NEW YORK – While hospitals and doctors' offices have mostly been using artificial intelligence-based tools for documentation and clerical tasks, experts in the space expect an incoming wave of applications for in-depth analysis of electronic health records, improved guidance on testing and treatment, and more efficient monitoring of care.
In total, it could amount to a new future for the clinical use of AI.
At Cedars-Sinai in Los Angeles, Yaron Elad, chief medical informatics officer and cardiologist at the medical center, said that it has long used clinical decision support tools to inform decisions on patient testing. Those tools, he said, may soon be replaced with AI-based algorithms that can incorporate more parameters and provide better guidance.
While a rules-based tool might recommend treatments with consideration for a patient's high cholesterol, family history of heart disease, and diagnosis of diabetes, for example, an AI-based model could provide deeper analysis of a patient's chart and identify a conglomeration of symptoms and risk factors, he said. If that analysis noted that CAT scan results years prior had identified calcified blood vessels, for instance, those findings might trigger additional testing or follow-up care that the patient should have with his or her primary care provider.
Clinician interest in AI-based tools is on the rise. The American Medical Association said earlier this year that 66 percent of physicians surveyed in 2024 reported using at least one AI-developed tool in practice, up from 38 percent in 2023. Most often they used the tools to document tasks such as taking notes on patient visits or writing discharge summaries.
The clinicians were optimistic, though, about the potential of the technologies for both clinical and administrative functions. In the survey of 1,183 physicians, 75 percent of respondents in 2024 said that AI-developed tools could help with efficiency, 72 percent said that the technologies could augment diagnostic abilities, and 62 percent said they could help to improve clinical outcomes.
Despite the high expectations for the potential of AI tools, surveyed members also expressed worries about such technologies, with only 35 percent of physicians saying they are more excited than concerned about AI usage, whereas 40 percent said they were equally excited and concerned and 25 percent said that they were more concerned.
According to the survey, physicians said that AI technologies need to incorporate feedback loops, address data privacy concerns, integrate into workflows, and come with adequate training and education for their use. Almost half of physicians said increased regulatory oversight also would increase their trust in AI tools.
At electronic health record provider Epic, executives said that nearly every healthcare organization in the US is using AI in some capacity, and about two-thirds are using generative AI to create blocks of text to summarize clinical information.
Phil Lindemann, VP of data and research at Epic, said that AI-developed algorithms are used with the firm's EHR system to summarize patient clinical histories and test results, identify which patients need extra reminders for appointments, draft replies to patient messages, find diagnosis codes, and draft appeals when payors deny claims for medically necessary services. Among Epic's current projects is one evaluating the use of AI to help doctors review a patient's medical history ahead of a visit. By summarizing, for example, recent data from the patient's cardiologist and lab results, the tools can save physicians time and reduce their mental workload, he said.
Meantime, computer technology firm Oracle said in October that it was planning to launch in 2025 a revamped electronic health record that incorporates generative AI-based summaries of patient conditions and medications and offers physicians access to additional summaries of patient treatments, side effects, and notes from previous visits. Suhas Uliyar, senior VP of product development at Oracle, said in an email that the upcoming Oracle Health EHR system will be used to analyze myriad sources of patient data including tools that leverage conversational AI, offer treatment recommendations, and automate administrative tasks.
Conversational AI is typically used to mimic human interactions by simulating conversation, and it has been employed in software such as chatbots or virtual agents.
The new EHR will offer patient-specific chart summaries, analyze patient data to provide more precise treatment recommendations, review testing recommendations, and let doctors find, for example, a patient's three most recent HbA1c levels using a natural language-based search, Uliyar said. The firm is already seeing encouraging results from its Oracle Health Clinical AI Agent that is used to summarize charts and automate note taking, with physicians reporting they spend nearly 30 percent less time on documentation.
"Meanwhile, in hospitals, AI tools can help triage patients, provide care pathway optimization, and improve clinical decision-making and operational efficiencies using predictive analytics," Uliyar said. He added that labs can streamline their integration of diagnostic data and automatically provide physicians with relevant clinical records.
Expanding beyond rules-based models
In the in vitro diagnostics testing space, the use of AI may be most common in digital pathology. AI-based algorithms have proliferated in digital pathology and formed the basis of tests for various conditions such as pulmonary disease and sepsis. In addition to analyzing individual test results to aid the diagnosis of disease, healthcare firms see the potential to use AI-developed tools to provide broader analysis of patterns in patient health records.
Stephan Fihn is co-leader of University of Washington Medicine's Predictive Analytics Committee, whose work includes investigating new ways to deploy AI in healthcare and medical research. He estimates that healthcare providers have now developed tens of thousands of algorithms designed to predict patient outcomes, although far fewer have been deployed so far for clinical use.
He said that UW Medicine has spent enormous amounts of time testing AI-based alert systems in Epic's EHR system and integrating them into workflows. The system's sepsis model, for example, is used to alert physicians and nurses to consider implementing the sepsis protocol, and careful tuning is needed to provide alerts when interventions could help patients without overwhelming clinicians with false alarms.
Elad said that most of Cedars-Sinai's clinicians began using in the last year an AI-based tool to draft responses to patient messages, with tight guardrails to prevent it from proposing any testing, diagnosis, or treatment of disease. Fihn said that patients have been sending their doctors far more text messages since the pandemic, and physicians have been using such AI-developed tools to lighten the load.
