Artificial intelligence (AI) and related technologies are rapidly gaining prominence across diverse industries, including healthcare. This progress holds tremendous potential for transforming patient care and optimizing administrative operations in healthcare organizations, encompassing providers, payers, and pharmaceutical companies.
Numerous research studies indicate that AI has the capacity to match or even surpass human performance in critical healthcare tasks like disease diagnosis. Presently, algorithms are already surpassing radiologists in detecting malignant tumors and assisting researchers in constructing cohorts for expensive clinical trials. However, due to several factors, we believe it will take several years before AI completely replaces humans across various medical processes. In this article, we explore the potential of AI in automating aspects of healthcare and highlight some of the challenges that hinder its rapid implementation in the field.
Types of AI of relevance to healthcare
Artificial intelligence encompasses a collection of various technologies rather than a single entity. While these technologies hold significant relevance to the healthcare sector, their applications and functionalities differ considerably. Below, we outline and explain specific AI technologies that play a crucial role in healthcare.
Machine learning – neural networks and deep learning
Machine learning, a popular form of artificial intelligence (AI), involves fitting models to data and training them to acquire knowledge. It is extensively used in various industries, including healthcare. Precision medicine, where treatment protocols are predicted based on patient attributes and contextual factors, is a common application of traditional machine learning in healthcare. Neural networks, a more complex type of machine learning, have been employed in healthcare research for several decades, aiding tasks like disease prediction. Deep learning, the most intricate form of machine learning, utilizes neural networks with multiple layers to predict outcomes. It has found applications in radiology for detecting cancerous lesions and in speech recognition for natural language processing. However, deep learning models often lack clear interpretability, making it challenging to explain their outcomes accurately.
Natural language processing
Since the 1950s, researchers in the field of artificial intelligence (AI) have been dedicated to understanding human language. This pursuit falls under the domain of Natural Language Processing (NLP), which encompasses various applications such as speech recognition, text analysis, translation, and more. NLP can be approached using two fundamental methods: statistical NLP and semantic NLP. Statistical NLP relies on machine learning techniques, particularly deep learning neural networks, and has led to notable advancements in recognition accuracy. For effective learning, it requires a substantial amount of language data known as a corpus.
Rule-based expert systems
During the 1980s, expert systems built upon collections of ‘if-then’ rules emerged as the dominant technology in the field of artificial intelligence (AI). These systems were widely adopted for commercial use during that era and subsequent periods. In healthcare, they found extensive application in the form of ‘clinical decision support’ systems over the past few decades, and their usage continues today. Many electronic health record (EHR) providers still incorporate a set of rules into their systems.
Expert systems rely on human experts and knowledge engineers to construct a series of rules within a specific knowledge domain. They perform effectively and are easily understandable up to a certain threshold. However, as the number of rules increases (often reaching several thousand) and conflicts arise among the rules, these systems tend to falter. Additionally, modifying the rules can be challenging and time-consuming, particularly when the knowledge domain undergoes changes. In the healthcare sector, there is a gradual shift towards approaches that leverage data and machine learning algorithms, gradually replacing the reliance on expert systems.
Physical robots
By now, physical robots have gained significant recognition, as evidenced by the annual installation of over 200,000 industrial robots worldwide. These robots are primarily employed in factories and warehouses, where they carry out predetermined tasks such as lifting, repositioning, welding, assembling objects, and even delivering supplies in hospitals. In recent times, robots have evolved to become more collaborative with humans, allowing easier training by guiding them through desired tasks. They are also becoming more intelligent, as additional AI capabilities are integrated into their operating systems. It is likely that physical robots will incorporate the same advancements in intelligence witnessed in other AI domains.
One notable application of robots in the medical field is surgical robots, which were first approved in the USA in 2000. These robots provide surgeons with enhanced capabilities, enabling improved precision, minimally invasive incisions, and suturing of wounds, among other benefits. Despite these advancements, human surgeons still make crucial decisions during surgical procedures. Robotic surgery is commonly utilized in various procedures such as gynecologic surgery, prostate surgery, and head and neck surgery, empowering surgeons with added capabilities and improving patient outcomes.
Robotic process automation
Robotic Process Automation (RPA) is a technology designed to execute structured digital tasks for administrative purposes, mimicking the actions of a human user following a predetermined set of rules or scripts. Unlike other forms of AI, RPA is cost-effective, easily programmable, and transparent in its operations. It should be noted that RPA does not involve physical robots but rather computer programs running on servers. By leveraging workflow, business rules, and integration with information systems’ presentation layer, RPA emulates a semi-intelligent user, performing tasks efficiently. In the healthcare industry, RPA finds application in repetitive tasks such as prior authorization, updating patient records, and billing. Additionally, when combined with technologies like image recognition, RPA can extract data from faxed images, enabling input into transactional systems.
Diagnosis and treatment applications
In the field of AI, there has been a long-standing focus on using technology for the diagnosis and treatment of diseases. Early rule-based systems like MYCIN showed promise but fell short in clinical practice due to their limited integration with medical workflows and record systems, as well as their inability to outperform human diagnosticians. IBM’s Watson, combining machine learning and natural language processing (NLP), gained attention for precision medicine but faced challenges in addressing specific cancer types and integrating into healthcare processes. Implementation issues persist in healthcare organizations, with rule-based systems lacking the precision of algorithmic machine learning systems that can handle vast amounts of genomic and omics-based data.
Despite these challenges, research labs and tech firms continue to develop AI approaches for disease diagnosis and treatment. Radiological image analysis has been a common focus, but other areas like retinal scanning and genomic-based precision medicine are emerging. Tech firms and startups are actively working on prediction models, image interpretation algorithms, and personalized treatment recommendations for cancer based on genetic profiles. Population health machine learning models are also used to predict disease risks and hospital readmissions, although data limitations can impact their effectiveness.
