PACU Delirium in Older Surgical Patients: Incidence, Nursing-Sensitive Correlates, and Outcomes in a Multicenter Chinese Cohort Study

Abstract

Background

Post-anesthesia care unit (PACU) delirium in older surgical patients is poorly characterized in Chinese multicenter settings, and whether nursing-sensitive PACU variables add explanatory value beyond conventional clinical risk factors has not been formally tested.

Methods

In this prospective cohort study across five Chinese tertiary hospitals, 2200 patients aged 65 years or older undergoing elective major noncardiac surgery (after excluding 60 with persistent deep sedation) were assessed for PACU delirium using a Richmond Agitation-Sedation Scale (RASS)-gated, repeated Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) protocol at three timepoints (10–20 and 40–60 minutes after admission, and before discharge). Two prespecified nested logistic models were compared in an explanatory framework: Model A (conventional clinical predictors) and Model B (Model A plus eight nursing-sensitive PACU variables). Complete-case analysis included 2,048 patients; models were compared using the likelihood ratio (LR) test, C-statistic, and integrated discrimination improvement (IDI).

Results

PACU delirium occurred in 422 of 2,200 patients (19.2%; 95% CI: 17.6%-20.9%) and was predominantly hypoactive (64.2%). In descending order of effect size, independent conventional correlates were known cognitive impairment, chronic kidney disease, ASA ≥ III, stroke/TIA, older age, anxiety, lower MoCA, and lower albumin. Adding the nursing-sensitive block modestly improved overall fit (LR test p = 0.037), but the discrimination gain was small (C-statistic 0.675 to 0.686; ∆AUC = +0.011) and the IDI was non-significant (IDI = +0.008, p = 0.384); urinary catheterization was the only independent nursing-sensitive correlate (OR 1.49; 95% CI: 1.18–1.89; p = 0.0009). PACU delirium was associated with all secondary outcomes, including ward delirium (OR 3.26), falls (OR 4.63), non-home discharge (OR 4.75), and prolonged hospitalization (geometric mean ratio 1.46), all p < 0.001.

Conclusion

PACU delirium affected approximately one in five older patients and was predominantly hypoactive. The nursing-sensitive block added statistically detectable but clinically limited discriminative value, with the signal concentrated in urinary catheterization. Structured, repeated PACU delirium screening-particularly for hypoactive presentations-and targeted catheter minimization are practical nursing priorities; external validation is required before any bedside risk prediction.

Introduction

Postoperative delirium (POD) is one of the most common neuropsychiatric complications after major surgery in older adults, characterized by acute and fluctuating disturbances in attention, awareness, and cognition.^1,2 Meta-analytic evidence indicates that POD affects approximately 11% to 51% of surgical patients depending on the population studied and the assessment strategy used, with pooled estimates near 20% in older adults undergoing noncardiac surgery.^3,4 The consequences are clinically important. Delirium is associated with prolonged hospitalization, increased short-term mortality, institutionalization, subsequent cognitive decline, and higher use of intensive care resources.^5,6 Hypoactive delirium deserves particular attention because it accounts for a large proportion of cases yet is frequently overlooked without structured screening; current European and American guidance therefore recommends routine delirium surveillance and multicomponent prevention in vulnerable older adults.^7,8

The post-anesthesia care unit (PACU) is the earliest practical setting in which delirium can be detected after surgery, but it requires careful conceptual distinction among three overlapping phenomena. Emergence agitation (or emergence delirium) is a brief, often hyperactive state confined to the first minutes of anesthetic recovery that typically resolves spontaneously; PACU delirium, as operationalized here, refers to a CAM-ICU-positive disturbance of attention and awareness that persists or recurs across the PACU stay after adequate arousal (RASS gating); and later postoperative delirium (POD) refers to the syndrome ascertained on the surgical ward over subsequent days.⁹ These entities lie on a continuum but are not interchangeable, and conflating them is a recurrent source of heterogeneity in the literature. PACU delirium is clinically meaningful because early cognitive disturbance may mark ongoing neurological vulnerability rather than a transient recovery-room artifact. Crucially, the PACU is a uniquely hostile environment for unstructured detection: residual anesthetic and analgesic sedation blunts the clinical picture, rapid patient turnover compresses observation time, and high nursing workload limits sustained cognitive assessment. These conditions disproportionately mask hypoactive delirium, in which patients are quiet and inattentive rather than agitated, providing a strong rationale for standardized, repeated CAM-ICU screening rather than reliance on clinical impression. Prior studies have reported widely varying PACU delirium rates, from 4.7% in general adult samples to 31% at admission in older patients and 37% in a single-center elderly cohort, reflecting differences in population, timing, and instrument choice.^10–12 Neufeld, Leoutsakos, Sieber, Joshi, Wanamaker, Rios-Robles and Needham¹³ further showed that many patients who later develop ward delirium have abnormal findings already detectable in the PACU. Together, these observations support the view that the recovery room represents not only a surveillance window but also a potential opportunity for early intervention.

The conceptual framework for delirium risk is the predisposing-precipitating model, in which baseline vulnerability interacts with perioperative insults.¹ Conventional predictors such as advanced age, cognitive impairment, frailty, ASA class, stroke history, chronic disease burden, surgical duration, and blood loss are well established in umbrella reviews and meta-analyses.^3,14 However, most of these factors are patient-level characteristics that are only weakly modifiable at the bedside. Existing perioperative delirium prediction models inherit this limitation: they are typically built on fixed preoperative or intraoperative variables, achieve only moderate discrimination (C-statistics commonly 0.65–0.75), are seldom externally validated, and rarely incorporate the dynamic, care-process variables that nurses actually control.

By contrast, nursing-sensitive PACU variables are directly embedded in clinical care processes and may therefore be more actionable. We define nursing-sensitive variables, following the nursing-sensitive outcomes framework of Doran,¹⁵ as exposures and states that nurses directly modify, monitor, or organize and that are plausibly linked to patient outcomes: nursing-modifiable exposures (hypothermia, pain, postoperative nausea and vomiting [PONV], shivering, and urinary catheterization), nursing-monitored patient states (hypoxemia), and system/process features (nurse-to-patient ratio and admission timing). Each has a biologically plausible mechanistic pathway to delirium: uncontrolled pain, hypothermia/shivering, and PONV act as nociceptive and physiological stressors that elevate catecholamine and inflammatory signaling; hypoxemia compromises cerebral oxygen delivery; indwelling urinary catheters cause discomfort, restrict early mobilization, and may precipitate occult infection or sleep disruption; and lower staffing and off-hours admission reduce the frequency and quality of reorientation and monitoring. The urinary catheter is especially notable because catheter avoidance is emphasized in the Hospital Elder Life Program (HELP), and prior ward-based studies have reported meaningful catheter-delirium associations.^16–18 Yet whether a block of nursing-sensitive PACU variables adds explanatory value beyond comprehensive conventional predictors has not been formally tested in a multicenter Chinese cohort.

This evidence gap matters because contemporary delirium prevention guidance increasingly recognizes the role of nursing assessment, communication, and multicomponent care bundles. The HELP intervention demonstrated that a structured nonpharmacological program could reduce delirium incidence, and a recent meta-analysis of nurse-led interventions similarly suggested benefit across adult settings.^16,19 Related bundled approaches in critical care, such as the ABCDEF bundle, have also linked structured delirium-oriented nursing care with improved outcomes.²⁰ However, these protocols were developed for the ward or ICU; validated nurse-led delirium prevention protocols tailored specifically to the PACU-with its time pressure, incomplete anesthetic recovery, and rapid turnover-are scarce globally and, to our knowledge, entirely lacking in Chinese multicenter practice.⁷ The issue is especially relevant in China because of the country’s very high surgical volume and rapidly aging population.²¹ Several context-specific features further heighten its importance: indwelling urinary catheters are routinely placed for major surgery and often arrive in the PACU by default rather than by strict indication; nurse-to-patient ratios in tertiary-hospital PACUs are frequently stretched beyond 1:2 during peak throughput; and no nationwide protocol mandates structured PACU delirium screening. Chinese meta-analyses suggest a POD burden broadly similar to global estimates, but much of the literature remains single-center, anesthesia-focused, and limited in its attention to nursing process variables or repeated validated PACU assessments.²²

We therefore conducted a multicenter prospective cohort study in older Chinese patients undergoing major noncardiac surgery. The study had four aims: first, to estimate the incidence and motor subtype distribution of PACU delirium using a repeated, RASS-gated CAM-ICU protocol; second, to test whether nursing-sensitive PACU variables add incremental associative value beyond conventional clinical predictors; third, to examine the association between PACU delirium and early adverse outcomes including ward delirium, ICU admission, falls, device self-removal, complications, and length of stay; and fourth, to derive an exploratory bedside risk index for future validation. We hypothesized that PACU delirium would be common, predominantly hypoactive, and driven mainly by conventional patient-level vulnerability factors, with only modest incremental explanatory contribution from the nursing-sensitive block. Because several nursing-sensitive exposures are measured during the same PACU stay in which the outcome is assessed, we explicitly framed the primary analysis as an explanatory incremental-association study rather than a temporally strict early-prediction model.

