AI in healthcare today covers a dizzying array of technologies with distinct capabilities. They serve different audiences too: One AI tool might answer patient questions, another helps clinicians interpret medical images, and a third assists researchers with drug discovery.
At the same time, many AI applications can run into the same underlying challenges. For example, both a patient-facing chatbot and a diagnostic imaging tool can bring up concerns related to data quality and potential algorithmic bias. To make sure AI plays a helpful role in healthcare, we need to evaluate the risks these tools have in common and the stakes attached to each use.
The many different healthcare applications of AI
AI is applied across many aspects of healthcare today, but its uses can be grouped into six broad roles:
1. Reducing administrative burden
Much of healthcare runs on administrative work. Clinicians and their teams document visits in electronic health records and complete prior authorizations. Other staff manage scheduling, billing, and records.
AI can help automate or simplify parts of this work by drafting clinical notes, summarizing unstructured data, assisting with coding and billing, routing requests, and performing many other routine tasks. This can lighten the workload for physicians and nurses at a time when they face high rates of burnout.
2. Detecting disease and predicting risk
A physician can only process so much information at once, but AI can analyze large volumes of images, laboratory results, clinical notes, genetic markers, and other data points to identify patterns. Finding subtle relationships in this data can help predict future health risks and detect abnormalities or early signs of disease for individual patients.
That said, AI still cannot replicate the full diagnostic process. Recent research found, for example, that large language machines (LLMs) performed better at identifying a final diagnosis than considering a range of possibilities to generate a differential diagnosis. While AI can serve as an important tool, human reasoning remains essential.
More broadly, population-level analyses performed by AI can help support infectious disease surveillance. Researchers have, for example, used AI technologies to help predict outbreaks by analyzing large amounts of social media data.
3. Personalizing treatment
Healthcare has traditionally relied on treatments that work for the average patient, but AI algorithms could help tailor care to each person's unique characteristics. Genetic information, medical history, biomarkers, and other clinical data could all be used to better diagnose individuals and match them with the most appropriate therapies. Tailoring treatments for the individual has the potential to improve outcomes while reducing unnecessary side effects and trial-and-error prescribing.
4. Assisting procedures and delivering treatment
AI is also being incorporated into technologies that interact physically with patients. For example, in robotic surgery, AI technologies can provide real-time guidance to help surgeons perform procedures more precisely. Researchers are also exploring microscopic robotic systems that could deliver drugs or genetic therapies directly to specific areas of the body, with the goal of making treatments more specific, minimizing systemic exposure.
5. Advancing medical research and drug development
Developing a new therapy can take years of research and testing. AI could help researchers move through that process more quickly by making it easier to spot promising possibilities in large amounts of data. AI may also speed clinical trials to get new treatments to patients.
6. Supporting patients beyond clinical settings
Patients are increasingly using AI tools to bring healthcare into their everyday lives. Many people already use virtual assistants and chatbots on healthcare websites and portals to get quick answers to routine questions. But AI can also support more continuous care by analyzing data from wearable devices for people managing chronic conditions. The collected data can also be shared with clinicians to identify changes that may need attention.
Common challenges behind healthcare AI
Although AI applications in healthcare vary widely, many bring up similar concerns:
Data and infrastructure
AI systems require access to large volumes of high-quality data, yet healthcare information today is often fragmented across incompatible systems. Many healthcare organizations will require better connected and more interoperable computing infrastructure to support AI reliably and at scale.
Reliability, “out of distribution” errors, and clinical utility
AI models don’t always perform as predicted across different healthcare settings or patient populations. They struggle with ambiguous clinical language or situations that differ from the data they were trained on. Models that have high accuracy rates in laboratory settings often fail to generalize to real-world hospital settings. Whether we’re talking about Epic’s 2022 sepsis detector or Google’s 2020 retina scanning model, differences between the lab and real world can cause dramatic differences in performance, highlighting the need for continuous monitoring and careful rollouts, particularly where clinical decision-making is involved.
Privacy, security, and accountability
Healthcare organizations must protect sensitive patient information while guarding against cybersecurity threats. They also need clear policies governing how data are collected and used, how AI systems are monitored over time, and who is responsible should an AI-assisted decision contribute to harm.
Fairness and access
AI systems can inherit biases from the data used to train them, potentially exacerbating racial and ethnic health disparities. Moreover, unequal access to AI technology may limit who benefits from technological advances.
Not all healthcare AI carries the same risk
Although many AI applications in healthcare share common challenges, their consequences can be very different. A wearable occasionally failing to record a step count is unlikely to have much consequence on a person’s wellbeing, but a monitoring system failing to detect a dangerous change in heart rhythm could be much more serious. Understanding AI in healthcare therefore requires looking beyond whether a particular risk exists to consider how serious its consequences are and what safeguards are needed to manage it.
Medication access offers a good example of how that principle can work in practice. Getting a patient started on a new therapy can require clinical teams to spend hours filling out prior authorization forms and reviewing supporting clinical information. At Cinnamon, automation and AI technologies help streamline much of that routine administrative work. These tools, however, never replace clinical judgment. The prescribing team is always required to review the information and remains responsible for submitting the request.
Whenever AI is used in healthcare, human responsibility and oversight must remain in place for any decision or action that could affect a patient’s health or safety. Safeguards should be matched to the stakes involved. The higher the stakes of a particular application, the more important strong data, careful validation, human oversight, and clear accountability become.
