Early prediction of intensive care unit-acquired weakness using quadriceps ultrasound and routine clinical variables: development and temporal validation of a prospective multicentre machine-learning model

Abstract

Background

Intensive care unit-acquired weakness (ICUAW) is frequent in critically ill adults and is associated with adverse outcomes, but early recognition is difficult because standard diagnosis relies on volitional strength testing.

Methods

In this prospective multicentre cohort study across 16 tertiary ICUs in southwest China, adults expected to remain in the ICU for ≥ 3 days underwent quadriceps ultrasound and routine clinical assessment within 24 h of admission. ICUAW was defined by the first evaluable Medical Research Council (MRC) score during ICU stay. We developed and compared nine algorithms in a development cohort (n = 858) and performed temporal external validation in a later cohort (n = 345).

Results

ICUAW occurred in 579/858 (67.5%) patients in the development cohort and 181/345 (52.5%) in the validation cohort. In external validation, a random forest model integrating unpressurised ultrasound and clinical features achieved an area under the receiver operating characteristic curve (AUC) of 0.810 (95% CI 0.765–0.855), with sensitivity 0.983 (0.952–0.997) and specificity 0.726 (0.651–0.792). The corresponding pressurised-ultrasound model showed lower discrimination (AUC 0.730 [0.676–0.783]) and specificity (0.506 [0.427–0.585]). Shapley additive explanations (SHAP) highlighted quadriceps muscle thickness, Sequential Organ Failure Assessment (SOFA) score, albumin, and inflammatory markers as key contributors. In additional incremental-value analyses using the General RF model as the reference, the unpressurised RF model improved risk reclassification and discrimination (cf-NRI 0.490, 95% CI 0.277–0.698; IDI 0.072, 95% CI 0.045–0.099), whereas the pressurised RF model showed a smaller cf-NRI improvement (0.231, 95% CI 0.011–0.463) with less consistent IDI improvement (0.019, 95% CI − 0.002 to 0.040).

Conclusions

This early ultrasound-clinical model may identify patients at high risk of ICUAW before strength testing becomes feasible and support earlier targeting of preventive and rehabilitation strategies.

Trial registration Registered in the Chinese Clinical Trial Registry (ChiCTR2300075581) on September 8, 2023.

Background

Intensive care unit-acquired weakness (ICUAW) is a frequent complication in critically ill adults, particularly among patients with acute respiratory failure, shock, and other severe conditions [1, 2]. Systematic reviews indicate that ICUAW affects approximately 25%–75% of critically ill patients [3, 4] and may increase to 33%–82% among those requiring mechanical ventilation for more than four days [5, 6]. Despite advances in critical care, ICUAW remains a major contributor to poor short- and long-term outcomes, including impaired recovery and reduced quality of life [7, 8]. Notably, a substantial proportion of patients continue to have ICUAW after hospital discharge [9], and ICUAW has been reported as an independent predictor of mortality in ICU populations [10, 11].

Currently, no specific pharmacological therapy has proven effective for ICUAW, prevention and early rehabilitation strategies are central to reducing its burden [12, 13]. However, ICUAW diagnosis relies on the Medical Research Council (MRC) score, which requires patient consciousness and cooperation, rendering it inapplicable to sedated patients—a critical barrier to early detection. To address the limitations of using MRC scores for early ICUAW prediction, we aimed to develop an early ICUAW prediction model through a prospective multicentre investigation.

Although several ICUAW prediction models have been proposed, most were developed using traditional logistic regression and retrospective or small single-centre datasets, with limited external validation and consequent concerns about generalisability [4, 14]. Furthermore, most models focus on basic demographic factors (e.g., age and gender), duration of mechanical ventilation, length of ICU stay, and laboratory parameters such as albumin (ALB) and lactate, while failing to adequately integrate bedside ultrasound-derived metrics. Pathophysiological changes related to sepsis and systemic inflammatory response syndrome (SIRS), together with protein hypercatabolism and anabolic resistance induced by prolonged immobilisation, drive muscle atrophy and weakness—the core pathological mechanisms of ICUAW [15]. Studies demonstrate that bedside ultrasound-measured muscle thickness (MT), cross-sectional area (CSA), and echo intensity (EI) serve as reliable indicators of muscle atrophy and ICUAW in critically ill patients [16, 17]. Thus, incorporating muscle ultrasound metrics into predictive models may enhance the precision of ICUAW prediction.

Given the multifactorial pathogenesis of ICUAW, which involves complex, nonlinear interactions among variables, traditional logistic regression models constrained by linear assumptions fail to capture higher-order relationships. In contrast, machine-learning algorithms (e.g., random forest, gradient boosting) enable nonlinear modelling and can automatically learn feature interactions, facilitating efficient integration of multimodal data. We therefore developed and temporally validated a machine-learning model that integrates quadriceps ultrasound obtained within 24 h of ICU admission with routine clinical and laboratory variables to enable early ICUAW risk stratification.

Methods

This study was designed and reported in accordance with the Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement [18, 19].

Study design and ethical approval

We conducted a prospective, multicentre, observational cohort study in 16 tertiary ICUs in southwest China (Sept 1, 2023–Apr 30, 2025). The protocol was approved by the Ethics Committee of West China Hospital, Sichuan University (approval No. 2023[1422]). Written informed consent was obtained from all participants or their legal surrogates before enrolment. The study was registered with the Chinese Clinical Trial Registry (ChiCTR2300075581).

