Abstract
Objective
Accurate prediction of pathological complete response (pCR) following neoadjuvant therapy (NAT) is critical for optimizing treatment in breast cancer. This study develops and validates an interpretable, cost-effective machine learning (ML) model integrating computed tomography (CT)-based body composition parameters with routine inflammatory and nutritional biomarkers to predict pCR.
Methods
In this retrospective single-center study (n = 189; January 2019–June 2023), patients were divided into training (n = 142) and independent temporal test (n = 47) sets. CT-based body composition parameters and blood test variables were analyzed. Independent predictors were identified via Least Absolute Shrinkage and Selection Operator and multivariate logistic regression. Eight ML algorithms were compared, and the optimal model was selected based on Area Under the Curve (AUC), calibration, and clinical utility. SHapley Additive exPlanations (SHAP) analysis visualized predictive contributions.
Results
Six independent predictors were identified: visceral adipose tissue density, skeletal muscle density, intramuscular adipose tissue content, albumin-to-alkaline phosphatase ratio, systemic inflammation response index, and molecular subtype. The eXtreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving an area under the curve (AUC) of 0.888 (95% CI [0.837–0.939]) in internal validation and 0.831 (95% CI [0.723–0.938]) in the independent test set. The model exhibited good calibration (Brier score = 0.180). SHAP analysis highlighted the contribution of host-related factors alongside tumor biology.
Conclusions
This interpretable ML model effectively integrates host-related body composition and inflammatory-nutritional markers to predict pCR. By utilizing routinely available data, this approach offers a practical, accessible tool for initial risk stratification, complementing existing imaging-based strategies and supporting personalized clinical decision-making.
Keywords: Breast cancer, Neoadjuvant therapy, Machine learning, Body composition, Pathological complete response
Introduction
Breast cancer remains one of the most prevalent malignancies worldwide and a leading cause of cancer-related mortality among women (Bray et al., 2024). With the continuous evolution of cancer treatment strategies, modern oncology has increasingly shifted toward personalized and risk-adapted therapeutic approaches (Sonkin, Thomas & Teicher, 2024). Neoadjuvant therapy (NAT) has become an integral component of breast cancer management, particularly for patients with stage II–III disease and selected high-risk early-stage subtypes, as it enables tumor downstaging, increases breast-conserving surgery rates, and provides an in vivo assessment of treatment response (Korde et al., 2021). Pathological complete response (pCR) after NAT is widely recognized as a clinically meaningful endpoint that correlates with favorable prognosis in specific breast cancer subtypes (Cortazar et al., 2014; Hirmas, Holtschmidt & Loibl, 2024). However, pCR rates vary across molecular subtypes and treatment regimens, and a substantial proportion of patients still fail to achieve pCR (Korde et al., 2021; Schmid et al., 2022; Romeo et al., 2021). Therefore, accurately predicting pCR before initiation of NAT is of substantial clinical importance, as it may guide treatment selection, optimize therapeutic intensity, and improve patient outcomes (Korde et al., 2021).
Recent research on pCR prediction has largely focused on imaging-based approaches, particularly dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), radiomics, and deep learning models derived from primary tumors, peritumoral tissues, and axillary lymph nodes (Sharafeldeen et al., 2025; Peng et al., 2022; Gamal et al., 2024; D’Anna et al., 2025; Iwase et al., 2021). Although these methods have shown promising predictive performance, they often rely on complex imaging pipelines (e.g., standardized acquisition, segmentation, and feature extraction), which may limit clinical accessibility and generalizability (Traverso et al., 2018; Park et al., 2019; Zwanenburg et al., 2020).
Importantly, treatment response in breast cancer is influenced not only by tumor-intrinsic biology but also by host-related factors such as systemic inflammation, immune status, and nutritional condition (Feeney et al., 2024; Yamamoto, Kawada & Obama, 2021). In parallel, body composition—particularly skeletal muscle and adipose tissue characteristics—has emerged as a clinically relevant marker reflecting metabolic reserve and treatment resilience (Iwase et al., 2021). Computed tomography (CT), routinely performed for staging in many patients, provides an objective and reproducible means of assessing body composition without additional cost or patient burden (Pickhardt, Summers & Garrett, 2021; Zhou et al., 2025a; Zhou et al., 2025b). Machine learning (ML) offers a powerful framework for integrating multidimensional clinical data and capturing complex, non-linear relationships among predictors (Deo, 2015). However, for real-world clinical adoption, predictive models must not only demonstrate accuracy but also provide interpretability to support individualized decision-making.
Against this background, the present study aims to develop and validate an interpretable ML model that integrates CT-based body composition parameters with systemic inflammatory and nutritional indicators to predict pCR in breast cancer patients undergoing neoadjuvant therapy. By leveraging routinely available host-related features and explainable ML techniques, this work seeks to complement existing imaging-centered approaches and provide a clinically practical tool for personalized treatment planning.
Materials and Methods
General information
This retrospective, single-center study was approved by the Ethics Review Committee of Jiangxi Cancer Hospital (Approval No: 2023ky126), and informed consent was waived. Given its retrospective design, the cohort is subject to potential selection and information bias; to mitigate these risks, we consecutively enrolled eligible patients using prespecified criteria and applied a temporal split into training and independent test sets to approximate real-world generalization. Personally identifiable information was anonymized during data analysis. We consecutively collected data from breast cancer patients meeting the inclusion criteria at Jiangxi Cancer Hospital between January 2019 and June 2023. All enrolled patients completed NAT cycles in accordance with Chinese Society of Clinical Oncology Guidelines (Li et al., 2024). Inclusion criteria were: (a) diagnosis of breast cancer confirmed by core needle biopsy before NAT with pathological immunohistochemical results; (b) complete blood tests and qualifying chest CT scan were performed within one week before NAT; (c) patients who underwent surgery at our hospital after NAT with Miller-Payne grading in pathology. Exclusion criteria were: (a) poor quality CT images; (b) patients using medications affecting platelet, lymphocyte, albumin, or alkaline phosphatase counts prior to NAT; (c) patients with systemic infections, autoimmune diseases, or hematologic disorders before NAT; (d) male breast cancer, bilateral breast cancer, or concurrent other malignancies. Figure 1 shows the detailed patient selection process.
Figure 1. Flowchart of included BC patients; BC, breast cancer.
Pathological examination
Before NAT, all patients underwent core needle biopsy; according to the American Society of Clinical Oncology (ASCO) and the College of American Pathologists (CAP) testing guidelines (Allison et al., 2020), if at least 1% of tumor cell nuclei are stained, the tumor is considered hormone receptor (HR) positive (Estrogen Receptor (ER)-positive or Progesterone Receptor (PR)-positive), with both ER and PR negative defined as HR-negative. Based on ASCO/CAP HER2 testing guidelines (Wolff et al., 2023), tumors with Immunohistochemistry (IHC) score 3+ or those scored 2+ with HER2 gene amplification confirmed by fluorescence in situ hybridization are considered HER2 positive. Ki67 proliferation index >20% indicates high expression, ≤20% indicates low expression. Tumors are classified into four subtypes based on HR and HER2 status: HR+HER2-, HR+HER2+, HR-HER2+, and HR-HER2-. Response to NAT was evaluated using Miller-Payne grading (Ogston et al., 2003), with pCR defined as the absence of residual invasive tumor cells in the primary breast lesion and no cancer metastasis in axillary lymph nodes post-surgery (ypT0/Tis ypN0).
