BPG is committed to discovery and dissemination of knowledge
Retrospective Cohort Study
Copyright: ©Author(s) 2026.
World J Clin Oncol. Nov 24, 2026; 17(11): 126582
Published online Nov 24, 2026. doi: 10.5306/wjco.126582
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)
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
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)


Write to the Help Desk