Robustness and Accuracy of Radiomics Models for Classifying IASLC Grading in Lung Adenocarcinomas: A Comprehensive Analysis of a Large Multicenter CT Database

Abstract

Introduction

Accurate preoperative prediction of the International Association for the Study of Lung Cancer (IASLC) grading is crucial for developing individualized management and surgical strategies in lung adenocarcinomas (LUAD). Computed tomography (CT) radiomics serves as an important imaging biomarker for classification tasks in LUAD. However, the robustness and accuracy of radiomics models remain subjects of ongoing debate.

Methods

In this study, we conducted an analytical comparison of two critical steps in radiomics: dimensionality reduction and feature selection, aiming to differentiate between Grade 1 and Grade 2–3 tumors according to the preoperative IASLC grading system for LUAD. 1) For dimensionality reduction, we sequentially combined the T-test, Pearson correlation, and Least Absolute Shrinkage and Selection Operator (LASSO), while considering principal component analysis (PCA) for comparison. 2) For feature selection, we utilized various machine learning (ML) techniques including Logistic Regression (LR), Random Forest (RF), Gradient Boosting (GBC), XGBoost, Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost). The diagnostic efficacy was evaluated using receiver operating characteristic curve (ROC) and the corresponding area under the curve (AUC).

Result: The AUC for PCA combined with various ML feature selection methods ranged from 0.502 to 0.719 in this classification task. In contrast, the AUC for the combined T-test, Pearson, and LASSO dimensionality reduction methods, along with various ML feature selection methods, significantly increased from 0.818 to 0.869. Among these, the LGBM achieved the highest performance, reaching an AUC of 0.869, while LR displayed the lowest performance with an AUC of 0.818.

Conclusion

We demonstrated that the T-test→Pearson→LASSO approach is more appropriate for radiomics feature dimensionality reduction compared to PCA. Additionally, we improved the commonly used LR feature selection method in medical research by employing the more advanced LGBM for distinguishing between Grade 1 and Grade 2–3 tumors in accordance with the preoperative IASLC grading system for LUAD.

Keywords

lung adenocarcinomas radiomics IASLC grading dimensionality reduction feature selection

Introduction

Lung adenocarcinoma (LUAD) is the most prevalent form of lung cancer, exhibiting a wide range of histological features. Approximately 80% to 90% ofLUAD cases present a mix of different histologic patterns.¹ The Pathology Committee of the International Association for the Study of Lung Cancer (IASLC) categorizes LUAD based on predominant subtypes (eg, solid, micropapillary, acinar, papillary, and lepidic) and high-risk patterns (including solid, micropapillary, and/or complex glandular components). Specifically, lepidic patterns with no or less than 20% high-risk components are classified as Grade 1. A predominant papillary or acinar component with less than 20% high-grade patterns is Grade 2, while any component with 20% or more high-grade patterns is Grade 3.^2–4

The overall survival and disease-free survival rates differ significantly among Grades 1–3, with Grade 1 showing a more favorable prognosis than Grades 2 and 3.^2–4 Consequently, Grades 2–3 often necessitate anatomic lobectomy and systematic lymph node dissection, whereas Grade 1 typically requires regular follow-up.^5,6 Thus, accurate preoperative prediction of IASLC grading in LUAD is vital for tailored disease management and treatment plans.

Radiomics, an emerging field within medical imaging, involves the extraction and analysis of quantitative features from radiographic images to characterize tumor heterogeneity and assess the biological characteristics of lung cancer.^7–9 This approach encompasses several stages: image preprocessing, feature extraction, dimensionality reduction, and feature selection. Previous studies have developed a computed tomography (CT)-based radiomics method for classifying the pathological types and IASLC grading of LUAD patients.^10–15 These studies particularly focused on using a combination of the T-test, Pearson correlation, and Least Absolute Shrinkage and Selection Operator (LASSO) (T-test→Pearson→LASSO (TPL)) for dimensionality reduction, followed by the radiomics score for classification based solely on logistic regression (LR).^10–15 Despite these advancements, enhancing the robustness and accuracy of radiomics models remains a significant challenge, especially during the stages of dimensionality reduction and feature selection.¹⁶

