ACM CHIL is the brainchild of Vector faculty member Marzyeh Ghassemi.
The inaugural ACM Conference on Health, Inference, and Learning (CHIL) kicks off Thursday, July 23rd, 2020. Originally scheduled to take place in Toronto, Canada, CHIL is now a virtual event, featuring keynotes from Yoshua Bengio (Mila, Université de Montréal), Elaine Nsoesie (Boston University), Sherri Rose (Stanford University – previously Harvard Medical School), Ruslan Salakhutdinov (Carnegie Mellon University), and Nigam Shah (Stanford University).
The brainchild of Vector Faculty Member and Canadian CIFAR AI Chair Marzyeh Ghassemi (University of Toronto), the conference builds on the success of last year’s ML4H Unconference and the Machine Learning for Health Workshop at NeurIPS. It targets a cross-disciplinary group of experts from both industry and academia, including machine learning clinicians and researchers working in areas like health policy, causality, fairness, clinical data-sharing platforms, and deployment.
“ACM CHIL is unique,” says Ghassemi. “We target innovative machine learning approaches, evaluation and deployments that are increasingly necessary as clinical machine learning moves from off-the-shelf adaptations to domain-relevant innovations.”
Along with Ghassemi serving as general chair, Vector is well represented among CHIL’s leadership. Faculty Member and Canadian CIFAR AI Chair Anna Goldenberg (Hospital for Sick Children, University of Toronto) is a member of the steering committee, Tasmie Sarker, part of Vector’s professional staff, is the conference’s logistics chair, and Vector Postdoctoral Fellow Shalmali Joshi serves as a communications chair. Vector Faculty Affiliates Laura Rossella (University of Toronto, Dalla Lana School of Public Health) and Avi Goldfarb (University of Toronto, Rotman School of Management) sit on the executive committee.
CHIL additionally features tutorials on Public Health Datasets for Deep Learning: Challenges and Opportunities, A Tour of Survival Analysis: from Classical to Modern, and Medical Imaging with Deep Learning along with a great line-up of proceedings, workshops, and a doctoral consortium. A number of papers and workshop papers were also co-authored by Vector researchers.
Vector-related papers accepted to CHIL 2020
Haoran Zhang, Amy X. Lu, Mohamed Abdalla, Matthew McDermott, Marzyeh Ghassemi
When a machine learning algorithm is trained on data that is fundamentally biased, it can result in a model that reflects those biases. In clinical applications, this could result in serious treatment disparities across subgroups. In this work, we investigate the bias that exists when state-of-the-art natural language processing models are used on clinical notes to predict a variety of clinically relevant tasks. Evaluating across protected attributes like gender, ethnicity, and insurance status, we find that there are many statistically significant performance gaps, with the model often performing better on the majority group. This demonstrates the need for rigorous evaluations of model biases before deployment in the clinical setting.
Shirly Wang, Matthew B. A. McDermott, Geeticka Chauhan, Marzyeh Ghassemi, Michael C. Hughes, Tristan Naumann
MIMIC-Extract introduces several data cleaning, processing, and aggregation steps that make the MIMIC-III database (a commonly used ICU dataset in the machine learning community) more accessible to researchers. These steps address several challenges in applying machine learning models to clinical data such as a high level of missingness and noises in the data. We also open sourced our code to facilitate reproducibility.
Vector-related workshop papers:
Learning Representations for Prediction of Next Patient State
Taylor Killian, Jayakumar Subramanian, Mehdi Fatemi, Marzyeh Ghassemi
My work that will be featured in a CHIL workshop focuses on establishing appropriate representations of the information gathered from observing patient health over time. Prior work, when investigating sequential treatment strategies for healthcare, has only considered immediate observations when choosing which treatment to administer. This is problematic as historical information does influence human doctors’ decisions. By thoughtfully combining observations of patient health over time, we expect to be able to provide more appropriate and reliable treatment suggestions from algorithmic aids.
A Comprehensive Evaluation of Multitask Representation Learning on EHR Data
Matthew McDermott; Bret Nestor; Wancong Zhang; Peter Szolovits; Anna Goldenberg; Marzyeh Ghassemi
(No summary available)