Predictors of Delirium Incidence, Time to Onset, and Recurrence in a Mixed Medical-surgical ICU Population: A Secondary Analysis Using Cox and CHIAD Decision Tree Modeling

Background: To address whether in intensive care unit (ICU) patients, which factors correlate with development of delirium (primary outcome), as well as more rapid delirium onset and recurrence (secondary outcomes). Methods: A retrospective secondary analysis of 4,200 patients was collected from two academic medical centers. Delirium was assessed with the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) in all patients. Univariate and multivariate Cox models, logistic regression analysis, and Chi-square Automatic Interaction Detector (CHAID) decision tree modeling were used to explore delirium risk factors. Results: Increased delirium risk as associated with exposed only to articial light (AL) hazard ratio (HR) 1.84 (95% CI: 1.66-2.044, P<0.001), physical restraint application 1.11 (95% CI: 1.001-1.226, P=0.049), and high nursing care requirements (>8 hours per 8-hour shift) 1.18 (95% CI: 1.048-1.338, P=0.007). Delirium incidence was inversely associated with greater family engagement 0.092 (95% CI: 0.014-0.596, P=0.012), low staff burnout and anticipated turnover scores 0.093 (95% CI: 0.014-0.600, P=0.013), non-ICU length-of-stay (LOS)<15 days 0.725 (95% CI: 0.655-0.804, P<0.001), and ICU LOS ≤ 15 days 0.509 (95% CI: 0.456-0.567, P<0.001). CHAID modelling indicated that AL exposure and age <65 years conveyed a high risk of delirium incidence, whereas SOFA score ≤ 11, APACHE IV score >15 and natural light (NL) exposure were associated with moderate risk, and female sex were associated with low risk. More rapid time to delirium onset correlated with baseline sleep disturbance (P=0.049), high nursing care requirements (P=0.019), and prolonged ICU and non-ICU hospital LOS (P<0.001). Delirium recurrence correlated with age>65 years (HR 2.198; %95 CI: 1.101-4.388, P=0.026) and high nursing care requirements (HR 1.978, 95% CI: 1.096-3.569), with CHAID modeling identifying AL exposure (P<0.001) and age >65 years (P=0.032) as predictive variables. Abbreviations: APACHE IV means Acute Physiology and Chronic Health Evaluation IV; SOFA means Sequential Organ Failure Assessment; MV means mechanical ventilator; LOS means length of stay; # noise related to the nursing stations, staff conversation in patients' bedside and medical devices; * statistically signicant. a As determined by having family at bedside for ≥ 2 hours daily; b As determined by the six-item cognitive impairment test (6-CIT) and > 8 score signicant as cognitive impairment; c As determined by the Pittsburgh Sleep Quality Index (PSQI) and PSQI score > 5 indicate worse sleep quality ; d As determined by the Charlson Comorbidity Index based on the International Classication of Diseases (ICD) that a score of zero indicates that no comorbidities were found and the higher the score shows comorbidity; e As determined by the ICU mobility score (IMS) is scored from 0 to 10, with a score of 0 to 4 meaning low mobility, 4 to 8 moderate mobility and a score between 8 and 10 meaning high mobility; f As determined by the anticipated turnover scale (ATS); g Noise related to the nursing stations, staff conversation in patients' bedside and medical devices, h As determined by requiring > 8 hours nursing care in an 8 hour shift. Non-ICU hospital LOS ICU LOS groups. as measured by the APACHE IV and SOFA score groups = 0.357) and 0.305), respectively. evaluation, SOFA: Sequential organ failure assessment, LOS: Length of stay, SD: Standard deviation, OR: Odds ratio, HR: Hazard ratio, CHAID: Chi-square Automatic Interaction Detector. MV: Mechanical ventilation, PSQI: Pittsburgh sleep quality index.