Clinical data collection and definition of variables
Clinical data include age, height, weight, the initial maximum tumor diameter (MTD), and body mass index (BMI), with BMI calculated as weight (kg) divided by height squared (m2). Pretreatment blood tests (within one week before NAT) provided white blood cell, neutrophil, monocyte, and lymphocyte counts, as well as albumin, globulin, bilirubin, and alkaline phosphatase levels. From these routine measures, we derived composite inflammatory and nutritional indices—including the neutrophil-to-lymphocyte ratio (NLR), derived NLR (dNLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), systemic immune-inflammation index (SII), systemic inflammation response index (SIRI), albumin-to-alkaline phosphatase ratio (AAPR), albumin-to-globulin ratio (AGR), prognostic nutritional index (PNI), albumin lymphocyte index (ALI), lymphocyte–albumin product (LA), and lymphocyte–monocyte score (LMS)—using standard formulas (Table S1). These indices were selected because they are routinely available prior to NAT and capture complementary aspects of host immune–inflammatory status and nutritional reserve that are biologically relevant to treatment tolerance and response.
CT scanning and body composition measurement
Within one week before NAT, all patients underwent unenhanced chest CT from the thoracic inlet to the adrenal glands. We quantified body composition at the L1 vertebral level because this slice was consistently available on routine pretreatment chest CT, enabling assessment without additional imaging or radiation. Prior studies suggest that L1-based muscle and adipose measures from chest CT can serve as practical surrogates for overall body composition; however, because body composition and CT protocols vary across populations and centers, the generalizability of L1-derived metrics should be interpreted cautiously and requires external validation. Details of the CT scanners and scanning settings are provided in Table S2. The images were accessed through the Picture Archiving and Communication System and analyzed using RadiAnt Viewer software. As an alternative to abdominal CT, a single CT image at the L1 vertebral level is a feasible method for assessing whole-body tissue quality (Zou et al., 2025); experienced radiologists identified the L1 level starting from the first thoracic vertebra, saving the cross-sectional DICOM images. Selected images were analyzed with semi-automated software Slicomatic 5.0, as shown in Fig. 2. The Hounsfield unit (HU) ranges for tissue components were defined as: subcutaneous adipose tissue (SAT) −190 to −30 HU; skeletal muscle (SM) −29 to +150 HU; intermuscular adipose tissue (IMAT) and visceral adipose tissue (VAT) −150 to −30 HU (Zou et al., 2025; Zhou et al., 2025a; Zhou et al., 2025b). L1-skeletal muscle (L1-SM) includes abdominal wall muscles, intercostal muscles, diaphragm, lumbar muscles, and paraspinal muscles (Zou et al., 2025). These regions were standardized by their height squared (m2), deriving SAT index (SATI), SM index (SMI), and IMAT index (IMATI). The mean HU value within each region was recorded to characterize tissue density, including SM density (SMD), VAT density (VATD), SAT density (SATD), and IMAT density (IMATD). Intramuscular adipose tissue content (IMAC) was evaluated by dividing the attenuation value of IMAT by that of SM (Zhou et al., 2025a; Zhou et al., 2025b; Zou et al., 2025). Total adipose tissue (TAT) area was estimated by summing SAT and VAT areas, with the VAT to SAT area ratio (VSR) calculated subsequently. Two radiologists with 9 years of experience independently assessed the body composition parameters in a blinded manner, and the mean of their measurements was used as final data.
Figure 2. Body composition segmentation at the L1 level based on CT images.
(A) Subcutaneous adipose tissue (SAT, red); (B) Visceral adipose tissue (VAT, green); (C) Intermuscular adipose tissue (IMAT, blue); (D) Skeletal muscle (SM, purple).
Variable selection and model building
Feature selection was conducted in the training set using a three-step strategy to balance interpretability and overfitting control, given the modest sample size. First, candidate variables were screened using univariable logistic regression with a liberal threshold (P < 0.10) to avoid prematurely excluding potentially relevant predictors. Second, LASSO regression with 10-fold cross-validation (via cv.glmnet) was applied to shrink coefficients, address multicollinearity, and select a parsimonious feature set; this cross-validation strategy ensured stable penalty parameter selection. Third, multivariable logistic regression was performed to identify independent predictors of pCR and report clinically interpretable effect-size estimates.
The detailed model building process is as follows:
-
(a)
Data division: Patients from January 2019 to July 2022 (142 cases) form the training set; those from August 2022 to June 2023 (47 cases) constitute the independent test set.
-
(b)
Comparison of multiple ML models: Based on the identified independent predictors, eight mainstream machine learning algorithms were constructed, including Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Gaussian Naive Bayes (GNB), and LightGBM. Model development was implemented in Python using the scikit-learn (0.22.1), lightgbm (3.2.1), and xgboost (1.2.1) libraries. We employed five-fold cross-validation to evaluate the training set, randomly partitioning the data into five subsets (80% for training and 20% for validation in each fold). Final model performance was assessed on the independent test set. The optimal model was identified based on the area under the receiver operating characteristic (ROC) curve (AUC) and calibration performance.
-
(c)
Visualization of the optimal model: Shapley Additive Explanations (SHAP) was applied to visualize feature importance and rankings (Zhou et al., 2025a; Zhou et al., 2025b), as well as to explain individual predictions, aiding in understanding the model’s decisions and identifying potential biases.
Statistical analysis
Cutoff values for body-composition parameters and blood biomarkers were derived only in the training set using ROC analysis with pCR as the endpoint, applying the Youden index (maximizing sensitivity + specificity) (Table S3). These prespecified cutoffs were then applied unchanged to the independent test set to prevent data leakage. Categorical variables are presented as frequencies (percentages) and compared using the chi-square test. Correlations among selected variables were assessed with Kendall’s tau. The 95% confidence intervals (CI) for the AUC were calculated using distinct approaches based on the validation step. For the independent test set, the 95% CI were estimated using a parametric normal approximation based on the standard error. For the internal 5-fold cross-validation, the 95% CI were derived from the standard error of the mean (SEM) of the AUC scores calculated across the five validation folds. The DeLong test was utilized specifically for evaluating the statistical significance of differences between the models’ ROC curves. Model calibration was evaluated using calibration plots, and clinical utility was assessed using decision curve analysis (DCA). Comparative model performance was assessed using the DeLong test, integrated discrimination improvement (IDI), and net reclassification improvement (NRI). All analyses were performed in Python (v3.5.6) and R (v3.6.3), with two-sided P < 0.05 considered statistically significant.
