FPH: FAIR PREDICTIONS IN HEALTH CARE

Summary of the context and overall objectives of the project

1) Problem/Issue Being Addressed

Diagnostic disparities exist when predictive algorithms are used in medical contexts. For instance, a certain cancer might be misdiagnosed differently among men and women due to a consistent algorithmic threshold. Similarly, limited data from ethnic minorities can skew diagnosis, favoring the dominant ethnic group.

2) Importance to Society

With the advent of AI in healthcare, ethical dilemmas have emerged, especially concerning potential biases in diagnostic tools. Ensuring AI-driven tools are unbiased is crucial as these decisions directly impact individuals’ well-being. It’s not just about using unbiased data but truly understanding what “fairness” means in this context.

3) Overall Objectives

The research project focuses on reviewing philosophical theories related to justice, fairness, and discrimination, understanding the intersections between probability theory and ethical considerations, exploring the epistemology of causality, and assessing real-world case studies to uncover biases in AI diagnostics in healthcare.

Project’s Conclusions

Drawing from the “fair equality of chances” principle by myself and colleagues (forthcoming in Economics and Philosophy), my research presents a fresh perspective on its relevance to healthcare. We identify intricate algorithmic biases that aren’t just discriminatory but can also mirror societal disparities on a broader scale. The findings underscore the necessity for a layered understanding of fairness in healthcare algorithms, recognizing the depth of bias and the ethical repercussions of maintaining status quo inequalities.

Work Performed and Main Results Achieved

1. Publications and manuscripts overview:

D2.1:

Article on the relevance of causality to discrimination, co-authored with colleagues from Polimi, published in Res Publica.
Conceptual map of fairness by prediction-based decisions in healthcare, currently submitted.

D2.2:

Exploration into the legitimacy of predictions as a justification for inequality, resulting in three collaborative articles:
- Two published in FAccT ’22.
- One forthcoming in Economics and Philosophy.
An additional paper on the same topic, co-authored with Clinton Castro, is currently under submission.

D2.3:

Development of a process model for a value-based design methodology. This work is available as a pre-print and is under revision for future publication.

2. Dissemination Activities:

D3.1: A reading group was successfully completed.

D3.2: The Didattica Innovativa courses underwent a change in audience. Instead of being directed at Polimi Students, a seminar was conducted at the Computational Cancer Biology lab of the Istituto Europeo di Oncologia. This seminar catered to early career researchers, emphasizing fairness in ML training.

D3.3: Guest lectures were executed in the computer ethics course by Viola Schiaffonati, effectively replacing the initial plan of instructing POLIMI students directly on fair ML.

D3.4: The course “Teaching Philosophy with Statistics” meant for high-school teachers was slightly modified, with details available in the outreach section below (D4.2-4.6).

In terms of global outreach, research papers were presented at renowned conferences and institutions, including the ACM FACCT Conference 2022 in Seoul, the Centre for Experimental-Philosophical Study of Discrimination Conference ’22 in Århus, and the Conference on Causality for Ethics and Society in Munich, among others. I co-organized the European Workshop on Algorithmic Fairness in 2022 and 2023 as one of the three co-chairs.

3. Outreach Activities:

Collaborated with Algorithmwatch for a public debate, producing a comprehensive blog article (D4.1-2).
Developed a module “inclusive machine learning” for the Responsible Innovators of Tomorrow, distributed via the EDx Platform (D4.3).
Developed open-source teaching materials on algorithmic discrimination and equity, available for download and discussed in an open event for educators (D4.2-6). All materials are open source and can be accessed at http://effediesse.mate.polimi.it.
Participated to two podcasts where I was interviewed about the manuscript D2.3 above: Ethical Machines by Reid Blackman, targeting industry leaders and The ReadME Project, targeting open source developers.

Progress Beyond State of Art and Potential Impacts

1. Progress Beyond the State of Art

In my deep dive into healthcare and the challenges posed by artificial intelligence (AI) and machine learning (ML), I have expanded my knowledge horizon by exploring the dimensions of algorithmic bias. Through an interdisciplinary lens, I have critically examined both individual- and group-level causality, unraveling the moral and ethical nuances of fairness, risk, and deservingness. My critique of existing fairness frameworks emphasizes their limitations and lays the groundwork for a more comprehensive approach to tackle the multifaceted nature of algorithmic bias. Drawing inspiration from the Fair Equality of Chances principle, as introduced by myself and colleagues, I have pioneered a unique perspective, showcasing how these biases, subtle yet profound, operate beyond overt discriminatory actions. My goal is clear: advocate for a holistic understanding of fairness in healthcare algorithms, emphasizing the recognition of biases and the moral quandaries they pose.

2. Impact

As I shared my specialized expertise on fairness in machine learning with POLIMI, it became a symbiotic relationship, enriching both the institution and strengthening my pedagogical skills. Drawing from my foundational background in ethics and aligning with my supervisor’s expertise in epistemology and probability, we fostered a multidisciplinary collaboration. This blossomed further through joint authorships with various researchers from diverse fields like mathematical statistics, computer science, and philosophy of science. Notably, collaborations with Dr. Francesco Nappo and Dr. Nicolò Cangiotti, among other esteemed peers from both European and US departments, have been instrumental in this integrative scholarly journey.

As I imparted my advanced competencies on fairness in machine learning to POLIMI, it was an exchange of knowledge, strengthening my teaching experience. My background in ethics, complemented by my supervisor’s strength in epistemology and probability, has paved the way for collaborative growth through the joint authorship of scientific articles.

While I faced resistance in traditional academic settings, my commitment has never waned. Venturing beyond academia, I am now harnessing my interdisciplinary knowledge to influence real-world scenarios, offering consultancy on fairness and AI to diverse stakeholders.

In reflection, my collaborations with Dr. Francesco Nappo and Dr. Nicolò Cangiotti, alongside other researchers from various European and US departments, have significantly contributed to our shared academic progress. Furthermore, the importance of AI ethics transcends academia. By offering AI ethics consultations to NGOs like Algorithmwatch and corporations, I’m directly influencing real-world applications. Additionally, recognizing the broader societal implications, I have made efforts to assist in governmental regulation activities by applying to serve as an expert for the ECAT office of the European Union. This endeavor represents my commitment to ensuring that advancements in AI are ethically sound and beneficial for all.