Ambient noise level and use of an alarm silence strategy were assessed using the TES 1352A sound level meter (SLM) device (TES Electrical Electronic Corp.,Taiwan) with a range of 30-130 decibel (dB). It has a 1.27 cm electret condenser microphone and accuracy of ± 1.5 dB (ref 94 dB@1KHz) [51]. For the most accurate estimate of what a patient would hear, the sound meter was placed adjacent to the patient's head (or on their pillow if out of the room) for measurements. Measurements were made by the patient's nurse three times daily (10 AM, 5 PM, and 10 PM).
As the World Health Organization (WHO) recommends noise levels in hospitals should be ≤ 40 dB during the day, and ≤ 35 dB during the night shift [52], days were categorized as noisy if ≥ 1 reading measured > 40 dB. Patients were then grouped according to environmental noise into more noise (≥ 50% days with measurements ≥ 40 dB) and less noise groups (< 50% dayswith measurements ≥ 40 dB).
At our center, we have a noise control policy to minimize sound by utilizing an alarm silence strategy which was accentuated during the study. The alarm load is minimized by setting alarms based on the patient's condition and planned reduction of unnecessary alarms. For example, alarms are silenced proactively when performing bedside procedures such as endotracheal tube suctioning, phlebotomy, and when handeling invasive lines. Alarms are then reset upon task conclusion. Additionally, ambient noise is reducedby muting personal phones, limiting unnesessary staff conversation in patient care and common areas. After the alarm silence strategy, noise intensity was measured again in the same locations using a sound meter in dB. If the level of noise was reduced < 40 dB, this item would be considered as positive for patients.
Categorical variables were expressed as counts (percentage), and continuous variables as mean ± standard deviation (SD). Patients were strate ed by the occurance or absence of delirium during the ICU LOS, and demographic and clinical characteristics were assessed using t-test with continuous variable and Chi-Square, or Fisher's Exact test (as appropriate) with categorical variables.
Univariate and multivariate Cox models were separately used to assess predictors of delirium incidenceas the time to delirium onset. In the Cox model, the time to delirium onset was the main predictor. In the multivariate analyses, the signi cant variables in a backward selection modeling (considering Pentry = 0.05 and Premoval = 0.10) were reported as hazard ratio (HR) with 95% CI, also; a multivariate linear regression was used to predict theoccurance of delirium. In addition, multivariate logistic regression was used to identify those factors exerting a statistically signi cant effect on the incidence of delirium recurrence by using backward method and the signi cant variables were reported as odds ratio with 95% CI.
Chi-square Automatic Interaction Detector (CHAID) decision tree analysis is a data mining technique which can be demonstrate the relationship between split variables and related factors in homogeneous population subgroups [54]. Moreover, CHAID enables one to deal with whole variables, partition consecutive data effectively, and make decision trees by using a forward stopping or pruning rule [55,56].
For CHAID decision tree analysis in this study, all parameters collected for delirium incidence and recurrence were used. The minimum parent and child nodes were determined as 100 and 50, respectively. "Nodes" are midpoints or terminal points after bifurcation according to each factor. The parent nodes are the nodes before bifurcation, and the child nodes are ones after bifurcation. Based on the result, a group of patients was divided into one of the terminal nodes (risk groups) with predictive probability calculated.
Signi cance was determined as an alpha of 0.05.

Results
A total of 4,200 subjects were included in the analysis. Demographic and clinical characteristics are presented in Table 1  Abbreviations: APACHE IV means Acute Physiology and Chronic Health Evaluation IV; SOFA means Sequential Organ Failure Assessment; MV means mechanical ventilator; LOS means length of stay; # noise related to the nursing stations, staff conversation in patients' bedside and medical devices; * statistically signi cant. a As determined by having family at bedside for ≥ 2 hours daily; b As determined by the six-item cognitive impairment test (6-CIT) and > 8 score signi cant as cognitive impairment; c As determined by the Pittsburgh Sleep Quality Index (PSQI) and PSQI score > 5 indicate worse sleep quality ; d As determined by the Charlson Comorbidity Index based on the International Classi cation of Diseases (ICD) that a score of zero indicates that no comorbidities were found and the higher the score shows comorbidity; e As determined by the ICU mobility score (IMS) is scored from 0 to 10, with a score of 0 to 4 meaning low mobility, 4 to 8 moderate mobility and a score between 8 and 10 meaning high mobility; f As determined by the anticipated turnover scale (ATS); g Noise related to the nursing stations, staff conversation in patients' bedside and medical devices, h As determined by requiring > 8 hours nursing care in an 8 hour shift. Abbreviations: APACHE IV means Acute Physiology and Chronic Health Evaluation IV; SOFA means Sequential Organ Failure Assessment; MV means mechanical ventilator; LOS means length of stay; # noise related to the nursing stations, staff conversation in patients' bedside and medical devices; * statistically signi cant. a As determined by having family at bedside for ≥ 2 hours daily; b As determined by the six-item cognitive impairment test (6-CIT) and > 8 score signi cant as cognitive impairment; c As determined by the Pittsburgh Sleep Quality Index (PSQI) and PSQI score > 5 indicate worse sleep quality ; d As determined by the Charlson Comorbidity Index based on the International Classi cation of Diseases (ICD) that a score of zero indicates that no comorbidities were found and the higher the score shows comorbidity; e As determined by the ICU mobility score (IMS) is scored from 0 to 10, with a score of 0 to 4 meaning low mobility, 4 to 8 moderate mobility and a score between 8 and 10 meaning high mobility; f As determined by the anticipated turnover scale (ATS); g Noise related to the nursing stations, staff conversation in patients' bedside and medical devices, h As determined by requiring > 8 hours nursing care in an 8 hour shift.
Non-ICU hospital LOS (P = 0.584) and ICU LOS (P = 0.552) was similar between groups. Illness severity as measured by the APACHE IV and SOFA score was similar between groups (P = 0.357) and (P = 0.305), respectively.
Patients with and without delirium differed signi cantly in terms of cognitive impairment at the time of admission. Baseline cognitive impairment signi cantly was higher in patients with delirium (33.7% vs. 7.4%, P < 0.001). Additionally, a greater portion of delirium patients (62.7%) were observed in the AL group (P < 0.001).Other characteristics did not differ signi cantly between groups (P > 0.05) Table 1.  score was > 11, then APACHE IV score was checked. If SOFA score was ≤ 11, then patient sex was assessed. Subjects were then strati ed according risk of delirium incidence into low (< 20%), moderate (20-30%), high (30-40%), and very high (> 40%) risk groups. The ndings suggest that AL and age < 65 years conveyed a high risk of delirium incidence, whereas SOFA score ≤ 11 and female sex were associated with low risk, and APACHE IV score (> 15) score and NL were associated with moderate risk.