Results
Patient population characteristics
We retrospectively collected data from 305 breast cancer patients receiving NAT, of whom 189 were included after screening; among them, 71 patients (37.57%) were aged ≤45 years, and 118 patients (62.43%) were older than 45 years; 95 patients (50.27%) had a BMI < 24 kg/m2, and 94 patients (49.73%) had a BMI ≥24 kg/m2; 163 patients (86.24%) had an initial tumor MTD ≤5 cm, while 26 patients (13.76%) had MTD > 5 cm, with 100 patients (52.91%) achieving pCR after NAT, and 89 patients (47.09%) did not achieve pCR. The demographic and clinical-pathological features of patients in the training and test groups were well balanced (P > 0.05), as shown in Table 1.
Table 1. Baseline characteristics of included BC patients.
| Variables | Dataset, No (%) | Univariate analysis | ||||
|---|---|---|---|---|---|---|
| Total sets (n = 189) | Training set (n = 142) | Test set (n = 47) | p-value* | Odds Ratio (95% CI) | p-value** | |
| MTD | 0.474 | |||||
| ≤ 5 cm | 163 (86.243) | 121 (85.211) | 42 (89.362) | |||
| >5 cm | 26 (13.757) | 21 (14.789) | 5 (10.638) | 0.626 (0.242, 1.618) | 0.333 | |
| Age | 0.356 | |||||
| ≤ 45 y | 71 (37.566) | 56 (39.437) | 15 (31.915) | |||
| >45 y | 118 (62.434) | 86 (60.563) | 32 (68.085) | 0.870 (0.443, 1.706) | 0.684 | |
| BMI | 0.834 | |||||
| <24 kg/m2 | 95 (50.265) | 72 (50.704) | 23 (48.936) | |||
| ≥ 24 kg/m2 | 94 (49.735) | 70 (49.296) | 24 (51.064) | 0.842 (0.436, 1.628) | 0.609 | |
| HR | 0.362 | |||||
| Negative | 75 (39.683) | 59 (41.549) | 16 (34.043) | |||
| Positive | 114 (60.317) | 83 (58.451) | 31 (65.957) | 0.354 (0.178, 0.707) | 0.003 | |
| Her2 | 0.836 | |||||
| Negative | 107 (56.614) | 81 (57.042) | 26 (55.319) | |||
| Positive | 82 (43.386) | 61 (42.958) | 21 (44.681) | 8.224 (3.842, 17.606) | <0.001 | |
| Ki67 | 0.869 | |||||
| Low | 50 (26.455) | 38 (26.761) | 12 (25.532) | |||
| High | 139 (73.545) | 104 (73.239) | 35 (74.468) | 2.977 (1.337, 6.628) | 0.008 | |
| Molecular subtype | 0.3 | |||||
| HR positive, HER2 negative | 68 (35.979) | 48 (33.803) | 20 (42.553) | |||
| HR positive, HER2 positive | 46 (24.339) | 35 (24.648) | 11 (23.404) | 12.779 (4.369, 37.376) | <0.001 | |
| HR negative, HER2 positive | 38 (20.106) | 27 (19.014) | 11 (23.404) | 25.771 (7.317, 90.775) | <0.001 | |
| HR negative, HER2 negative | 37 (19.577) | 32 (22.535) | 5 (10.638) | 5.168 (1.790, 14.924) | 0.002 | |
| Pathological type | 0.639 | |||||
| Invasive adenocarcinoma | 103 (54.497) | 76 (53.521) | 27 (57.447) | |||
| Non-specific invasive adenocarcinoma | 86 (45.503) | 66 (46.479) | 20 (42.553) | 0.744 (0.383, 1.442) | 0.381 | |
| SAT | 0.13 | |||||
| Low | 34 (17.989) | 29 (20.423) | 5 (10.638) | |||
| High | 155 (82.011) | 113 (79.577) | 42 (89.362) | 0.403 (0.172, 0.944) | 0.036 | |
| VAT | 0.967 | |||||
| Low | 92 (48.677) | 69 (48.592) | 23 (48.936) | |||
| High | 97 (51.323) | 73 (51.408) | 24 (51.064) | 0.507 (0.260, 0.989) | 0.046 | |
| IMAT | 0.584 | |||||
| Low | 139 (73.545) | 103 (72.535) | 36 (76.596) | |||
| High | 50 (26.455) | 39 (27.465) | 11 (23.404) | 1.853 (0.877, 3.912) | 0.106 | |
| SM | 0.376 | |||||
| Low | 98 (51.852) | 71 (50.000) | 27 (57.447) | |||
| High | 91 (48.148) | 71 (50.000) | 20 (42.553) | 0.505 (0.259, 0.984) | 0.045 | |
| SATD | 0.232 | |||||
| Low | 173 (91.534) | 128 (90.141) | 45 (95.745) | |||
| High | 16 (8.466) | 14 (9.859) | 2 (4.255) | 2.105 (0.668, 6.629) | 0.203 | |
| VATD | 0.056 | |||||
| Low | 141 (74.603) | 101 (71.127) | 40 (85.106) | |||
| High | 48 (25.397) | 41 (28.873) | 7 (14.894) | 2.822 (1.323, 6.022) | 0.007 | |
| IMATD | 0.903 | |||||
| Low | 106 (56.085) | 80 (56.338) | 26 (55.319) | |||
| High | 83 (43.915) | 62 (43.662) | 21 (44.681) | 0.581 (0.297, 1.138) | 0.113 | |
| SMD | 0.578 | |||||
| Low | 83 (43.915) | 64 (45.070) | 19 (40.426) | |||
| High | 106 (56.085) | 78 (54.930) | 28 (59.574) | 1.917 (0.979, 3.752) | 0.058 | |
| AAPR | 0.072 | |||||
| Low | 42 (22.222) | 36 (25.352) | 6 (12.766) | |||
| High | 147 (77.778) | 106 (74.648) | 41 (87.234) | 3.766 (1.616, 8.777) | 0.002 | |
| AGR | 0.629 | |||||
| Low | 115 (60.847) | 85 (59.859) | 30 (63.830) | |||
| High | 74 (39.153) | 57 (40.141) | 17 (36.170) | 0.533 (0.269, 1.055) | 0.071 | |
| NLR | 0.024 | |||||
| Low | 78 (41.270) | 52 (36.620) | 26 (55.319) | |||
| High | 111 (58.730) | 90 (63.380) | 21 (44.681) | 0.473 (0.236, 0.948) | 0.035 | |
| dNLR | 0.101 | |||||
| Low | 62 (32.804) | 42 (29.577) | 20 (42.553) | |||
| High | 127 (67.196) | 100 (70.423) | 27 (57.447) | 0.446 (0.213, 0.933) | 0.032 | |
| PLR | 0.625 | |||||
| Low | 148 (78.307) | 110 (77.465) | 38 (80.851) | |||
| High | 41 (21.693) | 32 (22.535) | 9 (19.149) | 2.153 (0.959, 4.834) | 0.063 | |
| PNI | 0.266 | |||||
| Low | 60 (31.746) | 42 (29.577) | 18 (38.298) | |||
| High | 129 (68.254) | 100 (70.423) | 29 (61.702) | 0.589 (0.285, 1.220) | 0.154 | |
