Keynote SpeakerZiad Obermeyer

Popular Talks by Ziad Obermeyer

• Bedside to bench: Reinventing medicine with AI

  AI is a powerful new tool for making sense of high-dimensional signals (images, waveforms, etc.) that humans struggle to process. In medicine, this is already starting to produce empirical discoveries that can drive theoretical insights. Rebooting this flow of ideas from ‘bedside to bench’—historically a key driver of progress, but… Read more

  AI is a powerful new tool for making sense of high-dimensional signals (images, waveforms, etc.) that humans struggle to process. In medicine, this is already starting to produce empirical discoveries that can drive theoretical insights. Rebooting this flow of ideas from ‘bedside to bench’—historically a key driver of progress, but one that has dried up in recent years—will give rise to major innovations in patient care, transform drug discovery, and create enormous economic value. It will also create a new science of medicine, powered by data and computation.

• (Machine) Learning About Sudden Cardiac Death from ECG Waveforms

  Every year, hundreds of thousands suffer sudden cardiac death. What makes these deaths so tragic is that many are preventable, with an implanted cardioverter defibrillator (ICD)—if only we could know who was at high risk before they died. Using a massive new dataset of ECGs linked to death certificates, we… Read more

  Every year, hundreds of thousands suffer sudden cardiac death. What makes these deaths so tragic is that many are preventable, with an implanted cardioverter defibrillator (ICD)—if only we could know who was at high risk before they died. Using a massive new dataset of ECGs linked to death certificates, we predict sudden cardiac death far better than current methods, both in a hold-out set of Swedish patients, and in a completely independent dataset from Taiwan. We also show that high-risk patients – and only high-risk patients – who receive ICDs have significantly lower mortality. Finally, we create a generative model of the ECG waveform to tie what the model is ‘seeing’ back to underlying cardiac electrophysiology.

• (Machine) Learning from Ground Truth: How to Focus AI on What Matters

  Despite enormous optimism around AI in health care, it’s easier to find examples of algorithms causing harm than doing good. I’ll argue that the root of these problems is the design principle underlying many of today’s AI tools: automation of (flawed, biased) human decision making. Drawing on my own work,… Read more

  Despite enormous optimism around AI in health care, it’s easier to find examples of algorithms causing harm than doing good. I’ll argue that the root of these problems is the design principle underlying many of today’s AI tools: automation of (flawed, biased) human decision making. Drawing on my own work, I’ll show several examples of how we can do better, by retraining AI to learn from patients and their outcomes rather than human judgment. Finding such ground truth outcomes is the key to AI that makes us better, instead of AI that automates our flaws.

• Dissecting algorithmic bias

  Algorithms can reproduce and even scale up racial biases. We demonstrate widespread bias in a widely used family of algorithms in health care, which predict health care costs as a proxy for health needs. Because Black patients face unequal access to care, they cost less than White patients with the… Read more

  Algorithms can reproduce and even scale up racial biases. We demonstrate widespread bias in a widely used family of algorithms in health care, which predict health care costs as a proxy for health needs. Because Black patients face unequal access to care, they cost less than White patients with the same needs. Algorithms predict this cost disparity accurately, and thus de-prioritize Black patients relative to their needs. This mechanism of bias—predicting a biased proxy—is widespread in health, as well as criminal justice, finance, and other social sectors where algorithms are applied. Crucially, this ‘label choice bias’ is fixable: retraining algorithms to predict less biased proxies can turn algorithms into a force for good, redistributing resources to those who need them and reducing disparities rather than perpetuating them.

• Data platforms

  Despite great optimism, actual applications of ML to health are in short supply. Using a concrete example, I’ll argue that the bottleneck is not financial or behavioral, but the scarcity of usable clinical data. I’ll also show how two data platforms, that allow safe and ethical access to health data,… Read more

  Despite great optimism, actual applications of ML to health are in short supply. Using a concrete example, I’ll argue that the bottleneck is not financial or behavioral, but the scarcity of usable clinical data. I’ll also show how two data platforms, that allow safe and ethical access to health data, can provide a new path forward for training and rigorous evaluation of ML algorithms. 

Biography

How to book Ziad Obermeyer as a keynote speaker