Yann LeCun’s Post

Yann LeCun - This is awesome. Has little to nothing to do with Meta (but am grateful for the start ups truly making a difference in this area), who, despite their new product re age on Instagram, ignores mental health issues amongst teens (especially girls) climbing, in preference to profits. Too little too late and mostly optics…

Twelve years ago I remember talking to a pathologist about the potential of computer vision technologies to revolutionize the field of medical image analysis. I'm glad to see it's coming to fruition.

Yann LeCun I am concerned about you making blank statements regarding use of AI algorithms in medicine prematurely. Approximately 40% of U.S. hospitals used Epic as their electronic medical record (EMR) system, covering nearly half of all hospital beds 2020-2023 during the Covid pandemic. During this time it became clear that The Epic Sepsis Model (ESM)based on biased and faulty AI algorithm failed to detect sepsis in 67% of patients who developed the condition. Failure to detect sepsis delayed treatment, increasing mortality by 4-8% per hour of delay. The missed cases by the AI ESM likely contributed to significant preventable deaths during the pandemic. To be exact: COVID-19 deaths in the U.S. (as of late 2023): ~1.2 million.If we conservatively assume that 20-30% of these deaths involved undetected sepsis the numbers linked to the use of a faulty AI algorithm at the times when it mattered the most are staggering. Between 240,000 and 360,000 deaths during the pandemic could have involved undetected sepsis, misclassified as direct COVID-19 fatalities. Furthermore while missing true sepsis cases, the model generated alerts for 18% of all hospitalized patients, overwhelming clinicians with false alarms and diverting resources.

AI has been used in medicine for many decades. MYCIN, an expert system CDSS (clinical decision support system) developed at Stanford in the early 1970s, outperformed human clinical specialists at Stanford in terms of correct diagnosis and treatment recommendations. https://en.wikipedia.org/wiki/Mycin Today, the application of AI in Medical Informatics is faster and more accurate, but this doesn't change the fact that the use of AI in CDSS has been with us all along.

Here's an important segment from the article (if you're hit with paywall): In 2022, the Department of Health and Human Services reported that about 6 percent of the more than 130 million people who go to emergency rooms every year are misdiagnosed. Of those, 2.6 million people are unnecessarily harmed because of their misdiagnosis. Roughly 400,000 are permanently disabled or die. Artificial intelligence has the potential to significantly reduce those tragedies. A recent study out of Boston comparing the performance of chatbot-assisted physicians in diagnosing patients to just chatbots themselves found that the bots performed considerably better. Given a patient’s case history and symptoms, the chatbot alone scored an average of 90 percent in correctly diagnosing their condition. Physicians using the technology scored only 76 percent on average — just marginally better than the 74 percent average for humans with no AI help at all. AI can also speed up care in emergency settings. One study found that hospitals that use AI to detect strokes from a patient’s brain scans were able to shave off almost 40 minutes before a surgeon can intervene. That’s precious time that can save lives.

Yann LeCun - Thanks to architectures like SAM2, computer vision is about to spark a new wave in radiology. Carez AI is developing prompt-based guidance for image-guided surgeries, enabling doctors to reduce harmful injections. Better pixels, less chemicals.  This is how innovation transforms: from research papers and lines of code to improved clinical outcomes— a patient returning to the arms of their family with a smile. If this isn’t value creation, I don’t know what is.

Interesting article! While not fully there yet, the day is not far off when AI is both: accurate at differential diagnosis and can be deployed at scale, bringing down the cost of medicine in remote areas and the third world. As AI engineers, we know that all the AI-based imaging models of the last couple of years have made huge strides, and they often match or beat human performance. On the other hand, there is an imperfect or unfinished implementation/deployment bridge from lab to reality in the hospitals, so all my medical friends remain only cautiously optimistic. I come from a remote part of India where it is well nigh impossible to find world-class medical specialists since they primarily work in big cities, and their services are beyond the affordability of most. If technologies like this could be deployed at an affordable cost and be available to general medical practitioners (physicians) in remote regions, it would be the most transformative consequence of the AI breakthroughs in our generation.

Great to see this being acknowledged! AI is making a real impact for patients and practitioners every day, and we are just at the beginning of this journey

In the early 90's my Uncle who worked in a lab screening for cervical cancer invited me to visit and take a look at the screening slides to see if the early image analysis systems I was developing could help identify the key indicators. After a few hours of coaching he asked ' Is it possible ?' Not yet I replied, even I can't see the indicators.

Yann We have developed an advanced AI model with 2 trillion parameters capable of analyzing up to 3,000 medical images simultaneously, including CT scans, MRIs, and various types of X-rays. Acting like a highly skilled radiologist, this model generates detailed 2-3 page diagnostic reports, providing comprehensive analyses that rival human expertise. Despite its groundbreaking potential, the medical community has largely opposed our innovation. However, we firmly believe that the day will come when such technologies are widely embraced, revolutionizing healthcare and setting new standards in diagnostic precision.