This was an observational cohort study. We present the following article in accordance with the TRIPOD reporting checklist. The selected study participants were 400 inpatients in an ICU with four wards in the affiliated hospital of a medical university in Heilongjiang Province from September 2021 to June 2022, and the data of 280 patients collected from September 2021 to March 2022 were used to develop the model. The criteria for inclusion and exclusion of study participants were shown in Table 2.

A total of 26 observational variables were included in this study, based on the predictors screened by the research group’s preliminary meta-analysis and expert consultation. According to the modeling sample size formula, the recommended sample size is 5–10 times the number of observed variables (16), considering the limitations of workforce and time in this study, the sample size was set to five times the number of variables. Moreover, the meta-analysis by Zang et al. (17) revealed an incidence of ICU-AW is 51.90%, considering the possible loss of 20% sampled patients, we aimed for a final sample size of ≥300 cases. A total of 400 ICU patients were enrolled actually. According to the logistic regression requirements for development of the model, patients were divided into the modeling group (n = 280) and the verification group (n = 120) in a ratio of 7:3. In the modeling group, 40 patients (14.29%) had ICU-AW, and 240 patients (85.71%) did not. In the validation group, 21 patients (17.5%) had ICU-AW and 99 patients (82.5%) did not.

The questionnaire was designed that included 26 risk factors. The results of our previous meta-analysis were included, and we consulted the opinions of experts in the ICU of our research center (including physicians and nurses with >10 years’ experience who were engaged in ICU-AW related research). Based on the preliminary meta-analysis results and expert opinions, we finally made this questionnaire for researchers to collect data. Patient datasets were collected through the Hospital Information System (HIS) and included data on a total of 26 risk factors in the following four categories see Table 4. All patient data were entered and reviewed by the researcher using Excel software. For data collection, all the researchers were trained uniformly to ensure consistency. The researchers participated in the entire data collection process and recorded it. The included participants were assessed once daily by a critical care nurse and a researcher using the MRC scale in an awake state. If the patient develops ICU-AW, the evaluation is discontinued; If the patient did not develop ICU-AW, evaluation was continued on day 2 by another investigator until the patient was transferred out of the ICU. After the study began, the researchers summarized the data collected each week and provided timely feedback and adjustments to problems that occurred during the data collection process.

IBM SPSS Statistics version 27.0 and MedCalc software were used for statistical analysis. Nominal variables are reported as numbers and proportions (%), and the chi-square test was used for group comparisons. Continuous data with a normal distribution are reported as the mean ± standard deviation (mean ± SD), and a t-test was used for group comparisons. Continuous data with a non-normal distribution are reported as medians and interquartile range M (P₂₅, P₇₅), and the non-parametric Wilcoxon rank sum test was used for group comparisons.

To develop the model, we tested the correlations between predictors (P < 0.05 in the univariate analysis) and ICU-AW using a binary logistic regression model (using step forward regression). The model was corrected using the Hosmer-Lemeshow goodness of fit test (P > 0.05 indicated good model fit), and the discriminatory capacity of the model was indicated by the receiver operating characteristic (ROC) area under the curve (AUC). After the Youden index was used to determine the optimal critical value of the ROC curve, the sensitivity and specificity of the prediction model were estimated. The model was internally validated using the bootstrap method with 500 replications. A normality test was conducted according to the scores of patients in the model validation group and the risk levels were categorized as follows: <50% (low risk), 50–75% (medium risk), and >75% (high risk).

The study population was divided into the ICU-AW group and no ICU-AW groups and univariate analysis of the clinical data of the two groups was undertaken. The results showed that there were statistically significant differences between the two groups for age, gender, body mass index, shock, length of ICU stay, use of norepinephrine, days of immobility, parenteral nutrition, mechanical ventilation time, and infectious diseases (P < 0.05) (Table 5).

The results are shown in Tables 6, 7. The results showed that gender (OR: 4.31, 95% CI: 1.682–11.042), shock (OR: 3.473, 95% CI: 1.191–10.122), mechanical ventilation time (OR: 1.592, 95% CI: 1.317–1.925), length of ICU stay (OR: 1.085, 95% CI: 1.018–1.156) and age (OR: 1.075, 95% CI: 1.036–1.115) were independent risk factors for ICU-AW in ICU patients. The prediction model formula is shown in Figure 3.

The Hosmer-Lemeshow test was χ² = 7.656 and the degree of freedom was eight, P = 0.468. A Calibration curve as shown in Figure 4. The ROC curve of the ICU-AW risk prediction model as shown in Figure 5. The area under the ROC curve of this model was 0.904, 95% CI (0.847, 0.961), the Youden index was 0.733, the optimal truncation value was 0.145, the sensitivity was 87.5%, and the specificity was 85.8%. See Figure 6, Table 8 for the area under the curve of each predictive variable in the model.

