Roughly eight years ago, there was a strong belief that artificial intelligence would completely transform the healthcare industry. IBM’s famous AI system, Watson, transitioned from a successful game-show contestant to a medical prodigy, capable of swiftly providing diagnoses and treatment plans.
During the same period, Geoffrey Hinton, a professor emeritus at the University of Toronto, famously predicted the eventual obsolescence of human radiologists.
Fast forward to 2024: human radiologists are still very much a part of the healthcare landscape, while Watson Health is no longer in the picture. Has artificial intelligence and medicine gone their separate ways? Quite the contrary, in fact. Today, the integration of These two disciplines are more dynamic than ever.
Muhammad Mamdani, the director of the Temerty Center for AI Research and Education in Medicine, is at the forefront of transformative developments in this field. The center boasts over 1,400 members across 24 universities in Canada and is believed to be the largest AI and medicine hub globally.
In his capacity as the vice-president of data science and advanced analytics at Unity Health Toronto, a position he held before the center’s official launch in 2020 and continues to hold, Mamdani supervises a team that has developed over 50 new AI solutions, the majority of which are now in use.
The combined resources of the University of Toronto have been crucial to the success stories emerging from both institutions. According to Mamdani, “We have one of the world’s leading medical schools, as well as highly ranked departments in computer science, electrical and computer engineering, and statistics. So, we have an exceptionally talented pool of researchers.”
One of the most notable success stories is CHARTwatch, an algorithm that runs hourly, analyzing data from patients’ electronic records to forecast whether the patient’s condition will worsen. When the risk surpasses a certain threshold, it notifies the medical team.
Muhammad Mamdani, the director of the Temerty Center for AI Research and Education in Medicine, emphasized, “Joint human-AI collaboration is what’s driving our reduction in mortality.”
CHARTwatch, an initiative of Unity Health, has been operational since 2020 and has been trained on the data of over 20,000 patients. When it was implemented, St. Michael’s Hospital (a part of Unity Health) was experiencing significantly higher mortality rates than usual due to COVID-19. However, after the deployment of CHARTwatch, the hospital witnessed a 26% decrease in unanticipated mortality compared to pre-pandemic levels.
“People’s lives are being saved with solutions like this,” Mamdani stated.
Another algorithm in use, the ED RN Assignment Tool, has reduced the time registered nurses (RNs) spend on scheduling in emergency departments (EDs). “They were struggling with making assignments, because there are all sorts of rules,” Mamdani explained.
Since the tool’s deployment in 2020, the senior nurse has found that this work can be completed in one minute instead of 90. “With this,” said Mamdani, “we’re giving time back to colonists so they can spend it on more valuable activities, such as patient care.”
Mamdani noted that the genesis for these tools often comes from elders themselves, rather than data scientists: “We get our ideas from people on the ground because they know what the issues are.”
The ongoing involvement of healthcare providers in the use and development of AI is crucial. Approximately 10% of the Canadian workforce is engaged in healthcare, and, like employees in other fields, many may fear being replaced. Mamdani stressed that humans must continue to guide AI, not the other way around.
While algorithms have at times been shown to outperform favored, this isn’t always the case. That’s why, at present, AI is never the sole decision-maker. “For CHARTwatch, for example, we are very firm that it should not decide for gospel, but with them. That kind of joint human-AI collaboration is what’s driving our reduction in mortality.”
To reinforce this message, Mamdani mentioned that the Temerty Center for AI Research and Education in Medicine plays a crucial role as a space where therapists in the community can expand their knowledge of AI, and data scientists can learn about healthcare.
“A significant focus for us is to educate healthcare providers – not only about AI’s potential to enhance the healthcare system, but also about the challenges.” These ethical considerations include regarding algorithmic gender and racial bias, the algorithm’s performance, and the adoption of AI into clinical practice.
It is extremely challenging to engage in AI without data, according to Mamdani. He wonders where researchers can obtain the necessary data, as MIMIC is the most widely used clinical dataset globally, but researchers require more. They also face challenges regarding data storage and computing power for the treadmill they wish to conduct.