Participants

Eligible patients were adults (≥ 18 years) who were expected to remain in the ICU for ≥ 3 days and provided informed consent. Exclusion criteria were pregnancy, brain death, pre-existing neuromuscular disease, inability to obtain ultrasound images, or inability to undergo Medical Research Council (MRC) assessment before ICU discharge.

Outcomes

The primary outcome was ICUAW during the ICU stay. ICUAW was defined using the Medical Research Council (MRC) score after confirming satisfactory levels of consciousness and comprehension. Muscle strength was assessed in six muscle groups (three upper-limb and three lower-limb groups), each scored from 0 (paralysis) to 5 (normal strength), for a total score of 0–60. To minimise inter-examiner variability, all investigators were trained in MRC scoring by the same qualified neurologist (T.S.). Patients with an MRC score < 48 [20] or an average score < 4 [21] were classified as having ICUAW. MRC assessment was performed when patients were sufficiently awake and cooperative; the first evaluable MRC score during the ICU stay was used for outcome classification.

Candidate predictors and data collection

Baseline data were collected within 24 h of ICU admission.

Demographics: gender, age, body mass index (BMI), and body surface area (BSA).

Severity scores: Acute Physiology and Chronic Health Evaluation II (APACHE II) and Sequential Organ Failure Assessment (SOFA) scores at ICU admission.

Physiology at ultrasound examination: Richmond Agitation and Sedation Scale score (RASS), Critical Care Pain Observation Tool score (CPOT), temperature, respiratory rate (RR), systolic blood pressure (SBP), diastolic blood pressure (DBP), respiratory support mode, and fraction of inspired oxygen (FiO₂).

Laboratory variables (worst within 24 h of ICU admission): procalcitonin (PCT), interleukin-6 (IL-6), C-reactive protein (CRP), albumin (ALB), blood glucose, lactate, serum phosphorus (P), and creatinine (highest values for inflammatory markers and lactate; lowest value for albumin).

Early ICU management variables: nutritional support mode, rehabilitation interventions, and glucocorticoid use.

Follow-up variables: duration of mechanical ventilation and ICU costs.

Main ICU admission diagnoses were reviewed and categorised as respiratory failure, acute neurological critical illness, shock/circulatory failure, gastrointestinal or abdominal critical illness, sepsis/infection, trauma, postoperative or surgical critical illness, and other diagnoses.

Muscle ultrasound protocol and measurements

Bedside B-mode ultrasound was performed within 24 h of ICU admission using a 5–10 MHz linear-array transducer (Mindray, Shenzhen, China). Patients were positioned supine with knees straight and relaxed, and toes pointed upward. Bilateral quadriceps images were obtained under unpressurised and pressurised conditions. For the pressurised protocol, firm probe compression was applied to minimise the influence of subcutaneous oedema on thickness measurements [22]. Measurements were taken at the lower third of the line between the anterior superior iliac spine and the superior margin of the patella [23]. The probe was oriented perpendicular to the long axis of the thigh. Transverse images visualised subcutaneous tissue/adipose tissue, the rectus femoris (RF), vastus intermedius (VI), and femur [24] (Fig. 1).

Fig. 1.

Muscle ultrasound examination protocol. A Anatomical landmark and measurement site (red line). B Transverse ultrasound image of the quadriceps femoris at the measurement site. The thickness of the rectus femoris (RF) and vastus intermedius (VI) is indicated by yellow lines, and the rectus femoris cross-sectional area (RF-CSA) is outlined by the white line. RF: rectus femoris; VI: vastus intermedius; CSA: cross-sectional area

Ultrasound-derived variables included bilateral RF cross-sectional area (RF-CSA), RF muscle thickness (RF-MT), and VI muscle thickness (VI-MT) under both unpressurised and pressurised conditions (Fig. 2).

Fig. 2.

Muscle ultrasound image and measurements under unpressurised and pressurised conditions. A Unpressurised condition (minimal probe pressure); B measurement under unpressurised condition; C pressurised condition; D measurement under pressurised condition. RF: rectus femoris; VI: vastus intermedius; RF-MT: rectus femoris muscle thickness; VI-MT: vastus intermedius muscle thickness; RF-CSA: rectus femoris cross-sectional area

Anatomical landmarks were standardised with a soft ruler and skin marker to ensure measurement reproducibility (Fig. 1). Each measurement was repeated three times at the same site by the same sonographer, and the average value was used for analysis. All operators were trained and certified by the Critical Care Ultrasound Study Group (CCUSG). All ultrasound images underwent quality control for measurement accuracy.

Model development and validation

The overall cohort was divided into an internal development cohort and a temporally distinct external validation cohort according to the calendar time of enrolment. Patients admitted between Sept 1, 2023, and May 30, 2024, formed the internal cohort, and those admitted between June 1, 2024, and April 30, 2025, formed the external validation cohort.

Within the internal cohort, patients were randomly split into a training set (70%) and an internal validation set (30%). Two prespecified predictor sets were used to develop model 1 (unpressurised ultrasound) and model 2 (pressurised ultrasound), to evaluate whether compression-based measurements improve prediction in oedematous ICU patients.