The 0.632 Bootstrap was used to re-sample for 500 times, and the AUC was 0.889. A total of 120 patients admitted to the ICU in the same hospital from March to May 2022 were selected as the model validation group. There were 72 males and 48 females. The age was 60.02 ± 15.33 years. According to the formula of this prediction model, ICU-AW was considered to occur when Z ≥ 0.145. ICU-AW occurred in 21 cases (17.5%). The sensitivity of the model was 71.4% (15 cases, with six cases misjudged). There were 99 cases without ICU-AW; 92 cases were predicted by the model, seven cases were misjudged, and the specificity was 92.9%. The overall accuracy of the model was 89.2%.

Normality tests were conducted according to the scores of patients in the model validation group, and the results showed a skewed distribution, P₂₅ = –4.42, P₅₀ = –3.30, P₇₅ = –1.61. The risk levels were divided as follows: <50% (predicted probability value <–3.30) was considered low risk; 50–75% (predicted probability value –3.30 to 1.61) was considered medium risk; and >75% (predicted probability value >–1.61) was considered high risk. Of the 120 patients in the validation group, 59 patients (49.2%) were classified into the low-risk group. There were 31 cases in the moderate risk group, accounting for 25.8%, and 30 cases in the high-risk group, accounting for 25%. A model score <–3.303 was defined as a low-risk group, and the goal of the health care provider was to reduce exposure to ICU-AW risk factors. A model score of >–3.303 and <–1.606 defined the medium-risk group, this group required closer monitoring, with a focus on related risk factors. A model score of >–1.606 was considered to indicate high-risk, and necessary measures were required to prevent the occurrence of ICU-AW in patients in this group.

As expected, the age of ICU-AW patients in this study was significantly higher than that of the no ICU-AW patients, suggesting that age is an independent risk factor for ICU-AW. As many countries around the world gradually enter an aging society, the average age of ICU patients increases, and patients will have problems such as decreased muscle reserve and muscle strength before they are admitted to ICU. However, the severity of the disease accelerates these changes, and even leads to the onset of muscle decay syndrome (30, 31), which may be the mechanism by which age influences the occurrence of ICU-AW. Related studies included age as one of the independent predictors of ICU-AW in their risk prediction models (8, 14), which is consistent with our study. In addition, a significant association was observed between gender and ICU-AW (OR: 4.310, 95% CI: 1.682–11.042). Through independent variable assignment, we found that the risk of ICU-AW in women was up to 4.31 times higher than that in men, which was consistent with other studies (18, 26, 32, 33). Nevertheless, whether gender is an exact independent risk factor for ICU-AW remains controversial, and the mechanism is unclear. Although an observational study did not find an association between sex hormones and ICU-AW (34), for the purpose of risk prevention and control, clinical staff should also regard women as a high-risk group for ICU-AW and focus on monitoring. Identifying immutable factors (e.g., age, sex) is crucial for early prediction and to understand the causal relationship between the relevant variables and ICU-AW.

Interestingly, shock was strongly associated with the development of ICU-AW as an independent predictor, which is related to the possibility that microcirculation disorders may lead to neuronal damage and axis mutations. A prospective study conducted by Anastasopoulos et al. (31) confirmed that shock is an important risk factor for ICU-AW in critically ill patients, which is consistent with the results of this study. However, few studies have reported the association between shock and ICU-AW in recent years. This lack of information is surprising because shock is not uncommon in ICU patients. This predictor (shock) is an important finding in understanding factors associated with ICU-AW.

Our study found that the duration of mechanical ventilation was positively correlated with ICU-AW, that is, the longer the duration of mechanical ventilation, the higher the incidence of ICU-AW. This finding is consistent with the results from various studies (10–12, 35, 36). Long-term mechanical ventilation can cause certain pathophysiological changes in the diaphragm, resulting in atrophy and dysfunction of the diaphragm. Within 48 h of mechanical ventilation, diaphragm injury was positively correlated with ventilation support, and diaphragm loss was most obvious in volumetric respiratory support mode (37). Ventilator-induced diaphragm dysfunction may also be associated with ICU-AW. We observed a significant correlation between the length of ICU stay and ICU-AW. Patients in ICU-AW group had spent significantly longer in ICU than those without ICU AW, which was consistent with the findings of Hermans and Van den Berghe (38). This may be due to the fact that ICU patients are prone to microcirculation disorders and malnutrition, leading to decreased muscle strength, and patients may show anxiety and depression during ICU treatment, which affects compliance with early rehabilitation. The risk of ICU-AW increases with the length of ICU stay. Although there is a lack of direct evidence that psychological factors are independent risk factors for ICU-AW, we hypothesized that psychological factors may influence patient compliance with early rehabilitation and exercise, leading to the occurrence of ICU-AW. This gives us an important inspiration that ICU medical staff should not neglect the psychological care of patients while providing specialized care, which is of inestimable value in improving the clinical outcome and reducing the occurrence of post-ICU syndrome. It also provides a new idea for further verifying the relationship between psychological factors and ICU-AW.

