Computed tomography-based nutritional associated nomogram on machine learning predicts survival outcomes in patients with resectable soft-tissue sarcoma

Abstract

BACKGROUND

Soft-tissue sarcomas (STS) are heterogeneous mesenchymal malignancies for which surgery remains the mainstay of curative treatment, yet recurrence and mortality rates remain substantial. Nutritional status and body composition measured via routine blood tests and computed tomography-derived metrics such as skeletal muscle and adipose tissue areas have emerged as important determinants of outcomes in cancer. Advanced machine-learning methods can integrate high-dimensional nutritional and radiologic variables to improve individualized survival prediction.

AIM

To identify prognostic value of nutrition-associated factors for patients with STS treated with excision and to construct a predictive model for nutritional assessment by traditional survival analysis and a random forest machine learning method.

METHODS

We retrospectively included 638 patients who were diagnosed with STS and underwent surgical excision from January 2009 to June 2018. Nutrition-associated indicators from peripheral blood tests and routine computed tomography were collected. The primary outcome was overall survival (OS). The secondary outcomes were progression-free survival and length of postoperative hospital stay. The random survival forest (RSF) analysis selected important variables for re progression-free survival and OS, and the random forest analysis selected important variables for length of hospitalization. Nomograms were constructed by the prognostic features to predict survival probabilities.

RESULTS

The RSF analysis identified stage, hospital duration, subtype, body mass index, and tumour size as important variables for OS. The RSF-based nomogram on nutritional indexes for various clinical outcomes showed consistent calibration capacities on calibration plots and great discriminative abilities on the C-index and Brier score.

CONCLUSION

Our study implicated the prognostic value of multiple nutritional assessment indexes for prediction of clinical outcomes in STS, and patients’ nutrition status need long-term surveillance and management.

Key Words: Soft-tissue sarcoma; Random survival forest; Systemic immune-inflammatory index; Controlling nutritional score; Malnutrition universal screening tool; Radiomic score; Nomogram; Prognosis

Core Tip: This study explored the prognostic value of various nutritional assessment indicators for patients with soft tissue sarcomas receiving surgical resection. We constructed the prediction model for short- and long-term prognoses via multi-sided variables with potential prognostication in patients with resectable sarcomas. We applied machine learning methods to increase the prediction abilities of our model for identification of individual survival risk. The sample size was nearly 700 larger than most studies about nutritional prognostic factors especially for sarcomas.

INTRODUCTION

Soft-tissue sarcomas (STS) originating from mesenchymal cells have a low incidence of morbidity compared with osteosarcomas[1] and various subtypes from different soft tissues. The consensus of the management of STS is that surgery with or without chemotherapy is the standard therapeutic regimen[2] for patients with localized and regional soft tissue sarcoma. Additional neoadjuvant or adjuvant chemotherapies are recommended for patients with advanced malignancies[3]. Recently, individualized treatment has presented good efficacy in terms of the long-term prognosis depending on the subtype[4]. However, the rates of progression and death are still high after systematic treatment. Therefore, we should adopt more easily available prognostic factors in addition to histologic features and set optimized strategies for patients to extend their survival time and improve their quality of life.

Currently, nomograms have been widely utilized in the construction of prognostic models predictive of multiple outcomes and may be applicable in further clinical practice. For soft tissue sarcoma, there have been several studies aiming to predict patient progression and survival[5-7] that have focused on clinical and histological features. However, indicators from other aspects are needed to perform a comprehensive evaluation. Nutritional assessment has been considered to be associated with the prognosis of patients with malignancies, and nutritional status may influence the completion of treatment[8]. Personal nutritional status affects immune-related responses[9] to diseases, so inflammatory indexes have been considered in the comprehensive evaluation of nutritional conditions. The counts of lymphocytes, platelets, neutrophils and monocytes are easy to obtain from peripheral blood and are used to calculate the neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, lymphocyte-to-monocyte ratio and systemic inflammatory index (SII), whose predictive values have been demonstrated in many kinds of malignancies, including STS[10-15]. SII, which includes the counts of neutrophils, platelets, and lymphocytes to compensate for the limitations of each parameter alone, is considered a reflection of patients’ inflammatory reaction. Recently, more haematological indexes have been included to construct new indicators, such as the prognostic nutritional index and controlling nutritional status (CONUT) score[16,17] both of which are significant in STS[18,19]. In this study, the CONUT score was taken as the main nutrition-associated factor, and the adoption of cholesterol, platelets, and plasma albumin could comprehensively express a patient’s nutritional status.