| SII | 0.019 | |||||
| Low | 55 (29.101) | 35 (24.648) | 20 (42.553) | |||
| High | 134 (70.899) | 107 (75.352) | 27 (57.447) | 0.522 (0.240, 1.135) | 0.101 | |
| LA | 0.291 | |||||
| Low | 100 (52.910) | 72 (50.704) | 28 (59.574) | |||
| High | 89 (47.090) | 70 (49.296) | 19 (40.426) | 0.598 (0.308, 1.163) | 0.130 | |
| ALI | 0.559 | |||||
| Low | 122 (64.550) | 90 (63.380) | 32 (68.085) | |||
| High | 67 (35.450) | 52 (36.620) | 15 (31.915) | 0.793 (0.400, 1.574) | 0.508 | |
| LMR | 0.374 | |||||
| Low | 78 (41.270) | 56 (39.437) | 22 (46.809) | |||
| High | 111 (58.730) | 86 (60.563) | 25 (53.191) | 1.397 (0.709, 2.750) | 0.334 | |
| LMS | 0.633 | |||||
| Low | 111 (58.730) | 82 (57.746) | 29 (61.702) | |||
| High | 78 (41.270) | 60 (42.254) | 18 (38.298) | 1.031 (0.530, 2.009) | 0.927 | |
| SIRI | 0.474 | |||||
| Low | 96 (50.794) | 70 (49.296) | 26 (55.319) | |||
| High | 93 (49.206) | 72 (50.704) | 21 (44.681) | 0.424 (0.216, 0.832) | 0.013 | |
| SATI | 0.514 | |||||
| Low | 76 (40.212) | 59 (41.549) | 17 (36.170) | |||
| High | 113 (59.788) | 83 (58.451) | 30 (63.830) | 0.574 (0.293, 1.127) | 0.107 | |
| VATI | 0.901 | |||||
| Low | 91 (48.148) | 68 (47.887) | 23 (48.936) | |||
| High | 98 (51.852) | 74 (52.113) | 24 (51.064) | 0.538 (0.276, 1.048) | 0.069 | |
| IMATI | 0.366 | |||||
| Low | 135 (71.429) | 99 (69.718) | 36 (76.596) | |||
| High | 54 (28.571) | 43 (30.282) | 11 (23.404) | 1.809 (0.876, 3.733) | 0.109 | |
| SMI | 0.982 | |||||
| Low | 145 (76.720) | 109 (76.761) | 36 (76.596) | |||
| High | 44 (23.280) | 33 (23.239) | 11 (23.404) | 0.318 (0.135, 0.746) | 0.008 | |
| VSR | 0.204 | |||||
| Low | 118 (62.434) | 85 (59.859) | 33 (70.213) | |||
| High | 71 (37.566) | 57 (40.141) | 14 (29.787) | 0.417 (0.208, 0.833) | 0.013 | |
| TAT | 0.192 | |||||
| Low | 27 (14.286) | 23 (16.197) | 4 (8.511) | |||
| High | 162 (85.714) | 119 (83.803) | 43 (91.489) | 0.340 (0.130, 0.886) | 0.027 | |
| IMAC | 0.956 | |||||
| Low | 120 (63.492) | 90 (63.380) | 30 (63.830) | |||
| High | 69 (36.508) | 52 (36.620) | 17 (36.170) | 1.866 (0.935, 3.724) | 0.077 | |
| pCR | 0.703 | |||||
| No | 100 (52.910) | 74 (52.113) | 26 (55.319) | |||
| Yes | 89 (47.090) | 68 (47.887) | 21 (44.681) | |||
Notes.
- BC
- breast cancer
- MTD
- maximum tumor diameter
- BMI
- body mass index
- HR
- hormone receptor
- Her2
- huan epidermal growth factor receptor-2
- SAT
- subcutaneous adipose tissue
- VAT
- visceral adipose tissue
- IMAT
- intermuscular adipose tissue
- SM
- skeletal muscle
- SATD
- subcutaneous adipose tissue density
- VATD
- visceral adipose tissue density
- IMATD
- intermuscular adipose tissue density
- SMD
- skeletal muscle density
- AAPR
- albumin-to-alkaline phosphatase ratio
- AGR
- albumin-to-globulin ratio
- NLR
- neutrophil-to-lymphocyte ratio
- PLR
- platelet-to-lymphocyte ratio
- PNI
- prognostic nutritional index
- SII
- systemic immune-inflammation index
- LA
- lymphocyte-albumin
- ALI
- advanced lung cancer inflammation index
- LMR
- lymphocyte-to-monocyte ratio
- LMS
- lymphocyte-to-monocyte score
- SIRI
- system inflammation response index
- SATI
- subcutaneous adipose tissue index
- VATI
- visceral adipose tissue index
- IMATI
- intermuscular adipose tissue index
- SMI
- skeletal muscle index
- VSR
- visceral-to-subcutaneous fat area ratio
- IMAC
- intramuscular adipose content
- pCR
- pathological complete response
Differences in data distribution between different cohorts.
Univariate logistic regression analysis.
Variables associated with pCR in the training set
Univariate LR analysis was performed on 35 indicators in the training cohort (Table 1). Seventeen indicators with P-values < 0.1 were included in LASSO regression analysis, with 10-fold cross-validation. Based on the minimum mean squared error at lambda = 0.012, 14 features with non-zero coefficients were selected: HER2, Ki67, Molecular subtype, SM, VATD, SMD, AAPR, AGR, PLR, SIRI, VSR, IMAC (Figs. 3A–3B). These variables were then incorporated into multivariate LR analysis, revealing that the independent predictors of pCR were: Molecular subtype with HR negative, HER2 negative (OR = 13.95; 95% CI [3.32–74.96]), VATD (OR = 4.21; 95% CI [1.16–19.20]), SMD (OR = 37.12; 95% CI [4.6–932.78]), AAPR (OR = 3.77; 95% CI [1.04–15.12]), SIRI (OR = 0.31; 95% CI [0.06–0.91]), and IMAC (OR = 37.78; 95% CI [4.37–1,003.41]) (Table 2). The correlation heatmap based on Kendall’s tau (Fig. 3C) showed that correlations among these six variables were all below 0.5, indicating no significant inter-variable correlations.
Figure 3. Selection of influencing factors using the LASSO regression model.
(A) LASSO coefficient profiles of 17 predictive factors. (B) Determination of the optimal penalty parameter λ in the LASSO regression model. (C) Kendall’s correlation analysis among independent predictive factors.