Developing reliable and highly accurate predictive models is essential for advancing radiomics methodologies. Machine learning (ML) classifiers are increasingly integrated with radiomic features to boost model performance.¹⁷ Various ML models, including LR, k-nearest neighbors (kNN), random forest (RF), and eXtreme Gradient Boosting (XGBoost), have been used in lung cancer research to predict gene mutations,¹⁸ assess pathological invasiveness,¹⁹ and predict outcomes.²⁰

In this study, we focused on systematically optimizing the radiomics workflow for IASLC grading by investigating both dimensionality reduction and machine learning strategies. Specifically, we compared supervised feature selection pipelines—combining T-test, Pearson correlation, and LASSO (T-test→Pearson→LASSO)—with unsupervised approaches such as Principal Component Analysis (PCA) to identify the most effective feature extraction scheme. Furthermore, we employed multiple supervised machine learning algorithms, including LR, RF, Gradient Boosting Classifier (GBC), XGBoost, Light Gradient Boosting Machine (LGBM), and Categorical Boosting (CatBoost), to comprehensively evaluate model robustness and generalizability. To the best of our knowledge, few studies have systematically compared the influence of different dimensionality reduction and machine learning frameworks on radiomics-based IASLC grading, particularly in differentiating Grade 1 from Grade 2–3 lung adenocarcinomas. By leveraging a large multicenter CT cohort, our study aims to enhance the reliability, interpretability, and clinical applicability of radiomics models for preoperative IASLC grading, thereby supporting more individualized management and treatment strategies in lung cancer.

Methods

Large Multicenter CT Database

In this study, we utilized a large multicenter CT database that included patients from three hospitals: Xiangtan Central Hospital (Center 1), the Affiliated Hospital of Southwest Medical University (Center 2), and the Affiliated Hospital of Guilin Medical University (Center 3). These patients received preoperative CT scans that revealed malignant pulmonary nodules and underwent surgical resection between January 2018 and June 2023, culminating in a pathological diagnosis of LUAD. Our study specifically aimed to enroll patients who met the following criteria: (i) a maximum tumor length not exceeding 30 mm; (ii) chest CT scans conducted within 7 days prior to surgery; and (iii) availability of CT scans with a thin slice thickness of less than 1.5 mm. The exclusion criteria included: (i) presence of multiple primary lung cancers or metastases, (ii) prior history of chemotherapy or radiotherapy, and (iii) concurrent existence of other malignancies.

Given the retrospective nature of our investigation, informed consent was obtained from all participants. The study was approved by the local ethics committees: the Medical Ethics Committee of Xiangtan Central Hospital (Approval No. 2021-07-009, July 19, 2021), the Clinical Trial Ethics Committee of the Affiliated Hospital of Southwest Medical University (Approval No. KY2020147, August 12, 2020), and the Medical Ethics Committee of the Affiliated Hospital of Guilin Medical University (Approval No. 2023YJSLL-121, December 22, 2023). To evaluate the proposed classification task utilizing a radiomic approach, all performance metrics were primarily assessed using the testing set. The screening flowchart of the enrolled patients is presented in Figure 1.

Figure 1.

The screening low chart of enrolled multicenter CT database.

CT Images Preprocessing and Lesion Segmentation

The CT scanning parameters employed in this study were consistent with those used in our prior research.^12,21 For the specific CT scanning parameters of our study, please refer to Supplemental Digital Content 1. We utilized ITK-SNAP software (version 4.0, www.itksnap.org) to perform a slice-by-slice delineation of the gross tumor volume along the tumor boundary, thereby generating a corresponding whole lesion mask. Initially, mask drawing was conducted on CT images using the lung window settings. A certified cardiothoracic radiologist identified the tumor border, and the mask was then manually segmented across all two-dimensional sections of the nodule, with the goal of minimizing the inclusion of adjacent vessels, bronchi, and pleura. Following this, another cardiothoracic radiologist with over a decade of experience in the field reviewed and adjusted the lesion delineation as necessary. Any discrepancies among the observers were discussed, and final masks were determined by consensus to minimize inter-observer variability and ensure the accuracy and reproducibility of the segmentation.

