TWTheresa Willem

Theresa Willem

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Theresa Willem has a total of 61 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Does Swarm Learning Solve Federation Opacity? The Need for Governance Beyond Decentralization

    by Marieke Bak, Stefanie Warnat-Herresthal, Theresa Willem, Bettina Zimmermann, Jonas Schulte-Schrepping, Lucas Secchim Ribeiro, Joachim Schultze, Stuart McLennan

    Hatherley and colleagues (2026) analyze a key problem in federated learning (FL) that they call “federation opacity”. They argue that FL creates a distinctive epistemic and ethical challenge becaus...


  • Open Access

    Should Dermatologists Recommend Direct‐to‐Consumer App‐Based Remote Diagnostics? An Ethical Analysis

    by Sonja Mathes, Sebastian Seurig, Theresa Willem, Marta Dechant, Stephan Traidl, Mina Kang, Pedro Marques Pinto, Annette Binder, Tilo Biedermann, Ulf Darsow

    Background: Dermatology patients still face barriers in accessing timely specialist care. As direct‐to‐consumer (DTC) apps for remote dermatological diagnostics proliferate, guidance is lacking. While promising efficiency and efficacy, their clinical—and ethical—legitimacy is not yet well established. Objectives: We aim to provide a comprehensive ethical analysis as an approach to structuring the field and to aid ethically grounded recommendations to patients. Methods: We applied a structured ethical analysis, as suggested by Marckmann, with (1) scenario description, (2) definition of normative criteria, (3) criterion‐wise assessment of all data, (4) synthesis of evaluations, and (5) practical recommendations, to systematically examine DTC dermatology apps through the lens of five bioethical/AI ethics principles: autonomy, beneficence, non‐maleficence, justice, and explicability. Results: Our analysis highlights that DTC apps may be endorsed in specific, rigorously vetted cases. Key conditions include formal regulatory approval, compliance with data protection standards, and provision of medically sound, transparent, and comprehensible outputs that must never undermine the standard of care, the integrity of health systems, or the trust‐based patient–physician relationship. In scenarios such as skin cancer detection, where AI performance is particularly strong, apps leveraging suitable models hold potential for ethical justification for standardized use, given their potential to bridge care gaps, reduce unnecessary referrals and harm, provided that bias is mitigated via robust dataset inclusion of marginalized groups and ongoing performance monitoring. Conclusions: We advocate a blended‐care model: apps are appropriate for monitoring and follow‐up, while initial diagnostics remain under professional oversight. As AI evolves, Human‐in‐the‐Loop systems offer a morally tenable path in areas of diagnostic uncertainty. Throughout, clinicians must remain vigilant of app limitations and inform patients of risk. Our ethical framework and recommendations aim to support the responsible augmentation of dermatologic practice with rapidly evolving DTC tools.

  • Open Access

    Consentful-by-design: a perspective on safeguarding data ownership from generative AI leveraging lessons from the healthcare domain

    by Akis Linardos, Theresa Willem, Alena Buyx, Showkot Hossain, Taeho Jung, Dimitrios Makris, Spyridon Bakas

    Research on Artificial Intelligence (AI) is lined with moral considerations. In healthcare, a high-risk field, sub-fields have emerged to mitigate AI-specific ethical issues such as fairness and transparency. However, similar considerations remain unaddressed beyond healthcare, and as generative AI tools (‘ GenAI ’) reach lay audiences, this neglect yields ethical concerns. The present work focuses on learning from ethical considerations in healthcare to mitigate challenges of GenAI . We structure our proposed mitigation strategies around three of the five established biomedical and AI ethics principles (autonomy, transparency, beneficence), highlighting the risks GenAI poses for intellectual property owners (scraping of copyrighted data, unpaid labor, plagiarism, fraud). We propose concrete ways to affirm these principles on GenAI, using biomedical AI examples and the emerging frameworks they have sparked in domains of data ownership, federated learning, and data provenance. This article comes at a pivotal time for AI, generalizing ethics-aware principles to GenAI to open new research avenues toward responsible AI.