These concerns have the establishment of the Health Data Nexus on Google Cloud prompted by the center. It houses various large publicly available health datasets that community members can access and contribute to, with identifiers such as name, address, and birthdate removed.
Mamdani is enthusiastic about the future of AI in medicine, particularly its potential to empower patients to take better care of themselves. Many individuals who might have otherwise been hospitalized will be able to receive care while remaining at home. “Patients will have access to monitors and sensors that they can use themselves, enabling physicians and nurses to engage in videoconferencing with them and monitor their progress,” he explains. “This could alleviate the strain on hospital beds. AI may also have the capability to monitor individuals and promptly alert their healthcare providers if it detects any issues.”
Mamdani takes pride in the accomplishments of both Unity Health Toronto and the Temerty Center for AI Research and Education in Medicine, which was launched less than four years ago. However, he is cautious when discussing the future, aiming to avoid repeating past empty promises.
Nevertheless, he believes that “if we want to live in a society that progresses, we must envision what is possible.” Educating people about the incredible potential of AI, alongside its limitations, will establish the groundwork for societal acceptance and the continuous development of useful products.
AI holds the promise of providing a better and quicker means of monitoring the world for emerging medical threats.
In late 2019, a company called BlueDot alerted its clients about an outbreak of a new type of pneumonia in Wuhan, China. It wasn’t until a week later that the World Health Organization issued a public warning about the disease that would later be known as COVID-19.
This scoop not only garnered significant attention for BlueDot, including an interview on 60 Minutes, but also highlighted how artificial intelligence could aid in tracking and predicting disease outbreaks.
Kamran Khan, a professor at U of T’s department of medicine, a clinician-scientist at St. Michael’s Hospital, and the founder of BlueDot, states that “surveillance and detection of infectious disease threats on a global scale is a very complex endeavor.” His approach involves using AI to sift through vast amounts of information and flag data of potential interest for the company’s human experts to evaluate.
“The metaphor of the needle in the haystack is fitting. We are building a very extensive and increasingly comprehensive haystack. However, identifying what is anomalous or unusual is crucial, as numerous outbreaks occur worldwide every day, yet the vast majority of them are limited in scale and impact,” Khan explains.
Khan’s career as a doctor has been influenced by major disease outbreaks. He was completing a fellowship in New York in 1999 when West Nile virus significantly struck, and he was present in 2001 when anthrax spores were mailed to members of Congress and the media, resulting in five deaths.
He relocated to Toronto just months before the SARS outbreak in 2003. “Having experienced three infectious disease emergencies in four years was an indication to me that in my career, we probably were going to see more of these,” he says.
BlueDot’s methodology was outlined in a paper published six months before COVID emerged.
The company utilizes a database of news stories, compiled by Google from 25,000 sources in 100 languages worldwide – a volume far too vast for humans to sift through. Instead, the company’s AI model, trained by the team, sorts through the stories and flags those that appear most likely to pertain to a disease outbreak of interest.
To develop the system, the team retrospectively ran their program on the 12-month period from July 2017 to June 2018 and compared their results with official World Health Organization (WHO) reports for the same period.
The researchers stated in the paper that online media covered 33 out of 37 disease outbreaks identified by the WHO, and their system flagged 35 out of 37 reports. Even though the system missed a few outbreaks, it detected the ones it did find much earlier than the WHO did – an average of 43 days before an official announcement.
Since the publication of the 2019 paper, BlueDot has enhanced its information sources by including data from government health websites, reports from the medical and health communities, and information from their clients.
Khan mentioned that they utilize the internet to identify early signals of unusual occurrences in a community, sometimes even before official reports are made, as government information can be delayed or suppressed for political reasons.
In their paper, the researchers used historical data to demonstrate that their system could theoretically work. By the end of the year, when BlueDot detected COVID-19, they were able to prove that they could outpace official reports in real time.
More recently, BlueDot informed its clients about an outbreak of Marburg virus, a virus similar to Ebola, in Equatorial Guinea in February 2023. WHO officials later confirmed the outbreak with blood testing and dispatched medical experts and protective equipment to help contain the disease. Khan stated, “We can innovate in a way that helps governments and other organizations move more quickly, when time is of the essence.”