Time to delirium onset
Multivariate linear regression analysis was conducted to identify those variables predictive of time to delirium onset. As shown in Table 3   Abbreviations: B means coe cient; SE means standard error; OR means odd ratio, which equals to the exponentiation of B coe cient; CI means con dence interval, a Categorized as mild (<4 hours), moderate (4-8 hours), or high (>8 hours) of nursing care in an 8-hour shift Figure 3depictsthe CHAID decision tree analysis for predictingdelirium recurrence in patients with delirium present on ICU admission (n = 320). This decision tree has a depth of 2 levels from the root node, with one intermediate node, and three terminal nodes. Each node contains three statistical values, category, percentage (%) and the number (n) of patients in this particular category. As shown in Fig. 3, the main variables associated with delirium recurrence were AL exposure (P < 0.001) and age > 65 years (P = 0.032).

Discussion
Delirium is a common and serious clinical syndrome characterized by uctuating cognitive dysfunction that affects 20-80% of ICU patients [57,58]. The risk of delirium relies on the interaction between predisposing and precipitating risk factors [23,29]. It is associated with increased short-and long-term morbidity and mortality [1-5, 7-9, 11-17]. Thus, a thorough understanding of mitigating and contributing factors is necessary to development of an accurate delirium prediction model for critically ill patients.
The incidence of delirium in this study (36.7%) was consistent with that of some published studies [59,60], but lower than some other cohorts [2,15]. The median time to ICU delirium onset was similar to other published studies [61,62]. Moreover, the seven variables identi ed on Cox regression analysis were similar to other published reports [60,63,64] including: light category (arti cial vs. natural), low level of family engagement (< 2 hours at bedside per day), high nurse burnout and anticipated turnover (ATS > 35), application of physical restraints, high nursing carerequirements (> 8 hours in 8 hours shift), ICU LOS > 15 days, and hospital LOS > 15 days. The ve varibles noted to be most predictive of developing delirium on CHAID decision tree modeling were AL group and age > 65 years (high risk), APACHE IV score > 15 (moderate risk), and SOFA score ≤ 11 and female sex (low risk). As it pertains to light exposure, loss of NL exposure is associated with circadian rhythm disturbances that may affect delirium incidence and outcomes in the critically ill [26,[65][66][67]. The connection of NL vs. AL light exposure and delirium incidence has been variably reported [26,[68][69][70]. This discrepancy may be related to differences in delirium de nition, screening method, NL category criteria, and sample size [26].
Beyond grouping by light exposure type, CHAID analysis further identi ed the female gender, SOFA > 11, and APACHE IV > 15 as a risk factors in the second and third layer of the decision tree model. These factors werelikely not detected in Cox regression analysis because of higher proportion of females in participants and the similar median score of APACHE IV and SOFA in two groups. In fact, one advantage of the CHAID decision tree is thatit can divide the population into subgroups with different characteristics and estimate the prevalence in each subgroup. While, regression analysis examines risk factors throughout the whole population and treats different factors equally [71]. However, we believed both models were clinically reasonable.
According to the Cox regression analysis, high nursing care and use of the physical restraint predisposed patients to 18% and 10% greater risk of delirium, respectively, and it is consistent with the other studies in this eld [72,73]. Physical restraints are often used for critically ill patients to ensure patient safety, ensure safety and prevent the removal of medical equipment (e.g., tracheal tubes) [74]. However, the use of physical restraints in different countries varies considerably. For example, the use of physical restraints in European general ICU populations ranges from 10-50%, 76% in Canada, and up to 87% in American surgical ICUs [75][76][77]. According to one meta-analysis, the prevalence of physical restraint use in Iranian medical-surgical ICUs was 47.6%, in keeping with the ndings of this analysis [78]. Similarly, physical restraint applications havepreviously been identi ed as an independent risk factor for development of ICU delirium [75,79]. As restraint use increased two-and three-fold, observed incidence of ICU delirium increased 2.38-and 3.62-fold respectively.