Methods

Study Design and Setting

This multicenter prospective observational cohort study was conducted across five Chinese tertiary teaching hospitals (S01–S05), each with a dedicated PACU of at least 10 monitored beds. Consecutive eligible patients were enrolled between January and December 2024. The study was designed to estimate the incidence and motor subtype distribution of PACU delirium, evaluate the incremental associative value of nursing-sensitive PACU variables beyond a strong conventional benchmark, and examine associations between PACU delirium and early adverse outcomes. The study followed STROBE guidance for cohort reporting;²³ TRIPOD concepts were used only for the nested model comparison and exploratory risk index components.

Participants

Eligible patients were aged 65 years or older, underwent elective major noncardiac surgery under general anesthesia, and were planned for postoperative PACU recovery followed by ward transfer. Preoperative delirium was screened on the day before surgery by trained study staff using the Chinese 3-Minute Diagnostic Confusion Assessment Method (3D-CAM-CN); a positive screen was an exclusion criterion. Other exclusion criteria were direct postoperative ICU admission, emergency surgery, cardiac/neurosurgical/transplant procedures, severe psychotic disorder precluding assessment, repeat eligible surgery during the same admission, and persistent deep sedation, defined as RASS <= −3 at all three scheduled PACU assessments. The deep-sedation rule was prespecified because RASS <= −3 precludes valid CAM-ICU testing; patients meeting it could neither be confirmed nor excluded as delirious and were therefore removed from the primary analysis set and addressed separately in bounding sensitivity analyses (see Statistical Analysis). Of 2,260 enrolled patients, 60 met the deep-sedation exclusion after surgery, leaving 2,200 patients in the full analysis set (Supplementary Figure S1). Because this exclusion could remove a vulnerable, high-risk subgroup and introduce mild selection bias, its potential impact was quantified in prespecified upper- and lower-bound sensitivity analyses that reintroduced these patients.

Data Collection

Data were recorded prospectively in a structured electronic case report form covering enrollment, preoperative assessment, intraoperative variables, PACU arrival, repeated PACU delirium assessments, ward delirium assessments on postoperative days (POD) 1–3, chart review, and hospital outcomes. The eCRF was organized into standardized forms spanning screening, baseline assessment, intraoperative data, PACU arrival, repeated PACU delirium assessment, ward delirium assessment, chart review through POD 5 or discharge, and final hospital outcomes. Real-time range checks and logic rules were built into the eCRF to reduce missingness and implausible values. Preoperative cognitive function was screened with the Beijing version of the Montreal Cognitive Assessment (MoCA-BJ), an open-access instrument freely available for clinical and research use after registration on the official MoCA website.^24,25 Following the published Chinese-version protocol, one point was added for participants with 12 or fewer years of formal education to mitigate education bias. Assessors completed standardized training and certification for the MoCA-BJ, RASS, CAM-ICU, and 3D-CAM-CN before study initiation. The full set of prespecified predictor variables, organized by conceptual domain and measurement timing, is presented in Table 1.

Table 1.

Predictor Variables by Domain and Measurement Timing

Domain	Variable	Measurement/Coding	Timing
Preoperative - Demographics	Age	Continuous (years)	Enrollment
	Sex	Binary (male/female)	Enrollment
	Education	Continuous (years of formal schooling)	Preoperative assessment
	Body mass index (BMI)	Continuous (kg/m2)	Preoperative assessment
Preoperative - Cognitive & Functional	Montreal Cognitive Assessment, Beijing version (MoCA-BJ)	Integer score 0–30 (1-point education adjustment for <= 12 years of schooling)	Preoperative assessment
	Clinical Frailty Scale (CFS)	Integer score 1–9	Preoperative assessment
	ASA physical status classification	Dichotomized as ≥ III vs I–II for modeling	Chart review
	Charlson Comorbidity Index (CCI)	Integer composite score	Chart review
Preoperative - Medical History	Hypertension	Binary (yes/no)	Chart review
	Diabetes mellitus	Binary (yes/no)	Chart review
	Coronary artery disease	Binary (yes/no)	Chart review
	Chronic obstructive pulmonary disease	Binary (yes/no)	Chart review
	Stroke or transient ischemic attack	Binary (yes/no)	Chart review
	Chronic kidney disease	Binary (yes/no)	Chart review
	Known cognitive impairment	Binary (yes/no)	Chart review / informant
	Prior delirium episode	Binary (yes/no)	Chart review / informant
	Visual impairment	Binary	Patient interview
	Hearing impairment	Binary	Patient interview
Preoperative - Psychosocial & Laboratory	Preoperative anxiety	NRS 0–10	Patient interview
	Sleep quality	NRS 0–10, higher = better	Patient interview
	Fasting duration (solids)	Continuous (hours)	Chart review
	Serum albumin	Continuous (g/L)	Most recent preoperative lab
	Hemoglobin	Continuous (g/L)	Most recent preoperative lab
Intraoperative	Surgical specialty	Multicategory	Operative record
	Surgery duration	Continuous (minutes)	Operative record
	Anesthesia maintenance type	Volatile/TIVA/combined	Anesthesia record
	Benzodiazepine administered	Binary (yes/no)	Anesthesia record
	Dexmedetomidine administered	Binary (yes/no)	Anesthesia record
	Duration of MAP < 65 mmHg	Continuous (minutes)	Anesthesia record
	Estimated blood loss > 500 mL	Binary (yes/no)	Operative record
	Minimum intraoperative temperature	Continuous (degrees C)	Anesthesia record
PACU Nursing-Sensitive Indicators	Temperature at PACU arrival	Continuous; hypothermia defined as < 36.0 degrees C	Nursing record
	Worst pain NRS in first 30 minutes	Integer 0–10	Nursing record
	Postoperative nausea or vomiting (PONV)	Binary (any episode)	Nursing record
	Shivering	Binary (any episode)	Nursing record
	Urinary catheter present at PACU arrival	Binary (yes/no)	Nursing record
	Hypoxemia	Binary (any SpO2 < 90% episode)	Nursing record
	Nurse-to-patient ratio	1:2 or better vs > 1:2	Unit staffing log
	Admission shift	Day shift vs evening/night	Nursing record

Primary Outcome Definition

The primary outcome was PACU delirium, defined as at least one positive CAM-ICU assessment during three scheduled PACU windows: T0 (10–20 minutes after arrival), T1 (40–60 minutes after arrival or the earliest assessable time after T0), and T2 (immediately before PACU discharge). The CAM-ICU was selected for the PACU because it is suitable for patients with fluctuating arousal, limited verbal interaction, and residual airway devices, whereas the 3D-CAM-CN was reserved for ward assessments on POD 1–3 when patients were fully interactive.²⁶ Arousal was evaluated first with the RASS. Patients with RASS <= −3 at a given timepoint were classified as not assessable and were reassessed within the time window when feasible; deep sedation alone was not classified as delirium. For assessable patients (RASS ≥ −2), the simplified Chinese CAM-ICU was administered according to standardized procedures. The CAM-ICU was specifically chosen because its four features (acute onset/fluctuation, inattention, altered level of consciousness, and disorganized thinking) can be elicited nonverbally in patients with residual airway devices or limited speech: inattention was assessed with the standardized Attention Screening Examination (letter or picture recognition) requiring only eye-opening or hand-squeeze responses, and disorganized thinking was assessed using yes/no questions and simple commands that intubated or hoarse patients could answer by gesture. Assessors were nurses and physicians independent of the intraoperative anesthesia team and were blinded to each patient’s prior PACU assessment results and to the conventional risk-factor data used in modeling; ward assessors (3D-CAM-CN) were likewise blinded to PACU delirium status.

Motor subtype was classified using the concurrent RASS at positive assessments: hyperactive (RASS +1 to +4), hypoactive (RASS −2 to 0), or mixed across timepoints, consistent with prior subtype work in critically ill adults.²⁷ Ward delirium on POD 1–3 was assessed with the 3D-CAM-CN, or CAM-ICU for ICU-transferred patients.

Secondary Outcome Definitions

Eight prespecified secondary outcomes spanning subsequent ward delirium, in-hospital safety events, resource utilization, and discharge disposition were assessed. Their operational definitions, data sources, and analytic scales are summarized in Table 2.

Table 2.