Table 2. Multivariate Cox regression analysis of independent predictors of pCR in the training set of BC.
| Variables | Training set (n = 142) | p-value | ||
|---|---|---|---|---|
| Odds Ratio | 95% CI | |||
| Molecular subtype | ||||
| HR positive, HER2 negative | Reference | |||
| HR negative, HER2 negative | 13.951 | 3.32, 74.957 | 0.001 | |
| VATD | ||||
| Low | Reference | |||
| High | 4.271 | 1.16, 19.204 | 0.039 | |
| SMD | ||||
| Low | Reference | |||
| High | 37.124 | 4.604, 932.777 | 0.005 | |
| AAPR | ||||
| Low | Reference | |||
| High | 3.774 | 1.041, 15.116 | 0.049 | |
| SIRI | ||||
| Low | Reference | |||
| High | 0.305 | 0.094, 0.905 | 0.037 | |
| IMAC | ||||
| Low | Reference | |||
| High | 37.781 | 4.371, 1,003.041 | 0.006 | |
Notes.
- BC
- breast cancer
- pCR
- pathological complete response
- HR
- hormone receptor
- Her2
- huan epidermal growth factor receptor-2
- VATD
- visceral adipose tissue density
- SMD
- skeletal muscle density
- AAPR
- albumin-to-alkaline phosphatase ratio
- SIRI
- system inflammation response index
- IMAC
- intramuscular adipose content
Model development and evaluation
Table 3 summarizes the performance of the eight machine learning models in both the internal validation and independent test sets. The ROC curves for the internal validation and test cohorts are depicted in Figs. 4A and 4B, respectively. Notably, the XGBoost model achieved the highest AUC values in both the internal validation set (AUC = 0.888; 95% CI [0.837–0.939]) and the independent test set (AUC = 0.831; 95% CI [0.723–0.938]). Figures 4C and 4D illustrate the calibration curves and DCA for the independent test set. XGBoost demonstrated the best calibration accuracy with the lowest Brier score of 0.180. Furthermore, DCA demonstrated that XGBoost provided a higher net clinical benefit compared to other models across a specific range of threshold probabilities.
Table 3. Performance of multiple machine learning models in internal validation and test sets.
| Model | AUC (95% CI) | Accuracy | Sensitivity | Specificity | PPV | NPV | F1 score |
|---|---|---|---|---|---|---|---|
| Internal validation set | |||||||
| XGBoost | 0.888 (0.837–0.939) | 0.796 | 0.779 | 0.811 | 0.791 | 0.800 | 0.785 |
| LightGBM | 0.872 (0.809–0.935) | 0.782 | 0.779 | 0.784 | 0.768 | 0.795 | 0.774 |
| LogisticRegression | 0.824 (0.789–0.859) | 0.754 | 0.721 | 0.784 | 0.754 | 0.753 | 0.737 |
| RandomForest | 0.870 (0.819–0.921) | 0.768 | 0.765 | 0.770 | 0.754 | 0.781 | 0.759 |
| DecisionTree | 0.840 (0.773–0.908) | 0.761 | 0.838 | 0.689 | 0.713 | 0.823 | 0.770 |
| SVM | 0.807 (0.783–0.831) | 0.711 | 0.677 | 0.743 | 0.708 | 0.714 | 0.692 |
| KNN | 0.762 (0.740–0.783) | 0.711 | 0.662 | 0.757 | 0.714 | 0.709 | 0.687 |
| GaussianNB | 0.800 (0.765–0.834) | 0.704 | 0.721 | 0.689 | 0.681 | 0.729 | 0.700 |
| Test set | |||||||
| XGBoost | 0.831 (0.723–0.938) | 0.681 | 0.714 | 0.654 | 0.625 | 0.739 | 0.667 |
| LightGBM | 0.795 (0.679–0.910) | 0.702 | 0.714 | 0.692 | 0.652 | 0.750 | 0.682 |
| LogisticRegression | 0.721 (0.592–0.849) | 0.681 | 0.714 | 0.654 | 0.625 | 0.739 | 0.667 |
| RandomForest | 0.779 (0.661–0.898) | 0.681 | 0.714 | 0.654 | 0.625 | 0.739 | 0.667 |
| DecisionTree | 0.698 (0.567–0.829) | 0.638 | 0.619 | 0.654 | 0.591 | 0.680 | 0.605 |
| SVM | 0.783 (0.665–0.901) | 0.723 | 0.714 | 0.731 | 0.682 | 0.760 | 0.698 |
| KNN | 0.683 (0.550–0.816) | 0.638 | 0.667 | 0.615 | 0.583 | 0.696 | 0.622 |
| GaussianNB | 0.701 (0.570–0.832) | 0.660 | 0.667 | 0.654 | 0.609 | 0.708 | 0.636 |
Notes.
- PPV
- positive predictive value
- NPV
- negative predictive value
Figure 4. Performance evaluation of machine learning models for predicting pCR.
(A) Receiver Operating Characteristic (ROC) curves of the eight machine learning models in the internal validation set. (B) ROC curves of the eight machine learning models in the independent test set. (C) Calibration curves of the models in the independent test set, with XGBoost demonstrating the best agreement between predicted and observed probabilities (lowest Brier score). (D) Decision Curve Analysis (DCA) in the independent test set, indicating that the XGBoost model provides superior net clinical benefit across a specific range of threshold probabilities compared to other models.
As shown in the DeLong Heatmap (Fig. 5A), although XGBoost exhibited the highest absolute AUC in the test set, pairwise comparisons using the DeLong test revealed no statistically significant differences between XGBoost and other major models (all p-values > 0.05). However, in terms of prediction probability improvement metrics—specifically IDI and NRI—XGBoost demonstrated a significant advantage (Figs. 5B–5C), suggesting superior risk stratification capabilities. Given the limited size of the independent test set, these performance estimates should be interpreted cautiously and require confirmation in larger external cohorts.
Figure 5. Statistical comparison and risk stratification improvement of the XGBoost model versus other algorithms.
(A) DeLong test heatmap showing pairwise P-values for AUC comparisons among models in the independent test set. (B) Integrated Discrimination Improvement (IDI) analysis, quantifying the improvement in prediction probabilities offered by XGBoost compared to other models. (C) Net Reclassification Improvement (NRI) analysis, demonstrating the significant advantage of the XGBoost model in correctly reclassifying patients into appropriate risk categories.
XGBoost model for SHAP
We computed global and local SHAP values for the XGBoost model to improve interpretability and potential clinical use. The SHAP summary plot (Fig. 6A) shows each feature’s contribution to the model’s predictions, and Fig. 6B ranks features by mean absolute SHAP values across the cohort. Two representative cases illustrate individual-level explanations: a patient with pCR (predicted probability 0.82; Fig. 6C) and a patient without pCR (0.28; Fig. 6D). SHAP values reflect model-based associations in this cohort and do not imply causality.
Figure 6. Summary analysis of SHAP values.
The SHAP importance plot (A) shows the weight distribution of the six most important features in the model, while the variable contribution plot (B) visually presents the positive (red) or negative (blue) effects of each feature on the predicted probability. Panels (C) and (D) use SHAP to visualize individual prediction results for cases with pathological complete response (pCR) and non-pCR after neoadjuvant chemotherapy, respectively: the baseline value represents the model’s base prediction probability, and f(x) indicates the final predicted probability.