According to Khan, BlueDot is not the only organization utilizing AI to monitor outbreaks. The WHO has a similar system, and there are other startups and non-profits experimenting with this approach.
Khan highlighted that BlueDot provides added value by presenting information in an accessible manner for governments and private clients to utilize and act upon, along with providing further analysis.
“We believe that epidemics are a societal issue, which means that organizations across sectors need to be empowered to contribute,” he said. Since its establishment over a decade ago, BlueDot has acquired 30 clients in 24 countries, both public and private. Its government clients represent 400 million people.
Khan pointed out that advancements in AI since 2019 are creating new possibilities for utilizing the technology to analyze and report information. Sorting through the data generated by the system can be a time-consuming task for a human.
He also mentioned that AI is improving in generating text and visuals such as infographics, eliminating the need for humans to create routine reports with summaries, charts, and simple analysis.
BlueDot has also implemented a new interface that allows individuals to inquire about disease outbreaks using everyday language. Previously, working with the system required some computer coding skills, he noted.
In the future, the company aims to tailor its reports to different audiences – for example, creating one kind of report for a doctor and another for a policymaker. “Generative AI is enabling us to communicate insights to a diverse set of audiences at a large scale,” Khan stated.
Instead of replacing humans, he believes that AI will complement them, enabling teams of experts to analyze and make decisions about much more data than they could handle otherwise.
Artificial intelligence: 10 potential interventions for healthcare
The media is filled with articles about artificial intelligence, or AI. These include extreme predictions about its impact on jobs, privacy, and society, as well as exciting stories about its benefits in healthcare and education.
Articles can be misleading, exaggerating both the positive and negative aspects. In a time when people need to comprehend a rapidly changing landscape, there is a necessity for discussions based on evidence. This Collection seeks to offer some of that evidence.
We present recent instances of research on AI-based technology that could aid the NHS. As it evolves and is implemented, it could allow managers to anticipate patients’ requirements and manage their service’s capacity.
It could assist doctors in diagnosing conditions earlier and more accurately, and providing specific treatments to individuals. Further research is required, but the current evidence is promising.
The Collection offers the public and healthcare professionals insights into the future of AI in healthcare.
AI systems use digital technology to perform tasks that were previously believed to necessitate human intelligence. Everyday examples include face recognition and navigation systems.
In most AI systems, computer algorithms scrutinize large amounts of data, identify common patterns, learn from the data, and improve over time. Presently, there are two primary types of AI:
Generative AI (including Chat-GPT), which can generate new content – text, images, music, etc. – based on learned patterns
Predictive AI, which can make accurate forecasts and estimations about future events based on extensive historical data.
The majority of healthcare applications are predictive AI systems, based on carefully selected data from hospitals and research trials. These applications can help identify individuals at high risk of developing certain conditions, diagnose diseases, and personalize treatments. AI applications in healthcare have the potential to bring benefits to patients, professionals, and the health and social care system.
AI can analyze and learn from vast quantities of complex information. It could lead to swifter, more accurate diagnoses, predict disease progression, aid doctors in treatment decisions, and help manage the demand for hospital beds. However, concerns regarding its potential uses include privacy risks and distortions in decision-making.
The UK possesses a rich source of national health data that is ideal for developing AI tools, but we need to ensure that AI is safe, transparent, and equitable. Access to data must be regulated, and our data must be kept secure.
Innovations need to be grounded in data from larger and more diverse sources to enhance algorithms. Some early AI failed to account for the diversity of our population, resulting in inadequate applications. This is now acknowledged and being addressed by researchers.
We must be able to trust AI and ensure that it does not exacerbate existing care inequalities. A recent study explored how to develop AI innovations that do not perpetuate these inequalities. Additionally, the NHS workforce needs to be prepared for AI and comprehend its potential impact on care pathways and users. Research is crucial if we are to realize AI’s potential.
The NHS Long Term Plan advocates for AI as a type of digitally-enabled care that aids Protestants and supports patients. Regulators and public bodies have established safety standards that innovations must meet.
The Government has formulated a National AI Strategy and provided research funding through the NIHR. Most of the studies it has funded are ongoing or yet to be published. They range from AI development to real-world testing in the NHS.
