prevalence and mortality burden of sickle cell disease, 2000–2021: a systematic analysis from the Global Burden of Disease Study 2021. Lancet Haematol 10, e585-e599 (2023).
Kato, G.J., et al. Sickle cell disease. Nat Rev Dis Primers 4, 18010 (2018).
Mody, A., et al. HIV epidemiology, prevention, treatment, and implementation strategies for public health. Lancet 403, 471–492 (2024).
UNAIDS. Global HIV & AIDS statistics — Fact sheet. (2022).
Abdool Karim, S.S., Abdool Karim, Q., Gouws, E. & Baxter, C. Global Epidemiology of HIV-AIDS. Infectious Disease Clinics of North America 21, 1–17 (2007).
Glaziou, P., Floyd, K. & Raviglione, M.C. Global Epidemiology of Tuberculosis. Semin Respir Crit Care Med 39, 271–285 (2018).
WHO. Global Tuberculosis Report 2023. (2023).
Global, regional, and national burden of hepatitis B, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Gastroenterol Hepatol 7, 796–829 (2022).
Brown, R., Goulder, P. & Matthews, P.C. Sexual Dimorphism in Chronic Hepatitis B Virus (HBV) Infection: Evidence to Inform Elimination Efforts. Wellcome Open Res 7, 32 (2022).
Tao, Y.T., et al. Global, regional, and national trends of syphilis from 1990 to 2019: the 2019 global burden of disease study. BMC Public Health 23, 754 (2023).
Chen, T., et al. Evaluating the global, regional, and national impact of syphilis: results from the global burden of disease study 2019. Sci Rep 13, 11386 (2023).
WHO. COVID-19 epidemiological update – 16 February 2024. (2024).
CDC. COVID-19 Stats: COVID-19 Incidence,* by Age Group† — United States, March 1–November 14, 2020§. (2021).
WHO. WHO COVID-19 dashboard. (2024).
Liu, Q., et al. Changes in the global burden of depression from 1990 to 2017: Findings from the Global Burden of Disease study. Journal of Psychiatric Research 126, 134–140 (2020).
Labaka, A., Goñi-Balentziaga, O., Lebeña, A. & Pérez-Tejada, J. Biological Sex Differences in Depression: A Systematic Review. Biol Res Nurs 20, 383–392 (2018).
McHugh, R.K., Votaw, V.R., Sugarman, D.E. & Greenfield, S.F. Sex and gender differences in substance use disorders. Clin Psychol Rev 66, 12–23 (2018).
Degenhardt, L., Stockings, E., Patton, G., Hall, W.D. & Lynskey, M. The increasing global health priority of substance use in young people. Lancet Psychiatry 3, 251–264 (2016).
Simha, A., et al. Effect of national cultural dimensions and consumption rates on stigma toward alcohol and substance use disorders. Int J Soc Psychiatry 68, 1411–1417 (2022).
Javaid, S.F., et al. Epidemiology of anxiety disorders: global burden and sociodemographic associations. Middle East Current Psychiatry 30, 44 (2023).
Solmi, M., et al. Incidence, prevalence, and global burden of schizophrenia - data, with critical appraisal, from the Global Burden of Disease (GBD) 2019. Mol Psychiatry (2023).
Cortese, S., et al. Incidence, prevalence, and global burden of ADHD from 1990 to 2019 across 204 countries: data, with critical re-analysis, from the Global Burden of Disease study. Mol Psychiatry 28, 4823–4830 (2023).
Sørensen, T.I.A., Martinez, A.R. & Jørgensen, T.S.H. Epidemiology of Obesity. Handb Exp Pharmacol 274, 3–27 (2022).
Blüher, M. Obesity: global epidemiology and pathogenesis. Nat Rev Endocrinol 15, 288–298 (2019).
Dai, H., et al. Global, regional, and national burden of ischaemic heart disease and its attributable risk factors, 1990–2017: results from the Global Burden of Disease Study 2017. Eur Heart J Qual Care Clin Outcomes 8, 50–60 (2022).
Khan, M.A.B., et al. Epidemiology of Type 2 Diabetes - Global Burden of Disease and Forecasted Trends. J Epidemiol Glob Health 10, 107–111 (2020).
Zhou, B., et al. Worldwide burden and epidemiological trends of tracheal, bronchus, and lung cancer: A population-based study. EBioMedicine 78, 103951 (2022).
The global, regional, and national burden of cirrhosis by cause in 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet Gastroenterol Hepatol 5, 245–266 (2020).

There is a conflict of interest T.L.T.W. and L.I.V. receive royalties for books published by ELSEVIER. I.K.K. receives funding for a collaborative project from Abbott Inc. She receives royalties for book chapters. Her spouse is an employee at Siemens AG and stockholder of Siemens AG and Siemens Healthineers. The remaining authors declare no competing interests.

Supplements.docx

Download PDF

Editorial decision: revise
23 Sep, 2024
Reviewer #3 agreed at journal
10 Sep, 2024
Review #2 received at journal
04 Sep, 2024
Reviewer #2 agreed at journal
28 Aug, 2024
Review #1 received at journal
20 Aug, 2024
Reviewer #1 agreed at journal
06 Aug, 2024
Reviewers invited by journal
01 Jul, 2024
Editor assigned by journal
25 Jun, 2024
Submission checks completed at journal
25 Jun, 2024
First submitted to journal
21 Jun, 2024
Unknown event
06 Jun, 2024

You are reading this latest preprint version

Demographic Inaccuracies and Biases in the Depiction of Patients by Artificial Intelligence Text-to-Image Generators

Status:

Version 1

Abstract

Figures

INTRODUCTION

RESULTS

Ratings and Inter-Rater Reliability

Accuracy of the Representation of Disease-Specific Demographic Characteristics

General Biases

Sex Differences in Stigmatized Diseases

Racial/Ethnical Differences in Stigmatized Diseases

DISCUSSION

METHODS

Text-to-Image Generators

Ratings

Statistical Analyses

Declarations

DATA AVAILABILITY

CODE AVAILABILITY

CONFLICT OF INTEREST

Funding:

AUTHOR CONTRIBUTIONS

References

Additional Declarations

Supplementary Files

Status:

Version 1