The patients’ physical conditions require overall assessment, and body mass index (BMI), as the most common tool, is easily interfered with by components of adipose tissue and skeletal muscle[20,21] and loses its accuracy. Radiological indexes may promote the consistency between the calculated and actual nutritional statuses[22-24], especially for the definition of sarcopenia. The total adipose tissue area, subcutaneous adipose tissue area, visceral adipose tissue area and skeletal muscle area adjusted by patient height represents individual body composition and has been found to be associated with patient prognosis[25-27]. In addition, the visceral adipose tissue to skeletal muscle ratio was found to be an independent prognostic factor in some kinds of malignancies[28,29].

The malnutrition universal screening tool (MUST) has become the most common method to assess the nutritional condition of inpatients[30], and its validity has been indicated by many British studies[31,32]. The criteria of MUST, including BMI, unintended weight loss and the duration of absent intake, indeed represent the short-term fluctuation of patients’ nutrition status according to previous studies. Although MUST has not been tested in Asian populations, it may be a practical tool to evaluate patients’ recent physical status. To our knowledge, no study has identified relationships between haematological, radiomic and physical nutrition-associated indexes and clinical outcomes. Meanwhile, the performance of some advanced machine learning methods has not been verified in the prediction of survival based on nutritional indicators. The random survival forest (RSF) algorithm has gained increasing attention because of its excellent predictive performance on high-dimensional variables and extensive applicability for highly correlated variables[33]. The RSF algorithm performs variable ranking, which reflects its ability to automatically predict outcomes[34]. RSF was applied to identify appropriate variables to construct a predictive model from survival data.

In this study, we aimed to explore the associations between incorporated indicators and progression-free survival (PFS), overall survival (OS) and length of postoperative hospitalization for patients with STS treated with surgical excision.

MATERIALS AND METHODS

Patients

We retrospectively collected data between January 2009 and December 2018 from patients with STS who underwent surgical resection at our hospital. The diagnosis of the patients was confirmed by pathological examination during procedures. Patients with complete data on clinical characteristics were eligible for inclusion, and the collected characteristics included age, sex, location of tumours, clinical stage, subtype, tumour size, weight and height. The staging process was completed by preoperative radiology examination and postoperative pathological evaluation based on the American Joint Committee on Cancer tumor node metastasis staging system and represented as histological stage (pathological tumor node metastasis). Due to the limited sample size of patients with stage III and IV disease, these patients were combined to form the non-localized subgroup for subsequent analysis. Haematological and radiomic information was available from peripheral blood tests performed less than two days before surgery and computed tomography (CT) images from less than one week before surgery. Information on patients’ weight loss, diets affected by acute disease, and conditions of basic diseases was accessed from patient medical history records. Complications including hypertension, diabetes mellitus, and other dysfunctions in the cardiovascular, respiratory, nervous and urinary systems were confirmed before surgical resection. Patients were excluded if any information mentioned above was missing. The nutrition-associated indicators were composed of haematological indexes, CT-associated indexes and physical state scores.

CT imaging and acquisition parameters

Abdominal CT studies used for body composition analysis were obtained as part of routine preoperative imaging within one week before surgery. Imaging protocols generally employed tube voltages of 120 kVp ranging 100-140 kVp, automatic tube current modulation with typical effective currents of 100-220 mA, and slice thicknesses of 3-5 mm for routine abdominal series. Reconstructions were performed using standard soft-tissue kernels, and axial soft-tissue were archived. CTs were performed with intravenous contrast in the portal venous phase in the majority of > 85% cases. A minority were non-contrast studies when contrast was contraindicated. Because Hounsfield unit (HU) values and the visibility of fat/muscle are potentially affected by acquisition factors, the present study standardized segmentation by using axial images at the third lumbar vertebra (L3) on the portal venous phase, applying fixed HU ranges for adipose (-190 HU to -30 HU) and skeletal muscle (-29 HU to 150 HU) for all scans, and normalizing area measures by height squared to produce indices (cm²/m²).