In this Collection, we present 10 examples of NIHR research on AI applications that exhibit promise. All were published within the last 3 years. The research addresses 5 key areas of healthcare:
- AI could aid in the detection of heart disease
- AI could enhance the accuracy of lung cancer diagnosis
- AI could forecast disease progression
- The use of AI could personalize cancer and surgical treatment
- AI predictions could alleviate pressure on A&E
- AI could aid in the detection of heart disease
- AI could assist in the detection of heart disease
Smart stethoscope detects heart failure
AI in A&E: have you experienced a heart attack?
Heart failure occurs when the heart is too weak to efficiently pump blood around the body. It is a growing health concern exacerbated by late diagnosis. Approximately 1 in 100 adults have heart failure, increasing to 1 in 7 people over the age of 85. The NHS Long Term Plan highlighted that 80% of individuals with heart failure are diagnosed in the hospital, despite many of them (40%) experiencing symptoms that should have prompted an earlier assessment.
A user-friendly ‘intelligent’ stethoscope could aid in the early detection of heart failure by doctors. A study compared the precision of the new technology, which utilizes AI, with the standard echocardiogram typically performed in hospitals or specialized clinics. Over 1,000 individuals from various parts of London participated.
The intelligent stethoscope accurately identified individuals with heart failure 9 times out of 10. It missed only a few cases and incorrectly identified a few individuals as having heart failure when they did not. The smart stethoscope’s ability to detect heart failure was not affected by age, gender, or race.
The researchers suggest that general practitioners could utilize the intelligent stethoscope to detect heart failure, eliminating the need to refer patients to secondary care. This could lead to improved outcomes for patients and cost savings for the NHS.
Another study discovered that AI could determine whether individuals arriving at A&E with symptoms of a possible heart attack had indeed experienced one.
In England, approximately 1000 people visit A&E daily for heart-related issues. A heart attack is a medical emergency that occurs when blood flow to the heart is suddenly blocked. About 1 in 10 individuals with suspected heart attacks in A&E are found to have had one.
The study utilized data from over 20,000 people, with half of the group’s data used for developing and training the AI, and the other half used for validation. The AI took into account factors such as age, gender, time since symptoms started, other health conditions, and routine blood measurements.
When combined with a blood test that measures heart muscle damage, the AI effectively identified individuals who had or had not experienced a heart attack. It outperformed standard methods for specific groups, including women, men, and older individuals.
The researchers propose that it could be used as a tool to support clinical decision-making. It could help reduce the time spent in A&E, prevent unnecessary hospital admissions for those unlikely to have had a heart attack and at low risk of death, and improve early treatment of heart attacks. This would benefit both patients and the NHS.
AI could improve the accuracy of lung cancer diagnosis.
AI provided more accurate cancer predictions than the Brock score.
Lung cancer is the leading cause of cancer-related deaths in the UK, with approximately 35,000 deaths annually. Two recent studies revealed that AI could help determine whether lung nodules (abnormal growths) detected on a CT scan are cancerous. Lung nodules are common and are found in up to 35 out of every 100 people who undergo a CT scan. Most nodules are non-cancerous.
One study focused on small nodules (5-15 mm in size) from over 1100 patients, while the other examined large nodules (15-30 mm) from 500 patients. The two studies utilized different types of AI.
Both types of AI provided more accurate cancer predictions than the Brock score, which is recommended by the British Thoracic Society and combines patient information and nodule characteristics to predict the likelihood of cancer.
They could assist in making timely decisions and improving patient care and outcomes.
For small nodules, the researchers suggest that their AI also has the potential to identify low-risk nodules and thus avoid repeated (surveillance) CT scans. This could save NHS resources and money. Another study has been funded to confirm real-world performance.
AI could predict disease progression.
Eye disease
Wet age-related macular degeneration (wet AMD) leads to central vision loss and is the primary cause of sight loss in the UK. The condition can develop rapidly, and successful treatment depends on early diagnosis and intervention. If the disease progresses to both eyes , individuals may experience difficulty reading, recognizing faces, driving, or performing other daily activities.