  • Open Access

    Suspended responsibility: The trouble with integrating researchers into shared-responsibility models for machine learning-supported decisions

    by Theresa Willem, Alena Buyx, Ruth Müller

    With the rise of machine learning-based decision support systems, the responsibility for potential errors is regularly questioned. Scholarly work regarding this issue has criticized a “responsibility gap,” that is, unattributable responsibility for the effects of machine learning-supported decisions. One proposition to close the responsibility gap is to include scientists, who lay the basis for the functioning of machine learning models in distributed responsibility models, assigning them a share of responsibility for potential errors. However, to date, responsibility models that include scientists have not gained a foothold in regulation. To provide an empirical basis for the discussion around novel responsibility models, this study provides a social scientific analysis of an interdisciplinary machine-learning consortium that works on machine learning-based decision support systems for healthcare – an area where errors have particularly fundamental consequences for individuals. We investigate researchers’ speculations about their responsibility for the downstream effects of their ML research results after translation to clinical practice. We find that researchers point to tensions in the scientific sector as well as to agential, local, and temporal shifts of their research outputs during translation to clinical practice as a major source of what we call “suspension of responsibility.” Our insights contribute to debates about novel responsibility models that are fair to patients, doctors, and scientists and can inform similar debates beyond healthcare.


  • Open Access

    Perceptions of Stigma Among Patients With Hepatitis B in Germany: Cross-Sectional Survey

    by Bettina M Zimmermann, Theresa Willem, Michael Rost, Nina Matthes, Alena Buyx

    Background: Many studies find associations between hepatitis B and stigma, but studies from the Western European context are lacking. Based on available studies, we hypothesized that younger age, higher education, male gender, higher privacy needs, and non-German mother tongue were positively associated with perceived hepatitis B–related stigma. Objective: This study aims to describe the prevalence of perceived social stigma among patients with hepatitis B in Germany and to assess what factors are associated with perceptions of hepatitis B–related stigma. Methods: Applying the short version of the Berger stigma scale, we surveyed 195 patients with hepatitis B about their perceptions of hepatitis B–related stigma, privacy needs, and demographic variables through a paper-based questionnaire. Venue-based recruitment of adult patients diagnosed with acute or chronic hepatitis B was implemented at 3 clinical centers in Germany. Patients who could not read German were excluded from the study. Results: From the 195 valid questionnaires, 45.1% (88/195) of participants identified as female, 36.6% (71/195) had a high school diploma, and 56.9% (111/195) reported a mother tongue other than German. The mean (SD) stigma score throughout the sample was 5.52 (6.02; range 0-24) and the median was 3.50 (IQR=9.75). Regression analysis revealed that non-German mother tongue, individual data privacy needs, and participants’ secrecy regarding their hepatitis B diagnosis independently predicted perceived hepatitis B-related stigma. More precisely, the higher the data privacy need and the more secret the hepatitis B diagnosis, the higher the perceived stigma, and perceived stigma was higher for patients with a non-German mother tongue. Age, gender, and education were no predictors of perceived stigma. Conclusions: The surveyed patients with hepatitis B in Germany reported lower levels of hepatitis B–related stigma than found in other studies conducted in Asian countries. The association with non-German mother tongue indicates an important cultural and social component in the perception of stigma. Community-based interventions and the sensibilization of health care professionals might help overcome perceptions of stigma among hepatitis B–affected populations.


  • Biases in machine-learning models of human single-cell data

    by Theresa Willem, Vladimir A. Shitov, Malte D. Luecken, Niki Kilbertus, Stefan Bauer, Marie Piraud, Alena Buyx, Fabian J. Theis

    Recent machine-learning (ML)-based advances in single-cell data science have enabled the stratification of human tissue donors at single-cell resolution, promising to provide valuable diagnostic and prognostic insights. However, such insights are susceptible to biases. Here we discuss various biases that emerge along the pipeline of ML-based single-cell analysis, ranging from societal biases affecting whose samples are collected, to clinical and cohort biases that influence the generalizability of single-cell datasets, biases stemming from single-cell sequencing, ML biases specific to (weakly supervised or unsupervised) ML models trained on human single-cell samples and biases during the interpretation of results from ML models. We end by providing methods for single-cell data scientists to assess and mitigate biases, and call for efforts to address the root causes of biases.


  • Open Access

    The Social Construction of Categorical Data: Mixed Methods Approach to Assessing Data Features in Publicly Available Datasets

    by Theresa Willem, Alessandro Wollek, Theodor Cheslerean-Boghiu, Martha Kenney, Alena Buyx