Definitions and Data Sources for Secondary Outcomes

Outcome	Definition	Data Source	Analysis Scale
Ward delirium (POD 1–3)	≥ 1 positive delirium screen during POD 1–3	F05 ward assessments	Binary
PACU length of stay	Time from PACU admission to PACU discharge (minutes)	F03 arrival time, F03 PACU discharge time	Continuous; log-transformed for regression
Postoperative hospital LOS	Time from surgery date to hospital discharge (days)	F07 hospital discharge date	Continuous; log-transformed for regression
Unplanned ICU admission	Any unplanned ICU transfer during the index hospitalization	F06 chart review	Binary
In-hospital falls	Any documented fall	F06 chart review	Binary
Self-removal of lines or tubes	Any self-removal of lines, catheters, drains, or tubes	F06 chart review	Binary
Major complications	Clavien-Dindo grade III or higher	F06 chart review	Binary
Non-home discharge	Discharge to any destination other than home	F07 discharge destination	Binary

These secondary outcomes were selected to reflect clinically meaningful short-term consequences of delirium that are relevant across perioperative disciplines. Ward delirium was included because it captures persistent or evolving postoperative neurocognitive vulnerability beyond the PACU. Falls and self-removal of lines or tubes were considered particularly relevant nursing-sensitive safety outcomes, whereas ICU admission, major complications, and non-home discharge reflect broader postoperative deterioration and recovery burden. PACU and hospital length of stay were included as utilization outcomes, recognizing that prolonged observation may partly reflect both illness severity and the clinical response triggered by delirium detection itself.

Statistical Analysis

Descriptive and Primary Analyses

The target sample size was 2,200 patients. Assuming a PACU delirium incidence of approximately 15--20% in older noncardiac surgical patients, this was expected to yield at least approximately 330–420 events. For the full Model B (approximately 30 estimated parameters), this corresponds to roughly 11–14 events per variable, exceeding the conventional minimum of 10 and consistent with the criteria of Riley et al for the sample size required to develop a stable multivariable prediction model with low overfitting and adequate shrinkage.²⁸ Baseline characteristics were summarized by delirium status using the t test or Mann--Whitney U-test for continuous variables and the chi-squared test for categorical variables, as appropriate. Because Table 3 reports numerous unadjusted baseline comparisons, these p-values are descriptive and hypothesis-generating; no formal correction for multiple comparisons was applied, and they were not used for variable selection (all predictors were prespecified). Readers should therefore interpret individual baseline p-values cautiously, and confirmatory inference rests on the prespecified adjusted models. Incidence was reported with Wilson 95\% confidence intervals. The decline in delirium prevalence across the three assessable timepoints (T0, T1, T2) was tested for a monotonic trend using the Cochran--Armitage trend test (Supplementary Methods 2).

Table 3.

Baseline Demographic, Clinical, and Perioperative Characteristics by PACU Delirium Status

Variable	No Delirium (n = 1,778)	Delirium (n = 422)	p Value
Age, years	72.6 ± 5.2	74.1 ± 5.5	< 0.001
Male sex	965 (54.3%)	228 (54.0%)	0.970
Education, years	8.5 ± 3.8	8.3 ± 3.9	0.074
BMI, kg/m2	24.4 ± 4.0	24.1 ± 4.1	0.168
MoCA total score	22.8 ± 3.1	21.6 ± 3.3	< 0.001
Clinical Frailty Scale score	3.5 ± 1.3	3.5 ± 1.4	0.350
ASA class (mean)	2.7 ± 0.7	2.9 ± 0.7	< 0.001
ASA class ≥ III	1,081 (60.8%)	304 (72.0%)	< 0.001
Charlson Comorbidity Index	1.5 ± 1.0	1.6 ± 1.1	0.025
Hypertension	1,061 (59.7%)	250 (59.2%)	0.863
Diabetes mellitus	507 (28.5%)	118 (28.0%)	0.882
Coronary artery disease	283 (15.9%)	60 (14.2%)	0.391
COPD	224 (12.6%)	51 (12.1%)	0.815
Stroke/TIA	187 (10.5%)	62 (14.7%)	0.020
Chronic kidney disease	148 (8.3%)	54 (12.8%)	0.006
Known cognitive impairment	101 (5.7%)	41 (9.7%)	0.004
Prior delirium episode	62 (3.5%)	21 (5.0%)	0.229
Visual impairment	380 (21.4%)	112 (26.5%)	0.024
Hearing impairment	297 (16.7%)	82 (19.4%)	0.236
Preoperative anxiety NRS	2.0 ± 2.3	2.4 ± 2.7	0.012
Sleep quality NRS	5.5 ± 2.0	5.3 ± 2.0	0.037
Albumin, g/L	38.1 ± 4.9	37.6 ± 5.2	0.050
Hemoglobin, g/L	125.2 ± 18.0	124.0 ± 18.2	0.382
Fasting duration (solids), h	12.0 ± 2.9	12.1 ± 2.9	0.355
Surgery duration, min	138.6 ± 48.9	140.7 ± 49.2	0.343
Benzodiazepine used	256 (14.4%)	61 (14.5%)	1.000
Dexmedetomidine used	642 (36.1%)	150 (35.5%)	0.883
MAP < 65 mmHg duration, min	7.5 ± 7.9	8.1 ± 8.4	0.212
Minimum intraoperative temp, degrees C	36.1 ± 0.4	36.1 ± 0.4	0.435
Temperature at PACU arrival, degrees C	36.5 ± 0.5	36.5 ± 0.5	0.316
Hypothermia at arrival (< 36 degrees C)	235 (13.2%)	61 (14.5%)	0.599
Urinary catheter at PACU arrival	937 (52.7%)	261 (61.8%)	< 0.001
Worst pain NRS (first 30 min)	3.0 ± 2.4	3.0 ± 2.6	0.530
PONV (any episode)	590 (33.2%)	148 (35.1%)	0.512
Shivering (any episode)	382 (21.5%)	87 (20.6%)	0.682
Hypoxemia (any SpO2 < 90%)	7 (0.4%)	3 (0.7%)	0.640
Nurse-to-patient ratio > 1:2	578 (32.5%)	129 (30.6%)	0.496

Notes: Continuous variables are presented as mean ± SD; categorical variables as n (%). The low observed hypoxemia rate reflects the stringent SpO2 < 90% threshold.

The primary analysis used two prespecified nested logistic regression models in an explanatory incremental-association framework. Model A represented the conventional clinical benchmark and included demographics, cognition/function, comorbidity/history, psychosocial/laboratory variables, intraoperative exposures, and site fixed effects. Model B added eight nursing-sensitive PACU variables: pain, hypothermia at arrival, hypoxemia, urinary catheter presence, PONV, shivering, nurse-to-patient ratio, and day-shift admission. Four of these variables - urinary catheter presence, PACU arrival temperature, staffing ratio, and admission shift - were known before or at the start of PACU assessment; the others were measured during the PACU stay. Because pain, PONV, shivering, and hypoxemia were recorded during the same PACU stay as delirium assessment, the analysis was interpreted as explanatory association rather than temporally strict prediction. All prespecified variables were retained and reported as adjusted odds ratios (ORs) with 95% confidence intervals (CIs).

Model A therefore provided a deliberately strong comparator rather than a sparse baseline model. In addition to demographic and cognitive variables, it included frailty, key medical history variables, psychosocial measures, laboratory data, benzodiazepine exposure, hypotension duration, surgery duration, anesthesia maintenance type, blood loss, and site terms. The goal was to minimize the risk that any apparent gain in Model B would simply reflect under-adjustment in the conventional benchmark. Center was modeled using fixed effects in the primary analysis to account for differences in case mix and practice patterns across hospitals. Because the small number of clusters (five centers) makes random-effects variance estimation imprecise, two supplementary approaches were used to confirm that the conclusions were not sensitive to the method of center adjustment: a generalized estimating equations (GEE) model with an exchangeable working correlation and site as the clustering unit, and a random-intercept generalized linear mixed model (Supplementary Methods 2). The complete set of predictors retained in each model and the eight nursing-sensitive PACU variables added in Model B are listed side by side in Table 4.

Table 4.

Specification of Model A and Model B

Predictor Block	Model A	Model B
Age	Y	Y
Sex	Y	Y
Education	Y	Y
MoCA	Y	Y
ASA ≥ III	Y	Y
Charlson Comorbidity Index	Y	Y
Clinical Frailty Scale	Y	Y
Stroke/TIA history	Y	Y
Prior delirium episode	Y	Y
Chronic kidney disease	Y	Y
Known cognitive impairment	Y	Y
Visual impairment	Y	Y
Hearing impairment	Y	Y
Preoperative anxiety NRS	Y	Y
Sleep quality NRS	Y	Y
Serum albumin	Y	Y
Benzodiazepine administered	Y	Y
MAP < 65 mmHg duration	Y	Y
Surgery duration	Y	Y
Anesthesia maintenance type	Y	Y
Blood loss > 500 mL	Y	Y
Site fixed effects	Y	Y
Worst pain NRS (first 30 min)		Y
Hypothermia at arrival		Y
Hypoxemia in PACU		Y
Urinary catheter at arrival		Y
PONV in PACU		Y
Shivering in PACU		Y
Nurse-to-patient ratio > 1:2		Y
Day-shift admission		Y

Variable selection was avoided by design: all prespecified predictors were forced into the models regardless of significance, so that the Model A versus Model B comparison reflects a contrast between conceptual predictor blocks rather than two data-driven models. Continuous predictors were checked for nonlinearity using restricted cubic splines (three knots) and likelihood ratio tests against the linear term; no meaningful departures from linearity were detected, so linear terms were retained for interpretability. Multicollinearity was assessed with variance inflation factors (VIF); all VIFs were below 2.0, indicating no problematic collinearity among predictors, including between conventional and nursing-sensitive blocks (Supplementary Methods 2).