Haematological indexes

We obtained results from the last peripheral blood test to calculate SII, C-reactive protein (CRP), and the CONUT score. The method of calculation was performed according to previously published articles[15-17,35,36]. SII = platelet count × neutrophil count/Lymphocyte count. SII was dichotomized by the optimal cut-off point from receiver operating characteristic (ROC) analysis in the training group (SII < 619.85, SII ≥ 619.85). CRP was divided into two groups according to the upper bound of our Hospital (CRP < 5.0, CRP ≥ 5.0). The CONUT score was calculated using data for serum albumin, total cholesterol, and total lymphocytes from peripheral blood: (1) Albumin concentrations ≥ 3.5, 3.0-3.49, 2.5-2.99, and < 2.5 g/dL were scored as 0 point, 2 points, 4 points, and 6 points, respectively; (2) Total lymphocyte counts ≥ 1600, 1200-1599, 800-1199, and < 800/mm³ were scored as 0 point, 1 point, 2 points, and 3 points, respectively; and (3) Total cholesterol concentrations ≥ 180, 140-179, 100-139, and < 100 mg/dL were scored as 0 point, 1 point, 2 points, and 3 points, respectively. The CONUT score was calculated as the sum of (1), (2), and (3). In this study, the CONUT score was divided into two groups, 0-2 and 3-12.

CT-associated indexes

The slice images of the 3^rd lumbar vertebra were obtained from routine abdominal CT and analyzed by ImageJ version 1.4.6. Region of interest measurements were applied to calculate the area (cm²) of total adipose tissue, subcutaneous adipose tissue, visceral adipose tissue and skeletal muscle by standard HU ranges of adipose and skeletal muscle (Figure 1). Segmentation followed a standardized procedure: The axial slice with both transverse processes visible and the L3 vertebral body clearly identified was selected, region of interests for muscle and fat compartments were manually delineated using the predefined HU thresholds (adipose: -190 HU to -30 HU; skeletal muscle: -29 HU to 150 HU). Two trained readers (reader 1, reader 2) independently segmented a random subset of 50 CT scans stratified across years and body habitus. Reader 1 repeated segmentations after a 4-week interval to assess intra-observer reproducibility. Inter-observer and intra-observer reproducibility were quantified by intraclass correlation coefficients (ICC) using two-way mixed-effects model and absolute agreement. An ICC ≥ 0.80 was considered excellent agreement. Each area normalized by height squared produced new indicators (cm²/m²): Total adipose index, subcutaneous adipose index (SAI), visceral adipose index (VAI) and skeletal muscle index (SMI). There were different standards for sex to define visceral obesity (male: Visceral fat area > 160 cm²; female: Visceral fat area > 80 cm²[37]), subcutaneous obesity (male: SAI ≥ 50.0 cm²/m²; female: SAI ≥ 42.0 cm²/m²[38]), and sarcopenia (male: SMI < 52.3 cm²/m² for BMI < 30 and SMI < 54.3 cm²/m² for BMI ≥ 30; female: SMI < 38.6 cm²/m² for BMI < 30 and SMI < 46.6 cm²/m² for BMI ≥ 30[39]).

Figure 1 The picture from ImageJ to draw the outlines of total adipose tissue. The white region fitted with the setting standard Hounsfield unit range of adipose tissue (-190 to -30) or skeletal muscle (-29 to +150). The white line depicted the boundaries of regions which would be measured by area (cm²). A: Area and skeletal muscle area; B: On the computed tomography image of 3^rd lumbar vertebra.

Physical state measurement

The assessment of physical state followed the MUST score to examine the patients’ risks of malnutrition. The MUST scores were calculated from the sum of the following three parts[40]: (1) BMI ≥ 20 kg/m², 18.5-20 kg/m², and < 18.5 kg/m² were scored as 0 point, 1 point, and 2 points, respectively; (2) Unintentional weight loss in the last 3-6 months < 5%, 5%-10%, and ≥ 10% were scored as 0 point, 1 point, and 2 points, respectively; and (3) No nutritional intake by mouth for > 5 days induced by acute disease was scored an additional 2 points. The patients were distributed into three MUST groups, low, medium and high risk, according to MUST scores of 0, 1 and 2+, respectively.

Outcomes

For survival analysis, the primary outcome was OS, defined as the duration from the day of surgery to the day of death or the last follow-up through telephone or mail, June 2018. The secondary outcomes were PFS and length of hospital stay after surgery. PFS was defined as the duration from the day of surgery to the day of progression or the last follow-up. The progression status was determined by two experienced oncology doctors according to clinical symptoms and laboratory and radiology examinations. For logistic regression, postoperative hospitalization was defined as the length between the days of surgery and discharge.

