Michael Muelly

38 statements · 1 topic

Featured statements

▶ Ep 2 · 6:29
I remember when I was a surgical intern at Penn State, we rolled out a simple sepsis alert based on four criteria of, you know, elevated white count or low white count, temperature, heart rate. And we actually rolled it out into deployment at Penn State. It didn't work very well, right, because it was such a simple, dumb rule set. It just went off all the time, right?
quote · Sepsis
▶ Ep 2 · 6:29
A simple sepsis alert system based on four criteria (elevated or low white count, temperature, heart rate) was rolled out at Penn State but failed because it was a simple rule set that went off all the time.
clinical · Sepsis

Nothing matches these filters — clear the search or widen the filters.

Michael's statements about Sepsis 38 statements

Open the Sepsis collection →

#APSA50: Artificial Intelligence

▶ Ep 2 · 1:46
epidemiological At the Radiological Society of North America annual meeting, there are about 100 to 150 AI startups presenting in the machine learning showcase, with some products available for purchase and deployed at healthcare systems. ↗
▶ Ep 2 · 2:24
clinical Yale has deployed AI for prioritization of head CTs with bleeds, cutting down the time until a radiologist reviews the study. ↗
▶ Ep 2 · 2:45
clinical Most current AI applications in radiology are in the operational realm, augmenting workflow rather than making diagnostic decisions. ↗
▶ Ep 2 · 3:03
opinion Radiology jobs are here to stay despite AI advances. ↗
▶ Ep 2 · 3:03
quote I'm not working at Google because I think my radiology job is going to go away, I think my radiology job is here to stay. ↗
▶ Ep 2 · 3:12
quote I would love for the chest X-rays to go away. Even that I'm not sure about, really. ↗
▶ Ep 2 · 3:35
opinion AI applications in healthcare offer tremendous opportunity to achieve a truly evidence-based medical system. ↗
▶ Ep 2 · 3:54
quote We've been talking about evidence-based medicine since the 70s, I think, right? And if you look at it, right, whatever you're taught during your surgical training is what you're going to do. Is there evidence behind it? Most of the time not. There is, it's typically very, very weak, right? ↗
▶ Ep 2 · 3:58
clinical What surgeons are taught during surgical training is what they will do in practice, and most of the time there is no evidence behind it, or the evidence is very weak. ↗
▶ Ep 2 · 4:43
clinical Laparoscopic surgery offers a natural opportunity for AI because it captures data digitally with video. ↗
▶ Ep 2 · 4:50
opinion Surgical robots are not currently using AI tools primarily for regulatory reasons, but will start to incorporate them once that hurdle is overcome. ↗
▶ Ep 2 · 5:28
clinical Trauma triaging is currently done based on very simple rules and is not very data driven. ↗
▶ Ep 2 · 5:42
opinion AI impact on the actual operating room (outside of operational tasks) will take a lot longer, if it happens at all. ↗
▶ Ep 2 · 6:29
clinical A simple sepsis alert system based on four criteria (elevated or low white count, temperature, heart rate) was rolled out at Penn State but failed because it was a simple rule set that went off all the time. ↗
▶ Ep 2 · 6:29
quote I remember when I was a surgical intern at Penn State, we rolled out a simple sepsis alert based on four criteria of, you know, elevated white count or low white count, temperature, heart rate. And we actually rolled it out into deployment at Penn State. It didn't work very well, right, because it was such a simple, dumb rule set. It just went off all the time, right? ↗
▶ Ep 2 · 7:06
clinical With deep learning, sepsis alerts can be much more sophisticated, and some places like Emory have sepsis alerts deployed into the workflow. ↗
▶ Ep 2 · 7:45
clinical Deep learning requires very large datasets because the neural networks learn just by example and need many examples with good variety of the data space. ↗
▶ Ep 2 · 8:11
clinical In radiology, whether a GE or Siemens device is used may affect what the algorithm learns. ↗
▶ Ep 2 · 8:19
clinical The current limitation for AI in healthcare is access to very large datasets, especially compared to other domains where billions of data points are available daily. ↗
▶ Ep 2 · 8:23
quote At Google, I like to tell our engineers who are used to getting billions of data points a day from search, and then they come and work on healthcare things. It's basically we have, you know, 7 to 8 billion people in the world right now, right? Most of them don't even have access to healthcare. You know, very few of them thankfully have anything wrong with them, right? The availability of data is very, very small. ↗
▶ Ep 2 · 8:45
clinical The more specific the disease or patient population (such as pediatric surgery), the harder it becomes to collect large, high-quality labeled datasets. ↗
▶ Ep 2 · 9:11
epidemiological Tesla has 100,000 to 200,000 cars driving around every day collecting data for training self-driving capability. ↗
▶ Ep 2 · 9:26
epidemiological Waymo has collected about 10 million miles of driving experience for self-driving cars. ↗
▶ Ep 2 · 9:34
clinical It is very difficult to generate large datasets in healthcare because data collection must fit into clinical workflow, cannot be easily simulated, and faces technical, organizational, and regulatory barriers. ↗
▶ Ep 2 · 10:26
quote Those models are very, very dumb, right? They learn whatever you show them. They're really good at that, right? They are fantastic at learning those associations from images. If you give them enough data, right, but they learn whatever association is there that may be useful to learn that. ↗
▶ Ep 2 · 10:26
clinical Deep learning models are very good at learning associations from images if given enough data, but they learn whatever association is present, including spurious ones. ↗
▶ Ep 2 · 10:45
clinical On CT scans, markers that exist outside the patient may tip the algorithm that the scan is from a particular hospital with different disease prevalence. ↗
▶ Ep 2 · 11:17
opinion Cloud-based data infrastructure is the way data centers in general and healthcare specifically will operate in the future, opening opportunities for access to larger datasets. ↗
▶ Ep 2 · 11:33
clinical Once data is centralized in the cloud, it becomes much easier to create frameworks that allow data sharing between institutions for research purposes, while still maintaining access controls. ↗
▶ Ep 2 · 11:44
quote Just because the data is in the cloud doesn't mean that anybody has access to it outside of the rights holder. So you still have to solve these other issues. But once the data is in a format and in a place where you can easily solve the technical challenge, then it becomes much easier to solve the regulatory organizational challenge. ↗
▶ Ep 2 · 12:26
clinical A model trained on data in the US may not work anywhere else due to differences in patient populations. ↗
▶ Ep 2 · 12:36
clinical If there is any racial or ethnic bias in training data, the AI model will have that bias as well. ↗
▶ Ep 2 · 13:00
opinion Training a machine learning model is actually a very small part of what needs to be done to deploy AI in healthcare. ↗
▶ Ep 2 · 13:05
clinical How the FDA will decide what is a safe medical device for AI applications remains an unsolved question. ↗
▶ Ep 2 · 13:21
opinion Even if AI models are not better than humans, they are much more consistent, which is their great promise. ↗
▶ Ep 2 · 13:33
clinical There is huge variance between how different radiologists read studies, sometimes incorrectly and sometimes due to differences of opinion. ↗
▶ Ep 2 · 13:53
clinical Similar variance exists among ophthalmologists and applies anywhere humans make decisions, including pediatric surgery. ↗
▶ Ep 2 · 14:04
opinion The great opportunity for AI is reducing variance and providing consistent care according to current knowledge. ↗