Nursing leadership and performance metrics in artificial intelligence integrated gastrointestinal surgical care: A narrative review

Table 1 Evidence synthesis of selected artificial intelligence studies relevant to gastrointestinal surgical nursing

Ref.	Setting/population	AI task	Validation and model performance	Implementation/clinical effectiveness	Applicability to nursing	Evidence relationship
Wang et al[5]	Systematic review and meta-analysis; 38 studies of gastrointestinal surgery	Prediction of postoperative outcomes	143 ML and 62 logistic-regression models; ML showed higher discrimination than LR (ΔAUC 0.07, 95%CI: 0.04-0.09); 44% of studies were at low risk of bias. Calibration and external validation were inconsistently reported across the underlying literature	Practical implementation and patient-level clinical effectiveness were not established	Relevant to GI surgical risk stratification and surveillance, but nursing-specific implementation and outcomes were not evaluated	Indirect
Bihorac et al[25]	Development/validation study; 51457 adults undergoing major inpatient surgery at a single center	Prediction of eight major postoperative complications and mortality	Repeated internal validation; complication AUCs 0.82-0.94 and mortality AUCs 0.77-0.83. Calibration was assessed. No external validation in this study	Designed for real-time perioperative use, but feasibility in clinical workflow required further prospective testing; clinical effectiveness was not demonstrated in this study	Potentially relevant to risk-based monitoring and escalation, but not GI-specific and nursing outcomes were not evaluated	Indirect
Bektaş et al[27]	Systematic review; 31 studies of colorectal surgery	Prediction of surgical outcomes using ML	Most studies were retrospective; external validation occurred in 4 studies (13%) and calibration was reported for 3 models (10%)	Routine clinical implementation and improvement in patient outcomes were not established	Directly relevant to colorectal surgical risk prediction; applicability to nursing workflow remains indirect because nursing-specific outcomes were not studied	Indirect
Bektaş et al[29]	Systematic review; 52 studies of hepatopancreaticobiliary surgery	Prediction of disease course, postoperative complications and other surgical outcomes	47/52 (90%) studies were retrospective and 5/52 (10%) prospective; external validation was generally absent	Authors identified external validation as necessary before clinical introduction; routine clinical effectiveness was not established	Relevant to procedure-specific risk assessment, but nursing implementation and nursing-sensitive outcomes were not evaluated	Indirect
Soria-Utrilla et al[34]	Retrospective observational study; 586 patients with colorectal cancer assessed before surgery	AI-assisted CT body-composition assessment	SMI discriminated low FFMI with AUC 0.82; sex-specific diagnostic cut-offs were derived within the study population. External validation was not reported	Diagnostic/morphofunctional assessment was evaluated; effects on downstream workflow or patient outcomes were not tested	May identify nutritional and functional vulnerability relevant to preoperative assessment and referral; direct nursing effects were not evaluated	Indirect
Cho et al[70]	Multihospital implementation study; 4 hospitals with different electronic record systems	One-hour fall-risk prediction with AI-powered decision support and tailored nursing recommendations	Locally optimized models were implemented across four sites; AUROC 0.8051-0.9581	Successfully integrated into different EMR systems. Nursing interventions increased, but fall rates did not significantly decrease	Provides direct evidence that AI can be incorporated into nursing workflow and outcomes tracking, but the setting was not gastrointestinal surgery	Direct nursing implementation evidence; non-GI setting

AI: Artificial intelligence; AUC: Area under the curve; AUROC: Area under the receiver operating characteristic curve; CT: Computed tomography; EMR: Electronic medical record; FFMI: Fat-free mass index; GI: Gastrointestinal; LR: Logistic regression; ML: Machine learning; SMI: Skeletal muscle index.