Discussion
Histopathological examination of surgical specimens has long been considered the gold standard for evaluating the efficacy of NAT in breast cancer; however, its time delay limits early monitoring and dynamic assessment of tumor response. This study is the first to integrate CT-based body composition parameters with blood biomarkers, specifically identifying six critical independent predictors of pCR: VATD, SMD, IMAC, AAPR, SIRI, and molecular subtype. Using these features, we compared the performance of eight ML algorithms, where results demonstrate that the XGBoost model is an effective tool showing high accuracy and stability. Finally, SHAP analysis was employed to visualize the entire prediction process, providing interpretability for clinical decision-making.
Regarding body composition, VATD was identified as a significant independent predictor. In our cohort, elevated VATD was associated with a higher likelihood of achieving pCR. While the exact biological mechanism remains to be fully elucidated, visceral adipose tissue is metabolically active, secreting adipokines such as leptin and adiponectin that may modulate the tumor microenvironment. Previous studies, such as those by Iwase et al. (2016), have similarly observed that higher visceral fat measures can correlate with better treatment response in specific breast cancer populations, potentially reflecting a nutritional reserve that supports treatment tolerance, rather than a direct anti-tumor mechanism.
In parallel, skeletal muscle quality—assessed via SMD and IMAC—played a crucial role in our predictive model. Our findings indicate that higher SMD and lower IMAC (reflecting reduced myosteatosis) are predictive of pCR. Low muscle density and high intramuscular fat content are hallmarks of sarcopenia and systemic depletion. This state often implies compromised metabolic function, which can affect drug pharmacokinetics and distribution. Consistent with Lee et al. (2021), who demonstrated that skeletal muscle depletion correlates with lower pCR rates, our data suggests that maintaining muscle quality may be a relevant factor in supporting neoadjuvant treatment efficacy, though these associations should be interpreted as prognostic rather than strictly causal.
Systemic inflammatory and nutritional balance, captured by AAPR and SIRI, also contributed independently to the model. We observed that higher AAPR and lower SIRI levels were associated with improved responses. AAPR integrates albumin (a marker of nutritional status) and alkaline phosphatase, while SIRI reflects the balance of pro-inflammatory cells (neutrophils, monocytes) versus immune-regulatory lymphocytes. Furthermore, it is crucial to recognize that these systemic inflammation markers do not exist in isolation but may be modulated by broader perioperative and treatment-related factors. Recent evidence, such as the meta-analysis on the potential effect of general anesthetics in cancer surgery (Li et al., 2023), highlights that anesthetic management and surgical stress can significantly influence inflammatory cytokines and potential metastatic risks. Situating our findings within this wider oncologic context suggests that the inflammation signals observed here might be part of a dynamic host response that could be optimized through comprehensive peri-treatment management.
Molecular subtype remains a fundamental driver of therapeutic response. Consistent with extensive literature (Liu et al., 2024; Shi et al., 2023), our analysis confirmed that Triple-Negative Breast Cancer status (HR-negative/HER2-negative) is a strong predictor of pCR compared to luminal subtypes. This association is likely driven by the higher proliferation rates and increased chemosensitivity characteristic of this aggressive phenotype, reinforcing the necessity of including tumor biology alongside host features in predictive modeling.
A key distinction of our study is the prioritization of clinical accessibility over complexity. We acknowledge that numerous recent studies utilizing advanced imaging modalities, such as DCE-MRI and deep learning radiomics, have reported predictive performance that occasionally exceeds that of our model (Peng et al., 2022; Gamal et al., 2024; D’Anna et al., 2025; Iwase et al., 2021). However, these approaches often require specialized acquisition protocols, complex segmentation pipelines, and significant computational resources, which may limit their generalizability and utility in resource-constrained settings. In contrast, our system offers specific advantages: it relies exclusively on routine non-contrast CT scans and standard blood tests, incurring no additional cost, radiation exposure, or logistical burden. This makes our approach particularly appropriate as a widely deployable screening tool for initial risk stratification in centers where advanced MRI-based AI analysis is not yet feasible.
To integrate these multidimensional features, XGBoost was selected as the optimal algorithm. As a gradient boosting decision tree method, XGBoost offers distinct advantages in handling structured clinical data and capturing non-linear interactions between host and tumor factors (Zhou et al., 2025a; Zhou et al., 2025b). Its regularization strategies effectively reduce overfitting, while the integration of SHAP values addresses the “black box” nature of ML models, providing clinicians with interpretable, individualized risk profiles.
Several limitations of this study should be acknowledged. First, as a single-center retrospective analysis, the potential for selection bias exists, and our findings require validation in larger, multi-center cohorts. Second, our model relies on pre-NAT baseline measurements; future research should incorporate longitudinal data to assess how dynamic changes in body composition and inflammation during therapy influence outcomes. Lastly, while we adjusted for key clinical variables, unmeasured confounders—including specific perioperative management details discussed above—could influence the observed association.
Conclusions
We developed an interpretable ML-based comprehensive model incorporating clinical-pathological features, body composition, and inflammation-nutrition indices to assess pCR after breast cancer NAT. By rigorously evaluating host-related factors alongside tumor biology, this study provides a practical, cost-effective tool that complements existing imaging-based strategies, supporting more personalized clinical decision-making.
Supplemental Information
In the raw data, the numerical codes of categorical variables were converted to their respective factors.
Acknowledgments
The authors express their thanks to the patients and their families, as well as all researchers and medical staff involved in this study.
Funding Statement
This work was supported by the Science and Technology Planning Project of the Jiangxi Provincial Health Commission (Project Nos. 202410394 and 202510509); the Science and Technology Planning Project of the Jiangxi Provincial Administration of Traditional Chinese Medicine (Project Nos. 2025021340 and 2025021565); the Open Research Fund of Jiangxi Cancer Hospital & Institute (Project No. KFJJ2023YB16); the “Five-level Progressive” Talent Cultivation Project of Jiangxi Cancer Hospital & Institute (Project No. WCDJ2024QH03); and the Start-up Fund for Doctoral Research of Jiangxi Cancer Hospital (Project No. BSQDJ2024007). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Contributor Information
Lan Liu, Email: liulan202306@163.com.
Yongjie Zhou, Email: zhou919986102@qq.com.
Additional Information and Declarations
Competing Interests
The authors declare there are no competing interests.
Author Contributions
Linhua Zhong conceived and designed the experiments, performed the experiments, analyzed the data, prepared figures and/or tables, authored or reviewed drafts of the article, and approved the final draft.
Qiao Zeng performed the experiments, prepared figures and/or tables, and approved the final draft.
Fei Zou performed the experiments, prepared figures and/or tables, and approved the final draft.
Mingxian Gong performed the experiments, prepared figures and/or tables, and approved the final draft.
Lan Liu conceived and designed the experiments, prepared figures and/or tables, and approved the final draft.
Yongjie Zhou conceived and designed the experiments, analyzed the data, prepared figures and/or tables, authored or reviewed drafts of the article, and approved the final draft.