Researchers believe that immobilization is the primary risk factor for ICU-AW (13). To our surprise, days of immobilization were found to be strongly associated with ICU-AW in univariate analysis believe that immobilization is the primary risk factor for ICU-AW (13). To our surprise, days of immobilization were found to be strongly associated with ICU-AW in univariate analysis (P < 0.001), but in logistic regression analysis, the number of days of immobilization did not appear to be an independent risk factor for ICU-AW and was not included in the model. This may be related to confounding factors and individual patient differences. Adding this factor (immobilization) to the model reduced its sensitivity to other covariables, so we did not include it in this model. Even so, ICU medical staff should carry out scientific and reasonable treatment and nursing and strictly manage the duration of physical restraint to reduce the risk of ICU-AW.

This model can predict the risk of ICU-AW at the beginning of admission to the ICU. After admission to ICU, the patient’s indicators were put into the model formula (Z = 0.081 × length of stay in ICU + 0.465 × days of mechanical ventilation + 1.245 × shock + 1.461 × gender + 0.072 × age-9.276). Patients were classified as low risk when model score <–3.303; When –3.303 <model scores <–1.606, it was classified as medium risk. Medical staff should pay enough attention to it and focus on monitoring various indicators to avoid the occurrence of ICU-AW in patients; When the model score is >–1.606, which is classified as high risk. Patients who have a higher probability of developing ICU-AW, and medical staff should pay high attention to it and actively intervene in patients.

Our study had certain advantages. First, the cohort study design is helpful to improve the universality of the risk prediction model (39), since it allowed us to carefully measure and record predictors and outcomes to improve their applicability to ICU patients. Second, in the process of data collection, unified diagnostic criteria were used to evaluate the occurrence of ICU-AW, with blind judgment of outcome indicators to avoid participative bias as much as possible. At the same time, the risk prediction model developed in this study was evidence-based to screen the related risk factors of ICU-AW. The included risk factors were scientific, comprehensive, and in line with clinical requirements. In this risk prediction model developed by logistic regression, although it is difficult to completely avoid the influence of confounding factors, we tried to reduce the influence of confounding factors as much as possible by explaining all major variables. Finally, the predictive variables included in the model are easy to measure, data acquisition is convenient, and clinical indicators can be obtained at an early stage of ICU admission, enabling ICU medical staff to identify the risk of ICU-AW as early as possible. This risk prediction model developed in this study has great performance and high predictive value, which is helpful for medical staff to implement scientific prevention strategies according to the predictive factors, to reduce the incidence of ICU-AW.

However, our study has some limitations. Firstly, due to the strict limitations of the inclusion criteria (requiring patients to be conscious), while the use of MRC scale is helpful for bedside muscle strength assessment, it may miss patients who are unconscious but have ICU-AW. Therefore, we believe that the incidence of ICU-AW may be underestimated (14.39% in the model group and 17.5% in the validation group in this study). The clinical assessment of ICU-AW is challenging, and one of the most critical factors limiting its assessment is altered patient consciousness, which can be due to a variety of reasons, including sedative drug use. We tried to overcome this problem by suspending sedative use prior to bedside muscular strength assessment. However, as other methods for diagnosing ICU-AW are invasive and expensive, it is necessary to use the MRC scale for bedside muscle strength assessment in clinical practice. Secondly, the risk prediction model constructed in this study is a static model, which can calculate the probability of ICU-AW after admission to ICU. However, since patients’ health may improve or deteriorate at any time during their stay in ICU, the probability of developing ICU-AW may also change; our model lacks consideration for changes in patients’ health status. Most risk prediction models for ICU-AW have this limitation, but even so, it makes sense to develop a risk prediction model that uses dynamic parameters as predictors, such as APACHE II scores. New risk factors for ICU-AW may emerge during ICU treatment, possibly during or after ICU admission, suggesting that the model needs to be updated in the future. Of course, this also provides an opportunity to further improve the performance of the model. Thirdly, COVID-19 patients were not taken into account in this risk prediction model. This is because COVID-19 patients were not treated in this research center but uniformly transferred to other designated hospitals. Therefore, no COVID-19 patients were collected during the data collection process. However, it also provides a new perspective for subsequent studies to investigate the association between COVID-19 patients and the development of ICU-AW. Fourthly, it is of great significance to establish a risk prediction model for predicting severe ICU-AW. Due to the insufficient sample size of patients with severe ICU-AW in this study, a prediction model suitable for predicting severe ICU-AW patients was not established, and the sample size will be further increased in the future. Finally, both the development and validation of the model were carried out in one hospital, and the clinical applicability of the results of this study in other countries or regions needs further evaluation. Large sample data are still needed to verify the model in the future, but the process of model validation is also the process of model optimization.