BPG is committed to discovery and dissemination of knowledge
Retrospective Cohort Study Open Access
Copyright: ©Author(s) 2026. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution-NonCommercial (CC BY-NC 4.0) license. No commercial re-use. See permissions. Published by Baishideng Publishing Group Inc.
World J Clin Oncol. Nov 24, 2026; 17(11): 126582
Published online Nov 24, 2026. doi: 10.5306/wjco.126582
Machine learning and logistic regression for ascertainable three-year recurrence status after soft tissue sarcoma surgery
Zhang-Fu Li, Bai-Cheng Yang, Shi-Qing Liu, Sheng-Ji Yu, Department of Orthopaedics, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China
Yi-Bo Liu, Qing Zhang, Department of Orthopedic Oncology Surgery, Beijing Jishuitan Hospital, Capital Medical University, Beijing 100035, China
ORCID number: Zhang-Fu Li (0000-0002-0112-1678); Sheng-Ji Yu (0000-0002-8994-0562).
Co-first authors: Zhang-Fu Li and Bai-Cheng Yang.
Co-corresponding authors: Qing Zhang and Sheng-Ji Yu.
Author contributions: Li ZF and Yang BC contributed to the study design, curation of the development cohort, the statistical performance and machine-learning analysis, and the manuscript drafting; Liu YB contributed to the collection, verification, and interpretation of the second-institution cohort; Liu SQ contributed to the data acquisition, variable definition, and outcome ascertainment; Liu SQ and Zhang Q contributed to the clinical interpretation and revision; Li ZF and Yang BC contributed to the preparation of the figures and tables; Zhang Q contributed to the supervision of the second-institution cohort; Yu SJ contributed to the supervision of the study, the methodological and clinical guidance. Zhang Q and Yu SJ contributed equally to this work as co-corresponding authors. Li ZF and Yang BC contributed equally to this work as co-first authors.
AI contribution statement: OpenAI Codex was used during manuscript revision for code assistance, language editing, and drafting support. All analyses, outputs, citations, and text were reviewed and verified by the authors, who take full responsibility for the final manuscript.
Supported by the National Natural Science Foundation of China, No. 82272964; and the National High Level Hospital Clinical Research Funding, No. 80102022625.
Institutional review board statement: This retrospective study used the approved multicenter soft tissue sarcoma database and was approved by the Independent Ethics Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, No. NCC2020C-341. The multicenter database framework covered de-identified data contributed by both participating institutions, including Beijing Jishuitan Hospital. The study was conducted in accordance with the Declaration of Helsinki.
Informed consent statement: The requirement for informed consent was waived by the approving ethics committee because this retrospective study used de-identified data.
Conflict-of-interest statement: The authors declare no conflicts of interest related to this work.
STROBE statement: The authors have read the STROBE Statement-checklist of items, and the manuscript was prepared and revised according to the STROBE Statement-checklist of items.
Data sharing statement: Patient-level data are not publicly available because of ethical and privacy restrictions. De-identified data may be available from the corresponding authors on reasonable request, subject to institutional ethics and data-protection approval. A submission-stage reproducibility supplement contains the analysis protocol, configuration registry, aggregate internal and second-institution results, input schema, lock files, software versions, non-identifiable calculator materials, and revision sensitivity outputs. Patient-level predictions and fold assignments are available only under controlled institutional approval.
Corresponding author: Sheng-Ji Yu, Professor, Department of Orthopaedics, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, No. 17 Panjiayuan Nanli, Chaoyang District, Beijing 100021, China. zlyygk@163.com
Received: August 10, 2026
Revised: September 24, 2026
Accepted: November 2, 2026
Published online: November 24, 2026
Processing time: 101 Days and 18.1 Hours

Abstract
BACKGROUND

Postoperative local-recurrence estimates could inform surveillance discussions after soft tissue sarcoma surgery, but machine-learning performance can be overstated when model selection is not separated from evaluation or when simpler regression models are not compared.

AIM

To develop and evaluate a model for binary 3-year local-recurrence status among patients with ascertainable status and compare it with prespecified logistic-regression models.

METHODS

This retrospective two-institution study analyzed 1035 development patients with ascertainable binary recurrence status at 3 years (331 recurrences) and 212 patients in a second-institution cohort (67 recurrences). The estimand was recurrence-status probability conditional on 3-year status being ascertainable, not cumulative incidence in all postoperative patients. The locked analytical package did not retain reason-specific exclusion counts or the source screening denominator for the second institution. Twelve predictors, 243 configurations across 14 model families, and three ensemble strategies were evaluated by five-fold outer cross-validation repeated five times with five-fold inner tuning. Mean Brier score was the primary selection criterion. The selected CatBoost model was fitted in the complete development cohort, frozen, and applied without recalibration in a protocol-frozen second-institution re-evaluation.

RESULTS

CatBoost had a mean internal area under the receiver operating characteristic curve (AUC) of 0.8126 and Brier score of 0.1584. In the second institution, its AUC was 0.8050 [95% confidence interval (CI): 0.7433-0.8668], Brier score 0.1615 (95%CI: 0.1331-0.1912), log loss 0.4884 (95%CI: 0.4146-0.5678), calibration intercept 0.148 (95%CI: -0.222 to 0.490), and calibration slope 0.985 (95%CI: 0.703-1.357). The CatBoost-minus-six-variable-logistic differences were -0.0083 for AUC (95%CI: -0.0480 to 0.0299) and 0.0043 for Brier score (95%CI: -0.0112 to 0.0198). A secondary full 12-variable logistic model also showed no material difference from CatBoost.

CONCLUSION

In the protocol-frozen second-institution re-evaluation, CatBoost retained discrimination but did not demonstrate incremental performance over logistic comparators. The study was not designed or powered to establish equivalence or non-inferiority. Prospective multicenter evaluation using time-to-event outcomes is required before clinical use.

Key Words: Soft tissue sarcoma; Local recurrence; Machine learning; Clinical prediction model; Nested cross-validation; Calibration; Decision curve analysis

Core Tip: We compared 243 configurations across 14 base-model families plus three prespecified ensembles by repeated nested cross-validation. The selected CatBoost model was frozen and applied without recalibration in a protocol-frozen second-institution re-evaluation. CatBoost retained an area under the receiver operating characteristic curve of 0.8050 but did not demonstrate incremental performance over either a six-variable or full 12-variable logistic model; this was not an equivalence or non-inferiority analysis. The browser interface is provided for reproducibility and independent evaluation, not for clinical decision-making.



INTRODUCTION

Soft tissue sarcomas (STS) are rare mesenchymal malignancies that account for approximately 1% of adult cancers and comprise more than 100 histological subtypes[1,2]. Curative-intent resection remains central to treatment, yet local recurrence can require further surgery or radiotherapy, compromise function, and accompany poorer long-term outcomes[3-5]. A postoperative risk estimate could support discussions about local surveillance intensity while preserving specialist judgment and patient preferences[3,6].

Existing tools address related but not identical outcomes. A postoperative nomogram estimates local recurrence after limb-sparing surgery[7]. Personalised sarcoma care was developed to support treatment decisions in high-grade extremity STS[8], whereas other nomograms and Sarculator primarily address recurrence, overall survival, or distant metastasis in different target populations[9-11]. Differences in outcome definition, case mix, treatment pathways, and predictor measurement can alter performance across institutions[12]. Evaluation should therefore include probability accuracy and comparison with simpler models, not discrimination alone.

