Can AI Catch What Doctors Miss?
#AI #medicine
A few days ago, I watched a great TED talk by Dr. Eric Topol, a cardiologist, that made me realise how mind-blowing and revolutionise AI is for our healthcare.
Here's overview of all the examples Dr. Eric Topol mentions in the video, where AI can be used to improve healthcare.
- Diagnostic Medical Errors: A Johns Hopkins study reveals these errors cause significant harm, leading to 800,000 Americans dying or becoming seriously disabled each year. This situation raises the question of whether AI can contribute to solving this problem.
- AI in Analyzing Retinal Images: AI demonstrated the ability to identify gender from retinal images. While human retinal experts had a 50% success rate (essentially a random guess) in determining gender from these images, AI achieved a remarkable 97% accuracy. This outcome was achieved by training the AI with a dataset of 100,000 images using supervised learning techniques.
- AI in Medical Imaging: AI does at least as well, if not better in detecting anomalies in X-rays, MRIs, CT scans etc. than expert physician.
- Colonoscopy and polyp detection: In 21 randomized trials focusing on colonoscopy and polyp detection, machine vision has consistently outperformed gastroenterologists in identifying polyps, especially notable as the day progresses. While the impact of this enhanced polyp detection on cancer progression remains uncertain, it highlights the remarkable ability of machine vision to detect details that might be missed by human experts.
- Deep learning models and disease detection in retina: Deep learning models have shown that computer vision can detect health issues in the retina that humans can't see. This includes markers for diseases like diabetes, high blood pressure, kidney, liver, and heart conditions, as well as early signs of Alzheimer's and Parkinson's diseases. These findings, based on large-scale image studies, suggest that in the future, retinal imaging might be a routine part of health checkups, offering a broad view of our overall health.
- ECG: The speaker, a cardiologist with three decades of experience, notes that modern electrocardiograms can reveal much more than traditional heart health indicators. They now identify patient details like age and sex, predict risks for stroke or atrial fibrillation, and diagnose conditions including diabetes, anemia, thyroid disorders, and kidney disease. This advancement marks a significant shift in the capabilities of ECG technology.
- Chest x-ray: Chest X-rays analyzed with machine vision can identify a patient's race, diagnose and monitor diabetes, and provide detailed information about heart health. Machine vision applied to chest X-rays now offers capabilities beyond the traditional scope of radiologists and cardiologists.
- Pathology: Machine vision, particularly through convolutional neural networks, is revolutionizing pathology by accurately identifying cancer's genomic mutations and structural variants from slides. This technology can also determine the tumor's origin and predict the patient's prognosis, capabilities that were previously uncertain or unknown in many cases.
- Keyboard liberations: "Keyboard liberation" in healthcare uses AI to automate clinical documentation, reducing data entry for doctors and enhancing patient care. This technological shift saves time and improves clinician-patient interactions, fostering a more patient-centric healthcare approach.
- Examples of success stories of AI diagnostic capabilitites:
- Six-year-old Andrew, after suffering from increasing pain, growth issues, and severe headaches, was diagnosed by ChatGPT with occulta spina bifida, a condition missed by 17 doctors over three years. The diagnosis revealed a tethered spinal cord, which was successfully treated with surgery. Andrew has since fully recovered and regained his health.
- A patient initially thought to be suffering from long COVID was correctly diagnosed by ChatGPT as having limbic encephalitis, after her sister entered her symptoms into the system. This correct diagnosis led to appropriate treatment, resulting in her significant recovery. This case is among 70 complex ones reviewed by the New England Journal of Medicine, where GPT-4 matched or surpassed expert clinicians in diagnostic accuracy.
Here's the TED talk: