Potential uses of AI for perioperative nursing handoffs: a qualitative study

Setting

Barnes-Jewish Hospital is a 1400 bed academic medical center in St Louis, Missouri where approximately 20,000 surgeries are performed annually. We focused on the Acute and Critical Care Surgery (ACCS) division, which performs approximately 1600 inpatient surgeries annually, primarily trauma and acute abdominal surgery. All postoperative patients (other than those directly admitted to intensive care) recover from anesthesia in the PACU, a 30-bed area. Four hospital units subsequently care for ACCS patients: two dedicated hospital wards and two high- dependency units which are shared with otolaryngology, abdominal organ transplant, and hepatobiliary services.

Surgical cases were selected for direct observation from the OR schedule based on the primary surgery service (ACCS). Patients likely to be admitted to high dependency units based on their procedures were also included. Interviews and direct observations were conducted under Washington University IRB approval (#201812137 and #202009066) with the consent of the PACU nurse to shadow their activities. Written consents were changed to verbal consents with electronic provision of study information during the COVID-19 pandemic.

Participants

Participants included the OR circulator nurse, anesthesia provider, surgery provider, PACU nurse, and wards nurse caring for these patients. Concurrently, a convenience sample of PACU and ward nurses was recruited from the above units (ACCS floor and high-dependency wards). Interviews were conducted until topic saturation.

Description of Perioperative Handoff Processes and Care Teams

The handoff process is illustrated in Figure 1. Prior to surgery, a preoperative holding area nurse completes a health status inventory in the Epic EHR and on a paper record (Appendix S1) passed to PACU. An informal handoff is completed between the preoperative nurse and OR circulating nurse. After surgery, the patient is transported to the PACU by a surgery resident or fellow, the OR circulating nurse, and an anesthesia provider (nurse anesthetist or anesthesiology resident). OR to PACU handoff follows a protocol (Appendix S1) where the circulating nurse, surgeon, and anesthesia provider each give handoff to the PACU nurse.

• Download figure
• Open in new tab

Figure 1:

Illustration of perioperative handoff stages

Once a patient is deemed ready to leave the PACU, handoff is given from the PACU nurse to the ward nurse either at the bedside (high dependency unit) or by phone call (other units). A guideline addresses the handoff between PACU and the wards (Appendix S1). Postoperative patients are jointly managed by fellows, resident physicians, nurse practitioners, and the attending surgeon. The nurse practitioner or resident implementing ward care is not directly involved in the surgery. For clarity’s sake, we will refer to that resident or nurse practitioner as the ward physician.

Data collection

Eligible patients were identified from the operating room status board between 9am and 5pm on weekdays, and all eligible patients were attempted for observation. A researcher approached PACU nurses before patient arrival for consent to shadow their handoff activities. The participants were observed, and unstructured notes on the OR-PACU handoff process were taken. For patients where the OR-PACU handoff was missed because of insufficient time to identify the patient’s bed assignment, the nurse was approached for consent to shadow the PACU-ward handoff. We asked the PACU nurse to alert us when a handoff was imminent. During handoff observation, we checked for completeness of information along the standard SBAR format, [27] roughly following the same content as prior PACU-specific handoff guides.[40] This (ref 40) checklist was included as a pilot for future studies related to our institutional context, not as primary data collection; we don’t report the resulting scores here. We also assessed if the handoff-giver seemed able to answer the receiver’s questions.

Recruitment for semi-structured interviews occurred during the observation sessions. Our interview questions were focused on handoff communication, patient assessment, physician communication, and potential roles for AI. Interviews were conducted over phone or voice- application (interview guides in Appendix S2). Interviews were recorded and transcribed verbatim; after selection of example quotations, stop words and verbal pauses were removed.

Analysis

Interviews and observation notes were double-coded by two researchers (King and Shambe) using a mixed inductive-deductive thematic analysis approach.[41] After each interview, the transcript was reviewed, and topical statements identified. Each statement was coded as relevant to OR-PACU, PACU-ward, or both handoffs. Open codes and subthemes were analyzed within handoff type with an inductive hierarchy and revisited to form into major themes. Statements were also coded for relevance to major study questions (Appendix S3), forming the deductive part of the analysis. Codes were compared between coders, and disagreements discussed until consensus. A third researcher (Abraham) was planned to resolve disagreements, but this was not necessary. The OR-PACU and PACU-ward coding trees were compared for similar themes that could be coalesced. Themes were reviewed and revised with Prof. Abraham, and statements not matching a theme reviewed for additional subthemes. After construction of the coding tree, statements were checked to validate their applicability to the higher-level themes. After 10 interviews, a first round of coding was done, and saturation assessed and found to be likely as most topics were addressed by multiple participants; no new topics were found to emerge after the 11^th interview.

A consolidated criteria for reporting qualitative research (COREQ)checklist [42] is included in appendix S4 as a qualitative research reporting framework along with some additional methods details.

Nurse understanding of patient condition guides care decisions

Participants stressed that their bedside presence allowed rapid detection and hopefully mitigation of complications. They universally agreed that their understanding of the issues facing a patient modified what signs and symptoms they were alert for (Table 1 line 3), what issues they communicated to the PACU or ward physician (Table 1 line 4), and what treatments they recommended. Several participants stated that although almost all treatment changes required a team discussion, their recommendations were likely to be considered or acted on.

Handoffs are important to nurse situational awareness; problem focus and information transfer may be improved by AI

Participants agreed that problem-focused handoffs with anticipatory guidance were extremely useful, but that many topics in handoff were not relevant or recited data without context (Table 1 line 6). Our observations matched this claim: when a problem or complication was under active management, it was likely to be explicitly discussed, but only a fraction of handoffs explicitly discussed potential problems. Closely related to this concern was a lack of shared priorities between the handoff giver and receiver. It was frequent for participants to describe receiving handoffs focusing on details they found to be irrelevant or unintelligible or to not value topics on which their counterparty asked questions (Table 1 line 7). Almost all participants agreed that if AI identified a patient at high risk for a complication, that this topic would be prioritized for discussion at handoff, and that those receiving handoff would ask follow-up questions regarding patient state and the current plan (Table 1 line 8).

Participants stressed that accurate handoff was a critical way to learn about the patient’s state, expectations for recovery, and needs especially in the high-turnover environment of PACU (Table 1 line 9). However, they acknowledged barriers where the documentation they relied on was incomplete (Table 1 line 10), the handoff-giver did not know the relevant information, or they did not understand what needed to be conveyed. Several participants commented on how AI risk prediction at handoff might mitigate mismatch between handoff givers and receivers. First, a high calculated risk could alert them that a known comorbidity was more severe than they expected (Table 1 line 11). Second, awareness that a patient was overall high-risk would prompt nurses to closely review all available data and prioritize careful patient evaluation (Table 1 line 12). Third, automatic identification of EHR data elements which increased the patients’ risk could mitigate data omissions, especially if that data was in an unusual location (Table 1 line 13). Although several participants gave examples of how they might relate data given at handoff to specific AI-identified problems (ameliorating the laundry-list type handoff of Table 1 line 7), none explicitly identified using the AI-identified problems to organize data.

AI may augment nurse care decision making and team communication

PACU handoff is a critical time for establishing joint plans and physician communication needs. We also noted this in our direct observations; surgery and anesthesiology physicians made requests for notification of symptom or physical status changes and occasionally gave explicit if- then anticipatory guidance. However, ward participants noted that physicians only rarely proactively contacted them, leaving nurses to deduce what issues required communication or nursing action (Table 1 line 15). In our observations, we noted the absence of ward physicians in the overwhelming majority of handoffs. Some participants noted that AI could help target post-handoff nurse-physician communication in two ways. First, if a patient had been identified as high risk, the barrier to contacting the physicians to discuss that topic would be lowered (Table 1 line 16). Second, the nurse’s view of patient risk might be difficult to communicate, and AI-based pattern matching would make this more concrete and easier to request physicians act on or personally evaluate (Table 1 line 16).

Participants noted incomplete physician documentation and other EHR information negatively affected their independent assessment of the patient (table 1 line 17). AI identification of alternative key data would then be valuable. Additionally, AI-identified risks for adverse events would allow the nurse to better target their assessment and monitoring independent of any effect on handoff (Table 1 line 18). Participants noted that AI identified elevated risks could allow them to target interventions within their scope of practice, such as fall prevention, delirium prevention, and pneumonia prevention (Table 1 line 19). Multiple participants endorsed the desire for more accurate prediction of patients likely to require higher nursing workload or ICU transfer, which they could use to allocate their resources.

User experience and information overload are likely barriers to using AI

Participants identified several barriers for nursing use of AI, largely centered around the user experience and potential for excessive information volume. First, because of the large number of different methods for accomplishing most tasks in Epic, participants did not recommend the same locations for finding AI risk prediction. Second, preferred visualizations also differed, with participants variously endorsing absolute risk estimates, relative risks, simplified high-medium- low risk flags, and plots. Several participants noted that existing clinical decision support and alerts already generate alarm fatigue, and that additional flags would likely be ignored (Table 1 line 23). Finally, participants noted the potential for information overload with more complex outputs (Table 1 line 24).

Other barriers to perioperative handoffs

Participants noted many other barriers and several facilitators to perioperative handoffs and patient care. These other findings, mostly reflecting local practices, are documented in appendix S5 as they are outside the scope of the goals of this study.

Limitations

Our study drew participants from a single center, which limits the range of experiences. The ward nurses worked in a small number of units, limiting the diversity of views. The setting was an academic medical center, so the views may not reflect the experiences of those in private practice. Our interview was semi-structured (but nearly structured), and participants were informed on the nature of our study. They may have endorsed ideas to be agreeable, but participants seemed to feel free to disagree.