Machine-learning methods can represent nonlinear associations and interactions, but complexity does not establish added clinical value[13-16]. Model searches are optimistic when selection and performance estimation use the same data, and empirical comparisons have not shown a consistent advantage over logistic regression[17-20]. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) + artificial intelligence (AI) recommends explicit definition of the target population, prediction time, outcome, predictors, evaluation data, and intended use, together with complete performance reporting[21]. Prediction model risk of bias assessment tool provides a separate framework for appraising risk of bias[22].

This study addressed two prespecified questions. First, what discrimination, prediction error, and calibration would a model selected from 243 configurations across 14 base-model families plus three ensemble strategies achieve in repeated nested evaluation and in a second institution after model freezing? Second, would the selected model improve on a fixed six-variable logistic benchmark? A secondary analysis compared CatBoost with a full 12-variable logistic model to separate algorithm choice from predictor-set size. Shapley additive explanations (SHAP) were used to describe the fitted model, and a browser-based research interface was provided for reproducibility testing.

MATERIALS AND METHODS
Study design and data sources

This retrospective prediction-model study used a development cohort from the National Cancer Center of China and a second-institution cohort from Beijing Jishuitan Hospital. The development cohort included surgically treated patients assessed between January 2010 and December 2021. Clinical, pathological, treatment, and follow-up data were obtained from institutional records. The intended prediction landmark was completion of definitive tumor resection, after postoperative pathology and margin information became available; the limitation in the temporal definition of the two treatment-status predictors is described below.

The second-institution cohort was assembled using the same eligibility, predictor, and binary outcome definitions. Before the current protocol was frozen, the same 212-patient cohort had been used in an earlier project iteration to evaluate a gradient-boosting model developed after Boruta feature selection performed in the development data. Second-institution outcomes were not used to perform that feature selection, to construct new predictors, or to tune the present model, but the investigators had seen the cohort’s case mix and earlier performance. The present assessment is therefore a protocol-frozen second-institution re-evaluation, not a fully blinded or previously unseen external validation.

For the current analysis, the statistical analysis plan and the 243-configuration registry were frozen on August 5, 2026. File-system timestamps and SHA-256 hashes document the sequence: Protocol and registry lock, development-only CatBoost selection and final fitting, model-artifact lock, and generation of second-institution predictions. No predictor, outcome, model definition, hyperparameter, recalibration step, or reported comparison was changed in response to the current re-evaluation results. The audit trail is summarized in Supplementary Table 1.

Participants and outcome

Eligible patients had histopathologically confirmed soft tissue sarcoma and underwent definitive surgery with curative intent. Exclusion criteria were non-sarcoma pathology, palliative or non-definitive local treatment, missing prespecified predictors or binary outcome, insufficient follow-up to establish 3-year status in patients without documented recurrence, or duplicate records. These rules produced complete-case, ascertainable-status analytical cohorts. An earlier cohort-assembly report documented 1651 development-institution records assessed and 1035 included; however, reason-specific exclusion counts and covariates for the 616 non-included records were not retained in the locked analytical package. For the second institution, the locked package retained only the 212 final analytical records, not a reproducible upstream screening denominator or excluded-patient dataset. Consequently, excluded-versus-included baseline comparisons could not be performed; the available participant-flow audit is reported in Supplementary Table 2.

The target outcome was documented local recurrence within 3 years after definitive surgery. Local recurrence was relapse at or adjacent to the primary surgical site, ascertained from surveillance imaging reports, clinic documentation, pathological confirmation when available, and institutional follow-up records. The same binary definition and coding rules were used at both institutions. Surveillance modality and frequency were clinician-directed rather than protocol-mandated, and the locked extracts did not preserve visit-level schedules, adjudicator identity, recurrence time, censoring time, death, transfer, or competing-event time. Range checks, category checks, duplicate-identifier checks, event-count checks, and cross-institution coding harmonization were completed before analysis. The estimand was therefore recurrence-status probability among surgically treated patients whose binary 3-year status was ascertainable, not cumulative incidence among all patients undergoing surgery.

Both cohorts were fixed retrospective samples, and no prospective sample-size target was used. Precision was evaluated through confidence intervals for second-institution performance estimates and paired differences. A nonsignificant difference was not interpreted as evidence of equivalence.

Candidate predictors and preprocessing

Twelve predictors were prespecified on the basis of clinical relevance[23,24] and routine availability: Age, sex, anatomical location, tumor depth[25], maximum tumor size, French Federation of Cancer Centers Sarcoma Group grade, surgical margin[26-29], radiotherapy, chemotherapy, bone or neurovascular invasion, unplanned surgery[30-33], and symptom duration. Histological subtype was not included because numerous uncommon subtypes would create sparse categories. The full predictor dictionary, coding, source, and intended measurement time are provided in Supplementary Table 3.

Radiotherapy and chemotherapy were binary status fields in the retrospective analytical exports. The exports did not preserve treatment dates or distinguish neoadjuvant treatment, a plan for adjuvant treatment, and completed postoperative treatment. Their availability at the intended postoperative landmark therefore cannot be verified for every patient, and the 12-predictor model should not be interpreted as a strict real-time landmark model based only on information known at surgery. The variables were retained because they were prespecified before the current analysis; a post hoc sensitivity analysis excluded both without changing the locked primary model.

After the eligibility and complete-case restrictions had been applied, neither locked analytical table contained missing predictor or outcome values. This statement describes the final exports and does not imply that the source clinical populations had no missing data. The locked package did not retain pre-exclusion missingness patterns, so multiple imputation or inverse-probability analyses of selection could not be performed. No imputation within the analytical cohorts, outcome oversampling, synthetic data generation, or outcome-driven feature selection was performed. Within every resampling pipeline, continuous predictors were standardized when required, non-CatBoost models used one-hot mappings learned from the training fold, and CatBoost used native categorical handling[34].

Candidate models and nested internal validation

The registry contained 243 configurations across 14 base-model families: Fixed six-variable logistic regression (1), fixed full 12-variable logistic regression (1), elastic-net logistic regression (35), random forest (18), extremely randomized trees (18), histogram gradient boosting (18), XGBoost (24), LightGBM (24), CatBoost (24), gradient boosting (18), radial-basis-function support vector machine (20), k-nearest neighbors (18), multilayer perceptron (18), and shrinkage linear discriminant analysis (6)[34-37]. Three additional ensemble strategies were prespecified: An equal-weight average of the three best families, a Brier-weighted average, and logistic stacking. Thus, Figure 1A and B depict 243 base-model configurations, whereas Figure 1C compares 17 strategies (14 tuned families plus three ensembles). The complete reconciliation is provided in Supplementary Table 4.

