A voice-based conversational artificial intelligence assistant successfully completed more than 86% of preprocedural telephone calls for patients undergoing cardiac catheterization and reduced the estimated nursing time required for patient preparation, according to a study published in npj Digital Medicine.

For endocrinologists evaluating conversational artificial intelligence (AI) tools, the study demonstrates the evolving role of large language model (LLM)-based voice assistants in supporting a range of clinical and administrative tasks.  

“Many specialties are facing a critical nursing shortage, and this study is highly promising in demonstrating the potential role of automation in reducing the burden of repetitive tasks such as preprocedural nursing calls,” Anoop Parameswaran, MD, MPH, a cardiologist at Mercy Hospital in Springfield, Missouri, told AACE Endocrine AI. “This approach also has the potential to expand into other preprocedural workflows.”  

Researchers at the Mount Sinai Health System evaluated “Sofiya,” a customized conversational LLM integrated into the cardiac catheterization laboratory workflow. The AI system was designed to provide procedure instructions, collect clinical information, answer common patient questions, and escalate issues requiring nursing review. Every AI-generated report and call transcript was subsequently reviewed and verified by a registered nurse before documentation entered the electronic health record (EHR). 

The prospective implementation study consisted of two sequential 90-day phases conducted between January 16 and July 17, 2025. During phase I, investigators optimized the AI system through rapid-cycle improvements developed collaboratively by clinicians and software engineers. Phase II evaluated real-world performance during routine clinical use under nurse supervision without additional system modifications. 

Overall, 1,431 patients scheduled for cardiac catheterization received 1,606 AI-assisted telephone calls during the 6-month study period. The patients had a mean age of 67 years, 28% were women, 52% were White, 17% were Black, 16% were Asian, and 13% identified as other ethnicities. Nearly three-quarters (73%) of the patients had previously been diagnosed with single- or multivessel coronary artery disease, and more than one-third had undergone a previous catheterization procedure. Baseline demographic and clinical characteristics were similar between the two study phases. 

Sofiya incorporated a customized conversational LLM trained on procedural and medical knowledge that could interpret varied patient responses, answer compound questions, and maintain natural conversation. Researchers constrained the model using domain-specific instructions, curated knowledge sources, and predefined escalation rules. The AI assistant was also instructed not to provide medication-related advice independently and instead to flag those questions for human follow-up. 

During each call, Sofiya verified patient identity, provided preprocedural instructions, collected key clinical information, answered patient questions or referred them to nursing staff when appropriate, and generated a structured report for nurse review before EHR upload.  

The primary study endpoint was the proportion of successfully completed calls, defined as those that reached the end of the script with patients answering all clinical questions and having no unresolved questions. Completed calls included those managed entirely by AI, those requiring protocol-driven nurse callbacks to clarify clinical findings, and those appropriately transferred to human staff because of patient preference or scheduling needs. 

During the initial optimization phase, Sofiya completed 696 of 806 calls (86%). Of those completed calls, 295 (37%) were resolved entirely by the AI system without requiring any nurse callback. Another 303 calls (38%) required only brief protocol-driven nurse follow-up, primarily to clarify medication use, positive clinical screening responses, or logistical issues. Ninety-eight calls (12%) appropriately transitioned to direct human interaction, including patients who declined to speak with AI, requested cancellation or rescheduling, or required communication through another person. 

Call completion remained stable after the optimization period. During phase II, Sofiya successfully completed 703 of 800 calls (88%), including 341 fully automated encounters (43%), 268 protocol-driven nurse callbacks (34%), and 94 appropriate transfers to nursing staff (12%).  

Weekly completion rates exceeded 90% during the final month after iterative refinements, including improved patient verification, advance text notifications, use of a hospital telephone number, and updates to the customized LLM.  

Incomplete calls declined modestly between study phases, from 110 (14%) during phase I to 97 (12%) during phase II. Most resulted from patients not answering or disconnecting before completing the questionnaire. AI-related failures also decreased from 48 calls (6%) to 24 calls (3%), largely because of fewer patient verification failures and technical improvements. 

Hallucinations were uncommon throughout the evaluation. During phase I, investigators identified six hallucinations (0.7%) that resulted in incomplete calls, corresponding to accurate responses to patient questions in 99.3% of calls. In one example, the AI incorrectly informed a patient that an overnight hospital stay would be required despite that information not appearing in its knowledge base. In another, the system provided an incorrect callback telephone number. Nurses reviewed these encounters, contacted the affected patients, and corrected the misinformation. No hallucinations leading to incomplete calls occurred during phase II. 

The researchers also assessed workflow efficiency using portal activity logs. Nurse-assisted preparation averaged 8.9 minutes per patient with the AI-supported workflow, compared with an estimated 20 minutes in the manual approach. Based on completed AI-assisted encounters, the researchers estimated annual savings of 26,869 nursing minutes, equivalent to approximately 37 12-hour nursing shifts.  

Among patients completing post-procedural surveys, the average satisfaction score increased from 95% in phase I to 98% in phase II. The highest ratings were for calls matching the procedure experience, aligning with patient expectations, and clearly explaining instructions.  

The study had several limitations. It included only English-speaking patients who did not require interpreters or caregiver-assisted communication, limiting generalizability to patients with limited English proficiency or those requiring family support. Speech recognition remained susceptible to background noise, accents, and variations in speaking style. Satisfaction surveys were administered only to patients who completed the AI interaction, introducing the potential for compliance bias because patients who declined AI or experienced technical failures were not surveyed. In addition, the patient satisfaction questionnaire had not been previously validated, and the estimate of manual nursing time was based on staff estimates rather than direct measurement. 

The researchers plan to expand the platform with multilingual capability, callback functionality, text messages containing directions and parking information, additional procedural workflows, EHR integration, and automated post-discharge follow-up. 

No conflicts of interest were reported. 

(Editor’s Note: The researchers noted that the study manuscript will undergo further editing before final publication, and there may be errors present that affect the content. All legal disclaimers apply.)