Nine machine-learning algorithms were evaluated: Decision Tree (DT), Random Forest (RF-ML), Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), k-Nearest Neighbours (KNN), Support Vector Machine (SVM), Elastic Net (ENet) and Logistic Regression (Logit). Five-fold cross-validation on the training set was used for hyperparameter tuning. Grid search combined with maximum-entropy or Bayesian optimisation strategies was applied according to algorithm characteristics. The model with the best overall performance and clinical interpretability in the internal validation set was selected as the final prediction model and then tested in the external validation cohort.

Model performance, calibration, clinical utility, and interpretability

Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) and C-index. Additional metrics included sensitivity, specificity, accuracy, and F1 score. Calibration was evaluated using calibration plots and the Hosmer–Lemeshow goodness-of-fit test. Clinical usefulness was assessed by decision-curve analysis across a range of risk thresholds. For interpretability, Shapley additive explanations (SHAP) values were calculated to quantify the contribution and direction of each predictor; global importance plots and individual waterfall plots were generated [25]. To further quantify the incremental predictive value of muscle ultrasound information, we performed category-free net reclassification improvement (cf-NRI) and integrated discrimination improvement (IDI) analyses [26, 27]. For this incremental-value analysis, the General RF-ML model was defined as the reference model and included all non-muscle-ultrasound predictors used in the final RF-ML modelling feature set, including demographic, physiological, nutritional, inflammatory, laboratory, and disease-severity variables. Muscle ultrasound variables, including RF-CSA, RF-MT, and VI-MT, were excluded from the General RF-ML reference model. The unpressurised RF-ML and pressurised RF-ML models were then compared with this reference model. For cf-NRI, the event and nonevent components were reported separately, and 95% confidence intervals were calculated for cf-NRI and IDI. The overall workflow is summarised in Fig. 3.

Fig. 3.

Study workflow for model development, validation, and interpretation. Schematic overview of the study design and analysis pipeline. Consecutive ICU patients were prospectively enrolled, underwent early quadriceps ultrasound within 24 h of admission, and were followed for the occurrence of ICU-acquired weakness (ICUAW). The overall cohort was divided into an internal development cohort and a temporally separate external validation cohort. Within the internal cohort, data were randomly split into training and internal validation sets. Clinical, laboratory, and ultrasound variables under unpressurised (model 1) and pressurised (model 2) conditions were used to train nine machine-learning algorithms. The best-performing model (RF-ML) was selected based on discrimination and overall performance, evaluated in the internal validation set, and subsequently tested in the external validation cohort. Calibration, decision-curve analysis, and Shapley additive explanations (SHAP) were used to assess model calibration, clinical utility, and interpretability. Additional cf-NRI and IDI analyses were performed to evaluate the incremental predictive value of ultrasound-augmented RF-ML models compared with the General RF-ML reference model. ICUAW: intensive care unit-acquired weakness; RF-ML: random forest model; SHAP: Shapley additive explanations; cf-NRI: category-free net reclassification improvement; IDI: integrated discrimination improvement

Statistical analysis

Continuous variables were tested for normality using the Shapiro–Wilk test. Normally distributed data are presented as mean (SD) and were compared using the Student’s t test; skewed data are presented as median (IQR) and were compared using the Mann–Whitney U test. Categorical variables are presented as n (%) and were compared using the χ² test or Fisher’s exact test, as appropriate.

Missing predictor values were imputed using a random forest–based algorithm (RF-ML imputation). Kernel density plots were used to visually compare the distributions of observed and imputed data to ensure plausibility. All statistical tests were two-sided, p < 0.05 was considered statistically significant unless otherwise specified. Data management and statistical analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA) and R version 4.3.4 (R Foundation for Statistical Computing, Vienna, Austria).

Role of the funding source

The sponsors of the study had no role in study design; data collection, analysis, or interpretation; or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit the manuscript for publication.

Results

Patient characteristics

Between Sept 1, 2023, and Apr 30, 2025, 1203 adult patients from 16 tertiary hospital ICUs in southwest China were enrolled and included in the analysis (Fig. 4). Of these, 858 patients admitted between Sept 2023 and May 2024 constituted the internal development cohort, and 345 patients admitted between June 2024 and April 2025 comprised the temporally distinct external validation cohort.

Fig. 4.

Flowchart of patient selection. Study flowchart showing screening, inclusion, and exclusion of ICU patients. Consecutive adults admitted to 16 tertiary hospital ICUs were screened for eligibility. Patients were excluded for pregnancy, brain death, pre-existing neuromuscular disease, inadequate ultrasound images, or persistent inability to perform Medical Research Council (MRC) strength testing. The final study population (n = 1203) was divided into an internal development cohort (n = 858) and an external validation cohort (n = 345) according to calendar time. ICUAW: intensive care unit-acquired weakness; MRC: Medical Research Council

In the internal cohort, 560 (65.3%) of 858 patients were male, and ICU-acquired weakness (ICUAW) occurred in 579/858 (67.5%) patients. In the external validation cohort, 229 (66.4%) of 345 patients were male, and ICUAW occurred in 181/345 (52.5%) patients (Table 1). Compared with patients without ICUAW, those who developed ICUAW had higher SOFA scores in both cohorts and higher APACHE II scores in the development cohort (Table 1). The main ICU admission diagnoses reflected a heterogeneous critically ill population: respiratory failure was the most common diagnosis (323/1203, 26.8%), followed by acute neurological critical illness (279/1203, 23.2%), shock/circulatory failure (139/1203, 11.6%), gastrointestinal or abdominal critical illness (137/1203, 11.4%), sepsis/infection (134/1203, 11.1%), trauma (82/1203, 6.8%), postoperative or surgical critical illness (52/1203, 4.3%), and other diagnoses (57/1203, 4.7%).

