We will see the rapid rise of complete AI simulations with AI agents operating as people; customers, employees, soldiers and patients. Computer games have long set the pace here and real world simulations are here, even of entire hospitals.
AI Town
Imagine a simulated town with 25 AI agents living, behaving, interacting and making decisions in a typical town with buildings, streets, parks, and other infrastructure. This AI town has been built by Stanford University. Their neat aim is to understand social dynamics, cooperation, conflict resolution and the emergence of social norms among the AI agents. It’s the sort of interdisciplinary research that is being stimulated by AI, pushing research boundaries in various fields such as urban planning, autonomous systems, and human-AI collaboration.
AI Hospital
A similar project is an AI hospital. This has tighter, more immediate and practical implications. The goal is to train physician ‘Agents’ to do their work in a controlled environment of a typical hospital, with realistic processes and context. I have discussed the importance of context LINK before and it matters.
Tsinghua University, in China, call it ‘Agent Hospital’ where ALL doctors, nurses, healthcare professionals and patients are driven by interacting AI agents. What takes real Doctors two years to complete, their AI doctors can treat in just a few days – that’s 10,000 patients! The whole point is to scale thereby reduce time, reduce costs, while maintaining or improving patient outcomes.
Their Doctor agents achieved 93.06 percent accuracy rate on the US Medical Licensing Exam questions, across the whole process of presentation, examination, diagnosis, treatment and follow-up. The hospital has consultation and examination rooms, with 14 doctors and four nurses, where the doctors do the consultation, diagnosis and treatment while the nurses concentrate on on-going support.
Research team leader Liu Yang, who is also the executive dean of Institute for AI Industry Research (AIR) and associate dean of the Department of Computer Science and Technology at Tsinghua University, believes this will be transformative in healthcare, not just for training in a safe and realistic environment, where theory has to be turned into practice, but also continuous professional development.
Future use
One can see how this could develop into real patients and a virtual AI hospital, with many of the processes within a real hospital, with real equipment, dealt with by AI agents. There are many intermediate stages and problems to overcome but one can see the general direction of travel. The one solution we are unlikely to see any time soon is the care and compassion that humans offer to patients. Even here we can, for the moment, imagine caring robots, who behave compassionately. One need not feel compassionate to behave compassionately. It is not as if all Doctors have perfect bedside manner.
The rewards are substantial:
• Healthcare available 24/7
• Quicker results
• Increased availability of healthcare in rural areas
• Access by the poor
• Decreased workload for healthcare systems
• Lower costs
The ultimate concept of a UNIVERSAL DOCTOR is surely a possibility, for the process of presentation, examination, diagnosis, treatment and follow-up, drawing on a breadth and depth of experience, theory and cases that no human Doctor could ever learn or hold in their head. Misdiagnosis, mistakes in treatment and poor follow up are still substantial problems in healthcare, as are costs along with shortages of Doctors and Nurses. They are hoping to implement some of their services later this year.
Clinical decision making
Let’s go to the next level of detail. Healthcare and learning are the two big winners in AI, said Bill Gates in early 2023. I’d add ‘general productivity’ and at my Keynote to the Scottish National Health Service at their annual conference, I tried to get across the idea that some very acute problems in healthcare are now being solved by AI. Medicine in an evidence-based profession and the evidence is coming forward across the entire healthcare front - AI is making a difference.
There are many ways of cutting the healthcare cake but the clinical decision-making process is a good place to start and recognisable to all:
1. Presentation
2. Diagnosis
3. Investigation
4. Treatment
We can add two other significant areas:
5. Productivity
6. Training
1. Presentation
Many years ago, I worked with Alan Langland, CEO of the NHS, then Vice-Chancellor at two Universities. He once told me of his regret at not having spent more on NHS Direct. That, he thought, was the sweet spot for improvement – the patient-healthcare interface. It is still a mess. Try getting a GP appointment (can you call back tomorrow at 8am?), our A&E departments are full of people who couldn’t get appointments, dentistry has all but disappeared in certain areas of the country. It is an interface that is full of inefficiencies and friction. I have used NHS Direct (now NHS 111 and the NHS website) several times and these services are superb in reducing friction, as are recent advances in online and telephone consultations. What was actually needed was something more sophisticated and I think new systems that provide dialogue with patients will come to the fore. People are already using GenAI to get advice or preliminary diagnoses.