Statistical analysis

All eligible subjects were randomly distributed into a training group and testing group at a ratio of 7:3. Categorical variables were compared between two groups by the χ² test. The continuous variables did not have normal distributions, so the comparison of these variables was completed by nonparametric tests and the Mann-Whitney test. Cox regression analysis was applied to explore the association between variables and OS or PFS. Variables with P values less than 0.05 in univariate analysis could enter the multivariate Cox proportional hazards model in the training cohort.

There were two methods to select prognostic factors to construct predictive models: Conventional Cox proportional hazards analysis and RSF analysis. Variables with P values less than 0.05 were considered independent prognostic factors in multivariate Cox proportional hazards analysis and integrated to construct a nomogram. The RSF analysis selected variables by the minimal depth methodology, which assumed that variables with high impact on the prediction were those that most frequently split nodes nearest to the root node, where they partitioned the largest samples of the population[41]. All variables were classified into the important group and the unimportant group for survival prediction by an optimistic threshold using the mean of the minimal depth distribution[42]. The important variables selected by RSF analysis were gathered to construct the another nomogram. These nutritional variables and other oncological features that may influence OS and PFS outcomes were collected to construct models to predict 3-year and 5-year survival rates. Postoperative hospitalization was divided into two kinds of outcomes, 0-14 days and 14+ days, based on the mean postoperative hospital stay. The association between variables and postoperative hospital days was determined by binary logistic regression analysis in the training cohort. The random forest analysis ranked variables by their contribution to the decrease in predictive error. The variables with larger contributions than the mean level were considered important variables for hospital duration. Thus, two predictive models were developed by multivariate logistic analysis and random forest analysis.

Since dichotomization of body-composition indices could reduce statistical power and produce arbitrary thresholds, the present study re-ran RSF and Cox proportional hazards models using CT-derived indices including SMI, VAI, SAI, and total adipose index as continuous variables. We also performed collinearity diagnostics using variance inflation factor (VIF) for all candidate variables. For contrast-phase effects, we repeated analyses restricted to contrast-enhanced portal venous phase CTs.

The predictive performance of developed models in present study was examined by the C-index, Brier score, and calibration plots from the validation and discrimination aspects in both the training and testing cohorts. The time-dependent area under the ROC curve values were calculated for PFS and OS, and the ROC curve was calculated for hospital days. A P value less than 0.05 was considered significant. All statistical tests were two-sided and completed by R version 3.6.1.

RESULTS

Demographic characteristics

A total of 638 patients with STS treated with surgery were included in the final statistical analysis and assigned to either the training cohort (n = 447) or the testing cohort (n = 191). The counts and proportions of clinical features are listed in Table 1. According to the results, the distributions of all features were not found to be significantly different, and this randomization was acceptable.

Table 1 Demographic features of patients with soft-tissue sarcomas receiving surgical resection, n (%).