Model Comparison and Secondary Analyses

Model comparison used the C-statistic, Brier score, bootstrap-corrected calibration slope, and the likelihood ratio (LR) test. Category-free net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated as supplementary metrics, and the methods and interpretation of a decision curve analysis are provided in Supplementary Methods 1. We treated NRI cautiously because it can exaggerate apparent improvement when baseline calibration is imperfect or event prevalence is moderate; IDI was therefore given greater interpretive weight alongside the delta-AUC.^29–33 Internal validation used 500 bootstrap resamples, with optimism estimated as the average difference between bootstrap apparent performance and performance when the same model was applied back to the original dataset.

For secondary outcomes, logistic regression was used for binary outcomes and log-linear models for PACU and hospital length of stay. Models adjusted for age, sex, ASA class, MoCA, surgery duration, center, CFS, CCI, stroke/TIA, and prior delirium. Rare binary outcomes were analyzed with Firth’s penalized logistic regression. To express effects on the absolute scale, crude event proportions by delirium status and absolute risk differences (with 95% confidence intervals) were reported alongside adjusted odds ratios for each binary secondary outcome. Given the very low observed hypoxemia event rate, a post-hoc power analysis was performed to quantify the study’s ability to detect an association between hypoxemia and PACU delirium.

Missing Data and Sensitivity Analyses

Primary modeling used complete-case analysis after excluding 152 patients with missing albumin, anxiety NRS, or sleep quality NRS values, leaving 2,048 patients. To assess whether data were plausibly missing at random, patients excluded for missing values were compared with complete cases on key baseline characteristics (age, sex, MoCA, ASA class, and PACU delirium status); no clinically meaningful differences were observed, including a similar delirium rate (Supplementary Methods 2). Multiple imputation by chained equations (20 datasets) was performed as a sensitivity analysis under a missing-at-random assumption. Additional prespecified sensitivity analyses excluded borderline sedation at T0, used only T1/T2 outcome definitions, restricted the cohort to major surgical specialties, excluded protocol deviations, reclassified deeply sedated excluded patients in upper- and lower-bound analyses, and excluded isolated T0-only positives. Because site heterogeneity might affect catheter use and delirium incidence, the main models included site fixed effects; a supplementary random-intercept analysis was also examined, with cautious interpretation given the small number of clusters (Supplementary Methods 2).³⁴

For transparency, all prespecified predictors were retained regardless of statistical significance. This choice was intended to preserve the a priori analytic framework, avoid instability from data-driven selection, and keep the comparison between Model A and Model B interpretable as a contrast between conceptual predictor blocks rather than between two separately optimized models.

Software

Analyses were performed in Python 3.12 using statsmodels, scikit-learn, SciPy, and matplotlib. A two-sided alpha of 0.05 was used.

Quality Assurance and Interrater Reliability

All sites used a standardized operations manual and underwent assessor certification before enrollment. At least 5% of delirium assessments were double-rated, with a prespecified kappa target of ≥ 0.80. Monthly audits reviewed range checks, temporal consistency, and missingness; discordant cases were adjudicated centrally by a blinded committee that included geriatric and psychiatric expertise.

Ethical Considerations

The study protocol was reviewed and approved on 18 December 2023 by the Medical Ethics Committee of Shanxi Provincial People’s Hospital (Taiyuan, Shanxi, China), which served as the central ethics review board for this multicenter study (approval number: 2023–312); approval thus preceded the start of enrollment in January 2024. In accordance with national regulations on multicenter clinical research in China, the four participating sites (S02–S05) accepted and recorded the central approval issued by the Medical Ethics Committee of Shanxi Provincial People’s Hospital prior to local enrollment. The study was conducted in accordance with the principles of the Declaration of Helsinki (as revised in 2013) and applicable national regulations on the ethical conduct of human research. Written informed consent was obtained from every participant preoperatively. Decisional capacity to consent was assessed preoperatively by trained study clinicians; patients with a positive preoperative delirium screen were excluded, and for the remaining patients with cognitive impairment whose capacity was judged limited, written informed consent was obtained from a legally authorized representative (proxy consent) in accordance with the approved protocol. All data were de-identified at the point of entry and stored on encrypted, access-controlled servers maintained by the coordinating center.

Results

Study Enrollment and Participant Characteristics

Of 2,260 enrolled patients, 2,200 formed the full analysis set after excluding 60 patients with persistent deep sedation at all three PACU assessments. Assessability improved over time: 2,122 patients were assessable at T0, 2,177 at T1, and 2,194 at T2. For primary modeling, 152 patients with missing albumin, anxiety NRS, or sleep quality NRS values were excluded, leaving 2,048 complete cases. Interrater reliability exceeded the prespecified threshold (CAM-ICU kappa 0.85; 3D-CAM-CN kappa 0.83), supporting consistency of outcome assessment across centers.

The cohort had a mean age of 72.9 ± 5.3 years, 54.2% were male, and 63.0% had ASA class ≥ III. Surgical case mix was broad and included general/abdominal, orthopedic, urologic, thoracic, gynecologic, head-neck, vascular, and hepatobiliary procedures, supporting the clinical heterogeneity expected in tertiary PACUs. Compared with patients without delirium, those with PACU delirium were older, had lower MoCA scores, more frequent ASA ≥ III, stroke/TIA, CKD, known cognitive impairment, higher anxiety, lower albumin, and more urinary catheterization on PACU arrival. In contrast, most other nursing-sensitive PACU variables showed little unadjusted separation between groups. Full baseline characteristics are provided in Supplementary Table S2; selected variables are summarized in Table 3. The unadjusted p-values in these baseline tables are descriptive and were not corrected for multiple comparisons; given the large number of variables tested, individual nominal associations should be regarded as hypothesis-generating, and formal inference rests on the prespecified adjusted models below.

Primary Outcome: Incidence and Motor Subtype Distribution of PACU Delirium

PACU delirium occurred in 422 of 2,200 patients (19.2%; 95% CI: 17.6%-20.9%). Site-specific incidence ranged from 15.8% to 23.1% (Figure 1b; Supplementary Table S1), and mixed-effects analysis suggested minimal between-hospital variation after adjustment (ICC = 0.0002; median odds ratio = 1.025; Additional Supplementary Analysis). The small random-effects variance indicated that most measurable variability in delirium occurrence was explained by patient-level rather than hospital-level factors.

Figure 1.

Primary Outcome: PACU Delirium - Associated Factors, Incidence, Subtypes, and Temporal Pattern. Panel (a): Forest plot of adjusted odds ratios (95% CI) from Model B, grouped by nursing-sensitive PACU variables (red) and conventional clinical variables (blue). Diamond markers indicate point estimates; horizontal lines represent 95% confidence intervals on a logarithmic scale. Filled diamonds denote variables with p < 0.05. Panel (b): PACU delirium incidence by center with Wilson 95% CI error bars; the dashed red line represents the overall cohort incidence (19.2%). Panel (c): Donut chart illustrating motor subtype distribution at the patient level (hypoactive 64.2%, mixed 19.2%, hyperactive 16.6%). Panel (d): PACU delirium rate at each assessment timepoint (T0, T1, T2), with the number of assessable patients annotated above each bar.

Among delirium cases, hypoactive delirium predominated (271/422, 64.2%), followed by mixed (19.2%) and hyperactive (16.6%) subtypes (Figure 1c). This distribution is clinically relevant because hypoactive presentations are less likely to be recognized by unstructured bedside observation. The predominance of non-agitated delirium supports the decision to use a structured repeated-assessment protocol rather than rely on clinical impression alone.

In an exploratory comparison of baseline characteristics across motor subtypes (Supplementary Table S5), patients with hypoactive delirium were modestly older and had slightly lower MoCA scores than those with hyperactive or mixed delirium, whereas urinary catheterization, ASA class, and comorbidity burden were broadly similar across subtypes. These subgroup contrasts were not statistically significant after accounting for the small per-subtype cell sizes and are presented as hypothesis-generating only. The motor-subtype distribution was also broadly consistent across the five centers, with hypoactive delirium predominating at every site (range 61.9%-66.0% of cases; Supplementary Table S6), arguing against a major institutional difference in subtype detection or case mix.

