Guest Lecture by Dr. Zheng Kaiping -National University of Singapore

Speaker audio begins at 03:02 due to an accidental mute. Title: Before the Model: Bias, Value, and Fidelity in Electronic Health Records Abstract: A medical record does not show everything that happens to a patient. People tend to visit a hospital when they feel unwell, and doctors order tests when they think the results will be useful. A patient with an acute illness may therefore have many records in a few days, while changes in a chronic condition may go unrecorded for months. What can an AI model learn from such data? This lecture starts with a simple example of how hospital visits and tests create bias in electronic health records. We will then look at two further questions: what we can learn from patients labeled as “negative” in a prediction task, and how we can spot records that may not accurately reflect a patient’s condition. Along the way, I will introduce data Shapley values through the basic idea of asking how much each example helps a model learn. The goal is to show why understanding and checking clinical data matters before using them to train healthcare AI, including deep learning models and medical language models. Biography: Dr. Kaiping Zheng is a Senior Research Fellow at the School of Computing, National University of Singapore. Her research lies at the intersection of data management, artificial intelligence, and healthcare. She develops methods to make clinical data more reliable and healthcare AI more trustworthy, with work published at leading conferences including KDD, SIGMOD, VLDB, ICML, and NeurIPS. She has extensive collaborations with clinicians across major hospitals in Singapore, connecting advances in data and AI with real clinical questions. She received her Ph.D. in Computer Science from NUS and was awarded the IMDA Excellence Prize for Best Ph.D. Thesis.