The future of AI in medicine | Conor Judge | TEDxGalway
TLDRConor Judge's TEDxGalway talk explores the potential of multimodal AI in revolutionizing medicine. He discusses the current imbalance in healthcare, where doctors spend more time collecting data than interacting with patients, and posits that AI could help restore this balance. Judge provides examples of single-model AI in healthcare, such as medical imaging analysis and disease prediction, before introducing multimodal AI that can process various data types. He emphasizes the importance of trust, explainability, and clinical trials in implementing AI safely, envisioning a future where AI enhances personalized and accessible healthcare, especially in underserved areas.
Takeaways
- 🎭 The speaker, Conor Judge, connects his childhood experience as a stammering detective in a play to his current role as a medical consultant and lecturer, emphasizing the importance of information gathering in both roles.
- 🔍 Conor highlights the imbalance in healthcare delivery, with doctors spending 70% of their time on data collection and only 30% on decision-making and patient communication, a problem exacerbated by electronic health records.
- 🤖 He introduces the concept of multimodal AI in medicine, which processes various data types like text, images, and numbers, similar to how humans use multimodal intelligence in medical practice.
- 👨⚕️ Conor discusses the potential of AI in healthcare, such as AI systems like Chest Link for X-ray triage and an AI model developed by University College London for diagnosing eye diseases and even predicting Parkinson's disease from retinal images.
- 📚 The speaker mentions the achievements of AI in passing the US medical licensing exam, with Med-PaM achieving an expert level score, showcasing the advancement of large language models in medical question answering.
- 🧠 Conor Judge advocates for the responsible use of medical AI to address the problems in healthcare, emphasizing the need for trust, explainability, and randomized clinical trials to ensure safe implementation.
- 🔑 Trust in AI for healthcare is a significant concern, with surveys showing that many patients would feel anxious if they knew their healthcare was AI-assisted, indicating the need for transparency and understanding.
- 🗝️ Explainability in AI is crucial to understand the decision-making process of the models, allowing healthcare professionals to make informed decisions rather than blindly following AI recommendations.
- 🧪 Randomized clinical trials are necessary for AI models to ensure their effectiveness and safety, just as they are for testing new medicines, to provide the highest level of evidence-based medicine.
- 👁️ The 'eyeball test' or the instinctive assessment of a patient's condition by looking at them is highlighted as a valuable human skill that should complement AI, suggesting a future where AI and human intuition work together.
- 🌐 Conor envisions a future where multimodal medical AI can democratize healthcare, providing insights and specialized care to remote areas, emphasizing the importance of compassion and the human-AI relationship in medicine.
Q & A
What was Conor Judge's role 26 years ago when he last stood on the stage of the town hall theater?
-Conor Judge was a 12-year-old boy participating in a drama competition for schools, playing the role of a detective in a play written by his best friend.
How much time does Conor Judge spend on collecting information about patients compared to making decisions based on that information as a medical consultant?
-Conor Judge spends 70% of his time collecting information about patients and only 30% of his time making decisions based on that information.
What is the imbalance Conor Judge refers to in healthcare delivery and how has technology made it worse?
-The imbalance Conor Judge refers to is the 70/30 split of time spent on data collection versus decision-making in healthcare. Technology, such as the introduction of electronic health records, has made this worse by increasing the administrative workload for doctors, thereby reducing the time they can spend face-to-face with patients.
What is multimodal AI and how does it differ from single model AI?
-Multimodal AI is an AI system that takes in and processes data in various forms such as text, images, and numbers. It differs from single model AI, which processes only one type of data, like images or text.
Can you describe the function of the AI system called Chest Link?
-Chest Link is a medical AI triage system that can autonomously analyze chest X-rays, looking for 75 different abnormalities. If none are found, it reports the X-ray as normal without human involvement. If an abnormality is detected, it refers the X-ray to a human radiologist for further reporting.
What is the significance of the AI model developed by researchers at University College London for diagnosing eye diseases?
-The AI model developed by researchers at University College London is significant because it can diagnose eye diseases and predict outcomes from conditions like macular degeneration. Impressively, it can also predict Parkinson's disease years before symptoms develop by analyzing the retina.
What is the role of Med-PaM, the medical large language model released by Google?
-Med-PaM is a medical large language model designed to perform medical question answering. It is the first AI model to pass the US medical licensing exam, demonstrating its ability to provide expert-level responses to medical queries.
How did Conor Judge use the multimodal version of Chat GPT to analyze an ECG and provide follow-up advice for a patient scenario?
-Conor Judge input an ECG image and a patient scenario into the multimodal version of Chat GPT. Although the ECG analysis wasn't perfect, the follow-up advice given by the AI was accurate and helpful for the patient's next steps.