Table 2 Artificial intelligence applications across perioperative phases and nursing implications in gastrointestinal and complex surgical care

Perioperative phase	AI-supported function	Surgical context	Nursing implications	Measurable nursing-sensitive indicators
Preoperative assessment	Risk stratification using demographic, laboratory, imaging, nutritional, and comorbidity data[5,25-30]	Pancreaticoduodenectomy; major hepatectomy; colorectal cancer surgery; multivisceral resection	Identification of patients requiring intensified education, nutritional support, prehabilitation, early stoma counselling, or planned higher-acuity postoperative monitoring	Completion of risk-based education; timely nutrition or physiotherapy referral; documentation of individualized monitoring plan; predefined escalation criteria
Preoperative imaging review	AI-assisted detection of sarcopenia, visceral obesity, frailty-related imaging markers, or anatomical complexity[34,35]	Hepatopancreatobiliary and oncological gastrointestinal surgery	Translation of imaging-derived risk into practical nursing care, including pressure-injury prevention, mobilization planning, respiratory preparation, and discharge support	Documentation of frailty or nutritional risk; early mobilization plan; referral to allied health professionals; pressure-injury risk reassessment
Operating room and immediate recovery	Prediction of hemodynamic instability, bleeding risk, respiratory compromise, or need for higher-acuity care[25,30,35]	Esophagectomy; major liver resection; complex abdominal surgery	Structured handover from operating room or recovery unit to ward or intensive care, with clear monitoring priorities	Completeness of AI-informed handover; time from recovery discharge to first ward reassessment; documentation of key intraoperative concerns
Early postoperative period	Detection of abnormal physiological patterns and early deterioration[5-9,25,30]	Anastomotic leak after colorectal surgery; pancreatic fistula; postoperative bleeding; sepsis; acute kidney injury	Prompt bedside reassessment, review of drain output, wound status, fluid balance, pain pattern, urine output, oxygen requirement, and escalation to surgical team	Time from alert to nursing reassessment; time from reassessment to medical escalation; proportion of alerts with documented response; missed or delayed escalation events
Postoperative ward care	Support for risk-stratified enhanced recovery pathways[10-13,47-49]	Colorectal surgery; pancreatic surgery; hepatobiliary surgery	Adjustment of monitoring frequency, mobilization targets, pain assessment, nutritional support, and discharge readiness according to risk	Adherence to individualized observation schedule; completion of ERAS-related nursing interventions; documented reasons for pathway deviation
Discharge planning	Prediction of readmission, delayed recovery, dehydration, infection, or poor self-management[5,25,27,30]	New ileostomy; complex wounds; drains at discharge; frail older patients; cancer surgery	Targeted patient education, confirmation of understanding, community nursing referral, stoma nurse involvement, and early follow-up planning	Completeness of discharge education; documentation of warning signs; referral to community or stoma services; early post-discharge contact; avoidable readmission

AI: Artificial intelligence; ERAS: Enhanced recovery after surgery.

Table 3 Nursing-sensitive performance domains for evaluating artificial intelligence integrated gastrointestinal and complex surgical care