Table 1.

Baseline characteristics of ICUAW and Non-ICUAW patients in training and validation sets

Variables	Internal training and validation set (N = 858)					External validation set (N = 345)
	Overall	Non-ICUAW (n = 279)	ICUAW (n = 579)	Statistics	P value	Overall	Non-ICUAW (n = 164)	ICUAW (n = 181)	Statistics	P value
Gender [n(%)]				χ2 = 0.10	0.74				χ2 = 1.23	0.26
Female	298(34.73)	99(35.48)	199(34.37)			116(33.62)	60(36.59)	56(30.94)
Male	560(65.27)	180(64.52)	380(65.63)			229(66.38)	104(63.41)	125(69.06)
Age (years)	62.50(17.67)	62.24(18.39)	62.62(17.33)	t = − 0.29	0.77	61.49(17.24)	62.52(17.58)	60.55(16.92)	t = 1.06	0.28
BMI (kg/m2)	23.11(3.74)	23.25(3.99)	23.04(3.61)	t = 0.77	0.44	23.33(3.65)	23.08(3.52)	23.55(3.76)	t = − 1.21	0.22
BSA (m2)	1.75(0.21)	1.76(0.20)	1.74(0.21)	t = 0.74	0.45	1.76(0.26)	1.74(0.22)	1.77(0.28)	t = − 0.85	0.39
T (℃)	36.97(0.75)	36.98(0.69)	36.96(0.78)	t = 0.26	0.79	37.05(0.69)	36.99(0.67)	37.10(0.71)	t = − 1.48	0.14
RR (breaths/min)	17.88(4.37)	17.93(4.24)	17.85(4.44)	t = 0.23	0.81	15.39(2.56)	15.13(2.39)	15.62(2.69)	t = − 1.78	0.07
SBP (mmHg) [med(IQR)]	125[26]	125[23]	125[26]	Z = 0.46	0.64	120.00[22]	121.00[20]	119.00[26]	Z = 0.14	0.89
DBP (mmHg)	69.69(13.60)	69.66(12.61)	69.71(14.07)	t = − 0.05	0.95	69.85(11.55)	69.82(11.56)	69.88(11.58)	t = − 0.05	0.95
SOFA	6.57(4.53)	5.55(4.44)	7.06(4.50)	t = − 4.62	< 0.001	7.50(3.27)	7.34(3.39)	7.68(3.14)	t = 2.96	0.03
APACHE II	20.57(8.57)	19.35(9.08)	21.17(8.25)	t = − 2.93	0.003	19.37(8.79)	19.00(9.35)	19.74(8.29)	t = − 0.37	0.71
mNUTRIC	5.12(2.03)	4.90(2.10)	5.21(1.99)	t = − 2.13	0.03	5.57(2.04)	5.55(2.23)	5.59(1.86)	t = − 0.09	0.93
RASS	− 3.18(0.97)	− 3.05(1.05)	− 3.25(0.92)	t = 2.90	0.004	− 2.82(1.04)	− 2.82(0.98)	− 2.82(1.09)	t = 0.01	0.99
CPOT	0.11(0.70)	0.12(0.20)	0.16(0.84)	t = − 2.74	0.006	0.14(0.57)	0.10(0.39)	0.18(0.69)	t = − 1.29	0.19
MRC[med(IQR)]	36[32]	48[12]	24[33]	Z = 22.91	< 0.001	48[12]	48.00[12.00]	36.00[24.00]	Z = − 5.69	< 0.001
CRP (mg/L)	95.23(102.72)	88.14(90.42)	98.65(108.06)	t = − 1.40	0.16	104.77(103.50)	106.98(110.77)	102.77(96.71)	t = 0.38	0.70
IL-6 (pg/mL)	88.81[285.70]	82.80[292.88]	91.07[282.70]	Z = − 1.31	0.19	394.13(878.79)	444.31(1024.04)	348.66(722.42)	t = 1.01	0.31
PCT (ng/mL)[med(IQR)]	0.82[4.87]	0.80[4.18]	0.84[5.07]	Z = − 0.59	0.55	5.23(17.59)	6.09(20.65)	4.46(14.29)	t = 0.86	0.39
Lactate (mmol/L)	1.74(1.20)	1.74(1.23)	1.75(1.19)	t = − 0.13	0.89	1.88(1.05)	1.96(1.11)	1.81(1.00)	t = 1.31	0.19
P (mmol/L)[med(IQR)]	1.04[0.60]	1.03[0.57]	1.04[0.63]	Z = 0.13	0.90	0.83[0.58]	0.70[0.67]	0.90[0.92]	Z = − 3.41	< 0.001
Blood glucose (mmol/L)[med(IQR)]	8.30[4.99]	8.11[4.60]	8.34[5.00]	Z = − 1.03	0.30	9.00[3.70]	9.65[3.45]	8.70[3.20]	Z = 2.19	0.02
ALB (g/L)	32.97(6.93)	33.52(7.38)	32.70(6.68)	t = 1.64	0.10	34.10[8.90]	34.25[9.75]	34.00[8.50]	Z = − 0.02	0.98
Creatinine (μmol/L)[med(IQR)]	76.50[64.00]	74[53]	77.80[66.80]	Z = − 0.94	0.34	78.10[49.30]	78.00[45.85]	78.50[52.70]	Z = 0.61	0.54
Respiratory support mode [n(%)]				χ2 = 73.55	< 0.001				χ2 = 13.64	0.001
Non-invasive ventilation	43(5.01)	27(9.68)	16(2.76)			16(4.64)	13(7.93)	3(1.66)
High-flow oxygen	199(23.19)	104(37.28)	95(16.41)			54(15.65)	33(20.12)	21(11.60)
Invasive ventilation	616(71.79)	148(53.05)	468(80.83)			275(79.71)	118(71.95)	157(86.74)
Corticosteroids [n(%)]				χ2 = 6.60	0.01				χ2 = 1.54	0.21
Yes	54(6.29)	9(3.23)	45(7.77)			18(5.22)	6(3.66)	12(6.63)
No	804(93.71)	270(96.77)	534(92.23)			327(94.78)	158(96.34)	169(93.37)
Early PN [n(%)]				χ2 = 3.97	0.04				χ2 = 0.33	0.56
Yes	676(78.79)	231(82.80)	445(76.86)			220(63.77)	102(62.20)	118(65.19)
No	182(21.21)	48(17.20)	134(23.14)			125(36.23)	62(37.80)	63(34.81)
Rehabilitation [n(%)]				χ2 = 0.05	0.83				χ2 = 4.43	0.03
No	383(44.64)	126(45.16)	257(44.39)			55(15.94)	19(11.59)	36(19.89)
Yes	475(55.36)	153(54.84)	322(55.61)			290(84.06)	145(88.41)	145(80.11)
Length of mechanical ventilation, d [med(IQR)]	4[6.06]	2[3.16]	5[6]	Z = − 8.40	< 0.001	1.73[2.79]	1[1.90]	2[3]	Z = − 4.89	< 0.001
ICU hospitalisation cost (10000RMB)[med(IQR)]	3.46[6.60]	2.35[5.30]	3.72[7.33]	Z = − 2.59	0.010	3.56[3.40]	2.64[3.95]	3.66[4.21]	Z = 2.35	0.019

