Transparency and equity in the implementation of AI in healthcare: the challenge of adequate representation

In the growing endeavor to integrate Artificial Intelligence (AI) models into clinical practice, a significant concern emerges in the scientific community: addressing biases is of paramount importance in a field where the guiding principle is "first, do no harm." The lack of diversity and proper representation of certain populations in the databases used for the development and evaluation of AI models has been recognized as one of the main reasons for disparities in algorithmic performance between groups [1,2]. This significantly affects populations in low- and middle-income countries, as they are often excluded from major public databases due, for example, to unequal access to healthcare and limited research resources [3]. Consequently, these populations suffer from what is known as "health data poverty," describing the inability of certain individuals or groups of individuals to benefit from innovations due to a scarcity of representative data [4]. This highlights the importance of addressing the gap in health data availability and representation to ensure that solutions and advances in the field benefit all communities equitably.

On the other hand, many of the databases used to create these AI-based tools have not been adequately described [5,6], hindering comprehensive analyses not only to investigate the existence of biases and understand their causes but also impeding the implementation of mitigation techniques if they arise.

For these reasons, we proposed to transparently collect and publish databases of medical images accompanied by demographic information that reflects the peculiarities of the Argentine population. We developed an 8-step standardized process to achieve this, visualized in Figure 1.

Digital Dermatology and the importance of image databases to enhance the application of AI in the specialty

One of the medical specialties extensively explored by AI is dermatology [7], where specialists heavily rely on visual analysis. Various imaging modalities are employed in daily practice for documentation, tracking, longitudinal comparison, analysis, referral, research, and teaching purposes [8]. Research teams have demonstrated that a convolutional neural network, a commonly used algorithm for automatic image processing, could achieve performance comparable to that of specialists in identifying skin cancer [9]. Simultaneously, a systematic review of the main dermatological image databases used in AI research revealed an uneven geographical distribution and, consequently, a limited representation of global population diversity [10]. This led us to embark on the collection, thorough review, and publication of dermatological images, forming a crucial component of a broader initiative encompassing the development, validation, and implementation of AI models as clinical decision support systems. 

The detailed process outlined in Figure 1 was followed to publish the first Argentine database of clinical and dermoscopic images of cutaneous lesions, along with metadata, in the International Skin Imaging Collaboration (ISIC) Archive [11]. This database was built at the Hospital Italiano de Buenos Aires (HIBA), located in Buenos Aires, Argentina. Images and data were extracted from the health information systems of this computerized center through the collaborative efforts of a multidisciplinary team. The paper published in Scientific Data [12] aims to provide detailed information on the processes of design, compilation and refinement of this database, including a description of its final content, in order to ensure the necessary transparency demanded by the community through various initiatives, guidelines and frameworks [13].

Multidisciplinary teams: key pillars for research and innovation

An essential aspect of undertaking this project was the formation of a multidisciplinary team, consisting of experienced dermatologists specialized in skin lesion assessment and dermoscopy, together with health informatics and bioengineers with in-depth knowledge of health information systems and experience in data governance and data science projects. While the dermatologists were responsible for case collection, image and metadata retrieval from electronic health records and their refinement, the computer scientists and engineers took on the task of designing the database, providing a user-friendly platform to improve the efficiency of data loading, and performing exploratory analysis of the resulting database.

Close collaboration between the teams was instrumental in achieving the goals set. Each team contributed their unique and valuable perspective, ensuring the creation of a complete, accurate and useful database. We emphasized the importance of diversity within the teams, not only in terms of profession, but also in terms of the demographic groups they represent. This diversity is crucial to avoid the unintentional introduction of biases inherent to individuals and to ensure proper consideration of the various segments that make up our society.

A first glance at the characteristics and uniqueness of the dataset

The dataset described in the article [12] was composed of information collected from 623 patients seen by expert dermatologists at HIBA. We included 1,616 images (1,270 contact-polarized dermoscopy images and 346 clinical images) captured from 1,246 lesions corresponding to the most frequent diagnoses observed at the institution. Moreover, the de-identified set of images was enriched with metadata including patient information such as age, sex, and skin tone, as well as details about lesion localization, diagnosis, and other relevant attributes.

In recent years, larger dermatologic image sets have become available, such as those published for the ISIC Challenges of 2019 [14-16] (N = 25,331 images) and 2020 [17] (N = 33,126 images), compared to the one presented here whose size resembles that of PAD-UFES-20 [18] (N = 2,298). While these datasets are composed entirely of dermoscopies [14-17] or clinical images [18], ours [12]  presents a combination of both imaging modalities. All collections share common features, including the publication of images without preprocessing, and variations in acquisition equipment, illumination conditions, resolution, and the presence or absence of artifacts. Melanocytic lesions emerge as the most frequent diagnoses across most databases, and all instances of malignant lesions were biopsy-proven. The metadata across these datasets exhibit similarities, as does the reporting rate for each attribute. Notably, only one dataset [18], apart from ours [12], integrated skin tone as additional information, with similar distributions in both cases—more records of intermediate tones (II and III) and fewer representations of the lightest (I) and darkest tones (IV-VI).