![](https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEgZalkGlb84eX57ZHv1yZfa8dXkVvnmSoXn_99pbqswhXbppZKBTZiC-tAfQonIL3vjrzXtRl7u2t1hidkTVim3rmxbQiNHNpTaYEZaW0tt1VkVvci1KvQvmwcyKbdkijbhzsJm7pGOiriaBLzjmFpApVAezL3TNbsm4B8T0YFmbE0vktv7HutmbQ/w400-h268/Karth.png)
In a study by Karthikesalingam (2023) on an AI system for diagnostic medical reasoning and conversations, an LLM chatbot was compared to physicians
The trained LLM had reasoning ability (an inference time chain-of-reasoning strategy and was trained with real-life conversations with professional actors pretending to be patients. The AI outperformed the board-certified Primary Care Physician (PCP)
Some research shows that chatbots, when they perform well, are actually preferred by patients, even seen as more empathetic, others show that using such tools show better results than human physicians. In an interesting study by Ayers (2023) a chatbot was preferred by patients 79% of the time and rated significantly higher for both quality and empathy.
Mei (2024) claims that using a Turing Test for behavioural and personality traits exhibited by AI: trust, fairness, risk-aversion, altruism & cooperation show that AI behaves like humans and learns. It was seen as more cooperative and altruistic.
Given that we are already seeing results, it is clear that this problem will be at least partially solved, still with some level of human intervention at some point. If so, the cost savings would be enormous.
AI embedded in wearables already provides biometric data to people covering a range of indicators. These are widening all the time. Neurolink and others are applying AI analysis to brain data through invasive and non-invasive devices, with measurable success in real patients.
There is still need a vast amount of social care, particularly for the elderly. Here a different form of AI is coming to the fore with robotics. That will take longer but progress has accelerated with companies like Figure and Optimus. The idea of a hospital, care-home or home companion is not so far-fetched.
2. Clinical decision making
On clinical decision making we are now seeing studies that show real progress in this pivotal skill. It takes a lot of time and money to train a physician. If we can augment this to even a small degree, huge savings can be achieved. Typical misdiagnosis rates in primary healthcare are between 5-15%. If inroads are made on reducing these figures, lives are saved and costs massively reduced. Imagine an AI that has a misdiagnosis rate of less than 3% - you’d be a fool to see a human doctor. Is that possible? The answer is surely yes.
Glass Health already deliver AI aided clinical decision making services and this filed is about to explode. AI-powered CDSS provide healthcare professionals with diagnostic suggestions based on the latest clinical guidelines and evidence from medical research. These systems analyse the patient's data and compare it against large, verified databases of medical knowledge to suggest possible diagnoses and recommend further tests or investigations if needed.
The Haim (2024) study, which compared DGPT-4 v experts with data from 100 stoke patients, showed that:
• GPT-4s & experts agreed in 80 cases
• In 5 it was right & expert wrong
• In 5 expert was right & AI wrong
• In 10 GPT-4 predicted mortality better than specialist AIs
Advanced Diagnostics comes into its own when AI algorithms assist in diagnosing diseases by analysing complex datasets, such as medical images, electronic health records (EHRs), and genetic information. These systems can identify patterns that may be missed by human judgement, offering high precision diagnostic suggestions.
This in turn can be used in Predictive Analytics, where AI can use historical data and real-time inputs to predict patient outcomes, helping clinicians make informed decisions about patient care, including predicting the risk of disease, likelihood of complications, and even potential responses to various treatments.