Clinicians also have been using "ambient listening" models to dictate conversations during patient visits and generate notes in the EHR.
"It does allow me to focus more on really taking care of the patient, really listening to them, and really having the opportunity to think about their care," Elad said.
He expects that future iterations of those scribes could hear a doctor tell a patient with recent chest pain that they need a stress test and a cholesterol panel and respond by automatically teeing up draft orders for the physician to sign.
Epic's Lindemann said that a hospital in Ohio has been using Epic's AI-based tools to review text-based notes from radiology findings to identify incidental findings that may require follow-up, and some reports have resulted in the earlier identification of cancers. Other customers have also been using AI-based tools with the company's health records to reduce readmissions by identifying the patients who are most likely to need follow-up visits or reminders to adhere to their medication schedules.
While rules-based sepsis protocols have long been available, he said that more recently developed AI-based tools have been used to save lives by identifying sepsis and health deterioration sooner, resulting in faster ICU admissions and stabilization.
Expanding analysis of EHRs
Shounak Majumder, gastroenterologist and researcher for the Mayo Clinic, said that clinicians' notes contain a wealth of unstructured information, and last year he and 10 other Mayo researchers coauthored an article in Pancreatology on the identification of pancreatic cancer risk factors from clinical notes using natural language processing.
Majumder said that such a tool can help primary care providers to identify which patients should receive risk-based cancer screening by finding notes about family history and prior genetic testing. It also could be used to help identify patients who would benefit from genetic testing.
"A tool like this brings that information to the surface and enables a physician to act on it," he said.
He noted that while screening and early detection improves outcomes, pancreatic cancer is usually not diagnosed until it is in an advanced stage. While MRI and endoscopic ultrasound are typically used for screening, he noted that multiple companies have been developing blood-based tests for biomarkers of early-stage pancreatic cancers.
Mainz Biomed and Liquid Biosciences, for example, announced last month that they had formed a collaboration to develop a blood-based early detection test for pancreatic cancer. Immunovia also recently said that it is preparing to launch this year an antibody-based ELISA for identifying early-stage pancreatic ductal adenocarcinoma.
Majumder also said that AI-based tools are good at finding and presenting information that would be relevant to a physician examining a patient in the ER presenting with debilitating abdominal pain. How these tools are used, though, will depend on whether physicians see them as a benefit or another burden.
Meanwhile, Epic's Lindemann said that the company is focusing resources on the development of AI-based "agents," or applications that can orchestrate a series of clerical and clinical tasks for a clinician. A physician could use an agent to identify patients who are due for colorectal cancer screening and recommend either colonoscopy or the use of Exact Sciences' Cologuard test. If the Cologuard results show an elevated risk of cancer, the agent could use the results to automatically alert the patient and propose times for a follow-up colonoscopy, he said.
"Think of a physician having a team of these virtual agents that are carrying out some of these really standard care pathways for cancer care, like colon cancer or breast cancer screening," he said.
Those agents also could be used for scheduling routine blood work or arranging follow-up messages to patients.
Elad said that Cedars-Sinai also has been working with a vendor to develop software that can generate narratives for patients on their lab results. Other projects involve augmenting the interpretation of radiology and electrocardiogram results.
Peter DeVault, VP of interoperability and genomics at Epic, said he hopes that AI-based tools will also help to improve access to genomic testing through better analysis of coverage requirements and clinical guidelines as well as alerts that tests are available. AI-based tools also could be used to support genome-wide and phenome-wide association studies that could lead to new predictive models that incorporate genotypic, phenotypic, and EHR data.
A note of caution
Some voices in healthcare have called for a cautious approach, though, to make sure that AI technologies live up to their potential.
Christian Rose and Jonathan Chen of Stanford University wrote last year in NPJ Digital Medicine that the electronic health record was once heralded as a way to reduce medical errors and improve efficiency, yet it has since increased administrative burden and burnout among clinicians. They called for caution to avoid repeating those mistakes with AI-developed tools.
Poorly designed interfaces disrupt workflows and frustrate clinicians while information overload, alert fatigue, and the complexity of EHR systems contribute to missed diagnoses and care inefficiencies. Extensive data entry requirements cut into patient care time and contribute to exhaustion, as well.
"While these technologies may help prevent errors in specific scenarios, their widespread use has inadvertently hindered patient safety — the very thing they were meant to improve," they wrote.
While the integration of AI-powered tools into EHRs is seen by some as the way to realize the benefits envisioned from EHRs, the authors said that user-centered design, data standards, ongoing refinements in response to user feedback, and user training are needed.
Elad said that how well AI-based suggestions are received by physicians depends on whether they are presented at the right time in the visit to provide relevant information. Alerts that are too early or late or occur during unrelated tasks contribute to alert fatigue.
Healthcare providers need to ensure that AI-developed models are trained on appropriate populations and designed to minimize the potential for "hallucinations," he said. Cedars-Sinai has developed AI councils for physicians, nurses, lab staff, and operations staff, and he noted that clinicians remain in charge of any decisions on patient care.
"While we're really excited about it, we also have to still keep our guard up and really put AI to the same kind of rigorous testing that we put any kind of new technology" through, Elad said.