    Background: In data-sparse areas such as health care, computer scientists aim to leverage as much available information as possible to increase the accuracy of their machine learning models’ outputs. As a standard, categorical data, such as patients’ gender, socioeconomic status, or skin color, are used to train models in fusion with other data types, such as medical images and text-based medical information. However, the effects of including categorical data features for model training in such data-scarce areas are underexamined, particularly regarding models intended to serve individuals equitably in a diverse population. Objective: This study aimed to explore categorical data’s effects on machine learning model outputs, rooted the effects in the data collection and dataset publication processes, and proposed a mixed methods approach to examining datasets’ data categories before using them for machine learning training. Methods: Against the theoretical background of the social construction of categories, we suggest a mixed methods approach to assess categorical data’s utility for machine learning model training. As an example, we applied our approach to a Brazilian dermatological dataset (Dermatological and Surgical Assistance Program at the Federal University of Espírito Santo [PAD-UFES] 20). We first present an exploratory, quantitative study that assesses the effects when including or excluding each of the unique categorical data features of the PAD-UFES 20 dataset for training a transformer-based model using a data fusion algorithm. We then pair our quantitative analysis with a qualitative examination of the data categories based on interviews with the dataset authors. Results: Our quantitative study suggests scattered effects of including categorical data for machine learning model training across predictive classes. Our qualitative analysis gives insights into how the categorical data were collected and why they were published, explaining some of the quantitative effects that we observed. Our findings highlight the social constructedness of categorical data in publicly available datasets, meaning that the data in a category heavily depend on both how these categories are defined by the dataset creators and the sociomedico context in which the data are collected. This reveals relevant limitations of using publicly available datasets in contexts different from those of the collection of their data. Conclusions: We caution against using data features of publicly available datasets without reflection on the social construction and context dependency of their categorical data features, particularly in data-sparse areas. We conclude that social scientific, context-dependent analysis of available data features using both quantitative and qualitative methods is helpful in judging the utility of categorical data for the population for which a model is intended.


  • Open Access

    Social Media Recruitment as a Potential Trigger for Vulnerability: Multistakeholder Interview Study

    by Nina Matthes, Theresa Willem, Alena Buyx, Bettina M Zimmermann

    Background: More clinical studies use social media to increase recruitment accrual. However, empirical analyses focusing on the ethical aspects pertinent when targeting patients with vulnerable characteristics are lacking. Objective: This study aims to explore expert and patient perspectives on vulnerability in the context of social media recruitment and seeks to explore how social media can reduce or amplify vulnerabilities. Methods: As part of an international consortium that tests a therapeutic vaccine against hepatitis B (TherVacB), we conducted 30 qualitative interviews with multidisciplinary experts in social media recruitment (from the fields of clinical research, public relations, psychology, ethics, philosophy, law, and social sciences) about the ethical, legal, and social challenges of social media recruitment. We triangulated the expert assessments with the perceptions of 6 patients with hepatitis B regarding social media usage and attitudes relative to their diagnosis. Results: Experts perceived social media recruitment as beneficial for reaching hard-to-reach populations and preserving patient privacy. Features that may aggravate existing vulnerabilities are the acontextual point of contact, potential breaches of user privacy, biased algorithms disproportionately affecting disadvantaged groups, and technological barriers such as insufficient digital literacy skills and restricted access to relevant technology. We also report several practical recommendations from experts to navigate these triggering effects of social media recruitment, including transparent communication, addressing algorithm bias, privacy education, and multichannel recruitment. Conclusions: Using social media for clinical study recruitment can mitigate and aggravate potential study participants’ vulnerabilities. Researchers should anticipate and address the outlined triggering effects within this study’s design and proactively define strategies to overcome them. We suggest practical recommendations to achieve this.


  • Open Access

    Embedded Ethics in Practice: A Toolbox for Integrating the Analysis of Ethical and Social Issues into Healthcare AI Research

    by Theresa Willem, Marie-Christine Fritzsche, Bettina M. Zimmermann, Anna Sierawska, Svenja Breuer, Maximilian Braun, Anja K. Ruess, Marieke Bak, Franziska B. Schönweitz, Lukas J. Meier, et al.

    Integrating artificial intelligence (AI) into critical domains such as healthcare holds immense promise. Nevertheless, significant challenges must be addressed to avoid harm, promote the well-being of individuals and societies, and ensure ethically sound and socially just technology development. Innovative approaches like Embedded Ethics, which refers to integrating ethics and social science into technology development based on interdisciplinary collaboration, are emerging to address issues of bias, transparency, misrepresentation, and more. This paper aims to develop this approach further to enable future projects to effectively deploy it. Based on the practical experience of using ethics and social science methodology in interdisciplinary AI-related healthcare consortia, this paper presents several methods that have proven helpful for embedding ethical and social science analysis and inquiry. They include (1) stakeholder analyses, (2) literature reviews, (3) ethnographic approaches, (4) peer-to-peer interviews, (5) focus groups, (6) interviews with affected groups and external stakeholders, (7) bias analyses, (8) workshops, and (9) interdisciplinary results dissemination. We believe that applying Embedded Ethics offers a pathway to stimulate reflexivity, proactively anticipate social and ethical concerns, and foster interdisciplinary inquiry into such concerns at every stage of technology development. This approach can help shape responsible, inclusive, and ethically aware technology innovation in healthcare and beyond.