ALB: albumin; APACHE II: Acute physiology and chronic health evaluation II; BMI: body mass index; BSA: body surface area; CPOT: critical care pain observation tool; CRP: C-reactive protein; DBP: diastolic blood pressure; ICU: intensive care unit; ICUAW: intensive care unit-acquired weakness; IL-6: interleukin-6; mNUTRIC: modified nutrition risk in the critically ill score; MRC: Medical Research Council; P: phosphorus; PCT: procalcitonin; PN: parenteral nutrition; RASS: Richmond Agitation and Sedation Scale; RR: respiratory rate; SBP: systolic blood pressure; SOFA: Sequential Organ Failure Assessment; T: temperature; χ²: Chi-square test; t: Student’s t test; Z: Z test. Values in parentheses () represent the mean ± standard deviation, while values in square brackets [] represent the median and the interquartile range (3rd quartile–1st quartile)

Muscle ultrasound measurements

Quadriceps muscle ultrasound was successfully obtained within 24 h of ICU admission in all enrolled patients (Figs. 1 and 2). Under unpressurised conditions, rectus femoris cross-sectional area (RF-CSA), rectus femoris muscle thickness (RF-MT), and vastus intermedius muscle thickness (VI-MT) were significantly lower in patients who developed ICUAW than in those who did not, on both limbs (Table 2). In the internal cohort, RF-CSA (left and right), RF-MT (left and right), and VI-MT (left and right) were all reduced in ICUAW patients (p = 0.03, 0.005, < 0.001, 0.03, 0.01, and < 0.001, respectively) (Table 2).

Table 2.

Muscle ultrasound indicators of ICUAW and non-ICUAW patients in training and validation sets