Ethics
The following information was supplied relating to ethical approvals (i.e., approving body and any reference numbers):
This study was conducted in accordance with the principles outlined in the Declaration of Helsinki and received approval from the Ethics Committee of Jiangxi Cancer Hospital (Ethics Approval No: 2023ky126), with a waiver of informed consent.
Data Availability
The following information was supplied regarding data availability:
The raw data and code are available in the Supplemental Files.
References
- Allison et al. (2020).Allison KH, Hammond M, Dowsett M, McKernin SE, Carey LA, Fitzgibbons PL, Hayes DF, Lakhani SR, Chavez-MacGregor M, Perlmutter J, Perou CM, Regan MM, Rimm DL, Symmans WF, Torlakovic EE, Varella L, Viale G, Weisberg TF, McShane LM, Wolff AC. Estrogen and progesterone receptor testing in breast cancer: ASCO/CAP guideline update. Journal of Clinical Oncology. 2020;38:1346–1366. doi: 10.1200/JCO.19.02309. [DOI] [PubMed] [Google Scholar]
- Bray et al. (2024).Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA—A Cancer Journal for Clinicians. 2024;74:229–263. doi: 10.3322/caac.21834. [DOI] [PubMed] [Google Scholar]
- Cortazar et al. (2014).Cortazar P, Zhang L, Untch M, Mehta K, Costantino JP, Wolmark N, Bonnefoi H, Cameron D, Gianni L, Valagussa P, Swain SM, Prowell T, Loibl S, Wickerham DL, Bogaerts J, Baselga J, Perou C, Blumenthal G, Blohmer J, Mamounas EP, Bergh J, Semiglazov V, Justice R, Eidtmann H, Paik S, Piccart M, Sridhara R, Fasching PA, Slaets L, Tang S, Gerber B, Geyer CJ, Pazdur R, Ditsch N, Rastogi P, Eiermann W, Von Minckwitz G. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis. Lancet. 2014;384:164–172. doi: 10.1016/S0140-6736(13)62422-8. [DOI] [PubMed] [Google Scholar]
- D’Anna et al. (2025).D’Anna A, Aranzulla C, Carnaghi C, Caruso F, Castiglione G, Grasso R, Gueli AM, Marino C, Pane F, Pulvirenti A, Stella G. Comparative analysis of machine learning models for predicting pathological complete response to neoadjuvant chemotherapy in breast cancer: an MRI radiomics approach. Physica Medica-European Journal of Medical Physics. 2025;131:104931. doi: 10.1016/j.ejmp.2025.104931. [DOI] [PubMed] [Google Scholar]
- Deo (2015).Deo RC. Machine learning in medicine. Circulation. 2015;132:1920–1930. doi: 10.1161/CIRCULATIONAHA.115.001593. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Feeney et al. (2024).Feeney G, Waldron R, Miller N, Malone C, Sweeney K, McLaughlin R, Lowery A, Barry K, Kerin M. Association of clinical biomarkers and response to neoadjuvant therapy in breast cancer. Irish Journal of Medical Science. 2024;193:605–613. doi: 10.1007/s11845-023-03489-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Gamal et al. (2024).Gamal A, Sharafeldeen A, Alnaghy E, Alghandour R, Saleh Alghamdi N, Ali KM, Shamaa S, Aboueleneen A, Elsaid Tolba A, Elmougy S, Ghazal M, Contractor S, El-Baz A. A novel machine learning approach for predicting neoadjuvant chemotherapy response in breast cancer: integration of multimodal radiomics with clinical and molecular subtype markers. IEEE Access. 2024;12:104983–105003. doi: 10.1109/ACCESS.2024.3432459. [DOI] [Google Scholar]
- Hirmas, Holtschmidt & Loibl (2024).Hirmas N, Holtschmidt J, Loibl S. Shifting the paradigm: the transformative role of neoadjuvant therapy in early breast cancer. Cancer. 2024;16:3236. doi: 10.3390/cancers16183236. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Iwase et al. (2021).Iwase T, Parikh A, Dibaj SS, Shen Y, Shrimanker TV, Chainitikun S, Kida K, Sapon ME, Sahin O, James A, Medrano A, Klopp AH, Ueno NT. The prognostic impact of body composition for locally advanced breast cancer patients who received neoadjuvant chemotherapy. Cancer. 2021;13:608. doi: 10.3390/cancers13040608. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Iwase et al. (2016).Iwase T, Sangai T, Nagashima T, Sakakibara M, Sakakibara J, Hayama S, Ishigami E, Masuda T, Miyazaki M. Impact of body fat distribution on neoadjuvant chemotherapy outcomes in advanced breast cancer patients. Cancer Medicine. 2016;5:41–48. doi: 10.1002/cam4.571. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Korde et al. (2021).Korde LA, Somerfield MR, Carey LA, Crews JR, Denduluri N, Hwang ES, Khan SA, Loibl S, Morris EA, Perez A, Regan MM, Spears PA, Sudheendra PK, Symmans WF, Yung RL, Harvey BE, Hershman DL. Neoadjuvant chemotherapy, endocrine therapy, and targeted therapy for breast cancer: ASCO guideline. Journal of Clinical Oncology. 2021;39:1485–1505. doi: 10.1200/JCO.20.03399. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lee et al. (2021).Lee BM, Cho Y, Kim JW, Ahn SG, Kim JH, Jeung HC, Jeong J, Lee IJ. Association between skeletal muscle loss and the response to neoadjuvant chemotherapy for breast cancer. Cancer. 2021;13:1806. doi: 10.3390/cancers13081806. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Li et al. (2024).Li J, Hao C, Wang K, Zhang J, Chen J, Liu Y, Nie J, Yan M, Liu Q, Geng C, Wang X, Wang H, Wang S, Wu J, Yin Y, Song E, Jiang Z. Chinese society of clinical oncology (CSCO) breast cancer guidelines 2024. Translational Breast Cancer Research. 2024;5:18. doi: 10.21037/tbcr-24-31. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Li et al. (2023).Li R, Mukherjee MB, Jin Z, Liu H, Lin K, Liu Q, Dilger JP, Lin J. The potential effect of general anesthetics in cancer surgery: meta-analysis of postoperative metastasis and inflammatory cytokines. Cancer. 2023;15:2759. doi: 10.3390/cancers15102759. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Liu et al. (2024).Liu S, Jiang C, Wu D, Zhang S, Qiao K, Yang X, Yu B, Huang Y. Development of predictive models for pathological response status in breast cancer after neoadjuvant therapy based on peripheral blood inflammatory indexes. BMC Women’s Health. 2024;24:560. doi: 10.1186/s12905-024-03400-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ogston et al. (2003).Ogston KN, Miller ID, Payne S, Hutcheon AW, Sarkar TK, Smith I, Schofield A, Heys SD. A new histological grading system to assess response of breast cancers to primary chemotherapy: prognostic significance and survival. Breast. 2003;12:320–327. doi: 10.1016/s0960-9776(03)00106-1. [DOI] [PubMed] [Google Scholar]