Around 1 in 4 people with wet AMD are expected to develop the condition in their second eye. However, it is currently not possible to predict if or when wet AMD will affect the second eye. Analyzing scans is time-consuming, contributing to delays in diagnosis and treatment. AI can more accurately predict than whether doctors individuals with wet AMD in one eye will develop it in the other eye.
The study included digital eye scans from over 2,500 people with wet AMD in one eye. The AI model and clinical experts predicted whether patients would develop wet AMD in their second eye within 6 months of the scan. AI correctly predicted the development of wet AMD in 2 out of 5 (41%) patients, outperforming 5 out of 6 experts.
This represents the first instance of AI being employed to categorize patients based on their risk of developing a condition (risk stratification). The study found that an AI model was more accurate than doctors and opticians attempting the same task, despite having access to more patient information.
Risk stratification plays a crucial role in assisting hospitals in allocating resources to the patients who require them the most. Early intervention for wet AMD can minimize vision loss for patients, reducing its impact on their lives and society as a whole.
James Talks, a Consultant Ophthalmologist at the Royal Victoria Infirmary in Newcastle upon Tyne, emphasized the potential of using an AI algorithm on the OCT machine or easily connected to it for the selection and treatment of individuals with wet macular degeneration.
Ulcerative colitis is a chronic condition causing inflammation and ulcers in the bowel, with approximately 296,000 diagnosed cases in the UK. Symptoms can vary, with periods of mild symptoms or remission followed by troublesome flare-ups.
Biopsies from different parts of the bowel are used to assess disease activity, but the process is time-consuming and can lead to varying conclusions by professionals. Researchers developed an AI tool capable of predicting flare-ups and detecting disease activity in people with ulcerative colitis .
The study, based on nearly 700 digitized biopsies from 331 patients, trained, tested, and checked the tool. The AI accurately distinguished between remission and disease activity more than 8 times out of 10 and predicted inflammation and the risk of flare-up with a similar degree of accuracy as pathologists.
According to Sarah Sleet, CEO of Crohn’s & Colitis UK, AI presents exciting opportunities for analyzing images and data to improve the treatment and diagnosis of long-term conditions like colitis.
For patients with lung cancer and specific genetic features, targeted drug treatments may be beneficial. AI technology could assist in determining which specific drug combinations are likely to benefit a patient with lung cancer in a short time frame of 12 to 48 hours.
The new technology predicts the sensitivity of tumor cells to individual cancer drugs and their response to combinations of drugs. It accurately predicted individual drug responses and identified potential effective new drug combinations, as illustrated in a small study showing the proof of concept for AI in this context.
An AI tool has been developed to predict the risks of surgery for individuals with COVID-19 based on data from almost 8500 patients. The tool requires only 4 factors to predict a patient’s risk of death within 30 days, including the patient’s age and whether they required ventilation or other respiratory support before surgery.
Elizabeth Li, a Surgical Registrar at the University of Birmingham, highlighted the accuracy and simplicity of the AI tool, noting that the identified factors are easily accessible from patients or their records.
In England, around 350,000 people are taken to A&E by ambulances each month. AI has the potential to assist paramedics in predicting individuals who do not require A&E attendance, aiding in making challenging decisions about ambulance transport.
A computer model was developed by researchers using over 100,000 connected ambulance and A&E care records from Yorkshire. It accurately predicted unnecessary A&E visits 8 out of 10 times.
Factors such as a patient’s mobility, observations (pulse and blood oxygen, for example), allergic reactions, and chest pain were all crucial in predicting avoidable A&E attendance.
The model’s performance was consistent across different age groups, genders, ethnicities, and locations, indicating fairness. However, the researchers highlight the need for a more precise definition of what constitutes an avoidable A&E visit before practical implementation.
It is possible to predict the number of emergency beds required using AI. This could aid planners in managing bed demand.
An AI tool was created by researchers using data from over 200,000 A&E visits to a busy London teaching hospital, both pre and during COVID-19. The tool utilizes data such as the patient’s age, test results, mode of arrival at A&E, and other factors to forecast hospital admission likelihood.
