Calendar

Abstract: AI for healthcare has the potential to revolutionize how we practice medicine. However, to do this in a fair and trustworthy manner requires special attention to how AI models work and their potential biases. In this talk, I will cover the considerations for building AI systems that improve healthcare.

Abstract:
For the development of ethical machine learning (ML) for precision medicine, it is essential to understand how values play into the decision-making process of developers. We conducted five group design exercises with four developer participants each (N=20) who were asked to discuss and record their design considerations in a series of three hypothetical scenarios involving the design of a tool to predict progression to diabetes. In each group, the scenario was first presented as a research project, then as development of a clinical tool for a health care system, and finally as development of a clinical tool for their own health care system. Throughout, developers documented their process considerations using a virtual collaborative whiteboard platform. Our results suggest that developers more often considered client or user perspectives after changing the context of the scenario from research to a tool for a large healthcare setting. Furthermore, developers were more likely to express concerns arising from the patient perspective and societal and ethical issues such as protection of privacy after imagining themselves as patients in the health care system. Qualitative and quantitative data analysis also revealed that developers made reflective/reflexive statements more often in the third round of the design activity (44 times) than in the first (2) or second (6) rounds. These statements included statements on how the activity connected to their real-life work, what they could take away from the exercises and integrate into actual practice, and commentary on being patients within a health care system using AI. These findings suggest that ML developers can be encouraged to link the consequences of their actions to design choices by encouraging “empathy work” that directs them to take perspectives of specific stakeholder groups. This research could inform the creation of educational resources and exercises for developers to better align daily practices with stakeholder values and ethical ML design.

Abstract
In this seminar I will be addressing the dire cancer survival outcomes in low- and middle-income countries (LMICs), with a particular focus on Sub-Saharan Africa. Cancer survival rates in Sub-Saharan Africa are alarmingly low. According to the World Health Organization, cancer deaths in LMICs account for approximately 70% of global cancer fatalities. In Nigeria, the five-year survival rate for breast cancer, one of the most common cancers, stands at a disheartening 10-30%, compared to over 80% in high-income countries. This stark disparity highlights the urgent need for sustained comprehensive cancer interventions in our region.

Abstract
Evaluating radiology reports is a challenging problem as factual correctness is extremely important due to the need for accurate medical communication about medical images. Existing automatic evaluation metrics either suffer from failing to consider factual correctness (e.g., BLEU and ROUGE) or are limited in their interpretability (e.g., F1CheXpert and F1RadGraph). In this paper, we introduce GREEN (Generative Radiology Report Evaluation and Error Notation), a radiology report generation metric that leverages the natural language understanding of language models to identify and explain clinically significant errors in candidate reports, both quantitatively and qualitatively. Compared to current metrics, GREEN offers: 1) a score aligned with expert preferences, 2) human interpretable explanations of clinically significant errors, enabling feedback loops with end-users, and 3) a lightweight open-source method that reaches the performance of commercial counterparts. We validate our GREEN metric by comparing it to GPT-4, as well as to error counts of 6 experts and preferences of 2 experts. Our method demonstrates not only higher correlation with expert error counts, but simultaneously higher alignment with expert preferences when compared to previous approaches.