Radiomics Feature Extraction

The original CT image and segmented mask were imported into Spyder (Python version 3.7.0) for further processing, with the pixel values of the images scaled to a range of 0 to 1. Subsequently, the volumes were resampled to a voxel size of 1×1×1 mm3 using a bicubic interpolation algorithm implemented with the ‘Numpy’ package. Following this step, radiomic features were extracted from the region of interest (ROI) on the CT image corresponding to the mask utilizing the ‘PyRadiomics' package.²² To capture high-order image patterns, in addition to the original images, we applied Wavelet filters (yielding 8 decompositions: LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH) and Laplacian of Gaussian (LoG) filters (with sigma values of 1.0, 2.0, 3.0, 4.0, and 5.0 mm). This comprehensive tool facilitated the extraction of a total of 1239 distinct radiomics features (the samples $x_{i}$ are all vectors that are $1 \times 1239$ in size), encompassing first-order statistics, shape-based metrics, gray level dependence matrix, gray level co-occurrence matrix, gray level run length matrix, gray level size zone matrix, and neighboring gray tone difference matrix. Due to the inclusion of CT images from multiple hospitals and various CT protocols, all radiomics features underwent normalization through a z-score transformation ( $z = \frac{x - μ}{σ}$ ). Specifically, the mean ( $μ$ ) and standard deviation ( $σ$ ) were calculated based on the training set, and these parameters were subsequently applied to standardize the features in the validation and testing sets.

Dimensionality Reduction

The extracted original CT radiomics features are highly correlated and are numerous. To reduce the dimensionality of the features and facilitate model training, we combined T-tests, Pearson correlation, and LASSO in the following order: T-test→Pearson→LASSO. Additionally, we employed unsupervised ML techniques, such as PCA.

T-test

The T-test, also known as the Student T-test, is a very common test method in statistical inference.²³ It is used when the statistic follows a normal distribution but the variance is unknown. The T-test uses t-distribution theory to deduce the probability of differences, thereby comparing whether the difference between two averages is significant. The T-test can be divided into one-sample test, independent two-sample test, and dependent T-test for paired samples.

The dependent T-test for paired samples is used to test the differences in data obtained from two matched groups of subjects or the data obtained from the same group of subjects under different conditions. The samples consisting of these two situations are related samples ( $p > 0.05$ ). If the difference between the two-paired sample $x_{i}$ and $x_{j}$ is $d_{i} = x_{i} - x_{j}$ independent and from normal distribution, whether the population expected value $μ$ of $d_{i}$ is $μ_{0}$ can be tested using the following statistic:

t = \frac{\bar{d} - μ_{0}}{s_{d} / \sqrt{n}},

(1)

where

i = 1, \dots, n

\bar{d} = \frac{\sum_{i = 1}^{n} d_{i}}{n}

is the mean of the differences,

s_{d} = \sqrt{\frac{\sum_{i = 1}^{n} {(d_{i} - \bar{d})}^{2}}{n - 1}}

is the standard deviation of the differences, and n is the number of pairs. The statistic t follows a distribution of

n - 1

degrees of freedom under the condition that the null hypothesis:

μ = μ_{0}

is true.

In this study, we performed a T-test to identify features with $p > 0.05$ , which were subsequently excluded. Finally, the samples $x_{i}$ are vectors with $x_{i}^{'} = 1 \times a$ in size.