The delirium rate declined across the PACU stay: 11.8% at T0, 10.3% at T1, and 9.1% at T2 among assessable patients (Figure 1d). A Cochran-Armitage test confirmed that this decline was a statistically significant monotonic trend across the three timepoints (z = −2.92, p = 0.004; Supplementary Methods 2); because the assessable denominators differ slightly across timepoints and the assessments are repeated within patients, this trend is best interpreted as descriptive of falling point prevalence during recovery rather than as resolution in a fixed cohort. Notably, 111 of 422 cases (26.3%) were detected only at T0, indicating that a later single assessment would miss a substantial fraction of cases. At the same time, a clinically meaningful number of patients were first identified after T0, supporting the value of serial rather than one-time screening. Adjusted associations between individual predictors and PACU delirium are visualized in the forest plot of Figure 1a, with detailed interpretation provided in the Explanatory Models section below.

Explanatory Models: Conventional Clinical vs Nursing-Sensitive PACU Predictors

Model A: Comprehensive Conventional Clinical Predictors

In the complete-case cohort (N = 2,048), Model A achieved a C-statistic of 0.675 (95% CI: 0.644–0.707), an optimism-corrected C-statistic of 0.646, a Brier score of 0.1434, and a bootstrap-corrected calibration slope of 0.829. These values indicate moderate discrimination with some coefficient optimism, which is expected in a relatively high-dimensional explanatory model. Significant conventional predictors were older age, lower MoCA, ASA ≥ III, stroke/TIA, CKD, known cognitive impairment, higher anxiety, lower albumin, and blood loss > 500 mL (inverse association); adjusted odds ratios with 95% confidence intervals for each predictor are reported in the Model A column of Table 5. The direction of association was stable and clinically plausible for all significant predictors except blood loss, which was interpreted cautiously. Because an inverse blood-loss association is biologically implausible and contrary to the established delirium literature, we conducted post-hoc exploratory analyses to probe it. The association was attenuated and lost significance when surgical specialty and surgery duration were added or when high-blood-loss cases were stratified by procedure type, suggesting that the apparent protective estimate reflects residual confounding by surgical and anesthetic management (eg, greater intraoperative monitoring, deliberate hemodynamic and transfusion management, and more frequent dexmedetomidine use in higher-blood-loss procedures) rather than a true biological effect. We therefore regard this coefficient as a statistical artifact and do not interpret it substantively.

Table 5.

Multivariable Logistic Regression Results for Model A and Model B

Predictor	Model A OR (95% CI)	Model A P	Model B OR (95% CI)	Model B P
Conventional clinical predictors
Age, per year	1.034 (1.01--1.06)	0.007	1.034 (1.01--1.06)	0.007
Female sex	1.01 (0.80--1.27)	0.950	1.00 (0.79--1.26)	0.996
Education, per year	0.982 (0.95--1.01)	0.237	0.981 (0.95--1.01)	0.211
MoCA score, per point	0.918 (0.88--0.96)	< 0.001	0.918 (0.88--0.96)	< 0.001
ASA class ≥ III	1.67 (1.30--2.14)	< 0.001	1.67 (1.30--2.14)	< 0.001
Charlson Comorbidity Index	0.98 (0.85--1.14)	0.819	0.98 (0.85--1.13)	0.807
Clinical Frailty Scale score	1.01 (0.93--1.11)	0.757	1.01 (0.93--1.11)	0.748
Stroke / TIA	1.66 (1.15--2.39)	0.006	1.70 (1.18--2.44)	0.005
Prior delirium episode	1.66 (0.95--2.91)	0.075	1.64 (0.93--2.89)	0.087
Chronic kidney disease	1.85 (1.17--2.93)	0.009	1.81 (1.14--2.88)	0.012
Known cognitive impairment	2.02 (1.30--3.14)	0.002	2.02 (1.29--3.15)	0.002
Visual impairment	1.26 (0.97--1.64)	0.088	1.26 (0.97--1.65)	0.088
Hearing impairment	1.12 (0.83--1.52)	0.439	1.10 (0.82--1.49)	0.526
Preoperative anxiety NRS	1.09 (1.04--1.14)	< 0.001	1.09 (1.04--1.14)	< 0.001
Sleep quality NRS	0.96 (0.90--1.01)	0.126	0.95 (0.90--1.01)	0.111
Serum albumin, per g/L	0.976 (0.954--0.999)	0.041	0.975 (0.952--0.998)	0.029
Benzodiazepine administered	1.07 (0.77--1.48)	0.684	1.05 (0.76--1.46)	0.774
MAP < 65 duration, per min	1.01 (1.00--1.02)	0.203	1.01 (1.00--1.02)	0.208
Surgery duration, per min	1.00 (1.00--1.00)	0.341	1.00 (1.00--1.00)	0.500
TIVA maintenance	1.00 (0.76--1.31)	0.988	1.00 (0.76--1.32)	0.994
Combined anesthesia	0.96 (0.73--1.27)	0.789	0.97 (0.73--1.29)	0.827
Blood loss > 500 mL	0.62 (0.41--0.94)	0.024	0.61 (0.40--0.92)	0.020
Nursing-Sensitive PACU Predictors
Worst pain NRS (first 30 min)	-	-	1.00 (0.96--1.05)	0.856
Hypothermia (< 36 degrees C)	-	-	1.20 (0.86--1.66)	0.282
Hypoxemia (SpO2 < 90%)	-	-	1.05 (0.21--5.33)	0.957
Urinary catheter present	-	-	1.49 (1.18--1.89)	0.0009
PONV (any episode)	-	-	0.98 (0.77--1.26)	0.897
Shivering (any episode)	-	-	0.81 (0.60--1.08)	0.154
Nurse-to-patient ratio > 1:2	-	-	1.11 (0.87--1.41)	0.417
Day-shift admission	-	-	0.86 (0.66--1.11)	0.248

Notes: Full coefficients are reported in Supplementary Table S4.

Abbreviations: OR, odds ratio; CI, confidence interval; TIVA, total intravenous anesthesia; NRS, Numeric Rating Scale.

Model B: Clinical Plus Nursing-Sensitive PACU Predictors

Model B, which added the eight nursing-sensitive PACU variables, increased the C-statistic to 0.686 (95% CI: 0.655–0.717), with an optimism-corrected value of 0.649, a Brier score of 0.1421, and a calibration slope of 0.778. All significant conventional predictors from Model A remained stable after addition of the PACU block, suggesting that the nursing-sensitive variables did not materially confound the conventional clinical architecture of risk.

Among the nursing-sensitive variables, urinary catheterization at PACU arrival was the only independent correlate of PACU delirium (OR 1.49, 95% CI: 1.18–1.89, p = 0.0009). The remaining seven nursing-sensitive variables were not individually significant, including pain, PONV, shivering, hypothermia, hypoxemia, staffing ratio, and day-shift admission. Adjusted odds ratios with 95% confidence intervals for both Model A and Model B, including the eight nursing-sensitive PACU predictors, are reported in Table 5 and graphically displayed in Figure 1a. Hypoxemia estimates were especially imprecise because the event rate under the prespecified SpO2 < 90% threshold was very low (10 events total; 7/1,778 vs 3/422). A post-hoc power analysis confirmed that with so few events the study had only approximately 7% power to detect even a moderate association (OR 2.0) at two-sided α = 0.05 (Supplementary Methods 2); the non-significant hypoxemia estimate should therefore be read as statistically uninformative rather than as evidence of no association. Multiple imputation yielded concordant results, supporting the robustness of the complete-case findings.

Incremental Associative Value of Nursing-Sensitive PACU Variables

Adding the nursing-sensitive PACU block improved overall model fit but only modestly changed discrimination. The apparent C-statistic increased from 0.675 to 0.686 (delta-AUC +0.011), and the optimism-corrected value changed from 0.646 to 0.649. The LR test was statistically significant (p = 0.037), but the IDI was not (p = 0.384), indicating limited clinical gain despite statistical evidence of added signal. The continuous NRI was positive and nominally significant; however, we emphasize at the outset that the NRI is well documented to overstate incremental value, particularly when baseline calibration is imperfect or event prevalence is moderate, and a significant NRI can occur even with negligible gains in discrimination. We therefore down-weight the NRI relative to the ΔAUC and the non-significant IDI, which together indicate that the added discriminative value was small. All comparative discrimination, calibration, and incremental-value metrics for the two models are summarized in Table 6, and the corresponding ROC curves, calibration plots, bootstrap calibration-slope distributions, and sensitivity-analysis AUCs are shown in Figure 2; supplementary DCA interpretation is provided in the supplement (Additional Supplementary Analysis).

Table 6.