Figure 1
Figure 1 Internal model search and repeated nested evaluation. A: Heatmap of all 243 prespecified hyperparameter configurations across 25 outer splits, comprising 6075 inner-validation results. Cell color represents the within-split percentile of mean inner-fold Brier score; B: Distribution of configuration-level mean inner-validation Brier scores within 14 model families; the diamond marks CFG-159; C: Mean outer-cycle Brier scores for 14 tuned base-model families and three prespecified ensembles. Error bars are SDs across five correlated repeated cycles and are not confidence intervals. Pale shading shows the prespecified descriptive one-standard-error grouping, calculated as the minimum mean Brier score plus the selected strategy’s SD divided by the square root of five; final selection remained based on the minimum mean Brier score.

Internal performance was estimated with stratified five-fold outer cross-validation repeated five times, producing 25 outer test folds. Hyperparameters were selected by stratified five-fold inner cross-validation within each outer training set. All preprocessing, tuning, component selection, and ensemble-weight estimation remained inside the outer training data. Each outer test fold was used only to generate held-out predictions. Identical outer and inner splits were used across strategies.

Model selection, final fitting, and second-institution evaluation

Mean outer-cycle Brier score was the primary selection measure. Secondary measures were area under the receiver operating characteristic curve (AUC), log loss, calibration intercept, calibration slope, observed-to-expected event ratio, and decision-curve net benefit. The fixed six-variable logistic model was selected as the primary conventional benchmark because it was parsimonious, clinically interpretable, and based on six variables identified in earlier development-only work before the current protocol-frozen comparison. It was not selected using the current second-institution outcomes. The secondary full 12-variable logistic model used the same predictors as CatBoost and therefore separated algorithm choice from predictor-set size. Logistic preprocessing was learned inside each resampling split, CatBoost used native categorical handling, and identical splits and evaluation metrics were used to make the comparisons as fair as model-specific preprocessing allowed.

CatBoost hyperparameters were re-tuned within the prespecified CatBoost registry by fixed stratified 10-fold cross-validation in the complete development cohort, again minimizing Brier score. The locked configuration (CFG-159) used 300 trees, depth 7, learning rate 0.03, L2 leaf regularization 5, and random strength 1. Both the six-variable and full 12-variable logistic models were fitted in the development cohort without use of second-institution outcomes. Model coefficients are provided in Supplementary Table 5.

All second-institution performance measures were calculated from predictions generated once by the development-fitted models. Bootstrap resampling did not retrain, recalibrate, or modify a model. AUC intervals used DeLong’s method for the primary two models. Other intervals and paired differences used 2000 patient-level bootstrap samples. Calibration was summarized by intercept, slope, observed-to-expected event (O/E) ratio, and a flexible cubic-spline curve (four quantile-positioned knots, degree 3) with pointwise bootstrap bands; eight quantile groups were superimposed descriptively. Decision curves were examined from thresholds 0.05 to 0.60. Because no surveillance-action threshold was clinically established, decision-curve results were considered exploratory[38-41].

Two post hoc analyses were added during revision. First, predictor distributions and CatBoost predicted probabilities were compared between institutions using standardized mean differences (SMDs). Second, to examine the treatment-timing concern without altering the primary model, CatBoost was refitted in the complete development cohort with the locked CFG-159 hyperparameters after excluding radiotherapy and chemotherapy; no external-data tuning or recalibration was performed. Its fixed second-institution predictions were evaluated with the same metrics and 2000 patient-level bootstrap samples. This sensitivity analysis was not used to replace or reselect the primary model.

Model interpretation and web implementation

Global feature attribution for the locked CatBoost model was quantified in the development cohort using SHAP. Mean absolute SHAP values ranked the average magnitude of each predictor’s contribution to the raw model output (log-odds scale)[42-44]. These post hoc attributions were not used for predictor or model selection and do not estimate causal or treatment effects.

A browser-based English-language research interface was made available at https://sts-recurrence-calculator.potatoli.chatgpt.site. It requires all 12 model inputs and returns the CatBoost estimate of binary 3-year recurrence-status probability conditional on ascertainable status. The locked CatBoost file SHA-256 was 1bf94cabf45ef62c476f32061617aa81cd86ab27cc8d2590759785833a76c324. Automated concordance testing across all 212 second-institution records showed a maximum absolute difference of 9.7 × 10-17 between interface-function predictions and locked reference probabilities. The interface requests no direct patient identifiers, but it has not undergone independent privacy, cybersecurity, access-control, data-retention, usability, regulatory, or prospective impact review. Identifiable information should not be entered, and the interface must not determine surveillance or treatment. A model card is provided in Supplementary Table 6.

Statistical software and reporting

Analyses used Python 3.9.6 with CatBoost 1.2.10, scikit-learn 1.6.1, XGBoost 2.1.4, LightGBM 4.6.0, SHAP 0.49.1, pandas 2.3.3, NumPy 2.0.2, and SciPy 1.13.1; primary second-institution confidence intervals were generated in R 4.6.1. Random seeds, fold assignments, candidate scores, held-out predictions, model files, data hashes, and software versions were retained. The report was structured using TRIPOD + AI for prediction-model studies and STROBE for the retrospective cohort description[21,45]. No patients or members of the public were involved in study design, conduct, or reporting.

RESULTS
Cohorts

The final complete-case, ascertainable-status development cohort comprised 1035 patients, of whom 331 (32.0%) had documented local recurrence within 3 years. The second-institution cohort comprised 212 patients, including 67 recurrences (31.6%). The locked analytical tables contained no missing predictor or outcome values after cohort restrictions. The available flow audit, including the absence of reason-specific exclusion data and the unrecoverable second-institution source screening denominator in the locked package, is shown in Supplementary Table 2. Table 1 summarizes the analyzed cohorts, and Figure 2 shows the model-development and evaluation workflow.

Figure 2
Figure 2 Protocol-frozen model-development and second-institution evaluation workflow. The complete development cohort was evaluated by repeated nested cross-validation. CatBoost was selected by the minimum mean Brier score, tuned in the complete development cohort, frozen, and applied unchanged to the second-institution cohort. This figure depicts the analytical workflow rather than participant screening. ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; SHAP: Shapley additive explanations.
Table 1 Characteristics of the analyzed development and second-institution cohorts, mean ± SD/n (%).
Characteristic
Development cohort (n = 1035)
Second-institution cohort (n = 212)
Age, years47.78 ± 13.6848.88 ± 13.57
Sex
    Female457 (44.15)90 (42.45)
    Male578 (55.85)122 (57.55)
Location
    Upper extremity150 (14.49)31 (14.62)
    Lower extremity504 (48.70)95 (44.81)
    Trunk381 (36.81)86 (40.57)
Depth
    Superficial307 (29.66)60 (28.30)
    Deep728 (70.34)152 (71.70)
Tumor size, cm5.95 ± 4.236.39 ± 4.12
FNCLCC grade
    1464 (44.83)92 (43.40)
    2246 (23.77)48 (22.64)
    3325 (31.40)72 (33.96)
Surgical margin
    Negative952 (91.98)190 (89.62)
    Positive83 (8.02)22 (10.38)
Radiotherapy
    No658 (63.57)110 (51.89)
    Yes377 (36.43)102 (48.11)
Chemotherapy
    No793 (76.62)154 (72.64)
    Yes242 (23.38)58 (27.36)
Bone or neurovascular invasion
    No802 (77.49)163 (76.89)
    Yes233 (22.51)49 (23.11)
Unplanned surgery
    No822 (79.42)178 (83.96)
    Yes213 (20.58)34 (16.04)
Symptom duration, months24.66 ± 13.2124.86 ± 13.84
Local recurrence within 3 years331 (32.0)67 (31.60)

Between-institution SMDs were below 0.12 for all baseline predictors except radiotherapy (absolute SMD: 0.238). The next largest differences were unplanned surgery (0.118) and tumor size (0.106). The descriptive SMD for CatBoost predicted probability was 0.078 (development apparent mean 0.312 vs second-institution mean 0.293). Full comparisons are presented in Supplementary Table 7.

