Three years ago, on March 11, 2020, the World Health Organization (WHO), announced that the coronavirus that causes covid was spreading worldwide and that the outbreak was officially a pandemic. Three years later, there have been an estimated 1.1 million deaths from covid in the US (US Centers for Disease Control and Prevention). I wrote this article at the start of 2022. Has anything changed in that time, and what have we learned?  

This pandemic has wreaked havoc in healthcare, but the human spirit yearns for hope with each global catastrophe. 

As we turn to a new year, it is a good opportunity to reflect on the current state of artificial intelligence in healthcare as well as the future possibilities of artificial intelligence for health and disease. 

Data science, machine and deep learning, artificial intelligence, and a panoply of technological tools have had an impact on medicine and healthcare in several domains, especially in medical imaging and decision support. As the Covid-19 pandemic demonstrated, however, these tools were not as successful as clinicians had hoped. This observation was probably more about deficiencies in healthcare data, databases, and information technology infrastructure than it is for AI itself. Despite this observation, expectations remain high that AI and its technological tools will deliver in the long term. 

An impressive portfolio of technological tools are now available in the domain of artificial intelligence in medicine. By far the most mature appears to be deep learning in the form of convolutional neural networks (CNN) in medical imaging. The Cambrian explosion of CNN tools have made progress in static imaging, but are now starting to make inroads into moving images such as ultrasound studies, endoscopic imaging, and even echocardiograms. 

Both machine and deep learning have also made progress in electronic medical records – in projects on readmission criteria or decision support – but these have not been nearly as productive as medical imaging due to the records being fragmented in location and complex in nature. 

In addition, there is promise in the area of drug design or repurposing in treatment for cancer patients and even for covid-19 patients during the pandemic as a result of machine and deep learning, especially with protein structure determination based on genomic sequencing. 

Natural language processing (NLP) capabilities with transformer architectures such as the generative pre-trained transformer 3 (GPT-3) have started to be considered for deployment in healthcare. This technological tool of NLP continues to advance at an exponential pace, and unsupervised learning also holds great promise for the discovery of new phenotypic expressions of disease subtypes and treatment responses. 

Healthcare is starting to embrace an older AI technology of robotic process automation (RPA) for administrative tasks that can be automated by algorithms rather than completed by humans, but for data scientists, this past decade has been a journey into healthcare with mixed dividends. While the aspiration to help improve patients’ lives and/or create a viable business venture was a driving force for artificial intelligence experts, the nuances of access to healthcare data and inadequacies of databases was a deterrent for some. For clinicians at all levels of education and training as well as practice, there is an escalating need to learn about the basics of AI as it is becoming more evident that those clinicians who understand AI will have a growing advantage over those who do not. 

Projecting into the future, there will be exciting developments in the diagnosis and treatment of medical conditions. There is exciting work on pushing AI “peripherally” to devices – even at the microprocessor level. This artificial intelligence of things, or AIoT, provides a portfolio of “intelligent” devices for the future of chronic disease management as well as population health strategies. In short, AI in healthcare will be in two directions: a centralized cloud for analytics and concomitantly a peripheral network with AI embedded in many devices and sensors. This will be the AI equivalent of a brain and peripheral nervous system. 

The limitations and nuances of existing electronic medical records in their current state demands a disruptive technology in the future. A promising technology is graph and hypergraph databases coupled with knowledge graphs to create a paradigm shift in how electronic medical records are structured and curated. Federated learning consists of edge devices with local data that can train their own copy of the model from a central server, and only the parameters/weights from these models (but not the data) are sent to the global model. Multimodal AI, such as combining perception and linguistic capabilities of machines, can increase the potential for AI to deal with the complexities of healthcare. 

In the area of medical education and clinical training, adding an AI dimension to extended reality can be termed intelligent reality. Along with this virtualization of clinical medicine and healthcare can be AI imbued in the virtual twin concept for both the patient as well as the health system. All of this demand for artificial intelligence will warrant the availability of quantum computing. For AI experts, there will be an increasingly dire need for more talent, especially at the PhD level, to work in healthcare, but an escalating amount of automated machine learning will be accessible. 

AI alone in medicine is not going to make an impact long term unless it is applied “intelligently” with human clinician insight and intuition to render it truly meaningful.  For clinicians, adoption will need to be accelerated to accommodate the technology that is available. A small cohort of clinicians will need to be champions of AI by learning a minimal amount of knowledge to be conversant with a data scientist. Creative uses of AI in the future can include embedding knowledge into the EHR while gaining continuing medical education credits. The ethical and legal aspects of healthcare AI continue to be widely and publicly discussed and debated. 

Decreasing the human burden of labeling medical images will be in the form of innovations in artificial intelligence such as few shots learning and generative adversarial networks that can enable more automated interpretation in the future and cognitive elements of artificial intelligence such as 1) Joseph Voss’ cognitive architecture (declarative and procedure learning and memory, perception, action selection, etc), 2) Geoff Hinton’s “capsule networks”, or 3) Jeff Hawkins’ “reference frames” described in his book A Thousand Brains: A New Theory of Intelligence, will need to be increasingly a broad motif in artificial intelligence in medicine and healthcare that will incorporate the insights, intuition, and intelligence of our clinicians. The phrase “artificial intelligence in medicine or healthcare” will no longer be used as it was decades before.

Lastly, there needs to be AI systems that can perform real-time AI. For this to occur, AI architectures will need to be even more robust and will need to include AI tools such as anytime algorithms, decision-theoretic meta-reasoning, and reflective architecture. These new AI tools will also need to incorporate the nuances of complexity and chaos theory as biomedical phenomena often have complex rather than complicated elements.

In our struggle against this menacing pandemic, we can look at the future with optimism and idealism, and with artificial intelligence as one of our essential resources in our portfolio. 

The future of artificial intelligence in healthcare and many other topics will be discussed at our in-person AIMed Global Summit scheduled for June 4-7th of 2023 in San Diego. Hope to see you before then as well as there! Book your place now! 

We at AIMed believe in changing healthcare one connection at a time. If you are interested in discussing the contents of this article or connecting, please drop me a line – [email protected]