Real-time decision support during surgeries or emergency care is also possible, offering critical information or suggestions based on the ongoing evaluation of patient data. This can be crucial in high-stakes environments where swift decision-making really matters.
Initially it is this ‘second layer’ of analysis or opinion that can help reduce errors in diagnosis, alerting healthcare professionals to things that were overlooked or offering reminders about best practice.
Importantly, as AI systems can continuously learn from new data and outcomes, adapting and improving their recommendations over time. It is always learning. This dynamic learning process ensures that the decision-making tools stay current with the latest medical research and practices.
3. Investigation
Radiographers and radiologists, those who operate the scanners and diagnose images are under heavy pressure in hospitals. Improvements in imaging through AI are now regularly appearing even in images, astonishingly even with no visible signs. AI algorithms are particularly effective in analysing medical images such as X-rays, CT scans, MRI scans, and ultrasound images. These algorithms can detect abnormalities such as tumours, fractures, or signs of diseases like pneumonia with high accuracy, often surpassing human performance in speed and accuracy. This can lead to quicker, more accurate diagnoses at lower costs, leading to better patient outcomes.
In pathology, AI can help analyse tissue samples, detecting abnormalities that suggest diseases such as cancer. AI systems can also assist pathologists by highlighting areas of interest in tissue slides, ensuring that subtle signs of disease are not overlooked.
AI can analyse complex laboratory test results, identifying patterns that might indicate specific conditions or diseases. This includes not just interpreting simple tests, but also handling complex data from genetic sequencing or biomarker analysis, helping to identify genetic disorders or predispositions to certain diseases.
In the wider sense, AI systems can integrate and analyse data from multiple sources, including electronic health records (EHRs), imaging studies, and wearable health devices. This holistic view can help healthcare providers gain a more comprehensive understanding of a patient's health status, leading to better diagnostic accuracy.
It can also predict the likelihood of a disease's presence or future risk based on historical data and current patient data. This can be particularly useful in chronic disease management, where early intervention can significantly alter the disease's trajectory and patient outcomes.
Then there is its use over time, especially for chronic conditions or in post-operative care, can optimise investigation, as AI can continuously analyse patient data to monitor for signs of disease progression or recovery. This ongoing monitoring can catch potential complications early, prompting timely interventions.
4. Treatment
Everything in healthcare is complex. Treatment is no exception. In truth, treatment can be ill-targeted, poorly managed, wasteful and expensive. AI can help on several fronts. Targeted treatments, especially drug choice and dosage are areas where critical choices can be made on cost and efficacy.
At the consumer level, wearables and other AI-enabled devices can continuously monitor patient health remotely. AI can analyse this data in real time to detect deviations from baseline health parameters, alerting healthcare providers to potential issues before they become severe.
Then there’s early detection predictive analytics, where AI can predict patient risks and disease progression by analysing data trends over time. This can help healthcare providers anticipate complications or disease recurrences, allowing for timely interventions and more proactive management of patient care.
Personalised medicine uses AI to analyse vast amounts of data from genetic tests, medical histories, and clinical outcomes to develop personalised treatment plans. This can help tailor therapies to the individual characteristics of each patient, increasing the effectiveness of treatments and reducing side effects.
Wider evidence-based treatment recommendations through systems that can analyse medical literature and clinical data to suggest the most effective treatment options based on the latest research and clinical guidelines would also be useful.
On the physical side, in surgical and procedural contexts, robotic systems can assist or even autonomously perform surgeries, sometimes with greater precision and control than human surgeons. This can lead to less invasive procedures, reduced recovery times, and better surgical outcomes.
New treatments, such as accelerated drug discovery have already been realised. AI can streamline the drug discovery and development process by predicting how different chemicals will react in the body, identifying potential drug candidates more quickly, and reducing the time and cost to bring new drugs to market. We have already seen huge success with Alphafold, saving 1 billion years of research by identifying the 3D structure of 200 million proteins. Deepmind has already entered the market with Alphaforl3 that looks at ALL modelcules. AI is fast becoming a useful tool in many areas of medical research. Research, in general, benefits from AI’s use at every stage of the research process.