Repeated nested internal validation

All 14 base-model families and three ensemble strategies generated complete outer-fold predictions in each of the five repeats. The 243 prespecified configurations produced 6075 inner candidate scores across 25 outer folds without an unmatched configuration. Repeated-validation results are summarized in Table 2 and Figure 1.

Table 2 Repeated nested internal-evaluation performance by modeling strategy, mean ± SD.
Strategy
AUC
Brier score
Calibration slope
CatBoost0.8126 ± 0.00520.1584 ± 0.00171.070 ± 0.080
Equal-weight top-3 ensemble0.8134 ± 0.00430.1586 ± 0.00161.086 ± 0.044
Brier-weighted top-3 ensemble0.8133 ± 0.00430.1586 ± 0.00161.086 ± 0.044
Gradient boosting0.8144 ± 0.00310.1589 ± 0.00111.117 ± 0.019
XGBoost0.8124 ± 0.00350.1595 ± 0.00111.054 ± 0.027
Logistic top-3 stack0.8117 ± 0.00440.1608 ± 0.00230.962 ± 0.038
Histogram gradient boosting0.8092 ± 0.00260.1613 ± 0.00120.912 ± 0.018
Random forest0.8058 ± 0.00530.1620 ± 0.00241.134 ± 0.073
RBF support vector machine0.8005 ± 0.00450.1634 ± 0.00220.976 ± 0.051
Extra trees0.8015 ± 0.00560.1637 ± 0.00301.067 ± 0.090
Elastic-net logistic0.7994 ± 0.00170.1640 ± 0.00091.042 ± 0.011
LightGBM0.8023 ± 0.00480.1642 ± 0.00180.837 ± 0.035
Multilayer perceptron0.7955 ± 0.00870.1647 ± 0.00260.867 ± 0.119
Full 12-variable logistic0.7998 ± 0.00210.1647 ± 0.00110.907 ± 0.013
Shrinkage LDA0.7974 ± 0.00210.1657 ± 0.00110.878 ± 0.015
Six-variable logistic benchmark0.7948 ± 0.00130.1663 ± 0.00050.948 ± 0.009
k-nearest neighbors0.7629 ± 0.00770.1802 ± 0.00181.012 ± 0.178

CatBoost had the lowest mean Brier score (0.1584; SD: 0.0017), with mean AUC: 0.8126 (SD: 0.0052), log loss: 0.4813 (SD: 0.0051), calibration intercept: 0.011 (SD: 0.030), calibration slope: 1.070 (SD: 0.080), and O/E ratio: 1.005 (SD: 0.015). The six-variable logistic benchmark yielded mean Brier score: 0.1663 (SD: 0.0005), AUC: 0.7948 (SD: 0.0013), and log loss: 0.4959 (SD: 0.0010). These SDs describe variability across correlated repeated cycles and are not confidence intervals. Complete internal metrics are given in Supplementary Table 8.

For the paired internal comparison, each patient’s five out-of-fold predictions were averaged before calculating model differences. CatBoost-minus-six-variable-logistic differences were 0.0232 for AUC [95% confidence interval (CI): 0.0084-0.0394] and -0.0099 for Brier score (95%CI: -0.0163 to -0.0039). These patient-level estimates differ from differences between repeat-level means in Table 2. CatBoost was selected by the minimum mean Brier score and ranked among the three best families in 24 of 25 outer folds. The choice was finalized before the reported second-institution predictions were generated.

Locked model and SHAP attribution

Fixed 10-fold tuning in the complete development cohort selected CFG-159. The final model artifact, predictor schema, data hash, package versions, and model hash were frozen. Apparent performance after fitting the complete cohort was not used as a validation estimate. The global SHAP analysis is shown in Figure 3 and Supplementary Table 9. French Federation of Cancer Centers Sarcoma Group grade had the largest mean absolute SHAP value (0.772), followed by bone or neurovascular invasion (0.478), unplanned surgery (0.476), tumor size (0.294), and anatomical location (0.229). These values summarize model attribution on the raw-output scale and do not indicate causal importance.

Figure 3
Figure 3 Global Shapley additive explanations summary for the locked CatBoost model in the development cohort. A: Mean absolute Shapley additive explanations (SHAP) values rank average attribution magnitude; B: Each point represents one patient and shows the direction of a predictor’s contribution to model log-odds. For continuous and ordered predictors, blue indicates lower values and orange indicates higher values. Binary predictors are coded from absence/Lower category to presence/higher category. Tumor-location colors denote upper extremity, lower extremity, and trunk. SHAP attribution is descriptive of the fitted model and is not causal. FNCLCC: French Federation of Cancer Centers Sarcoma Group; SHAP: Shapley additive explanations.
Protocol-frozen second-institution re-evaluation

Applied without recalibration to the second institution, CatBoost achieved AUC: 0.8050 (95%CI: 0.7433-0.8668), Brier score: 0.1615 (95%CI: 0.1331-0.1912), log loss: 0.4884 (95%CI: 0.4146-0.5678), calibration intercept: 0.148 (95%CI: -0.222 to 0.490), calibration slope: 0.985 (95%CI: 0.703-1.357), and O/E ratio: 1.079 (95%CI: 0.885-1.276) (Table 3; Figure 4A and B). The slope point estimate was close to 1, but its interval and the intercept interval were wide.

Figure 4
Figure 4 Protocol-frozen second-institution evaluation of CatBoost and the six-variable logistic benchmark in 212 patients. A: Receiver operating characteristic curves; shaded regions are pointwise 95% patient-bootstrap bands, whereas the reported area under the receiver operating characteristic curve (AUC) intervals use DeLong’s method; B: Cubic-spline calibration curves using four quantile-positioned knots and degree 3, with pointwise 95% bootstrap bands, eight quantile-grouped observed proportions, and predicted-risk distributions; C: Exploratory decision curves from threshold 0.05 to 0.60; D: Paired AUC difference is CatBoost minus logistic; Brier advantage is logistic minus CatBoost. Positive values therefore favor CatBoost in both rows. All curves arose from fixed predictions without recalibration or retraining. AUC: Area under the receiver operating characteristic curve; CI: Confidence interval.
Table 3 Protocol-frozen second-institution performance of CatBoost and logistic comparators (95%CI).
Model
AUC
Brier score
Log loss
Calibration intercept
Calibration slope
O/E ratio
CatBoost0.8050 (0.7433-0.8668)0.1615 (0.1331-0.1912)0.4884 (0.4146-0.5678)0.148 (-0.222 to 0.490)0.985 (0.703-1.357)1.079 (0.885-1.276)
Six-variable logistic0.8134 (0.7533-0.8734)0.1572 (0.1310-0.1866)0.4770 (0.4116-0.5509)-0.044 (-0.408 to 0.305)0.978 (0.710-1.333)0.979 (0.811-1.151)
Full 12-variable logistic0.8104 (0.7488-0.8719)0.1580 (0.1297-0.1863)0.4800 (0.4066-0.5549)-0.059 (-0.432 to 0.272)0.921 (0.691-1.259)0.973 (0.806-1.129)