Performance domain	Definition	Clinical relevance	Suggested indicators
Workflow efficiency[14-16,47-49]	The extent to which AI supports prioritization without adding unnecessary digital burden	Surgical nurses often manage multiple unstable patients, drains, wounds, medications, mobilization targets, and discharge tasks during the same shift	Time spent retrieving risk information; duplicate documentation steps; task completion rates; staff-reported usability; time from alert visibility to review
Response-time optimization[44,45,47-49]	The interval between AI-generated concern and meaningful clinical action	Delays in response may worsen outcomes in sepsis, anastomotic leak, bleeding, respiratory compromise, or organ dysfunction	Time from alert to bedside reassessment; time from nursing concern to surgical review; time from review to intervention; proportion of high-priority alerts acted on within target time
Early complication recognition[5-9,25-30,44,45]	Earlier identification of deterioration before crisis-level presentation	Complications after colorectal, pancreatic, and hepatobiliary surgery often develop through subtle trends rather than one obvious sign	Time from first abnormal trend to escalation; proportion of complications preceded by documented nursing concern; unplanned ICU transfers; failure-to-rescue events
Adherence to risk-stratified pathways[10-13]	Alignment of nursing care with the patient’s predicted and evolving risk profile	High-risk patients may require intensified monitoring, earlier allied health input, or deviation from standard enhanced recovery pathways	Completion of individualized observation plans; adherence to escalation thresholds; timely nutrition, physiotherapy, wound, or stoma referral; documented pathway deviations
Documentation accuracy[17-22]	Completeness and consistency of clinically relevant nursing documentation following AI-supported alerts or risk classifications	AI systems and clinical teams both depend on accurate documentation of observations, actions, and reasoning	Proportion of AI-supported alerts with documented nursing reassessment and response; completeness of predefined clinical fields relevant to the identified risk; documentation of escalation and subsequent action; missing or incomplete required entries, assessed through structured record audit
Interdisciplinary communication[47-49]	Timely and documented transfer of clinically relevant AI-derived risk information between nurses and other professionals involved in the patient’s care	AI-derived risk information is clinically useful only if it reaches the right professionals in time	Proportion of relevant handovers containing AI-derived risk information; proportion of escalations with documented acknowledgment by the receiving team; time from nursing escalation to team acknowledgment; documentation of agreed actions or follow-up
Cognitive workload and alert burden[50-53]	The effect of AI tools on mental workload, interruption, and alarm fatigue	Excessive or poorly explained alerts may reduce trust and delay response to clinically important signals	Number of alerts per shift; proportion of actionable alerts; staff-rated alert usefulness; false or non-actionable alert frequency; missed alerts
Patient safety outcomes[44-46]	Patient-level outcomes potentially influenced by AI-supported nursing surveillance and coordination	The final aim of AI implementation is safer care, not simply more prediction	Delayed deterioration events; postoperative sepsis; unplanned ICU admission; readmission; medication errors; prolonged length of stay; patient-reported discharge preparedness

The performance domains presented in this table represent a conceptual and practice-oriented framework proposed by the authors to support future implementation and evaluation of artificial intelligence integrated surgical nursing. They should not be interpreted as a formally validated measurement instrument. AI: Artificial intelligence; ICU: Intensive care unit.

Table 4 Implementation risks and nursing leadership safeguards in artificial intelligence-supported surgical pathways

Implementation risk	How it may appear in practice	Potential consequence	Nursing leadership safeguard
Alert fatigue	Frequent low-priority alerts during already busy ward shifts	Important signals may be ignored or delayed	Define alert priority levels; monitor alert volume; remove non-actionable alerts; collect staff feedback
Unclear accountability	Alert visible to several professionals, but no one is clearly responsible for first response	Delay in reassessment or escalation	Assign first-response responsibility; define response time; document ownership of action
Over-reliance on AI	Staff wait for digital confirmation before escalating bedside concern	Clinical deterioration may be underestimated when no alert is generated	Reinforce that nursing concern overrides absence of alert; include this principle in training
Under-use of AI	Alerts are ignored because they are poorly understood or not integrated into workflow	Technology has little real-world clinical effect	Provide scenario-based training; integrate alerts into handover and ward routines
Poor interpretability	Nurses receive a risk score without explanation of relevant contributing factors	Low trust, inconsistent response, or inappropriate escalation	Require clinically understandable alert displays; include contributing variables where possible
Documentation burden	AI adds new forms or fields without reducing existing duplication	Increased workload and incomplete records	Align documentation templates with nursing workflow; remove duplicate fields; audit documentation quality
Inequitable performance	Model performs less well in frail, older, complex, or underrepresented patient groups	Risk may be underestimated or overestimated in vulnerable patients	Monitor subgroup performance; encourage clinical override; review unexpected deterioration events
Weak handover integration	Risk information remains in the electronic record but is not communicated during transitions	Receiving nurses may not know what to monitor	Add AI-derived risk status to structured perioperative handover
Lack of staff confidence	Nurses are unsure whether they are expected to trust, question, or act on alerts	Variable adoption and inconsistent practice	Use case-based education, simulation, and feedback after real alerts
Unmeasured implementation	Institution installs AI tool but does not evaluate nursing workflow or patient safety impact	No evidence that care delivery improved	Use a nursing-sensitive audit dashboard before and after implementation

AI: Artificial intelligence.