Wellbeing may see out with the realm of AI, but we have websites such as Psychologist and Replica that have millions of hits, often delivering credible CBT therapy. It would seem that the anonymity, availability 24/7 and convenience of the experience is what people like.
5. Productivity
I won’t dwell on this but increasing general administrative and clinical administration is probably the quickest win. AI can speed up and automate routine tasks such as data entry through text to speech, along with preliminary data analysis, referencing patient data and so on, freeing up medical professionals to focus on more complex and nuanced aspects of diagnosis and patient care.
Take just one example; screening for clinical trials. The Unlu study (2024) looked at costly & time-consuming screening, using GPT-4 to identify inclusion & exclusion
The results showed that GPT-4 was closely aligned with expert clinicians (GPT4 97.9% -100% v Study staff 91.7% - 100%), with GPT4 performing better on inclusion criteria of “symptomatic heart failure”.
At the next level, AI can assist in resource allocation within healthcare facilities by predicting patient inflows and identifying which patients require immediate attention versus those who can safely wait. Productivity increases through AI tools and systems can also help optimise hospital workflow and resource management, ensuring that resources are allocated efficiently and that patients receive timely care. This can improve outcomes and reducing hospital stays, improving both patient outcomes and operational efficiency.
6. Training
AI is like water in training. It soaks into every aspect of learning - learner engagement (hundreds of millions using it), learners support (used by learners to accelerate their learning), performance support (learning in the workflow) and assessment.
When I did ‘AI in learning’ work for Health Education England, I worked with Alan Ryan, the CEO. He said something interesting, that “all jobs in healthcare are essentially apprenticeships”. That’s a more profound statement than many realise and it stuck with me. Healthcare professionals learn much on the job and need to update their knowledge and skills regularly. At the moment this is all rather haphazard. We’re involved in AI learning projects delivered to hospitals. This is where the action is, not on using AI to design and carpet bomb people with more ‘courses’.
We can also see progress in real organisations, like Moderna, where a suite of 400 chatbots have been implemented across the organisation to increase productivity and quality of output. The rise of expert multimodal chatbots is now here with GPT4o. The multimodality of this technology also lifts it out of the all-too-common text world, as medicine is a profoundly visual subject. Virtual patients as avatars that you can engage in dialogue are already with us, lots more is to come in terms of skills training.
In a world where we have sometimes extreme shortages of nurses, doctors and healthcare professionals, automation is a solution for some tasks but so is faster and more effective training.
Conclusion
Healthcare is already benefitting from AI in almost every area of medical endeavour. This is likely to accelerate, with new ways of dealing with inputs of data and forms of presentation, as well as faster and more accurate investigation, less misdiagnosis, and certainly more targeted and better treatments. Additional benefits accrue in general productivity and training. The era of AI healthcare is upon us and the direction of travel is clear – the UNIVERSAL DOCTOR.
Bibliography
Ayers et al. 2023. Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Platform
Haim, A., Katson, M., Cohen-Shelly, M., Peretz, S., Aran, D. and Shelly, S., 2024. Evaluating GPT-4 as a Clinical Decision Support Tool in Ischemic Stroke Management. medRxiv, pp.2024-01.
Karthikesalingam A. and Natarajan V., Research Leads, Google Research (2023) AMIE: A research AI system for diagnostic medical reasoning and conversations.
Mei, Q., Xie, Y., Yuan, W. and Jackson, M.O., 2024. A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), p.e2313925121.
Unlu, O. et al.., 2024. Retrieval Augmented Generation Enabled Generative Pre-Trained Transformer 4 (GPT-4) Performance for Clinical Trial Screening. medRxiv, pp.2024-02