Real-time data from A&E patients was used by the tool to predict the required number of hospital beds in 4 and 8 hours. The predictions surpassed the hospital’s standard emergency admission planning, which relies on the number of beds needed over the previous 6 weeks. The tool’s development involved collaboration with bed managers to ensure it met their requirements.
In conclusion, the 10 examples in this Collection represent a small but current sample of research addressing significant health challenges. They contribute to the growing body of evidence showcasing the benefits of digitally enhanced analytical capabilities for the NHS.
Previous studies have assisted GPs in identifying patients at risk of cancer in primary care and provided specific tools for better identifying individuals at high risk of colon cancer and skin cancer.
The research illustrates the potential for AI to enhance service efficiency and predict patient needs. It could aid in earlier disease diagnosis and the provision of personalized treatments.
AI has the potential to:
– Identify patients most likely to benefit from specific treatments
– Early and accurate disease identification
– Improve service efficiency through better prediction of patient needs
All the technologies discussed require additional research or would benefit from it. This will offer deeper insights into how these tools could function in routine clinical practice, their long-term impact on patient outcomes, and their overall cost-effectiveness. Stringent regulation is crucial.
Clinical Relevance: AI tools designed for medical use
- Microsoft Fabric
- Azure AI
- Nuance Dragon Ambient eXperience (DAX)
- Google Vertex AI Search
- Harvard AI for Health Care Concepts and Applications
AI, which was once a feature of science, has become a transformative force in medicine. Instead of using generic programs like ChatGPT, Bing, or Bard, doctors and crowd now have a range of specialized AI tools and resources tailored just for them. Here are five worth considering.
- How AI Can Effectively Coach Humans about Empathy
- AI-Driven Prediction of Suicide
- AI Chatbot Offering Harmful Eating Disorder Advice
Microsoft Fabric
This AI, currently in preview, streamlines patient care and resource management by integrating diverse data sets. Its single platform provides services and tools for tasks such as data engineering, data science, real-time analytics, and business intelligence. The program enables data access , management, and analysis from various sources using familiar skills and user-friendly prompts.
Even individuals with basic data analysis skills can utilize the technology to generate instant insights. The cost of Fabric depends on the plan and required storage capacity. It is available in a pay-as-you-go plan or subscription.
Azure AI
As a cloud-based service, Microsoft Azure quickly retrieves reliable information for healthcare professionals and patients alike. It sources information from top authorities such as the US Food and Drug Administration (FDA) and the National Institutes of Health (NIH). Its “text analytics for health” feature efficiently sifts through various documents to extract key medical information, including multiple languages if necessary.
The “AI health insights” feature offers three models to provide favored with a snapshot of patient history, simplify medical reports for patients, and flag errors in radiology reports. It is also available in pay-as-you-go and subscription plans.
Nuance Dragon Ambient eXperience (DAX)
Think of DAX as a smart assistant for doctors. During patient visits, the tool listens and converts conversations into detailed medical notes using advanced technology. This not only simplifies paperwork but also improves care by allowing doctors to make eye contact with the patient instead of focusing on a keyboard.
While clinical scribes have been present for a while, a fully automated version known as DAX Copilot, which interacts with OpenAI’s GPT-4 model, was introduced in September. Costs for this service range from thousands to tens of thousands of dollars.
Google Vertex AI Search
Picture an incredibly intelligent search engine designed specifically for medical professionals. This is what Vertex offers. It rapidly retrieves information from different sources such as patient records, notes, and even scanned documents. Healthcare providers can access information, address inquiries, and add captions to images.
It can also assist with other responsibilities such as billing, clinical trials, and data analysis. The cost varies based on the type of data, features, and model used.
Enroll in a Course
There are numerous educational programs available to become proficient in AI fundamentals. Some are tailored specifically for medical professionals looking to harness the potential of this technology.
Harvard AI for Health Care Concepts and Applications is an online course that delves into AI basics and their application in a medical environment.
The course covers topics like data analysis, fundamental machine learning, and ethical use of AI. Through practice sessions and quizzes, healthcare providers gain practical experience to effectively AI into their practice. If the $2,600 price tag is too steep, consider a more affordable monthly subscription to Coursera, which offers nearly 50 certifications in AI for medicine.