The six-variable logistic benchmark achieved AUC: 0.8134 (95%CI: 0.7533-0.8734), Brier score: 0.1572 (95%CI: 0.1310-0.1866), log loss: 0.4770 (95%CI: 0.4116-0.5509), calibration intercept: -0.044 (95%CI: -0.408 to 0.305), calibration slope: 0.978 (95%CI: 0.710-1.333), and O/E ratio: 0.979 (95%CI: 0.811-1.151). CatBoost-minus-logistic differences were -0.0083 for AUC (95%CI: -0.0480 to 0.0299; paired DeLong P = 0.677) and 0.0043 for Brier score (95%CI: -0.0112 to 0.0198). The data therefore provided no evidence of incremental performance, although the intervals did not establish equivalence.

In the secondary same-predictor-set comparison, full 12-variable logistic regression achieved AUC: 0.8104 (95%CI: 0.7488-0.8719), Brier score: 0.1580 (bootstrap 95%CI: 0.1297-0.1863), and log loss: 0.4800 (95%CI: 0.4066-0.5549). CatBoost-minus-full-logistic differences were -0.0054 for AUC (bootstrap 95%CI: -0.0438 to 0.0307) and 0.0035 for Brier score (95%CI: -0.0121 to 0.0192). Thus, changing the algorithm while retaining all 12 predictors did not demonstrate an incremental second-institution advantage.

The decision curves crossed across the examined range, and neither model showed a consistent descriptive net-benefit advantage (Figure 4C). Because no surveillance threshold was clinically established for this use case, these findings are exploratory. Figure 4D defines the plotted effects as CatBoost minus logistic for AUC and logistic minus CatBoost for Brier score, so positive values consistently favor CatBoost.

In the post hoc treatment-variable sensitivity analysis, CatBoost excluding radiotherapy and chemotherapy achieved an AUC of 0.7984 (bootstrap 95%CI: 0.7291-0.8595) and Brier score of 0.1640 (95%CI: 0.1358-0.1940). Relative to the locked 12-predictor CatBoost model, the AUC difference was -0.0067 (95%CI: -0.0193 to 0.0053) and the sensitivity-minus-locked Brier difference was 0.0026 (95%CI: -0.0019 to 0.0072). Thus, removing the two treatment-status variables did not materially change the conclusion, although this post hoc analysis does not resolve confounding by indication or establish a leakage-free prediction landmark (Supplementary Table 10).

DISCUSSION

The frozen CatBoost model retained an AUC of 0.8050 in the second institution. Its calibration-slope point estimate was close to 1, but confidence intervals for slope and intercept were wide; the data were compatible with approximate calibration but did not establish precise or reproducible calibration. CatBoost did not demonstrate superior performance over either the six-variable benchmark or the full 12-variable logistic model. Model selection, performance in another institution, and incremental clinical value therefore remained distinct questions.

Repeated nested validation separated tuning from held-out evaluation. Each patient contributed held-out predictions in five outer cycles, while preprocessing, hyperparameter tuning, selection of ensemble components, and ensemble fitting remained within the corresponding training data. This design reduces selection-induced optimism relative to a single split[20]. The small SDs across repeated cycles should nevertheless not be read as independent-sample uncertainty because the same patients were reused. CatBoost was selected by the lowest mean Brier score in development data alone.

The protocol-frozen second-institution comparison placed the internal improvement in context. Point estimates for CatBoost-minus-six-variable-logistic AUC and Brier score changed direction, but their confidence intervals included no difference. The full 12-variable logistic comparison led to the same substantive conclusion, indicating that the finding was not explained only by the smaller benchmark predictor set. These results agree with empirical comparisons in which machine learning has not consistently outperformed logistic regression for clinical prediction. The study’s principal value is therefore the controlled assessment of whether additional algorithmic complexity produced incremental performance, not a claim of machine-learning superiority. Similar point estimates do not establish equivalence or non-inferiority.

SHAP described how the fitted CatBoost model distributed prediction across its inputs. These attributions cannot separate causal effects from correlation, treatment selection, measurement patterns, or unmeasured disease severity. Radiotherapy and chemotherapy are preferentially administered to selected patients, often those with aggressive biology, unfavorable anatomy, higher stage, or other high-risk features; consequently, any model attribution for these variables is vulnerable to confounding by indication. Clinicians must not infer from their SHAP values that either treatment promotes recurrence, prevents recurrence, or lacks efficacy. The post hoc model excluding both treatment fields produced similar overall performance but does not convert observational attribution into causal evidence.

Performance measures answered different questions. AUC measures ranking; Brier score and log loss quantify prediction error; and calibration intercept, slope, observed-to-expected ratio, and the flexible curve address different aspects of probability agreement. Decision curves crossed, and no surveillance action or threshold had been prespecified clinically. A threshold could be clinically meaningful only after it is explicitly linked to an action, such as multidisciplinary review, shorter follow-up intervals, or additional site-directed imaging, and after the relative consequences of false-positive and false-negative escalation are agreed. The present data do not justify a universal cutoff or a specific surveillance schedule[3,46]; the numerical net benefits in Supplementary Table 11 are provided to support future protocol design, not current care.

The browser interface facilitates reproducibility and independent evaluation of the locked model. A model hash, feature schema, fixed test set, and patient-level concordance test provide checks against silent model drift. Public accessibility does not confer clinical validity. The interface has not undergone prospective impact, usability, security, or regulatory evaluation and must not be used to determine surveillance or treatment[21,22,43,44,47-49].

Methodological strengths include a prespecified search space, identical resampling splits across 246 strategies, complete nesting of preprocessing and tuning, development-only selection, model locking, fixed-prediction bootstrap evaluation, and paired comparison with both parsimonious and same-predictor-set logistic models. Discrimination, Brier score, log loss, calibration parameters, O/E ratio, and exploratory decision curves are reported. Fold assignments, held-out predictions, software versions, data hashes, coefficients, test vectors, and the locked model were retained.

Several limitations constrain interpretation. Both cohorts were retrospective and drawn from high-volume tertiary sarcoma centers. The same second-institution cohort had been examined in earlier project work, including evaluation of a prior gradient-boosting model, so investigators had prior knowledge of its case mix and performance; protocol freezing prevents post-lock changes but cannot erase bias from historical exposure. Complete-case and ascertainable-status restrictions may select patients with better documentation, longer follow-up, or different recurrence risk. Reason-specific exclusion counts, excluded-patient covariates, the second-institution screening denominator, visit-level surveillance, and time-to-event data were not retained in the locked package, precluding excluded-vs-included comparisons, multiple-imputation selection analyses, and competing-risk analyses. The binary endpoint therefore does not estimate cumulative incidence in all postoperative patients. Treatment dates were not preserved, so radiotherapy and chemotherapy may encode post-landmark care and are vulnerable to both timing leakage and confounding by indication. The 67 second-institution events produced substantial uncertainty around calibration and model differences. Histological subtype, imaging, detailed pathology, and molecular features were not modeled; subgroup and fairness performance were not established; and input errors or category shifts may degrade interface predictions.