Variable	All subjects (n = 638)	Training cohort (n = 447)	Testing cohort (n = 191)	P value
Age (year), median (IQR)	49 (24)	49 (24)	47 (26)	0.524
Age (year)				0.259
< 65	528 (82.8)	365 (81.7)	163 (85.3)
≥ 65	110 (17.2)	82 (18.3)	28 (14.7)
Gender				0.780
Male	342 (53.6)	238 (53.2)	104 (54.5)
Female	296 (46.4)	209 (46.8)	87 (45.5)
Location of primary tumor				0.140
Trunk	501 (78.5)	344 (77.0)	157 (82.2)
Limb	137 (21.5)	103 (23.0)	34 (17.8)
Subtype				0.454
Liposarcoma	127 (19.9)	88 (19.7)	39 (20.4)
Rhabdomyosarcoma	69 (10.8)	56 (12.5)	13 (6.8)
Synovial sarcoma	69 (10.8)	48 (10.7)	21 (11.0)
Undifferentiated sarcoma	108 (16.9)	73 (16.3)	35 (18.3)
Leiomyosarcoma	111 (17.4)	77 (17.2)	34 (17.8)
Fibrosarcoma	154 (24.1)	105 (23.5)	49 (25.7)
Stage				0.437
I	334 (52.4)	234 (52.3)	100 (52.4)
II	216 (33.9)	156 (34.9)	60 (31.4)
III + IV	88 (13.8)	57 (12.8)	31 (16.2)
Maximum diameter of tumor, (cm)				0.435
< 10.0	403 (63.2)	278 (62.2)	125 (65.4)
≥ 10.0	235 (36.8)	169 (37.8)	66 (34.6)
BMI rank				0.165
Underweight	75 (11.8)	57 (12.8)	18 (9.4)
Normal	426 (66.8)	301 (67.3)	125 (65.4)
Overweight	129 (20.2)	82 (18.3)	47 (24.6)
Obese	8 (1.3)	7 (1.6)	1 (0.5)
SII				0.765
< 619.85	423 (66.3)	298 (66.7)	125 (65.4)
≥ 619.85	215 (33.7)	149 (33.3)	66 (34.6)
CRP				0.838
< 5.0	518 (81.2)	362 (81.0)	156 (81.7)
≥ 5.0	120 (18.8)	85 (19.0)	35 (18.3)
CONUT score				0.792
0-2	551 (86.4)	385 (86.1)	166 (86.9)
3-12	87 (13.6)	62 (13.9)	25 (13.1)
MUST group				0.355
Low risk	407 (63.8)	279 (62.4)	128 (67.0)
Medium risk	102 (16.0)	71 (15.9)	31 (16.2)
High risk	129 (20.2)	97 (21.7)	32 (16.8)
Subcutaneous adiposity				0.240
No	383 (60.0)	275 (61.5)	108 (56.5)
Yes	255 (40.0)	172 (38.5)	83 (43.5)
Visceral adiposity				0.093
No	313 (49.1)	229 (51.2)	84 (44.0)
Yes	325 (50.9)	218 (48.8)	107 (56.0)
Sarcopenia				0.523
No	315 (49.4)	217 (48.5)	98 (51.3)
Yes	323 (50.6)	230 (51.5)	93 (48.7)
CT-derived index, median (IQR)
TAI	89.67 (52.30)	87.92 (54.07)	92.30 (50.27)	0.259
SAI	46.47 (30.28)	45.22 (31.66)	51.20 (27.91)	0.194
VAI	36.34 (31.77)	36.14 (31.18)	37.85 (33.30)	0.350
SMI	45.35 (11.43)	44.90 (11.63)	46.35 (10.83)	0.056
Complication				0.422
No	481 (75.4)	341 (76.3)	140 (73.3)
Yes	157 (24.6)	106 (23.7)	51 (26.7)
Outcomes, median (IQR)
Length of hospitalization (day)	12 (12)	12 (12)	10 (10)	0.093
Progression-free survival (month)	25 (37)	26 (37)	25 (34)	0.943
Overall survival (month)	35 (33)	36 (32)	32 (32)	0.129

P value are obtained from χ² tests for categorical variables and Mann-Whitney tests for continuous variables because of its non-normal distribution. Interquartile range describing the range between the first quartile and the third quartile for variables with non-normal distribution representing their discreteness. IQR: Interquartile range; BMI: Body mass index; SII: Systematic inflammatory index; CRP: C-reaction protein; CONUT: Controlling nutritional score; CT: Computed tomography; TAI: Total adipose tissue index; SAI: Subcutaneous adipose tissue index; VAI: Visceral adipose tissue index; SMI: Skeletal muscle index; MUST: Malnutrition universal screening tool.

As for reproducibility of CT-derived measurements on the subset of 50 randomly selected scans, inter-observer ICCs were: Skeletal muscle area ICC = 0.898 [95% confidence interval (CI): 0.845-0.951], visceral adipose area ICC = 0.892 (95%CI: 0.836-0.948), subcutaneous adipose area ICC = 0.902 (95%CI: 0.842-0.962), total adipose area ICC = 0.874 (95%CI: 0.824-0.924). Intra-observer ICCs for reader 1 were: Skeletal muscle area ICC = 0.914, visceral adipose area ICC = 0.908, subcutaneous adipose area ICC = 0.897, total adipose area ICC = 0.885.

PFS and OS

The nutritional assessment tools and other clinical features of tumours were included in the univariate and multivariate Cox regression analyses for PFS and OS (Supplementary Table 1). Subtype, stage, tumour size, SII, CRP, CONUT score, and complications were entered into the multivariate Cox proportional hazards model for PFS (P < 0.05), and subtype, stage, tumour size, and length of hospitalization were ultimately included in the multivariate Cox regression model for OS. Subtype was identified as the only independent factor for PFS, and patients with fibrosarcoma had significantly better PFS than patients with other soft tissue sarcomas. Subtype, clinical stage, and length of hospital stay were found to be risk factors for OS. When CT indices were included as continuous variables, SMI and VAI showed modest associations with OS in univariate Cox models (SMI HR per 10 cm²/m² = 1.573, P = 0.04; VAI hazard ratio per 10 cm²/m² = 2.042, P = 0.01), but these associations attenuated after adjustment for stage, subtype and length of hospitalization. VIF values for CT indices and BMI ranged from 1.2 to 3.8, indicating moderate correlation but not extreme multicollinearity. RSF variable importance measures in continuous-variable runs ranked CT-derived SMI and VAI below clinical stage and tumor subtype for OS. Results were consistent when analyses were restricted to contrast-enhanced portal venous phase scans, supporting that CT indices in this cohort were not independent predictors after adjusting for established clinical prognosticators.