Additionally, the presence of family at bedside for > 2 hours per day (reported as family engagement) was identi ed as a potential mitigating factor for ICU delirium this study, similar to other published reports [80,81]. This raises questions about the role that family may play in the care of a critically-ill loved one and presents an opportunity for inquiry as ICU visitation policies have been restricted in many cases during the current COVID-19 pandemic. Current evidence suggests that this may potentially be accomplished in the con nes of traditional visiting hours, rather more exible visitation policies that may contribute to staff burnout [82,83].
Healthcare provider turnover is an important indicator for care qualityand is widelyused as a measure for health-care system analysis. Burnout and provider turnover may disrupt patient care quality and continuity [41][42][43]84]. However, whether provider burnout is linked to patient development of ICU delirium remains unclear. In the current study, provider burnout and intent for job turnover was assessed as regards to its correlation to development of ICU delirium. This study found that delirium risk was higher in patients whose providers had higher rates of burnout and anticipated turnover as measured by ATS scores (HR 0.093, 95% CI: 0.014-0.600, P = 0.013).
To identify factors predictive of delirium recurrence amongst those patients with delirium present on ICU admission, backward logistic regression analysis and CHAID decision tree modeling identi ed exclusive AL light exposure and age > 65 years as major risk factors in the present study.Similar to prior studies, hospitalization in a room without NL exposure was associated with a 3.24-fold increase in delirium recurrence [26], whereas age > 65 years increased delirium recurrence by 2.19-fold [62]. This may not be entirely surprising, as the elderly may be more susceptible to the effects of metabolic disturbances, hypoxemia, and other stresses imposed by the critically ill state [62]. It remains unclear whether the high levels of nursing requirements associated with increased delirium recurrence are merely a re ection of patients with more severe illness or delirium, or whether it correlates with an as-yet unmeasured risk factor.
This report details the largest study of its type on ICU delirium. More than twenty related factors were analyzed using two different prediction model methods. Nevertheless, this study is not without limitations. First, our prediction model method requires knowledge of the patient's medical history. In some cases, this may be limited by recall bias, or non-availability of information. Second, it's related to the inherent limitations of an observational study design.

Conclusion
Development of ICU delirium correlated with application of physical restraints, high nursing care requirements, prolonged ICU and non-ICU hospital length-of-stay, exposure exclusively to arti cial (rather than natural) lighting, less family engagement, and greater staff burnout and anticipated turnover scores. ICU delirium occurred more rapidly in patients with baseline sleep disturbance, and recurrence correlated with presence of delirium on ICU admission, exclusive arti cial light exposure, and high nursing care requirements. Several of these factors are suitable for further study and intervention including natural light exposure, minimizing physical restraint application, and most notably the potential impacts of provider burnout and intent to turnover on patient development of ICU delirium.

Declarations
Ethics approval and consent to participate: The study was approved by the Investigative Review Board at the participating academic medical centers. Study participation was optional for respondents. Informed consent was obtained from the patient, legal guardian or healthcare surrogate or designated healthcare proxy.

Consent for Publication:
Informed consent was obtained from the patient, legal guardian or healthcare surrogate and allowed for both study participation and publication of de-identi ed aggregate results. There is no data contained within the manuscript from which individual patients or participants may be identi ed.
Availability of data and material: The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.  A CHAID decision classi cation tree analysis to predict delirium among participants.

Figure 3
A CHAID decision classi cation tree analysis to predict delirium recurrence in patients with delirium at the admission time