Pearson Correlation Coefficient

Pearson correlation coefficient is a statistical metric that measures the strength and direction of a linear relationship between two random variables.²⁴ The Pearson correlation coefficient between two variables is defined as the quotient of the covariance and the standard deviation between the two variables:

ρ_{x_{i}^{'}, x_{j}^{'}} = \frac{c o v (x_{i}^{'}, x_{j}^{'})}{σ_{x_{i}^{'}} σ_{x_{j}^{'}}} = \frac{E [(x_{i}^{'} - \bar{x_{i}^{'}}) (x_{j}^{'} - \bar{x_{j}^{'}})]}{σ_{x_{i}^{'}} σ_{x_{j}^{'}}},

(2)

where

c o v (x_{i}^{'}, x_{j}^{'})

is the covariance of

x_{i}^{'}

and

x_{j}^{'}

\bar{x_{i}^{'}}

and

σ_{x_{i}^{'}}

are the sample mean value and sample standard deviation for the sample of

x_{i}^{'}

, respectively.

The above formula defines the overall correlation coefficient, which is usually represented by $ρ$ . To estimate the covariance and standard deviation of the sample, Pearson correlation coefficient can be obtained, which is commonly represented by r as follows:

r = \frac{\sum_{i = 1}^{n} (x_{i}^{'} - \bar{x_{i}^{'}}) (x_{j}^{'} - \bar{x_{j}^{'}})}{\sqrt{\sum_{i = 1}^{n} {(x_{i}^{'} - \bar{x_{i}^{'}})}^{2}} \sqrt{\sum_{i = 1}^{n} {(x_{j}^{'} - \bar{x_{j}^{'}})}^{2}}},

(3)

r

can also be estimated by the mean of the z-score of the sample points

(x_{i}^{'}, x_{j}^{'})

, resulting in an expression equivalent to the above:

r = \frac{1}{n - 1} \sum_{i = 1}^{n} (\frac{x_{i}^{'} - \bar{x_{i}^{'}}}{σ_{x_{i}^{'}}}) (\frac{x_{j}^{'} - \bar{x_{j}^{'}}}{σ_{x_{j}^{'}}}),

(4)

where

\frac{x_{i}^{'} - \bar{x_{i}^{'}}}{σ_{x_{i}^{'}}}

is the z-score for the sample of

x_{i}^{'}

In this study, we conducted Pearson's correlation analysis to eliminate redundant features that showed a correlation coefficient $r \geq 0.90$ , following previous studies.^25,26

Finally, the samples $x_{i}^{'}$ are vectors with $x_{i}^{″} = 1 \times b$ in size.

LASSO Regression

One of the most important problems in ML is over-fitting. During the minimization of the loss function, our focus on reducing the loss value on the training set can lead us to neglect the model's generalization capabilities. Adding regularization terms is a very good way to solve the over-fitting problem. Common regularization terms are L1 regularization and L2 regularization, which correspond to LASSO regression and Ridge regression respectively in linear regression.

The basic idea of LASSO is to minimize the residual sum of squares under the constraint condition that the absolute sum of the regression coefficients is less than a constant, so that some regression coefficients that are strictly equal to 0 can be generated, and an interpretable model can be obtained.

The mathematical expression is as follows:

ω := a r g min_{ω} {| | y_{i} - x_{i}^{″} ω_{i} | |}, s . t . | | ω_{i} | | \leq h,

(5)

where

h > 0

is the adjustment parameter, and the compression of the overall regression coefficient can be realized by controlling the adjustment parameter h.

This mathematical expression is also equivalent to minimizing the following penalty least squares:

ω := a r g min_{ω} {| | y_{i} - x_{i}^{″} ω_{i} | |^{2} + λ | | ω_{i} | |_{1}},

(6)

where

λ

corresponds to h and can be converted to each other.

The main advantage of LASSO method is that it compresses the variables with large parameter estimates less, while the variables with small parameter estimates are compressed to 0, and the parameter estimation of LASSO analysis has continuity, which is suitable for model selection of high-dimensional data.^27,28 The determination of h value can be estimated using the Least-angle regression (LARS).²⁹ LARS is similar to forward stepwise regression. At each step, it finds the feature most correlated with the target. When there are multiple features having equal correlation, instead of continuing along the same feature, it proceeds in a direction equiangular between the features. Finally, the samples $x_{i}^{″}$ are vectors with $x_{i}^{″} = 1 \times c$ in size.