Data provenance differs among these datasets, as the ones published by ISIC [14-17] include patient images from countries such as Australia, Austria, the United States, and Spain, whereas PAD-UFES-20 [18] originated in Brazil, and ours [12] in Argentina. In this sense, the sociodemographic and epidemiological differences between our populations encouraged the conduct of this work to ensure the representativeness of the Argentinean people and their particularities in dermatological image archives. We also hope to encourage other Latin American countries to become involved in this type of endeavor to facilitate the translation of new technologies to clinical settings in the region. 

Opening paths for an Open Data program

As far as our knowledge extends, this is the first database of clinical and dermoscopic images of skin lesions compiled and made available to the public by a highly complex hospital in Hispanoamerica. While the database can be combined with larger image collections and used to train AI algorithms, its value lies in processes of evaluation and validation, as well as in the comparison of different AI systems. Furthermore, we consider this a first step toward the establishment of a collaborative database involving different medical centers in the country and even with other healthcare providers in the region to create a truly diverse initiative that ensures access to cutting-edge technology for our population.

Additionally, we are actively processing institutional databases from other specialties, enhancing an internally developed DICOM de-identification application, and designing a data governance program to optimize the management of various institutional data types, including clinical records, genetic information, signals, and medical images. These efforts aim to increase the representation of our country's population in databases used to create AI-based technologies, seeking to mitigate potential algorithmic biases that could affect our citizens when these tools are translated into the clinical environment.

References

1. Seyyed-Kalantari, L., Zhang, H., McDermott, M. B., Chen, I. Y., & Ghassemi, M. (2021). Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12), 2176-2182.
2. Ricci Lara, M. A., Echeveste, R., & Ferrante, E. (2022). Addressing fairness in artificial intelligence for medical imaging. nature communications, 13(1), 4581.
3. Chen, I. Y., Pierson, E., Rose, S., Joshi, S., Ferryman, K., & Ghassemi, M. (2021). Ethical machine learning in healthcare. Annual review of biomedical data science, 4, 123-144.
4. Ibrahim H, Liu X, Zariffa N, Morris AD & Denniston AK. Health data poverty: an assailable barrier to equitable digital health care.  Lancet Digit. Health. 3(4), e260-e265 (2021).
5.  Yi, P. H., Kim, T. K., Siegel, E., & Yahyavi-Firouz-Abadi, N. Demographic reporting in publicly available chest radiograph data sets: Opportunities for mitigating sex and racial disparities in deep learning models. J. Am. Coll. Radiol. 19(1), 192-200 (2022). 
6. Abbasi-Sureshjani, S., Raumanns, R., Michels, B. E., Schouten, G., & Cheplygina, V. Risk of training diagnostic algorithms on data with demographic bias. In Interpretable and Annotation-Efficient Learning for Medical Image Computing: Third International Workshop, iMIMIC 2020, Second International Workshop, MIL3ID 2020, and 5th International Workshop, LABELS 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4–8, 2020, Proceedings 3 (Springer International Publishing).  183-192  (2020). 
7. Young, A. T., Xiong, M., Pfau, J., Keiser, M. J., & Wei, M. L. Artificial intelligence in dermatology: a primer. J. Invest. Dermatol. 140(8), 1504-1512 (2020).
8. Milam, E. C., & Leger, M. C. Use of medical photography among dermatologists: a nationwide online survey study. J. Eur. Acad. Dermatol. Venereol. 32(10), 1804-1809 (2018).
9. Esteva, A., et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 542(7639), 115-118 (2017). 
10. Wen, D., et al. Characteristics of publicly available skin cancer image datasets: a systematic review. Lancet Digit. Health. 4(1), e64-e74 (2021).
11. Hospital Italiano de Buenos Aires - Skin Lesions Images (2019-2022). International Skin Imaging Collaboration https://doi.org/10.34970/587329 (2023).
12. Ricci Lara, M. A., Rodríguez Kowalczuk, M. V., Lisa Eliceche, M., Ferraresso, M. G., Luna, D. R., Benitez, S. E., & Mazzuoccolo, L. D. (2023). A dataset of skin lesion images collected in Argentina for the evaluation of AI tools in this population. Scientific Data, 10(1), 712.
13. Ganapathi, S., et al. Tackling bias in AI health datasets through the STANDING Together initiative. Nature Med. 28(11), 2232–2233 (2022). 
14. Tschandl, P., Rosendahl, C., & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data. 5, 180161 (2018). https://doi.org/10.1038/sdata.2018.161
15. Codella, N. C., et al. Skin Lesion Analysis Toward Melanoma Detection: a Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), hosted by the International Skin Imaging Collaboration (ISIC). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018) Pp. 168-172 (IEEE, 2018).
16. Combalia, M., et al. Bcn20000: Dermoscopic lesions in the wild. Preprint at https://arxiv.org/abs/1908.02288 (2019).
17. Rotemberg, V., Kurtansky, N., Betz-Stablein, B. et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci. Data. 8, 34 (2021). https://doi.org/10.1038/s41597-021-00815-z
18. Pacheco, A. G., et al. PAD-UFES-20: A skin lesion dataset composed of patient data and clinical images collected from smartphones. Data Brief. 32, 106221 (2020).