Model Comparison: Discrimination, Calibration, and Incremental Value Metrics

Metric	Model A	Model B	Difference
C-statistic (AUROC)	0.675	0.686	+0.011
Optimism-corrected C-statistic	0.646	0.649	+0.003
Bootstrap-corrected calibration slope	0.829	0.778	−0.051
Brier score	0.1434	0.1421	−0.0013
Pseudo-R-squared	0.0663	0.0745	+0.0082
AIC	1,921.3	1,920.9	−0.4
BIC	2,073.2	2,117.8	+44.6
LR test chi-squared (df = 8)	-	-	16.37 (p = 0.037)
Continuous NRI (total)	-	-	+0.232 (p < 0.0001)
IDI	-	-	+0.008 (p = 0.384)

Notes: Complete-case analysis: N = 2,048.

Abbreviations: AUROC, area under the receiver operating characteristic curve; AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; LR, likelihood ratio; NRI, Net Reclassification Improvement; IDI, Integrated Discrimination Improvement.

Figure 2.

Model Performance Comparison: Discrimination, Calibration, and Sensitivity Analysis. Panel (a): Receiver operating characteristic curves for Model A (blue) and Model B (red), with the area under the curve shaded. The annotation box reports delta-AUC, likelihood ratio test p value, and IDI. Panel (b): Calibration plot showing observed outcome proportions against predicted probabilities across deciles for both models; the dashed diagonal represents perfect calibration. The inset histogram displays the predicted probability density distributions. Panel (c): Bootstrap calibration slope distributions for Models A and B (500 resamples); the dashed horizontal line indicates perfect calibration (slope = 1.0). Panel (d): Horizontal bar chart of C-statistics from seven sensitivity analyses (S1–S6, including upper and lower bounding analyses for deeply sedated patients); the dashed red vertical line indicates Model B’s primary analysis AUC (0.686).

The overall pattern across metrics was therefore internally consistent: the nursing-sensitive block carried some collective explanatory information, but the absolute increment in patient-level discrimination was small. The near-overlap of the two ROC curves in Figure 2a reinforces this point visually, and the calibration plot in Figure 2b confirms that both models tracked the diagonal closely up to the 30% predicted-probability range. The modest reduction in Brier score and the decline in bootstrap-corrected calibration slope from 0.829 to 0.778 (Figure 2c) further suggest that the larger model added limited practical separation while increasing susceptibility to overfitting. From an explanatory standpoint, the key message is that nursing-sensitive PACU variables may contain detectable signal, but most of that signal in this cohort was concentrated in urinary catheterization rather than in the full block.

The contrast between the statistically significant LR test and the non-significant IDI is important for interpretation. The LR test indicates that the added block improved overall fit to the observed data, whereas IDI and delta-AUC address whether that improvement translated into meaningfully better separation of patients with and without delirium. In the present study, those latter improvements were small. This distinction supports a restrained interpretation: the nursing-sensitive block was not uninformative, but neither did it materially transform the model’s ability to classify individual patients.

Exploratory Associations Between PACU Delirium and Secondary Outcomes

In exploratory adjusted analyses, PACU delirium was associated with every prespecified secondary outcome (Table 7 and Figure 3). The strongest associations were observed for non-home discharge (OR 4.75), falls (OR 4.63), self-removal of lines or tubes (OR 4.06), and ward delirium (OR 3.26). PACU delirium was also associated with unplanned ICU admission, major complications, longer PACU stay, and longer postoperative hospital stay.

Table 7.

Association Between PACU Delirium and Secondary Outcomes

Outcome	Delirium (n = 422)	No Delirium (n = 1778)	Absolute Difference (95% CI)	Adjusted OR or GMR (95% CI)	p Value
Ward delirium (POD 1--3)	105 (24.9%)	143 (8.0%)	+16.9 pp (12.6--21.2)	OR 3.26 (2.45--4.34)	< 0.001
Unplanned ICU admission	29 (6.9%)	42 (2.4%)	+4.5 pp (2.0--7.0)	OR 2.88 (1.77--4.69)	< 0.001
In-hospital falls	20 (4.7%)	20 (1.1%)	+3.6 pp (1.5--5.6)	OR 4.63 (2.46--8.71)	< 0.001
Self-removal of lines / tubes	20 (4.7%)	22 (1.2%)	+3.5 pp (1.4--5.6)	OR 4.06 (2.18--7.56)	< 0.001
Major complications (CD ≥ III)	72 (17.1%)	144 (8.1%)	+9.0 pp (5.2--12.8)	OR 2.57 (1.88--3.51)	< 0.001
Non-home discharge	113 (26.8%)	130 (7.3%)	+19.5 pp (15.1--23.9)	OR 4.75 (3.56--6.34)	< 0.001
PACU length of stay (median, min)	106	66	+40 min (median)	GMR 1.56 (1.50--1.62)	< 0.001
Hospital length of stay (median, days)	11.6	7.9	+3.7 d (median)	GMR 1.46 (1.39--1.54)	< 0.001

Notes: Models adjusted for age, sex, ASA class, MoCA, surgery duration, center, CFS, CCI, stroke/TIA, and prior delirium. Rare outcomes were analyzed with Firth’s penalized logistic regression. pp, percentage points. Absolute differences for binary outcomes are crude between-group differences in event proportions with Wald 95% confidence intervals; for length-of-stay outcomes the unadjusted median difference is shown.

Figure 3.

Secondary Outcomes: Forest Plot and Length-of-Stay Distributions by PACU Delirium Status. Panel (a): Forest plot of adjusted odds ratios and geometric mean ratios (95% CI) for the association between PACU delirium and secondary outcomes. Blue markers represent binary adverse outcomes (logistic regression); red markers represent length-of-stay outcomes (log-linear regression). The dashed vertical line at 1.0 indicates the null value. Panel (b): Boxplot with jittered scatter points comparing PACU length of stay (minutes) between patients without delirium (blue) and with delirium (red). Panel (c): Boxplot with jittered scatter points comparing postoperative hospital length of stay (days) between groups.

The raw event rates and absolute risk differences (ARDs) reinforce the clinical relevance of these estimates while keeping them on an interpretable scale. Ward delirium occurred in 24.9% of patients with PACU delirium versus 8.0% without (ARD +16.9 percentage points), and non-home discharge in 26.8% versus 7.3% (ARD +19.5 points). Major complications differed by +9.0 points (17.1% vs 8.1%), unplanned ICU admission by +4.5 points (6.9% vs 2.4%), falls by +3.6 points (4.7% vs 1.1%), and self-removal of lines or tubes by +3.5 points (4.7% vs 1.2%); the corresponding ARDs with 95% confidence intervals are listed in Table 7. Patients with PACU delirium also had a median PACU stay of 106 minutes compared with 66 minutes in those without delirium, and a median postoperative hospital stay of 11.6 versus 7.9 days. Because these analyses are observational, PACU delirium should be understood as a risk marker that identifies a higher-risk group rather than as a proven cause of these downstream events.

The magnitude of the non-home discharge association was especially notable More than one quarter of patients with PACU delirium were discharged somewhere other than home, compared with only 7.3% of those without delirium. This pattern suggests that delirium identified in the PACU may capture a broader syndrome of postoperative vulnerability that extends beyond cognitive status alone. Similarly, the elevated rates of falls and device self-removal indicate that even brief early postoperative delirium has implications for nursing workload, safety monitoring, and inpatient complication prevention.

Sensitivity Analyses

Model B remained stable across sensitivity analyses, with C-statistics ranging from 0.672 to 0.692 (Table 8 and Figure 2d). Results were therefore robust to alternative handling of early sedation, exclusion of protocol deviations, restriction to major specialties, reclassification of persistently deeply sedated patients, and exclusion of isolated T0-only positives.

Table 8.

Sensitivity Analyses: Model B C-Statistic Under Alternative Analytic Conditions

Analysis	N	Events (%)	C-Statistic
Primary analysis (reference)	2,048	392 (19.1%)	0.686
S1: Exclude borderline sedation at T0	1,274	275 (21.6%)	0.689
S2: Outcome from T1 + T2 only	2,048	311 (15.2%)	0.688
S3: Major surgical specialty subgroup	1,234	242 (19.6%)	0.684
S4: Exclude protocol deviations	1,986	383 (19.3%)	0.692
S5a: Re-include deep sedation as delirium (upper bound)	2,260	482 (21.3%)	0.672
S5b: Re-include deep sedation as non-delirium (lower bound)	2,260	422 (18.7%)	0.683
S6: Exclude isolated T0-only positives	2,048	311 (15.2%)	0.691

Notes: The primary analysis and sensitivity analyses S1–S4 and S6 use the complete-case modeling set as the denominator. Analyses S5a and S5b were instead performed on the full analysis set with the 60 persistently deeply sedated patients reintroduced (N = 2260), which accounts for the larger sample in those two rows.

This consistency strengthens confidence in two aspects of the study design. First, the RASS-gated repeated-assessment framework was not unduly driven by a small subgroup of early borderline or deeply sedated patients. Second, the modest explanatory performance of the full model was not materially altered by reasonable changes in outcome definition or sample restriction. The sensitivity analyses therefore support the stability of the main inference even though the study was not designed as a formal prediction model.