Future work should prioritize prospective multicenter evaluation with time-to-event outcomes, explicit handling of censoring, missingness and competing risks, prespecified recalibration rules, subgroup assessment, and decision-impact analysis. Clinical use would require a clearly defined user, decision, prediction time, and action threshold. The relevant question is whether a model improves decisions or outcomes beyond a simpler tool, not whether a larger algorithm search raises internal AUC.

CONCLUSION

In a protocol-frozen second-institution re-evaluation among patients with ascertainable binary 3-year recurrence status, the locked CatBoost model retained discrimination but did not demonstrate improvement over either the prespecified six-variable benchmark or a full 12-variable logistic model. This study was not designed or powered to establish equivalence or non-inferiority. Prospective multicenter evaluation with a verified prediction landmark, time-to-event outcomes, clinically justified action thresholds, and formal decision-impact assessment is required before any model is used to guide care.

References
1.  Gamboa AC, Gronchi A, Cardona K. Soft-tissue sarcoma in adults: An update on the current state of histiotype-specific management in an era of personalized medicine. CA Cancer J Clin. 2020;70:200-229.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 172]  [Cited by in RCA: 396]  [Article Influence: 66.0]  [Reference Citation Analysis (2)]
2.  von Mehren M, Kane JM, Agulnik M, Bui MM, Carr-Ascher J, Choy E, Connelly M, Dry S, Ganjoo KN, Gonzalez RJ, Holder A, Homsi J, Keedy V, Kelly CM, Kim E, Liebner D, McCarter M, McGarry SV, Mesko NW, Meyer C, Pappo AS, Parkes AM, Petersen IA, Pollack SM, Poppe M, Riedel RF, Schuetze S, Shabason J, Sicklick JK, Spraker MB, Zimel M, Hang LE, Sundar H, Bergman MA. Soft Tissue Sarcoma, Version 2.2022, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw. 2022;20:815-833.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 401]  [Cited by in RCA: 302]  [Article Influence: 75.5]  [Reference Citation Analysis (2)]
3.  Gronchi A, Miah AB, Dei Tos AP, Abecassis N, Bajpai J, Bauer S, Biagini R, Bielack S, Blay JY, Bolle S, Bonvalot S, Boukovinas I, Bovee JVMG, Boye K, Brennan B, Brodowicz T, Buonadonna A, De Álava E, Del Muro XG, Dufresne A, Eriksson M, Fagioli F, Fedenko A, Ferraresi V, Ferrari A, Frezza AM, Gasperoni S, Gelderblom H, Gouin F, Grignani G, Haas R, Hassan AB, Hecker-Nolting S, Hindi N, Hohenberger P, Joensuu H, Jones RL, Jungels C, Jutte P, Kager L, Kasper B, Kawai A, Kopeckova K, Krákorová DA, Le Cesne A, Le Grange F, Legius E, Leithner A, Lopez-Pousa A, Martin-Broto J, Merimsky O, Messiou C, Mir O, Montemurro M, Morland B, Morosi C, Palmerini E, Pantaleo MA, Piana R, Piperno-Neumann S, Reichardt P, Rutkowski P, Safwat AA, Sangalli C, Sbaraglia M, Scheipl S, Schöffski P, Sleijfer S, Strauss D, Strauss S, Sundby Hall K, Trama A, Unk M, van de Sande MAJ, van der Graaf WTA, van Houdt WJ, Frebourg T, Casali PG, Stacchiotti S; ESMO Guidelines Committee, EURACAN and GENTURIS. Soft tissue and visceral sarcomas: ESMO-EURACAN-GENTURIS Clinical Practice Guidelines for diagnosis, treatment and follow-up(☆). Ann Oncol. 2021;32:1348-1365.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 855]  [Cited by in RCA: 820]  [Article Influence: 164.0]  [Reference Citation Analysis (3)]
4.  Issels RD, Lindner LH, Verweij J, Wessalowski R, Reichardt P, Wust P, Ghadjar P, Hohenberger P, Angele M, Salat C, Vujaskovic Z, Daugaard S, Mella O, Mansmann U, Dürr HR, Knösel T, Abdel-Rahman S, Schmidt M, Hiddemann W, Jauch KW, Belka C, Gronchi A; European Organization for the Research and Treatment of Cancer-Soft Tissue and Bone Sarcoma Group and the European Society for Hyperthermic Oncology. Effect of Neoadjuvant Chemotherapy Plus Regional Hyperthermia on Long-term Outcomes Among Patients With Localized High-Risk Soft Tissue Sarcoma: The EORTC 62961-ESHO 95 Randomized Clinical Trial. JAMA Oncol. 2018;4:483-492.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 301]  [Cited by in RCA: 243]  [Article Influence: 30.4]  [Reference Citation Analysis (0)]
5.  Liang Y, Guo T, Hong D, Xiao W, Zhou Z, Zhang X. Time to Local Recurrence as a Predictor of Survival in Patients With Soft Tissue Sarcoma of the Extremity and Abdominothoracic Wall. Front Oncol. 2020;10:599097.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 2]  [Cited by in RCA: 12]  [Article Influence: 2.0]  [Reference Citation Analysis (0)]
6.  Rothermundt C, Whelan JS, Dileo P, Strauss SJ, Coleman J, Briggs TW, Haile SR, Seddon BM. What is the role of routine follow-up for localised limb soft tissue sarcomas? A retrospective analysis of 174 patients. Br J Cancer. 2014;110:2420-2426.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 57]  [Cited by in RCA: 81]  [Article Influence: 6.8]  [Reference Citation Analysis (0)]
7.  Cahlon O, Brennan MF, Jia X, Qin LX, Singer S, Alektiar KM. A postoperative nomogram for local recurrence risk in extremity soft tissue sarcomas after limb-sparing surgery without adjuvant radiation. Ann Surg. 2012;255:343-347.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 107]  [Cited by in RCA: 139]  [Article Influence: 9.9]  [Reference Citation Analysis (1)]
8.  van Praag VM, Rueten-Budde AJ, Jeys LM, Laitinen MK, Pollock R, Aston W, van der Hage JA, Dijkstra PDS, Ferguson PC, Griffin AM, Willeumier JJ, Wunder JS, van de Sande MAJ, Fiocco M. A prediction model for treatment decisions in high-grade extremity soft-tissue sarcomas: Personalised sarcoma care (PERSARC). Eur J Cancer. 2017;83:313-323.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 44]  [Cited by in RCA: 97]  [Article Influence: 10.8]  [Reference Citation Analysis (0)]
9.  De Sanctis R, Zelic R, Santoro A. Nomograms predicting local and distant recurrence and disease-specific mortality for R0/R1 soft tissue sarcomas of the extremities. Front Oncol. 2022;12:941896.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1]  [Cited by in RCA: 3]  [Article Influence: 0.8]  [Reference Citation Analysis (0)]
10.  Callegaro D, Miceli R, Bonvalot S, Ferguson P, Strauss DC, Levy A, Griffin A, Hayes AJ, Stacchiotti S, Pechoux CL, Smith MJ, Fiore M, Dei Tos AP, Smith HG, Mariani L, Wunder JS, Pollock RE, Casali PG, Gronchi A. Development and external validation of two nomograms to predict overall survival and occurrence of distant metastases in adults after surgical resection of localised soft-tissue sarcomas of the extremities: a retrospective analysis. Lancet Oncol. 2016;17:671-680.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 211]  [Cited by in RCA: 378]  [Article Influence: 37.8]  [Reference Citation Analysis (4)]