- Park et al. (2019).Park JE, Park SY, Kim HJ, Kim HS. Reproducibility and generalizability in radiomics modeling: possible strategies in radiologic and statistical perspectives. Korean Journal of Radiology. 2019;20:1124–1137. doi: 10.3348/kjr.2018.0070. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Peng et al. (2022).Peng Y, Cheng Z, Gong C, Zheng C, Zhang X, Wu Z, Yang Y, Yang X, Zheng J, Shen J. Pretreatment DCE-MRI-based deep learning outperforms radiomics analysis in predicting pathologic complete response to neoadjuvant chemotherapy in breast cancer. Frontiers in Oncology. 2022;12:846775. doi: 10.3389/fonc.2022.846775. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Pickhardt, Summers & Garrett (2021).Pickhardt PJ, Summers RM, Garrett JW. Automated CT-based body composition analysis: a golden opportunity. Korean Journal of Radiology. 2021;22:1934–1937. doi: 10.3348/kjr.2021.0775. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Romeo et al. (2021).Romeo V, Accardo G, Perillo T, Basso L, Garbino N, Nicolai E, Maurea S, Salvatore M. Assessment and prediction of response to neoadjuvant chemotherapy in breast cancer: a comparison of imaging modalities and future perspectives. Cancer. 2021;13:3521. doi: 10.3390/cancers13143521. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Schmid et al. (2022).Schmid P, Cortes J, Dent R, Pusztai L, McArthur H, Kummel S, Bergh J, Denkert C, Park YH, Hui R, Harbeck N, Takahashi M, Untch M, Fasching PA, Cardoso F, Andersen J, Patt D, Danso M, Ferreira M, Mouret-Reynier MA, Im SA, Ahn JH, Gion M, Baron-Hay S, Boileau JF, Ding Y, Tryfonidis K, Aktan G, Karantza V, O’Shaughnessy J. Event-free survival with pembrolizumab in early triple-negative breast cancer. New England Journal of Medicine. 2022;386:556–567. doi: 10.1056/NEJMoa2112651. [DOI] [PubMed] [Google Scholar]
- Sharafeldeen et al. (2025).Sharafeldeen A, Taher F, Alghamdi NS, Alnaghy E, Alghandour R, Ali KM, Shamaa S, Gamal A, Ghazal M, Contractor S, El-Baz A. A comprehensive deep learning system with MGRF modeling for predicting breast cancer response to neoadjuvant chemotherapy. IEEE Access. 2025;13:128654–128672. doi: 10.1109/ACCESS.2025.3590649. [DOI] [Google Scholar]
- Shi et al. (2023).Shi Z, Huang X, Cheng Z, Xu Z, Lin H, Liu C, Chen X, Liu C, Liang C, Lu C, Cui Y, Han C, Qu J, Shen J, Liu Z. MRI-based quantification of intratumoral heterogeneity for predicting treatment response to neoadjuvant chemotherapy in breast cancer. Radiology. 2023;308:e222830. doi: 10.1148/radiol.222830. [DOI] [PubMed] [Google Scholar]
- Sonkin, Thomas & Teicher (2024).Sonkin D, Thomas A, Teicher BA. Cancer treatments: past, present, and future. Cancer Genetics. 2024;286–287:18–24. doi: 10.1016/j.cancergen.2024.06.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Traverso et al. (2018).Traverso A, Wee L, Dekker A, Gillies R. Repeatability and reproducibility of radiomic features: a systematic review. International Journal of Radiation Oncology Biology Physics. 2018;102:1143–1158. doi: 10.1016/j.ijrobp.2018.05.053. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wolff et al. (2023).Wolff AC, Somerfield MR, Dowsett M, Hammond M, Hayes DF, McShane LM, Saphner TJ, Spears PA, Allison KH. Human epidermal growth factor receptor 2 testing in breast cancer: ASCO-College of American Pathologists guideline update. Journal of Clinical Oncology. 2023;41:3867–3872. doi: 10.1200/JCO.22.02864. [DOI] [PubMed] [Google Scholar]
- Yamamoto, Kawada & Obama (2021).Yamamoto T, Kawada K, Obama K. Inflammation-related biomarkers for the prediction of prognosis in colorectal cancer patients. International Journal of Molecular Sciences. 2021;22:8002. doi: 10.3390/ijms22158002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zhou et al. (2025a).Zhou Y, Zhao J, Zou F, Tan Y, Zeng W, Jiang J, Hu J, Zeng Q, Gong L, Liu L, Zhong L. Interpretable machine learning models based on body composition and inflammatory nutritional index (BCINI) to predict early postoperative recurrence of colorectal cancer: multi-center study. Computer Methods and Programs in Biomedicine. 2025a;269:108874. doi: 10.1016/j.cmpb.2025.108874. [DOI] [PubMed] [Google Scholar]
- Zhou et al. (2025b).Zhou Y, Zuo Z, Zhao J, Tan Y, Deng J, Wei X, Li H, Gong L, Liu L, Zhong L. Development and validation of time-to-event machine learning models for predicting disease-free survival in patients with locally advanced colorectal cancer: a multicenter cohort study. Annals of Surgical Oncology. 2025b;33:1288–1300. doi: 10.1245/s10434-025-18815-3. [DOI] [PubMed] [Google Scholar]
- Zou et al. (2025).Zou F, Zhao J, Zhong L, Gong L, Jiang J, Hu J, Zeng W, Liu L, Zhou Y. CT-based body composition and inflammatory nutritional biomarker nomogram for predicting early postoperative recurrence of non-small cell lung cancer: a multicenter study. Journal of Thoracic Disease. 2025;17:8046–8062. doi: 10.21037/jtd-2025-1211. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Zwanenburg et al. (2020).Zwanenburg A, Vallieres M, Abdalah MA, Aerts H, Andrearczyk V, Apte A, Ashrafinia S, Bakas S, Beukinga RJ, Boellaard R, Bogowicz M, Boldrini L, Buvat I, Cook G, Davatzikos C, Depeursinge A, Desseroit MC, Dinapoli N, Dinh CV, Echegaray S, El NI, Fedorov AY, Gatta R, Gillies RJ, Goh V, Gotz M, Guckenberger M, Ha SM, Hatt M, Isensee F, Lambin P, Leger S, Leijenaar R, Lenkowicz J, Lippert F, Losnegard A, Maier-Hein KH, Morin O, Muller H, Napel S, Nioche C, Orlhac F, Pati S, Pfaehler E, Rahmim A, Rao A, Scherer J, Siddique MM, Sijtsema NM, Socarras FJ, Spezi E, Steenbakkers R, Tanadini-Lang S, Thorwarth D, Troost E, Upadhaya T, Valentini V, Van Dijk LV, Van Griethuysen J, Van Velden F, Whybra P, Richter C, Lock S. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. 2020;295:328–338. doi: 10.1148/radiol.2020191145. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
In the raw data, the numerical codes of categorical variables were converted to their respective factors.
Data Availability Statement
The following information was supplied regarding data availability:
The raw data and code are available in the Supplemental Files.