What are the three key elements needed to implement multimodal AI safely in healthcare according to Conor Judge?
-The three key elements needed to implement multimodal AI safely in healthcare are trust, explainability, and randomized clinical trials. Trust is essential to overcome patient anxiety about AI in healthcare, explainability helps to understand the AI's decision-making process, and randomized clinical trials ensure the AI's effectiveness and safety.
What is the importance of the 'eyeball test' in medicine and how does Conor Judge see it fitting into the future of multimodal AI?
-The 'eyeball test' refers to the initial visual assessment of a patient's condition, which has been shown to be more accurate than some sophisticated models. Conor Judge sees this fitting into the future of multimodal AI by incorporating a picture or video of the patient into the AI model, allowing for a more personalized and efficient healthcare experience.
What is Conor Judge's vision for the future of healthcare with the integration of multimodal medical AI?
-Conor Judge envisions a future where multimodal medical AI makes healthcare more efficient, personalized, and accessible, especially in remote areas of low and middle-income countries that lack access to specialized care. The integration of AI should prioritize compassion and understanding, allowing doctors to spend more time with patients and improve their health outcomes.
Outlines
😀 From Childhood Drama to Medical Mysteries
The speaker reminisces about their time as a 12-year-old boy with a stammer, participating in a drama competition. They played a detective solving a fictional hotel's robbery mystery. Fast forward 26 years, they are now a medical consultant and a senior lecturer, still solving mysteries but in the context of diagnosing patients. The imbalance in healthcare, with 70% of time spent on data collection and 30% on patient interaction, is highlighted, worsened by technology like electronic health records designed for billing rather than efficiency.
🩺 AI in Medical Imaging and Diagnosis
The speaker introduces ChestLink, an AI system approved to autonomously report on chest X-rays by identifying 75 abnormalities. It autonomously classifies normal X-rays, referring only abnormal ones to human radiologists. Another example is an AI model trained on retinal images to diagnose eye diseases and predict Parkinson’s disease before symptoms appear. However, this AI cannot replace compassionate care. Lastly, MedPaLM, a Google-developed medical language model, is highlighted for passing the US medical licensing exam, showing AI’s potential in medical tasks.
🧠 Trust, Explainability, and Clinical Trials in AI
The speaker emphasizes the importance of trust, explainability, and randomized clinical trials for the safe implementation of multimodal AI in healthcare. Surveys show patients' anxiety about AI in medical treatment and fears of rapid AI integration. Explainable AI is crucial to understand AI’s decisions, avoiding confirmation bias. Randomized clinical trials, the gold standard in medicine, are necessary for AI models. The speaker envisions a future where AI includes patient images or videos for more accurate diagnoses, improving healthcare efficiency and accessibility, especially in underserved regions.
Mindmap
Keywords
💡AI in medicine
💡Data analytics
💡Multimodal AI
💡Medical triage
💡Machine learning
💡Natural language processing
💡ECG
💡Task sharing
💡Explainability
💡Randomized clinical trials
💡Compassion in healthcare
Highlights
Conor Judge's return to the stage after 26 years, from a young actor to a medical consultant and lecturer.
The analogy between solving a mystery in a play and diagnosing patients in medicine.
The 70-30 ratio of time spent on data collection versus decision-making in healthcare.
The impact of electronic health records on the doctor-patient relationship and efficiency.
The introduction of multimodal AI as a potential solution to healthcare inefficiencies.
Definition and explanation of multimodal AI in the context of medical practice.
The role of AI in triaging chest X-rays with the system 'Chest Link' by OxyPn.
AI's ability to diagnose eye diseases and predict outcomes from retina images.
The surprising capability of AI to predict Parkinson's disease from retinal images.
The importance of using AI in conjunction with trained healthcare professionals.
Google's Med-PaM, the first AI to pass a US medical licensing exam.
The advancement of Med-PaM 2 with an expert-level score on the medical licensing exam.
The demonstration of multimodal AI's capabilities with a patient scenario and ECG analysis.
The comparison of Med-PaM M's radiology report with human radiologists' reports.
The three essential components for implementing multimodal AI safely: trust, explainability, and clinical trials.
The need for trust in AI within healthcare as revealed by a US survey.
The concept of explainable AI and its importance in understanding AI decisions.
The necessity of randomized clinical trials for AI models in medicine.
The role of the 'eyeball test' in medicine and its surprising accuracy compared to AI models.
The vision of integrating patient images or videos into multimodal AI for more personalized healthcare.
The potential of multimodal medical AI to democratize access to specialized care globally.
The closing emphasis on the importance of compassion, understanding, and the human-AI relationship in healthcare.