11.  Callegaro D, Miceli R, Bonvalot S, Ferguson PC, Strauss DC, van Praag VVM, Levy A, Griffin AM, Hayes AJ, Stacchiotti S, Pèchoux CL, Smith MJ, Fiore M, Tos APD, Smith HG, Catton C, Szkandera J, Leithner A, van de Sande MAJ, Casali PG, Wunder JS, Gronchi A. Development and external validation of a dynamic prognostic nomogram for primary extremity soft tissue sarcoma survivors. EClinicalMedicine. 2019;17:100215.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 30]  [Cited by in RCA: 59]  [Article Influence: 8.4]  [Reference Citation Analysis (0)]
12.  Kobbeltvedt MR, Lobmaier I, Spreafico M, Callegaro D, Miceli R, Kizilaslan F, Swanson D, Hompland I, Pasquali S, Fiocco M, van de Sande MAJ, Gronchi A, Boye K. A comparison of the risk prediction models PERSARC and Sarculator in patients with localized soft tissue sarcoma of the extremities and trunk wall. ESMO Open. 2025;10:105517.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in RCA: 6]  [Reference Citation Analysis (0)]
13.  Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44-56.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 6739]  [Cited by in RCA: 4727]  [Article Influence: 675.3]  [Reference Citation Analysis (16)]
14.  Crombé A, Spinnato P, Italiano A, Brisse HJ, Feydy A, Fadli D, Kind M. Radiomics and artificial intelligence for soft-tissue sarcomas: Current status and perspectives. Diagn Interv Imaging. 2023;104:567-583.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 28]  [Reference Citation Analysis (0)]
15.  Zhang Z, Zhao Y, Canes A, Steinberg D, Lyashevska O; written on behalf of AME Big-Data Clinical Trial Collaborative Group. Predictive analytics with gradient boosting in clinical medicine. Ann Transl Med. 2019;7:152.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 56]  [Cited by in RCA: 184]  [Article Influence: 26.3]  [Reference Citation Analysis (2)]
16.  Ji GW, Jiao CY, Xu ZG, Li XC, Wang K, Wang XH. Development and validation of a gradient boosting machine to predict prognosis after liver resection for intrahepatic cholangiocarcinoma. BMC Cancer. 2022;22:258.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 32]  [Cited by in RCA: 31]  [Article Influence: 7.8]  [Reference Citation Analysis (0)]
17.  Barreñada L, Dhiman P, Timmerman D, Boulesteix AL, Van Calster B. Understanding overfitting in random forest for probability estimation: a visualization and simulation study. Diagn Progn Res. 2024;8:14.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 13]  [Cited by in RCA: 45]  [Article Influence: 22.5]  [Reference Citation Analysis (0)]
18.  Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019;110:12-22.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1585]  [Cited by in RCA: 1352]  [Article Influence: 193.1]  [Reference Citation Analysis (7)]
19.  Iwagami M, Inokuchi R, Kawakami E, Yamada T, Goto A, Kuno T, Hashimoto Y, Michihata N, Goto T, Shinozaki T, Sun Y, Taniguchi Y, Komiyama J, Uda K, Abe T, Tamiya N. Comparison of machine-learning and logistic regression models for prediction of 30-day unplanned readmission in electronic health records: A development and validation study. PLOS Digit Health. 2024;3:e0000578.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 15]  [Reference Citation Analysis (0)]
20.  Varma S, Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics. 2006;7:91.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 869]  [Cited by in RCA: 1093]  [Article Influence: 54.7]  [Reference Citation Analysis (0)]
21.  Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, Ghassemi M, Liu X, Reitsma JB, van Smeden M, Boulesteix AL, Camaradou JC, Celi LA, Denaxas S, Denniston AK, Glocker B, Golub RM, Harvey H, Heinze G, Hoffman MM, Kengne AP, Lam E, Lee N, Loder EW, Maier-Hein L, Mateen BA, McCradden MD, Oakden-Rayner L, Ordish J, Parnell R, Rose S, Singh K, Wynants L, Logullo P. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1587]  [Cited by in RCA: 2634]  [Article Influence: 1317.0]  [Reference Citation Analysis (17)]
22.  Wolff RF, Moons KGM, Riley RD, Whiting PF, Westwood M, Collins GS, Reitsma JB, Kleijnen J, Mallett S; PROBAST Group†. PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Ann Intern Med. 2019;170:51-58.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 2040]  [Cited by in RCA: 2020]  [Article Influence: 288.6]  [Reference Citation Analysis (9)]
23.  Weskamp P, Ufton D, Drysch M, Wagner JM, Dadras M, Lehnhardt M, Behr B, Wallner C. Risk Factors for Occurrence and Relapse of Soft Tissue Sarcoma. Cancers (Basel). 2022;14:1273.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 3]  [Cited by in RCA: 19]  [Article Influence: 4.8]  [Reference Citation Analysis (0)]
24.  Sambri A, Caldari E, Montanari A, Fiore M, Cevolani L, Ponti F, D'Agostino V, Bianchi G, Miceli M, Spinnato P, De Paolis M, Donati DM. Vascular Proximity Increases the Risk of Local Recurrence in Soft-Tissue Sarcomas of the Thigh-A Retrospective MRI Study. Cancers (Basel). 2021;13:6325.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 5]  [Cited by in RCA: 11]  [Article Influence: 2.2]  [Reference Citation Analysis (0)]
25.  Lin JS, Coleman L, Voskuil RT, Malik A, Mayerson JL, Scharschmidt TJ. Local recurrence rates of superficial versus deep soft tissue sarcoma. Arch Orthop Trauma Surg. 2024;144:2967-2973.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 2]  [Reference Citation Analysis (0)]
26.  Gundle KR, Kafchinski L, Gupta S, Griffin AM, Dickson BC, Chung PW, Catton CN, O'Sullivan B, Wunder JS, Ferguson PC. Analysis of Margin Classification Systems for Assessing the Risk of Local Recurrence After Soft Tissue Sarcoma Resection. J Clin Oncol. 2018;36:704-709.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 189]  [Cited by in RCA: 170]  [Article Influence: 21.3]  [Reference Citation Analysis (0)]
27.  Bilgeri A, Klein A, Lindner LH, Nachbichler S, Knösel T, Birkenmaier C, Jansson V, Baur-Melnyk A, Dürr HR. The Effect of Resection Margin on Local Recurrence and Survival in High Grade Soft Tissue Sarcoma of the Extremities: How Far Is Far Enough? Cancers (Basel). 2020;12:2560.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 18]  [Cited by in RCA: 40]  [Article Influence: 6.7]  [Reference Citation Analysis (0)]
28.  Wittenberg S, Paraskevaidis M, Jarosch A, Flörcken A, Brandes F, Striefler J, Kaul D, Roohani S, Khakzad T, Märdian S, Rau D. Surgical Margins in Soft Tissue Sarcoma Management and Corresponding Local and Systemic Recurrence Rates: A Retrospective Study Covering 11 Years and 169 Patients in a Single Institution. Life (Basel). 2022;12:1694.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 20]  [Cited by in RCA: 11]  [Article Influence: 2.8]  [Reference Citation Analysis (0)]