The RSF analysis selected important variables, including subtype, stage, CRP, and SII for PFS (Supplementary Figure 1A) and stage, hospital duration, subtype, BMI, and tumour size for OS (Figure 2A). The predicted survival from RSF analysis in terms of PFS (Supplementary Figure 1A) and OS (Figure 2B) was determined for patients in the training dataset. Nomogram plots were generated based on variables selected by multivariate Cox analysis and RSF analysis to show the predicted 3- and 5-year survival rates [PFS (Supplementary Figure 1B) and OS (Figure 3)] for patients with STS receiving surgical resection in the training group. The calibration curves of the 3-year and 5-year survival probabilities showed consistency between the predicted results and the actual observations for PFS (Supplementary Figure 1C) and OS (Figure 4). The C-index and Brier score reflected the predictive performance in terms of the accuracy of calibration and discrimination for PFS and OS (Table 2). The models constructed by the variables from the RSF analysis had better predictive accuracy than those constructed by the variables from the multivariate Cox analysis for both PFS and OS according to the C-index and Brier score. The time-dependent ROC analysis also confirmed the advantages of variable selection by RSF analysis compared with conventional Cox proportional hazards analysis for both PFS (Supplementary Figure 1D) and OS (Figure 5) at various time points.

Figure 2 Visualization of random survival forest analysis results for overall survival in the training cohort. A: Variable ranking with minimal depth method of random survival forest analysis on overall survival (OS) in the training cohort. The variables in the dashed box were identified as important for OS prediction. The dashed vertical line was the optimistic threshold using the mean of the minimal depth distribution which classified variables with minimal depth lower than this threshold as important in prediction of outcomes; B: Random forest predicted survival for each patient in the training cohort on OS. The lines with dark grey corresponded to censored individuals, and light grey curves corresponded to individuals occurring death events. BMI: Body mass index; CRP: C-reactive protein; MUST: Malnutrition universal screening tool.

Figure 3 Nomograms for predicting 3-year and 5-year overall survival probabilities in patients with soft-tissue sarcomas receiving surgical resection. A: Nomogram on overall survival constructed by important variables from random survival forest analysis for patients with soft-tissue sarcomas receiving surgical resection to predict 3-year and 5-year survival probabilities; B: Nomogram on overall survival constructed by significant variables from multivariate Cox analysis for patients with soft-tissue sarcomas receiving surgical resection to predict 3-year and 5-year survival probabilities. BMI: Body mass index.

Figure 4 Calibration curves for the nomogram presenting agreement between predicted and observational survival probabilities of overall survival for patients with soft-tissue sarcomas receiving surgical resection. The gray line of Y = X represents a perfect predictive power by an ideal model. The fit goodness with this diagonal line coincided with the model’s predictive performance. A: Calibration plot for comparison between nomogram predicted 3-year survival rates and actual observation for overall survival (OS) in the training cohort; B: Calibration plot for comparison between nomogram predicted 5-year survival rates and actual observation for OS in the training cohort; C: Calibration plot for comparison between nomogram predicted 3-year survival rates and actual observation for OS in the testing cohort; D: Calibration plot for comparison between nomogram predicted 5-year survival rates and actual observation for OS in the testing cohort. RSF: Random survival forest.

Figure 5 Time-dependent area under the receiver operating characteristic curves of predictive models for overall survival in training and testing cohorts. A: Time-dependent area under the receiver operating characteristic curves for predictive models on overall survival for the training cohort; B: Time-dependent area under the receiver operating characteristic curves for predictive models on overall survival for the testing cohort. AUC: Area under the receiver operating characteristic curves.

Table 2 Prediction accuracy of nomogram.