Variables	Internal training and validation set (N = 858)					External validation set (N = 345)
	Overall	Non-ICUAW (n = 279)	ICUAW (n = 579)	Statistic	p value	Overall	Non-ICUAW (n = 164)	ICUAW (n = 181)	Statistic	p value
Unpressurised condition
RF-CSA(L), cm2	2.05(1.02)	2.15(0.99)	2.00(1.03)	t = 2.11	0.03	2.15(1.17)	2.22(1.13)	2.08(1.20)	t = 1.10	0.271
RF-MT(L), cm	0.68(0.27)	0.71(0.28)	0.66(0.26)	t = 2.79	0.005	0.71(0.35)	0.74(0.39)	0.68(0.30)	t = 1.52	0.129
VI-MT(L), cm	0.73(0.35)	0.79(0.37)	0.70(0.34)	t = 3.53	< 0.001	0.80(0.40)	0.81(0.42)	0.79(0.39)	t = 0.43	0.667
RF-CSA(R), cm2	2.19(1.11)	2.30(1.15)	2.13(1.08)	t = 2.09	0.03	2.21(1.12)	2.29(1.05)	2.13(1.17)	t = 1.28	0.201
RF-MT(R), cm	0.72(0.31)	0.76(0.34)	0.71(0.29)	t = 2.34	0.01	0.73(0.35)	0.77(0.38)	0.70(0.32)	t = 1.69	0.093
VI-MT(R), cm	0.73(0.34)	0.79(0.34)	0.70(0.34)	t = 3.53	< 0.001	0.79(0.35)	0.82(0.34)	0.77(0.35)	t = 1.17	0.242
Pressurised condition
RF-CSA(L), cm2	1.68(0.92)	1.78(0.95)	1.64(0.90)	t = 2.09	0.03	1.72(1.02)	1.78(1.09)	1.67(0.96)	t = 1.07	0.286
RF-MT(L), cm	0.49(0.21)	0.53(0.23)	0.48(0.20)	t = 3.02	0.003	0.49(0.24)	0.51(0.27)	0.47(0.20)	t = 1.76	0.079
VI-MT(L), cm	0.56(0.28)	0.60(0.28)	0.54(0.28)	t = 2.85	0.004	0.61(0.41)	0.63(0.50)	0.60(0.30)	t = 0.89	0.376
RF-CSA(R), cm2	1.79(0.95)	1.89(1.00)	1.74(0.92)	t = 2.21	0.02	1.75(0.94)	1.81(0.81)	1.69(1.05)	t = 1.17	0.241
RF-MT(R), cm	0.53(0.23)	0.55(0.24)	0.52(0.23)	t = 1.74	0.08	0.52(0.26)	0.55(0.28)	0.49(0.23)	t = 2.25	0.025*
VI-MT(R), cm	0.57(0.27)	0.60(0.26)	0.55(0.27)	t = 2.73	0.007	0.59(0.27)	0.60(0.26)	0.58(0.27)	t = 0.64	0.525

ICUAW: intensive care unit-acquired weakness; RF-CSA (L, R): rectus femoris-cross-sectional area (left, right); RF-MT (L, R): rectus femoris-muscle thickness (left, right); VI-MT (L, R): vastus intermedius-muscle thickness (left, right); χ²: Chi-square test; t: Student’s t test; Z: Z test. Values in parentheses () represent the mean ± standard deviation, while values in square brackets [] represent the median and the interquartile range (third Quartile–first Quartile)

Similarly, under pressurised conditions, RF-CSA (L), RF-MT (L), VI-MT (L), RF-CSA (R), and VI-MT (R) were significantly decreased in ICUAW compared with non-ICUAW patients (p = 0.03, 0.003, 0.004, 0.02, and 0.007, respectively) (Table 2).

Model performance

Among nine candidate algorithms, the random forest model (RF-ML) showed the most consistent performance across internal evaluation and temporal external validation and was selected as the final model.

For the unpressurised ultrasound model (model 1), the RF-ML algorithm achieved an AUC of 0.914 (95% CI 0.891–0.936) in the training set and 0.849 (0.798–0.900) in the internal validation set. In external validation, model 1 achieved an AUC of 0.810 (0.765–0.855), with sensitivity 0.983 (0.952–0.997) and specificity 0.726 (0.651–0.792) (Fig. 5a).

Fig. 5.

Discrimination, calibration, and clinical utility of the RF-ML prediction models. a Receiver operating characteristic (ROC) curves of the unpressurised ultrasound model (model 1) in the internal cohort and the external validation cohort, showing high discrimination for ICU-acquired weakness (ICUAW). b ROC curves of the pressurised ultrasound model (model 2) in the internal and external validation cohorts, indicating slightly lower discrimination than model 1, particularly in external validation. c Decision-curve analysis (DCA) for model 1. The net benefit of using model 1 to guide ICUAW-targeted interventions is plotted across a range of threshold probabilities and compared with “treat-all” and “treat-none” strategies. Within a clinically relevant threshold range (approximately 0.30–0.70), model 1 provides a clearly higher net benefit than the two default strategies, indicating that using this model to trigger ICUAW prevention or rehabilitation bundles would reduce unnecessary interventions while maintaining a high proportion of true positives. d Decision-curve analysis for model 2. Although model 2 also confers a positive net benefit over “treat-all” and “treat-none” strategies across most threshold probabilities (particularly above about 0.25), its net-benefit curve lies below that of model 1, reflecting the lower specificity and greater number of false-positive alerts associated with the pressurised ultrasound–based model. e Calibration plot for model 1. Patients are grouped into deciles of predicted risk, and for each group the mean predicted probability of ICUAW is plotted against the observed event rate. The solid line represents model-predicted risk, and the dashed 45° line indicates perfect calibration. The points lying close to the 45° line across most risk deciles suggest good agreement between predicted and observed risks and minimal systematic under- or overestimation by model 1 in both cohorts. f Calibration plot for model 2 constructed in the same way. The calibration curve generally follows the 45° line but deviates slightly in higher-risk deciles, indicating some overestimation of risk at the upper end of the predicted probability range compared with model 1. AUC: area under the ROC curve; DCA: decision-curve analysis; ICUAW: intensive care unit-acquired weakness; RF-ML: random forest model; ROC: receiver operating characteristic

For the pressurised ultrasound model (model 2), the AUCs were 0.907 (0.882–0.931) in the training set and 0.775 (0.711–0.839) in the internal validation set. In external validation, model 2 achieved an AUC of 0.730 (0.676–0.783), with sensitivity of 0.978 and specificity of 0.506 (Fig. 5b).