29.  Yurtbay A, Aydın Şimşek Ş, Cengiz T, Bariş YS, Say F, Dabak N. The Impact of Surgical Margin Distance on Local Recurrence and Survival in Patients with Soft Tissue Sarcoma. Medicina (Kaunas). 2025;61:289.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1]  [Cited by in RCA: 6]  [Article Influence: 6.0]  [Reference Citation Analysis (1)]
30.  Larios F, Gonzalez MR, Ruiz-Arellanos K, Aquilino E Silva G, Pretell-Mazzini J. Is Unplanned Excision of Soft Tissue Sarcomas Associated with Worse Oncological Outcomes?-A Systematic Review and Meta-Analysis. Cancers (Basel). 2024;16:443.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 24]  [Cited by in RCA: 20]  [Article Influence: 10.0]  [Reference Citation Analysis (0)]
31.  Sacchetti F, Alsina AC, Morganti R, Innocenti M, Andreani L, Muratori F, Scoccianti G, Totti F, Campanacci DA, Capanna R. Re-excision after unplanned excision of soft tissue sarcoma: A systematic review and metanalysis. The rationale of systematic re-excision. J Orthop. 2021;25:244-251.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 33]  [Cited by in RCA: 30]  [Article Influence: 6.0]  [Reference Citation Analysis (0)]
32.  Scoccianti G, Innocenti M, Frenos F, Muratori F, Sacchetti F, Beltrami G, Capanna R, Campanacci DA. Re-excision after unplanned excision of soft tissue sarcomas: Long-term results. Surg Oncol. 2020;34:212-217.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 5]  [Cited by in RCA: 14]  [Article Influence: 2.3]  [Reference Citation Analysis (0)]
33.  Fromm J, Klein A, Mentrup F, Lindner LH, Nachbichler S, Holzapfel BM, Goller SS, Knösel T, Dürr HR. Unplanned Resections of Soft Tissue Sarcomas-Necessity of Re-Resection? Cancers (Basel). 2024;16:1851.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in RCA: 6]  [Reference Citation Analysis (0)]
34.  Prokhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A. CatBoost: Unbiased boosting with categorical features. Adv Neural Inf Process Syst. 2018;31:6639-6649.  [PubMed]  [DOI]  [Full Text]
35.  Zhang Z. Introduction to machine learning: k-nearest neighbors. Ann Transl Med. 2016;4:218.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 798]  [Cited by in RCA: 389]  [Article Influence: 38.9]  [Reference Citation Analysis (0)]
36.  Chen TQ, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ‘16). Assoc Comput Mach. 2016;785-794.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 12755]  [Cited by in RCA: 11091]  [Article Influence: 1109.1]  [Reference Citation Analysis (4)]
37.  Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu TY. LightGBM: A highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst. 2017;30:3146-3154.  [PubMed]  [DOI]
38.  Alba AC, Agoritsas T, Walsh M, Hanna S, Iorio A, Devereaux PJ, McGinn T, Guyatt G. Discrimination and Calibration of Clinical Prediction Models: Users' Guides to the Medical Literature. JAMA. 2017;318:1377-1384.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 1323]  [Cited by in RCA: 1272]  [Article Influence: 141.3]  [Reference Citation Analysis (7)]
39.  Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, Pencina MJ, Kattan MW. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology. 2010;21:128-138.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 4068]  [Cited by in RCA: 3896]  [Article Influence: 243.5]  [Reference Citation Analysis (9)]
40.  Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26:565-574.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 4224]  [Cited by in RCA: 4577]  [Article Influence: 228.9]  [Reference Citation Analysis (12)]
41.  Vickers AJ, van Calster B, Steyerberg EW. A simple, step-by-step guide to interpreting decision curve analysis. Diagn Progn Res. 2019;3:18.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 901]  [Cited by in RCA: 893]  [Article Influence: 127.6]  [Reference Citation Analysis (1)]
42.  Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17:e70056.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 675]  [Cited by in RCA: 420]  [Article Influence: 210.0]  [Reference Citation Analysis (2)]
43.  Loh HW, Ooi CP, Seoni S, Barua PD, Molinari F, Acharya UR. Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011-2022). Comput Methods Programs Biomed. 2022;226:107161.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 55]  [Cited by in RCA: 275]  [Article Influence: 68.8]  [Reference Citation Analysis (3)]
44.  Nohara Y, Matsumoto K, Soejima H, Nakashima N. Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Comput Methods Programs Biomed. 2022;214:106584.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 37]  [Cited by in RCA: 459]  [Article Influence: 114.8]  [Reference Citation Analysis (3)]
45.  von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370:1453-1457.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Cited by in Crossref: 15623]  [Cited by in RCA: 16097]  [Article Influence: 847.2]  [Reference Citation Analysis (12)]
46.  Rutkowski P, Lugowska I. Follow-up in soft tissue sarcomas. Memo. 2014;7:92-96.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 13]  [Cited by in RCA: 21]  [Article Influence: 1.8]  [Reference Citation Analysis (0)]
47.  Amann J, Blasimme A, Vayena E, Frey D, Madai VI; Precise4Q consortium. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Inform Decis Mak. 2020;20:310.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 1509]  [Cited by in RCA: 895]  [Article Influence: 149.2]  [Reference Citation Analysis (7)]
48.  Collins GS, Dhiman P, Ma J, Schlussel MM, Archer L, Van Calster B, Harrell FE Jr, Martin GP, Moons KGM, van Smeden M, Sperrin M, Bullock GS, Riley RD. Evaluation of clinical prediction models (part 1): from development to external validation. BMJ. 2024;384:e074819.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 435]  [Cited by in RCA: 510]  [Article Influence: 255.0]  [Reference Citation Analysis (11)]
49.  Efthimiou O, Seo M, Chalkou K, Debray T, Egger M, Salanti G. Developing clinical prediction models: a step-by-step guide. BMJ. 2024;386:e078276.  [RCA]  [PubMed]  [DOI]  [Full Text]  [Full Text (PDF)]  [Cited by in Crossref: 265]  [Cited by in RCA: 355]  [Article Influence: 177.5]  [Reference Citation Analysis (2)]
Footnotes

Peer review: Externally peer reviewed.

Peer-review model: Single blind

Specialty type: Oncology

Country of origin: China

Peer-review report’s classification

Scientific quality: Grade A, Grade B, Grade B

Novelty: Grade A, Grade B, Grade B

Creativity or innovation: Grade B, Grade B, Grade B

Scientific significance: Grade A, Grade B, Grade B

P-Reviewer: Guo SB, MD, PhD, China; Primadhi RA, Associate Professor, MD, PhD, Indonesia; Wang XD, MD, PhD, China S-Editor: Yang CY L-Editor: A P-Editor: Wang WB

Write to the Help Desk