The sensitivity results are also informative in a practical sense. When the outcome was redefined using only T1 and T2, discrimination remained essentially unchanged, suggesting that the overall correlational structure was not dependent on the earliest screen alone. Likewise, excluding isolated T0-only positives slightly increased the C-statistic, which offers some reassurance that the main findings were not dominated by possible emergence-related misclassification. Bounding analyses that reintroduced persistently deeply sedated patients yielded somewhat attenuated performance, as expected, but did not alter the overall substantive interpretation.

The imputed-data sensitivity analysis adds a further layer of reassurance. Although complete-case analysis was retained as primary because missingness affected only a minority of patients and the prespecified PACU variables were complete, the multiple-imputation results were closely aligned with the complete-case findings (Supplementary Methods 2). In particular, the catheter association and the small incremental advantage of Model B were preserved. This concordance suggests that the main conclusions are unlikely to be artifacts of excluding patients with missing albumin, anxiety, or sleep-quality data.

Similarly, the supplementary random-intercept analysis indicated that the catheter association was not simply a product of center-level catheterization practices (Additional Supplementary Analysis). Although the small number of hospitals limits the precision of variance-component estimation, the fixed-effects and random-effects results pointed in the same direction. Together with the low ICC, this pattern supports the interpretation that the main findings are not driven primarily by one outlying center or a highly idiosyncratic local workflow.

Exploratory Bedside Risk Index

An exploratory bedside risk index was derived from the statistically significant Model B predictors (older age, lower MoCA, ASA ≥ III, stroke/TIA, chronic kidney disease, known cognitive impairment, higher preoperative anxiety, lower albumin, blood loss > 500 mL, and urinary catheterization; Supplementary Methods 1); the full set of weighted score components and the explicit linear-score equation are reported in Supplementary Table S3A. Because the blood loss > 500 mL term retained the biologically implausible inverse weight discussed in the Explanatory Models (Model A) results above, it contributes negatively to the score and must not be read as clinically protective; it was kept solely to mirror the significant Model B predictors, and its presence further underscores that the index is exploratory. Under leave-one-center-out validation with predictor selection repeated de novo in each training set, discrimination was modest and varied across held-out centers (AUC range 0.604–0.668; median 0.640; Supplementary Table S3B). Calibration was also unstable across centers, and the median AUC of approximately 0.64 lies well below the 0.75–0.80 range generally considered a minimum for individual-level risk stratification. Because this discrimination is substantially lower than that of the full explanatory models and varies appreciably between sites, the index is reported strictly as hypothesis-generating and is not intended for bedside clinical decision-making without prospective external validation.

Discussion

This multicenter study yielded three main findings. First, PACU delirium affected roughly one in five older Chinese surgical patients and was predominantly hypoactive. Second, adding eight nursing-sensitive PACU variables to a strong conventional model improved fit statistically but only modestly improved discrimination, with urinary catheterization emerging as the only independent nursing-sensitive correlate. Third, PACU delirium was consistently associated with adverse in-hospital outcomes and is best regarded as a risk marker for a higher-risk recovery trajectory. Overall, the findings suggest that the most defensible nursing contributions in the PACU are structured detection and timely communication; catheter minimization is a biologically plausible but as-yet-unproven target. Importantly, because the nursing-sensitive block as a whole did not add clinically meaningful discrimination, our findings do not by themselves justify broad changes to nursing practice based on that block, and we are careful below not to conflate a statistically significant improvement in model fit with clinically useful predictive value.

Incidence and Detection

The observed incidence is consistent with the broader literature and supports the view that PACU delirium is a common early postoperative event in older adults. Our estimate is close to pooled Chinese postoperative delirium estimates and sits within the broad incidence range reported internationally for PACU delirium, although exact comparisons are complicated by differences in case mix, timing, and diagnostic tools.^9,12,22 The negligible ICC suggests that after adjustment, patient-level vulnerability dominated over hospital-level variation, though the precision of such variance estimates is limited with only five centers.³⁴

The predominance of the hypoactive form has direct clinical implications because quiet, inattentive patients are easy to miss in a high-throughput recovery environment. This pattern aligns with previous subtype studies in ICU and hospital cohorts, where hypoactive or predominantly hypoactive presentations were also common.^27,35 The repeated-assessment finding is equally important: more than one quarter of delirium cases were detected only at T0, so a single later screen would have underestimated burden. That observation echoes prior work showing that delirium fluctuates and that single timepoint assessment underestimates true burden.^9,36 RASS gating reduced, but could not completely eliminate, overlap between early delirium and residual anesthetic effects.

At the same time, our incidence estimate should not be interpreted as directly interchangeable with studies using different instruments or one-time assessment strategies. Studies relying on routine observation, Nu-DESC screening, or late postoperative ascertainment are likely to capture overlapping but not identical clinical phenomena. The present estimate therefore most directly reflects delirium detectable during structured serial recovery-room assessment in older adults after major noncardiac surgery. That distinction matters for interpretation because the PACU is a transition environment in which both residual anesthetic effects and evolving postoperative delirium may coexist.

Incremental Value of Nursing-Sensitive Variables

The conventional clinical model showed moderate discrimination, comparable to the modest performance typical of many perioperative delirium association models.¹⁴ Even after adding a broad range of baseline and intraoperative variables, much of the variance in delirium remained unexplained, underscoring the multifactorial nature of the syndrome. Our model performance is also lower than that reported for dedicated ICU delirium prediction tools, but the settings, timing, and available predictors differ substantially. The nursing-sensitive block produced only a small incremental gain. Although the LR test suggested statistically detectable additional signal, the ΔAUC (+0.011) and non-significant IDI indicate limited clinical improvement in discrimination. These two questions are distinct: the LR test asks whether the block improves fit to the observed data, whereas the ΔAUC and IDI ask whether it improves the separation of individual patients with and without delirium. In this cohort the answer to the first was a qualified yes and to the second essentially no. A statistically significant LR test with a C-statistic remaining near 0.69 does not constitute evidence of clinically actionable predictive value, and we therefore do not recommend adopting the nursing-sensitive block as a risk-prediction instrument or using it to justify broad practice change. This interpretation is consistent with methodological cautions surrounding NRI and related incremental-value metrics.^29–33

Several factors may explain the modest incremental value of the PACU block. First, some PACU variables overlap temporally with the delirium assessment window and may therefore behave as concurrent correlates rather than antecedent predictors. Second, standardized PACU nursing protocols may narrow between-patient variability in pain management, warming, or routine monitoring. Third, the very low hypoxemia event rate under the prespecified threshold limited power for that variable. Fourth, the relatively short PACU observation window may be insufficient for some nursing-sensitive exposures to demonstrate measurable associations.

Urinary catheterization at PACU arrival was the only independent nursing-sensitive correlate. This finding is clinically relevant because catheter use is potentially modifiable and was recorded before the first delirium assessment. The result is directionally consistent with earlier studies and with delirium-prevention frameworks that emphasize minimizing unnecessary tethering devices.^16–18 However, confounding by indication is a substantial concern and we explicitly caution against a causal reading. Patients who arrive in the PACU with an indwelling catheter tend to have undergone longer and more complex procedures, received larger fluid volumes, and have greater baseline frailty and comorbidity-all of which independently raise delirium risk. Although our models adjusted for many of these factors, residual confounding by unmeasured surgical complexity, intraoperative fluid balance, and the clinical reasoning that prompted catheter placement cannot be excluded. We therefore interpret catheterization as a potentially modifiable risk marker that generates a testable hypothesis for prospective catheter-minimization trials, rather than as proof that the catheter itself precipitates delirium.

It is also noteworthy that the other seven nursing-sensitive variables were not independently associated with PACU delirium in the adjusted model. This does not imply that pain control, thermal management, PONV prevention, or staffing are clinically unimportant. Rather, in the present cohort, those variables either showed limited variability, were measured concurrently with the outcome, or contributed little independent signal once major baseline vulnerability markers were accounted for. In other words, the absence of statistical independence in this explanatory framework should not be mistaken for evidence that these nursing domains are irrelevant to patient care.

Conventional Predictors and Secondary Outcomes

The significant conventional predictors in the final model - older age, lower MoCA, ASA ≥ III, stroke/TIA, CKD, known cognitive impairment, higher anxiety, and lower albumin - are broadly consistent with established delirium literature.¹⁴ Several of these variables plausibly reflect baseline neurological vulnerability, physiological reserve, or broader medical frailty. Preoperative anxiety may also represent a clinically meaningful psychological vulnerability marker. The inverse association with blood loss > 500 mL was unexpected and is more likely to reflect residual confounding, coding limitations, or correlated operative management rather than a protective biological effect.