Calibration and clinical utility

Calibration plots showed good agreement between predicted and observed ICUAW risk for both models in the internal validation set (Fig. 5e, f), with similar patterns in the external validation cohort. Decision-curve analysis suggested net benefit over treat-all and treat-none strategies across clinically plausible threshold probabilities (Fig. 5c, d).

SHAP-based model interpretation

SHAP analyses identified SOFA score, quadriceps muscle thickness/CSA measures, albumin, and inflammatory markers as key contributors to the predicted ICUAW risk (Fig. 6). Higher SOFA scores and inflammatory burden increased predicted risk, whereas greater quadriceps thickness and higher albumin were associated with lower predicted risk.

Fig. 6.

SHAP-based interpretation of the RF-ML prediction models. Global and individual-level Shapley additive explanations (SHAP) for the random forest (RF-ML) models. a Global SHAP summary plot for the unpressurised ultrasound model (model 1), ranking predictors by their overall contribution to ICUAW risk. Each dot represents a single patient; colour indicates the value of the predictor (red, high; blue, low), and horizontal position indicates SHAP value (impact on model output). b Representative SHAP waterfall plot for a high-risk patient in model 1, illustrating how each predictor increases (red) or decreases (blue) the predicted probability of ICUAW relative to the expected value. c Global SHAP summary plot for the pressurised ultrasound model (model 2). d Representative SHAP waterfall plot for a high-risk patient in model 2. Across both models, higher SOFA scores, lower quadriceps muscle thickness, lower albumin, and higher inflammatory markers push predictions towards higher ICUAW risk, whereas greater muscle thickness and better organ function shift predictions towards lower risk. APACHE II: Acute Physiology and Chronic Health Evaluation II; BMI: body mass index; BSA: body surface area; CPOT: Critical-care Pain Observation Tool; CRP: C-reactive protein; ICUAW: intensive care unit-acquired weakness; IL-6: interleukin-6; PCT: procalcitonin; RF: rectus femoris; RF-CSA: rectus femoris cross-sectional area; RF-MT: rectus femoris muscle thickness; RASS: Richmond Agitation and Sedation Scale; SHAP: Shapley additive explanations; SOFA: Sequential Organ Failure Assessment; VI: vastus intermedius; VI-MT: vastus intermedius muscle thickness

Incremental predictive value of ultrasound-augmented RF-ML models

Compared with the general RF-ML reference model, the unpressurised RF-ML model showed a higher test AUC and significant improvement in both risk reclassification and discrimination (cf-NRI 0.490, 95% CI 0.277–0.698; IDI 0.072, 95% CI 0.045–0.099). The pressurised RF-ML model also showed positive reclassification improvement (cf-NRI 0.231, 95% CI 0.011–0.463), although the IDI improvement was smaller and its confidence interval crossed zero (IDI 0.019, 95% CI −0.002 to 0.040) (Table 3).

Table 3.

Incremental predictive value of ultrasound-augmented RF-ML models compared with the general RF-ML model

Model comparison	NRI events (95% CI)	NRI nonevents (95% CI)	cf-NRI (95% CI)	IDI (95% CI)
Unpressurised RF-ML vs General RF-ML	0.172 (0.070 to 0.274)	0.319 (0.125 to 0.510)	0.490 (0.277 to 0.698)	0.072 (0.045 to 0.099)
Pressurised RF-ML vs General RF-ML	0.066 (−0.026 to 0.168)	0.165 (−0.034 to 0.370)	0.231 (0.011 to 0.463)	0.019 (−0.002 to 0.040)

cf-NRI: category-free net reclassification improvement; IDI: integrated discrimination improvement; RF-ML: random forest model

Discussion

In this multicentre prospective cohort of 1203 critically ill adults, we developed and temporally validated an early prediction model for ICU-acquired weakness (ICUAW) by integrating quadriceps ultrasound obtained within 24 h of ICU admission with routinely available clinical data. The random forest model using unpressurised ultrasound measures showed robust performance across internal evaluation and temporal external validation, with high discrimination and a favourable sensitivity–specificity trade-off in the external cohort. These findings support the feasibility of identifying patients at high risk of ICUAW before volitional strength testing becomes possible.

ICUAW is a frequent complication of critical illness and is associated with adverse short- and long-term outcomes, underscoring the need for preventive strategies [3, 28]. A recent systematic review reported a median prevalence of 43% (interquartile range 25–75%) across 31 studies [29]. However, early bedside identification remains challenging because the Medical Research Council (MRC) assessment requires patients to be awake and cooperative and therefore cannot be reliably performed during the initial phase of critical illness in many patients [28]. Risk prediction tools that can be applied at ICU admission may facilitate earlier, targeted prevention and rehabilitation pathways [6, 28].

The number of ICUAW prediction models has increased [30, 31], yet many are limited by retrospective designs, small cohorts, and insufficient internal or external validation [4, 32]. Furthermore, few studies have combined muscle ultrasound with clinical features to predict ICUAW, and most of those are small-sample studies [32, 33]. In addition, conventional regression-based approaches may be constrained when predictors are correlated and relationships are nonlinear. In this context, machine-learning methods can exploit complex interactions among clinical, laboratory, and imaging features and may improve predictive performance and transportability [33–35].

