A voice-based artificial intelligence system completed most pre-procedural calls for patients awaiting cardiac catheterisation while maintaining nurse oversight and high satisfaction among respondents. Findings from a prospective two-phase implementation at Mount Sinai Health System in New York were published in npj Digital Medicine. The assistant provided logistical and procedural instructions, collected clinical information and answered supported questions or redirected patients to a nurse. Over six months, it made 1,606 calls to 1,431 patients, with completion rates remaining above 86% as system errors declined.
Must Read: Clinical AI Moves from Prediction to Action
Automating Routine Preparation
Before implementation, a registered nurse called each patient the day before the procedure to provide arrival, parking, fasting and post-procedure information. The nurse also collected information on medicines, allergies, dialysis, diabetes, fever and other relevant clinical factors, answered questions and entered a note into the electronic health record. The AI-assisted workflow transferred the repetitive parts of this process to the voice system while retaining clinical review and escalation.
The assistant initiated calls according to an administrative schedule and automatically redialled patients when necessary. After identity verification, it delivered basic information, collected responses to a defined clinical questionnaire and addressed questions using customised medical and procedural knowledge. It could pause while patients wrote down details, repeat information and wait when requested. Medication advice remained outside its permitted scope, and related questions were directed to nursing staff.
Each call generated a transcript, summary and clinical variables for review on an online portal. Calls were classified by status so nurses could prioritise incomplete encounters and those requiring verification before reviewing fully completed calls. A nurse checked every encounter for accuracy, with particular attention to clinical data, made callbacks when clarification was needed and verified the generated report before it entered the electronic health record. No AI-assisted encounter proceeded without registered nurse review. Nurses received individual training from experienced catheterisation laboratory colleagues, and most were able to use the system independently within one or two days.
Completion Improves Through Iterative Changes
The first 90-day phase combined clinical use with rapid system improvements. During this period, 806 calls were made and 86.4% reached the end of the script with all questions completed. The weekly completion rate rose from 58.3% to 92% as changes were introduced. These included simpler identity verification, improved date-of-birth validation, advance text notifications and use of a hospital number. The customised language model was updated.
In the following 90-day period, nurses led routine operation without active optimisation. Of 800 calls, 87.9% were completed. Fully automated calls requiring no nursing callback increased from 36.6% in the first phase to 42.6% in the second. Around one-third of calls in each phase required a protocol-driven callback to clarify a response, medication use or a patient question. A smaller proportion was transferred to human handling because of refusal to engage with AI, logistical or scheduling requests, another person answering or a request to speak with staff.
Incomplete calls accounted for 13.7% in the first phase and 12.1% in the second. Patient-related causes included unanswered calls and hang-ups. AI reliability issues fell from 6.0% to 3.0%, with temporary portal downtime responsible for most technical problems. Six unsupported or incorrect responses led to incomplete calls during the first phase, including incorrect information about an overnight stay and a wrong callback number. Nurses reviewed the affected calls and corrected the information. No such incidents led to incomplete calls during routine operation.
Nurse Oversight Remains Central
The AI-assisted preparation process averaged 8.9 minutes per patient compared with an estimated 20 minutes for the previous manual approach. Based on completed self-service calls and protocol-driven callbacks, the calculated annual saving was equivalent to 37.3 twelve-hour nursing shifts. The system therefore reduced time spent on repetitive communication and documentation while nurses continued to manage clinical clarification, exceptions and final validation.
Patient feedback was favourable among those who completed the call and questionnaire. Average satisfaction increased from 94.7% during optimisation to 98.1% during routine operation. High scores covered clarity, preparation, ease of understanding, comfort, detail and the handling of questions. However, the satisfaction questionnaire was developed locally and had not been validated. Only patients who reached the end of the call were surveyed, excluding those who refused AI engagement or encountered technical problems, which may have favoured positive responses.
The implementation also required institutional review of security, compliance, legal, ethical and AI-related risks. All calls remained subject to nurse attestation, and the system used curated knowledge, escalation rules and restricted instructions to limit inaccurate responses. The voice interface did not require a smartphone or internet connection, but speech recognition could be affected by noise, accents, speaking style and vocabulary. English-speaking patients who could take the call without an interpreter, family member or caregiver were included, limiting the applicability of the findings to people requiring supported communication or with limited English proficiency.
Voice-based AI completed most pre-procedural cardiac catheterisation calls during both optimisation and routine use, while registered nurses retained responsibility for review, escalation and report verification. Iterative changes improved patient engagement and reduced system errors, and the workflow shortened the estimated time required for preparation. Patient satisfaction was high among those completing the process, although the survey and eligibility criteria limit broader interpretation. The implementation supports automation of repetitive instructions and data collection, but also shows that clinical oversight, clear escalation boundaries, technical monitoring and attention to communication needs remain integral to safe operational use.
Source: npj Digital Medicine
Image Credit: iStock
References:
Kini A, Vengrenyuk A, NP DP et al. (2026) Conversational artificial intelligence for pre-procedural patient preparation: implementation, validation and patient satisfaction. npj Digit Med. https://doi.org/10.1038/s41746-026-02959-x
