An artificial intelligence (AI) tool can rapidly and accurately identify stroke patients contraindicated for thrombolytic drugs.

The results of a retrospective cohort study showed that the tool demonstrated high sensitivity and specificity in identifying contraindications to thrombolysis among 388 patients evaluated for acute stroke at a large US health system. 

“This and other AI tools are going to make the doctor’s job a lot safer, doing things that AI is much better at than physicians such as going through tons of charts,” study investigator and vascular neurologist, Bing Yu Chen, MD, Cleveland Clinic, Ohio, told Medscape Medical News. 

However, he added that the goal is not to replace physicians but to complement their work.

“Ultimately, the physician is always in control, the captain of the ship. AI is going to provide information with the physician making the ultimate decision,” Chen added.

The findings were presented at the International Stroke Conference (ISC) 2026 meeting. 

Catching a “Near Miss” 

Manual review of electronic health records (EHRs) to screen for contraindications to thrombolysis during stroke evaluation is labor intensive, time consuming, and prone to error. Among the more than 20 possible contraindications are anticoagulant use, a history of intracranial hemorrhage, and recent surgery.

The investigators developed an AI tool that uses a large language model (LLM) to scan EHRs in real time and flag contraindications. The system analyzed the 30 most recent provider notes preceding the stroke evaluation.

Administering thrombolytic therapy when there are contraindications can result in serious adverse outcomes. Chen cited a recent case involving a patient who presented to the emergency department with stroke symptoms within the therapeutic time window. 

Although the patient initially appeared to be an appropriate candidate for thrombolysis, the AI tool identified a recent gastrointestinal bleed, prompting reconsideration of treatment. 

If thrombolysis had been administered, the patient likely would have experienced significant bleeding, Chen said, who characterized this case as a “near miss.”

The investigators evaluated the tool in 388 stroke cases (median patient age, 69 years) across 16 Cleveland Clinic sites in Ohio. Of these individuals, 56.7% were female and 77.6% were White. All sites used the same EHR system. 

A total of 98 contraindications were identified among the 388 patients, with some individuals presenting with more than one. These included anticoagulant use, a history of unprovoked intracranial hemorrhage, ischemic stroke within the previous 3 months, and major surgery or serious trauma within 14 days.

The majority of patients (78.4%) had no contraindications, a proportion that “is similar to what the literature suggests,” Chen noted.

False Positives 

Of the 98 contraindications, the AI tool correctly identified 93, yielding a sensitivity of 94.9% (95% CI, 88.5-98.3).

Because the system missed five contraindications, Chen emphasized that clinicians should not blindly trust large language AI and should always double-check patient charts.

Those missed cases may be addressable by incorporating laboratory results and other external data, but Chen cautioned that adding too much information carries risks. An overloaded system, he noted, could become slower and potentially less accurate.

The study showed the tool had a negative predictive value of 100.0% (95% CI, 99.9-100.0). This, said Chen, means that if the AI determines a patient has no contraindications, it’s pretty much certain that’s the case.

The tool also demonstrated high specificity (99.1%; 95% CI, 98.9-99.3) and overall accuracy (99.0%; 95% CI, 98.8-99.2).

However, the system generated a substantial number of false positives, with a positive predictive value of 50.5% (95% CI, 43.1-58.0).

Chen explained that the system is designed to err on the side of caution and avoid missing potential issues, which can lead it to flag contraindications that are not clinically relevant. For example, the AI might identify a patient with an aneurysm, but a chart review could show that the aneurysm is too small to be of concern.

“AI sometimes flags something because it’s unsure, and that creates a false positive,” said Chen.

Advantages, Limitations 

However, the AI tool indicates the source of each flag, including the date and author of the note, making the findings easily verifiable by a physician, Chen said. He added that the system could be further trained to reduce the false-positive rate, an area the team plans to address in future work.

Secondary analyses did not identify differences in the tool’s accuracy by gender, race, or ethnicity, although Chen noted that the dataset was likely underpowered to detect such differences.

A key advantage of the tool is its speed. On average, the app required just 14 seconds to review an individual chart and generate an output report. The system is also inexpensive, with an annual cost of about $500, roughly $0.07 per patient.

The investigators plan to design a trial to assess whether the AI tool reduces physician workload, shortens door-to-needle time, and ultimately helps prevent complications in patients, Chen said.

Chen said he could envision the tool being applied in other clinical areas, including cardiology, trauma, and emergency medicine.

He also outlined several study limitations. In addition to the low positive predictive value, the AI tool currently analyzes only chart notes and does not incorporate laboratory data or radiology reports, which could identify additional contraindications. 

The study was conducted at a single center, limiting generalizability to institutions with different EHR systems, and the researchers evaluated only one LLM rather than multiple AI platforms.

Potential Diagnostic Challenges 

Commenting on the study for Medscape Medical News, Mark Alberts, MD, chief of neurology at Hartford Hospital in Connecticut, said the study provides a substantial amount of useful information and that the AI program appears capable of efficiently identifying many clinically important measures. 

He said the AI program appears capable of efficiently identifying many clinically important measures. However, he cautioned that decisions about whether to administer thrombolytic therapy in the setting of suspected acute stroke are complex and involve multiple considerations. 

He described real-world scenarios that can complicate time-sensitive decisions about thrombolytic therapy. For example, clinicians may need to determine whether neurologic deficits reflect an evolving stroke or a complex migraine, whether a metabolic abnormality is mimicking stroke, or whether a subtle CT finding represents acute ischemia or a small brain tumor. He added that it remains unclear whether, or how, the current AI program would address these diagnostic challenges.