PACU delirium was also associated with every prespecified secondary outcome, including ward delirium, falls, self-removal of lines or tubes, non-home discharge, and prolonged stay. Several non-exclusive pathways could link an early delirium episode to these later events. First, delirium directly impairs attention, judgment, and safety awareness, plausibly increasing falls and self-removal of lines or tubes. Second, a quiet hypoactive episode in the PACU may go uncommunicated at handoff, so that ward staff are not alerted to heightened surveillance needs, allowing evolving ward delirium and safety events to go undetected. Third, delirium frequently coincides with reduced early mobilization and functional deconditioning, which prolong hospitalization and reduce the likelihood of discharge home. Finally, part of the association is almost certainly shared vulnerability: the same frailty, cognitive, and physiological reserve deficits that predispose to PACU delirium independently predict complications and non-home discharge. The association with subsequent ward delirium is consistent with prior perioperative work suggesting that PACU delirium may function as an early sentinel event.¹² Likewise, the observed relationship with falls is consistent with broader evidence linking delirium to inpatient fall risk. These findings reinforce the clinical relevance of early PACU delirium detection, even if explanatory models remain only moderately discriminative. In practical terms, delirium detected in the PACU should trigger communication and heightened surveillance rather than be dismissed as a transient recovery-room abnormality.

The association with prolonged PACU and hospital stay also has resource implications. Some of the excess PACU time may reflect the fact that patients identified as delirious appropriately receive closer observation before discharge from the unit. Even so, the consistent association across multiple outcomes suggests that PACU delirium is not merely a measurement artifact. Instead, it appears to identify a group of patients at elevated near-term clinical risk who may benefit from targeted observation, handoff, and nonpharmacological prevention strategies during the early postoperative period.

Clinical Implications

The practical implications are straightforward, but it is essential to distinguish detection from prediction. Our recommendation for structured screening rests on the high observed burden of delirium-especially hard-to-recognize hypoactive cases-and on the validated detection performance of the CAM-ICU, not on the discriminative accuracy of our explanatory or exploratory risk models, which was only moderate and is insufficient for individual-level risk stratification. In other words, we advocate systematically measuring delirium that is already present, not predicting which patients will develop it. First, older PACU patients should undergo structured delirium screening rather than reliance on clinical impression alone, in line with contemporary guidance from perioperative and geriatric societies.^7,8 Second, screening should occur at least twice, once early in PACU recovery and once before discharge. Third, PACU delirium status should be incorporated into handoff to the ward because it identifies patients at higher risk of subsequent delirium and other adverse outcomes. Fourth, catheter minimization deserves targeted quality-improvement and interventional evaluation. More broadly, these findings support nurse-led multicomponent surveillance and communication approaches similar in principle to those tested in HELP and other delirium-oriented care bundles.^16,19,20

From an implementation perspective, the study suggests a practical workflow rather than a burdensome new program. A brief arousal check followed by structured CAM-ICU assessment at predefined PACU timepoints is feasible, reproducible, and aligned with the observed temporal pattern of cases in this cohort. A positive PACU screen could then trigger three immediate nursing actions: enhanced observation in the PACU, explicit communication during ward handoff, and avoidance of potentially unnecessary lines or restraints where clinically safe. Such a workflow would be consistent with current evidence emphasizing early recognition, reorientation, and minimization of modifiable precipitating factors.

Strengths and Limitations

Strengths include the multicenter design, repeated validated assessments, focus on nursing-sensitive PACU variables, and prespecified model comparison strategy with multiple sensitivity analyses. The use of a strong conventional benchmark reduced the risk that apparent incremental value would be inflated by an under-specified comparator. Limitations include the observational design, temporal overlap between some PACU exposures and the delirium assessment window, moderate discrimination and some overfitting, missing data in several preoperative variables, restriction to tertiary Chinese hospitals, and limited power for hypoxemia because of its low event rate. Long-term cognitive outcomes were not captured, and residual confounding by unmeasured perioperative variables remains possible.

Several additional limitations warrant emphasis. First, despite repeated assessment, some very early positive CAM-ICU results may still have reflected emergence-related phenomena rather than fully established delirium, even though the sensitivity analyses were reassuring on this point. Second, because the study was designed as an explanatory association analysis rather than a prediction tool development project, the models should not be used for bedside individual risk stratification. Third, although complete-case and imputed analyses were concordant, missingness in self-reported and laboratory variables could still have introduced some bias. Fourth, the tertiary-hospital setting may limit generalizability to smaller hospitals, lower-resource PACUs, or non-Chinese populations. Finally, the outcome horizon was short, so the study cannot address whether PACU delirium independently predicts longer-term cognitive decline or functional deterioration.

An additional interpretive limitation is that some ostensibly modifiable nursing-sensitive variables may themselves be downstream expressions of the same perioperative vulnerability that gives rise to delirium. For example, pain, PONV, or shivering may not function as independent precipitating causes in a brief PACU window; instead, they may co-occur with delirium because both reflect the intensity of surgical stress, anesthetic recovery, or baseline frailty. This conceptual overlap reinforces why the present analysis is best understood as explanatory association rather than causal attribution. It also explains why preserving temporal ordering is likely to be critical in future mechanistic or interventional studies.

Future Directions

Future work should externally validate the exploratory risk index, test catheter-minimization and screening interventions prospectively, and use landmark or incident-outcome designs that better separate exposure timing from delirium onset. Longer follow-up is also needed to determine whether PACU delirium independently predicts later cognitive decline. Additional work could also explore whether more granular intraoperative measures, including depth-of-anesthesia indices or hemodynamic trajectories, improve explanatory performance beyond the variables available in the present cohort.

Future research should also evaluate implementation questions. These include whether routine repeated PACU delirium screening is feasible at scale in busy units, how screening results should best be integrated into handoff workflows, and whether targeted nurse-led responses after a positive PACU screen reduce downstream complications. Cost-effectiveness analyses and pragmatic quality-improvement studies would be especially valuable for translating observational findings into perioperative practice.

Another important next step is refinement of variable timing. Landmark analyses focused on incident delirium after an initial negative screen could better distinguish antecedent predictors from concurrent correlates. That approach may be especially useful for PACU variables such as pain, oxygenation, PONV, and shivering, which are clinically meaningful but difficult to interpret causally when exposure and outcome are measured in the same short interval. More detailed exposure definitions, including alternative hypoxemia thresholds or cumulative burden measures, may also reveal associations that were not detectable under the current coding scheme.

Conclusion

PACU delirium affected approximately one in five older Chinese adults undergoing major noncardiac surgery, was predominantly hypoactive, and was associated with multiple adverse outcomes. Conventional clinical vulnerability markers explained most of the measurable signal. Adding nursing-sensitive PACU variables yielded a statistically detectable but clinically limited improvement, with urinary catheterization as the only independent nursing-sensitive correlate. The clearest clinical message is therefore that structured, repeated PACU delirium screening-together with explicit PACU-to-ward communication-is clinically important and feasible, whereas the incremental value of the nursing-sensitive variable block as a predictive instrument is modest. Urinary catheter minimization is a plausible, testable target rather than a proven intervention, and the exploratory risk index requires prospective external validation before any bedside use.

Funding Statement

This research received no external funding.

Data Sharing Statement

The data that support the findings of this study are not publicly available because they contain potentially identifiable patient information. De-identified data may be available from the corresponding author upon reasonable request and subject to institutional approval and applicable data-protection requirements.

Ethics Approval and Consent to Participate

This multicenter study was reviewed and approved on 18 December 2023 by the Medical Ethics Committee of Shanxi Provincial People’s Hospital (Taiyuan, Shanxi, China; approval number: 2023-312), which acted as the central institutional review board for all five participating tertiary hospitals; ethics approval preceded the start of enrollment in January 2024. In line with the regulations governing multicenter clinical research in China, the four collaborating sites (S02–S05) formally accepted and recorded the central ethics approval issued by the Medical Ethics Committee of Shanxi Provincial People’s Hospital prior to the start of local enrollment; no separate institutional approval numbers were issued at the collaborating sites under this central-review arrangement. All study procedures were conducted in strict accordance with the principles of the World Medical Association’s Declaration of Helsinki (as revised in 2013) and with applicable national laws and institutional ethical standards governing research involving human participants. Written informed consent was obtained from every patient preoperatively before any study-specific procedures were undertaken. Decisional capacity was assessed preoperatively by trained study clinicians; for patients with limited decisional capacity, written informed consent was obtained from a legally authorized representative (proxy consent), as permitted by the approved protocol. All participant data were de-identified prior to analysis and stored on encrypted, access-controlled servers maintained by the coordinating center.

Author Contributions

Xuesong Ma: Conceptualization, Methodology, Formal analysis, Writing - Original Draft. Manhong Zhang: Investigation, Data curation, Validation, Writing - Review & Editing. Fangfang Wang: Supervision, Methodology, Project administration, Writing - Review & Editing. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare no conflicts of interest in this work.