Feature Selection

We employed multiple supervised ML algorithms for feature selection, including LR, RF, GBC, XGBoost, LGBM, and CatBoost. The selected radiomics features, after dimensionality reduction, were individually entered into the models^30,31:

LR is a statistical model that uses a logistic function to model a binary dependent variable. It's widely used for classification tasks, where it estimates the probability that a given input point belongs to a certain class.³²

RF is an ensemble learning method that constructs multiple decision trees during training and outputs the class that is the mode of the classes for classification of the individual trees. It inherently performs feature selection by evaluating the importance of each feature based on its contribution to the decision-making process across all trees.³³

GBC is a powerful ensemble technique that builds models in a stage-wise fashion. Each new model attempts to correct the errors of the previous models and can handle both binary and multi-class classification problems. Feature importance can be derived from the contribution of each feature to the overall model performance across the boosting stages.³⁴

XGBoost is an optimized distributed gradient boosting library designed for efficiency, flexibility, and portability. It implements a scalable tree boosting system and is widely recognized for its performance in various ML competitions. Feature importance in XGBoost can be assessed based on the gain of each feature when splitting the data.³⁵

LGBM is a gradient boosting framework that employs tree-based learning algorithms and is designed to be more efficient than traditional gradient boosting methods. It utilizes techniques such as Gradient-based one-side sampling and exclusive feature bundling to reduce the number of data instances and features, respectively, without sacrificing accuracy.³⁶

CatBoost is an open-source gradient boosting library that can directly handle categorical features during training, significantly improving model performance for datasets with many categorical variables. It implements several techniques to prevent overfitting, such as ordered boosting. Feature importance in CatBoost can be evaluated based on the influence of each feature on the model's predictions.³⁷

The diagnostic efficacy was evaluated using receiver operating characteristic curve (ROC) and the corresponding area under the curve (AUC). Additionally, we adopted four metrics to assess classification performance: Accuracy, Sensitivity, Specificity, and Precision, which are defined in equation 7.

A c c u r a c y = \frac{T P + T N}{T P + T N + F P + F N},

S e n s i t i v i t y = \frac{T P}{T P + F N},

(7)

S p e c i f i c i t y = \frac{T N}{T N + F P},

P r e c i s i o n = \frac{T P}{T P + F P},

where TP, TN, FP and FN denote the number of true positives, true negatives, false positives and false negatives, respectively.

The Overall Diagnostic Pipeline

We summarize the diagnostic pipeline of this study in Figure 2 and present the theoretical framework in Algorithm 1. The process begins with the segmentation of the input CT image, followed by the extraction of radiomics features. To reduce correlation and redundancy among these original features, we apply the T-test→Pearson→LASSO method. Using the final selected features, we construct a diagnostic model using the optimal ML algorithm. For the binary classification task, which specifically distinguishes between Grade 1 and Grade 2–3 tumors based on the preoperative IASLC grading system for LUAD, we use the training set to train the model, the validation set to tune parameters and select the optimal configuration, and finally, we test the model on the test set.

Figure 2.

Pipeline for the radiomics methodology: dimensionality reduction and feature selection processes.

Implementation Details

All data preprocessing and feature selection procedures were implemented in Python (version 3.7.0). The ‘SciPy’ package was used to perform the T-test and Pearson correlation analyses during feature selection. The ‘scikit-learn’ package was used to perform LASSO regression and several ML classification methods.^38,39

Statistical Analysis

All statistical analyses and modeling were performed using R (version 3.6.3; R Foundation for Statistical Computing). The comparison of diagnostic efficacy ROC curves and calculation of Accuracy, Sensitivity, Specificity, and Precision were conducted using the reportROC package. Comparisons of baseline characteristics across the training, validation, and testing sets were performed using the tableone package.