Variable	C-index				Brier score
	Three-year		Five-year		Three-year		Five-year
PFS	Cox selection	RSF selection	Cox selection	RSF selection	Cox selection	RSF selection	Cox selection	RSF selection
Training cohort	0.635 (0.576-0.694)	0.659 (0.600-0.718)	0.653 (0.574-0.733)	0.695 (0.614-0.776)	23.4 (13.5-33.2)	22.9 (13.3-32.5)	22.9 (10.5-35.2)	22.2 (10.2-34.2)
Testing cohort	0.601 (0.504-0.698)	0.675 (0.583-0.767)	0.615 (0.482-0.749)	0.702 (0.586-0.817)	23.6 (17.0-30.1)	22.3 (16.0-28.6)	23.5 (14.7-32.3)	21.6 (13.6-29.7)
OS
Training cohort	0.573 (0.511-0.634)	0.581 (0.519-0.642)	0.660 (0.595-0.724)	0.657 (0.592-0.723)	22.4 (20.0-24.9)	22.3 (19.8-24.8)	18.8 (16.8-20.9)	18.7 (16.7-20.8)
Testing cohort	0.474 (0.378-0.569)	0.473 (0.377-0.568)	0.564 (0.444-0.684)	0.584 (0.466-0.703)	25.8 (22.3-29.3)	25.6 (22.1-29.0)	19.4 (16.0-22.9)	19.2 (15.8-22.6)

Prediction accuracy of nomogram using different variable selection methods for progression-free survival and overall survival in the training and testing sets of patients with soft-tissue sarcomas receiving surgical resection. OS: Overall survival; PFS: Progression-free survival; RSF: Random survival forest.

Hospital duration

Based on the results of univariate logistic analysis (P < 0.05), age at diagnosis, location of primary tumour, subtype, stage, tumour size, complications and some nutritional assessment indexes were included in the multivariate binary logistic regression model. The axial location of primary tumours and poor basic condition were independent factors for long postoperative hospital stays. Otherwise, the subtype significantly influenced the short-term prognosis (Supplementary Table 2).

The random forest analysis selected subtype, BMI, MUST group, and histological stage as important variables for hospital duration (Supplementary Figure 2A). The predictive errors decreased as trees were added, and we identified significant improvement in predictive errors before 100 trees were built in the forest (Supplementary Figure 2A). We constructed a nomogram based on the variables chosen by random forest analysis and binary logistic analysis (Supplementary Figure 2B). The calibration curves of risks displayed favourable agreement in the training cohort and moderate consistency in the testing cohort between the predicted results and the actual observations (Supplementary Figure 2C). The Brier scores were 22.4 and 22.1 in the training set and 23.2 and 22.6 in the testing set for the model constructed by the logistic regression-selected variables and the model constructed by the random forest-selected variables, respectively. The ROC analysis verified the advantages of variable selection by random forest analysis with a larger area under the ROC curve than binary logistic analysis for hospital duration (Supplementary Figure 2D).

DISCUSSION

In this study, we evaluated the relationships between three kinds of nutrition-associated tools from peripheral blood, routine CT and overall physical state and different clinical outcomes, including survival status, OS, PFS and postoperative hospital days, by traditional and advanced statistical methods for patients with STS treated with surgery. The prognostic variables selected by two methods, Cox regression analysis and RSF analysis, for PFS and OS were integrated to construct predictive models whose accuracy was evaluated from the aspects of calibration and discrimination. The models developed by RSF-selected variables showed better predictive performance than those developed by Cox-selected variables.

Nomograms have been used to construct predictive models containing standard oncological variables for patients with soft tissue sarcomas. In our study, these variables were also included. Subtype was found to be significantly associated with OS, PFS and hospitalization duration by multivariate Cox regression. Histological grade was considered another prognostic variable for OS, which was consistent with the findings of a previous study[6]. There have been few studies based on the Chinese population to identify the prognosis of soft tissue sarcoma compared with the large samples from open databases, such as the Surveillance, Epidemiology, and End Results database, focused on the United States population. Our study provided evidence about the predictive value of standard oncological variables for soft tissue sarcoma.