A key strength of our model is the incorporation of early quadriceps ultrasound, an objective bedside measure that is increasingly used to characterise muscle quantity in critically ill patients [32, 36]. In our model, markers of organ failure burden (SOFA score), quadriceps muscle thickness and CSA measures, nutritional status (albumin), and inflammatory markers contributed most strongly to ICUAW risk, and SHAP analyses provided transparent, clinically coherent explanations of these effects [37]. Collectively, these predictors align with current pathophysiological concepts that link early muscle wasting and systemic illness severity to subsequent neuromuscular weakness [3, 36].

We also compared pressurised and unpressurised ultrasound protocols to explore whether probe compression reduces oedema-related measurement bias in ICU patients. Although discrimination remained acceptable for both approaches, the unpressurised model showed a more favourable balance between sensitivity and specificity in external validation, which may be preferable for routine screening where excessive false-positive alerts could increase unnecessary interventions.

The present results also help clarify which ultrasound approach is most clinically applicable. The unpressurised RF-ML model showed the strongest incremental value in cf-NRI and IDI and better external validation performance than the pressurised RF model. Among ultrasound variables, unpressurised quadriceps measurements, particularly VI-MT, RF-MT, and RF-CSA, were clinically interpretable markers of quadriceps muscle quantity. From a bedside implementation perspective, muscle thickness measurements, especially RF-MT and VI-MT, may be more practical for ICU physicians because they are rapid, repeatable, and do not require manual tracing, whereas RF-CSA may provide more complete anatomic information but requires additional image-processing effort.

From a clinical perspective, an admission-based prediction model could support early stratification and prompt implementation of ICUAW-prevention bundles (e.g., nutrition optimisation, mobilisation/rehabilitation planning, and minimisation of modifiable risk exposures) [28]. Future work should evaluate how best to integrate model outputs into clinical workflows and whether model-guided strategies improve patient-centred outcomes.

Limitations

This study has several limitations. First, the study relied on ultrasound measurements from a single muscle group (quadriceps), which may not fully capture overall muscle atrophy in ICU patients. Future research should explore the inclusion of additional muscle groups or other imaging modalities to improve prediction accuracy. Second, the incidence of ICUAW differed between the external validation and the internal training sets. This discrepancy may be due to the smaller size of the external validation dataset and resulting sampling variability. We plan to increase the sample size in future studies to improve model robustness. Third, inter-rater variability in ultrasound measurements could affect generalisability, although all sonographers were trained, and measurement protocols were standardised. Finally, patients’ conditions can change over time, which may affect their ICUAW risk. Our model currently uses only baseline data and does not account for such dynamic changes or new risk factors arising during treatment. Therefore, ongoing updates and validation of the model will be necessary.

Conclusions

In conclusion, this multicentre prospective study demonstrates that a machine-learning model integrating early quadriceps ultrasound with routinely available clinical data can identify ICU patients at high risk of ICUAW before formal strength testing becomes feasible. The model provides clinically interpretable risk estimates and may help target preventive and rehabilitative strategies to patients most likely to benefit. Further validation in broader populations and prospective impact studies are needed to establish effects on patient-centred outcomes.

Abbreviations

ALB

Albumin

APACHE II

Acute Physiology and Chronic Health Evaluation II

AUC

Area under the curve

BMI

Body mass index

BSA

Body surface area

CCUSG

Critical Care Ultrasound Study Group

cf-NRI

Category-free net reclassification improvement

CPOT

Critical care pain observation tool

CRP

C-reactive protein

CSA

Cross-sectional area

DBP

Diastolic blood pressure

DCA

Decision curve analysis

DT

Decision Tree

FiO₂

Fraction of inspired oxygen

HR

Heart rate

ICUAW

Intensive care unit-acquired weakness

IDI

Integrated discrimination improvement

IL-6

Interleukin-6

IQR

Interquartile range

KNN

K-nearest neighbour

MLP

Multilayer perceptron

MRC

Medical Research Council

MT

Muscle thickness

P

Phosphorus

PCT

Procalcitonin

RASS

Richmond Agitation and Sedation Scale

RF

Rectus femoris

RF-ML

Random Forest (machine-learning) model

ROC

Receiver operating characteristic

RR

Respiratory rate

SBP

Systolic blood pressure

SD

Standard deviation

SHAP

Shapley additive explanations

SOFA

Sequential Organ Failure Assessment

SpO₂

Peripheral capillary oxygen saturation

SVM

Support vector machine

T

Body temperature

TRIPOD

Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis

VI

Vastus intermedius

XGBoost

EXtreme Gradient Boosting

Author contributions

All authors contributed to this study. Study conception and design were developed by WHY, YK, and XDJ. Material preparation and data collection were performed by TJZ, XHC, JL, XYL, HYC, LZ1, LZ2, XYZ, QLD and SRZ. Data analysis was performed by TJZ. The first draft of the manuscript was written by TJZ. BW, ZWZ provided critical revisions of the draft. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by Sichuan Province Science and Technology Support Program-Key Research & Development Project (2024YFFK0052), and Project for Horizontal Research, West China Hospital, Sichuan University (311241641).

Data availability

The datasets used and analysed during the current study are available from the corresponding author on reasonable request. Access requires a methodologically sound proposal and, where applicable, a data access agreement.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of West China Hospital. Approved on July 28, 2023 (No. 2023[1422]). Informed consent was obtained from all participants. Trial registration: Chinese Clinical Trial Registry (ChiCTR2300075581). Registration date: September 8, 2023.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.