SII indicates the systemic inflammatory reaction in patients with soft tissue sarcoma and presented consistent results with those of previous studies[12,15]. CRP assesses patients’ immune conditions and acute response to malignancies within a short time[43], and its relationship with long-term prognosis was confirmed in both nonmetastatic and metastatic sarcomas in our study. Recently, the association between inflammatory reactions and progressive nutritional decreases has commonly been accepted[44], and intensive immune action may cause excessive consumption, inducing sarcopenia and even cachexia. The CONUT score has been considered a more comprehensive assessment tool than pure inflammatory indexes. Previous studies have proven the strong association between the CONUT score and OS[18], which indicated the independent prognostic value of the CONUT score for OS in the Chinese population. Indeed, the inclusion of serum albumin, total lymphocyte count and total cholesterol increased its agreement with accepted nutritional assessment tools such as subjective global assessment and full nutritional assessment[45] and made it more practical than complex questionnaires.

The body composition of skeletal muscle and adipose tissue provided a better representation of actual nutritional conditions compared with BMI. The third lumbar vertebra slice from sectional CT images can be used to sufficiently measure the components of skeletal muscle and adipose tissue in the Asian population[46]. The use of skeletal muscle mass has not been approved in patients with sarcomas. However, its prognostic value has been confirmed in many kinds of carcinomas[47,48]. Visceral adipose tissue has been demonstrated to have a significant association with long-term survival in STS[49,50]. Nevertheless, our study did not identify prognostic value for body composition-associated indicators after normalization by patient height, which may eliminate interference from these confounding factors.

Although prior oncology studies have demonstrated prognostic associations for CT-derived sarcopenia and adiposity, in our cohort these indices did not remain independent predictors in multivariable Cox or RSF models. Several factors may explain this negative finding. First, collinearity between BMI and CT-derived adiposity measures can reduce the independent contribution of CT indices when included together. Our VIF diagnostics indicated moderate correlation. The threshold selection for sarcopenia and visceral obesity varies across studies and populations that relying on published cutoffs often derived from Western cohorts might not capture prognostically relevant distributions for our predominantly Asian cohort. Otherwise, disease-specific factors such as tumor subtype and stage might exert stronger prognostic influence in STS than body-composition measures, such that CT indices add limited incremental discrimination beyond established oncologic variables. Moreover, heterogeneity in CT acquisition over the 2009-2018 inclusion period could introduce measurement variability despite our standardization procedures. However, sensitivity analyses restricted to portal venous-phase studies yielded similar results. Finally, dichotomization of continuous measures reduces statistical power; our analyses using continuous CT measures showed modest univariate associations which were attenuated after adjustment. These findings suggest that CT-derived body-composition measures might be context-dependent prognosticators in STS and underscore the need for externally validated population-specific thresholds and prospective studies integrating longitudinal nutritional trajectories.

MUST has been used as a prospective tool for the detection of the risk of malnutrition for inpatients in the European population[44]. The MUST score is composed of three parts (BMI, intentional weight loss and short-term absence of intake from the mouth), each of which has been considered an independent prognostic factor. The practicability of the MUST score in the Asian population has remained unknown, with insufficient evidence from cross-sectional studies[51,52]. The MUST score evaluates patients’ overall physical state in combination with inflammation-associated indicators and other nutrition-associated indexes to establish nutritional interventional strategies and decrease patients’ risk of malnutrition[53]. Controversial arguments exist regarding whether patients should receive immune-enhanced or immune-deregulated diets to decrease dispensable energy consumption[54]. The time to adjust nutritional intake based on systemic nutritional assessment tools is still not clear, so more clinical studies are needed to find the appropriate time and threshold.

There were plenty of strengths in this study. We performed RSF analysis to select important variables with a strong ability to predict outcomes by automated examination. In our study, different kinds of nutrition-associated indexes had high correlations with each other, and the variable selection by multivariate Cox proportional hazards analysis led to unavoidable errors in determining whether these nutritional indexes should be included in the Cox regression model together. The RSF analysis was capable of coping with complicated internal structures for highly correlated variables, which was appropriate in this study. The main limitation of this study was that it was a retrospective study with inevitable recall bias. Although the results of laboratory and radiology examinations were objective and hard to interfere with, the records for MUST, especially the diet change in a short time, were defective, the errors of which were hard to correct and were affected by different recorders. In addition, there were inconsistencies between the depicted adipose and skeletal muscle areas on sectional CT images and the actual body components attributed to the wide HU range of fat and muscle tissue referred to in this study.

CONCLUSION

In conclusion, we identified the independent prognostic value of nutrition-associated indexes from routine examinations for patients with STS treated with resection. We also assessed the prognostic value of variables and integrated predictive models with appropriate methods, including traditional survival analysis and advanced machine learning methods. These indicators can be used as references for the preoperative interventions of patients’ nutritional status and to guide subsequent therapy to improve outcomes.