Results

Patient Characteristics

The complete CT database, comprising 1333 scans, was systematically categorized into three distinct subsets: a training set consisting of 800 scans (60%), a validation set comprising 200 scans (15%), and a testing set containing 333 scans (25%). In total, the dataset consisted of 487 Grade 1, 759 Grade 2, and 87 Grade 3 scans. Among them, 292 Grade 1, 447 Grade 2, and 61 Grade 3 scans were assigned to the training set; 73 Grade 1, 115 Grade 2, and 12 Grade 3 scans to the validation set; and 122 Grade 1, 197 Grade 2, and 14 Grade 3 scans to the testing set.

Algorithm 1 Radiomics Model Framework: T-Test, Pearson, LASSO for Dimensionality Reduction, and ML for Feature Selection.

Input: The training samples {(

y_{i}

x_{i}

) | i = 1,…,n}. Inference:

1: Tumor segmentation; // §

2: Radiomics feature extraction: features

x_{i}

; // §

3: Dimensionality reduction: // §

1: T-test:

t = \frac{\bar{d} - μ_{0}}{s_{d} / \sqrt{n}}

features

x_{i}^{'}

; // §

2: Pearson correlation coefficient:

r = \frac{\sum_{i = 1}^{n} (x_{i}^{'} - \bar{x_{i}^{'}}) (x_{j}^{'} - \bar{x_{j}^{'}})}{\sqrt{\sum_{i = 1}^{n} {(x_{i}^{'} - \bar{x_{i}^{'}})}^{2}} \sqrt{\sum_{i = 1}^{n} {(x_{j}^{'} - \bar{x_{j}^{'}})}^{2}}}

(3), features

x_{i}^{″}

; // §

3. LASSO:

ω := a r g min_{ω} {| | y_{i} - x_{i}^{″} ω_{i} | |^{2} + λ | | ω_{i} | |_{1}}

(6), features

x_{i}^{″}

; // §

4: Feature selection:

{\hat{y}}_{i} = f_{M L} ({x^{″'}}_{i})

; // §

Output: Diagnostic prediction

{\hat{y}}_{i}^{(K)}

Table 1 presents a comprehensive overview of patient demographics, tumor characteristics, and grade distribution, demonstrating that the study cohort is representative across essential clinical variables. Statistical analysis revealed no significant differences ( $p > 0.05$ ), indicating a well-balanced distribution of tumor grades and stages, thereby affirming the comparability and representativeness of the included patient population.

Table 1.

Comprehensive Comparison of Training, Validation and Testing Sets.

Variables	Total (n = 1333)	Training Set (n = 800)	Validation Set (n = 200)	Testing Set (n = 333)	P value
label, n (%)					0.999
Grade1	487 (36.5)	292 (36.5)	73 (36.5)	122 (36.6)
Grade2–3	846 (63.5)	508 (63.5)	127 (63.5)	211 (63.4)
CT density, n (%)					0.651
pGGN	627 (47)	385 (48.1)	97 (48.5)	145 (43.5)
SN	334 (25.1)	196 (24.5)	51 (25.5)	87 (26.1)
SSN	372 (27.9)	219 (27.4)	52 (26)	101 (30.3)
Location, n (%)					0.421
LLL	178 (13.4)	95 (11.9)	33 (16.5)	50 (15)
LUL	354 (26.6)	225 (28.1)	46 (23)	83 (24.9)
RLL	230 (17.3)	133 (16.6)	31 (15.5)	66 (19.8)
RML	99 (7.4)	60 (7.5)	16 (8)	23 (6.9)
RUL	472 (35.4)	287 (35.9)	74 (37)	111 (33.3)
Margin, n (%)					0.295
Ill-defined	317 (23.8)	202 (25.2)	42 (21)	73 (21.9)
Well-defined	1016 (76.2)	598 (74.8)	158 (79)	260 (78.1)
Shape, n (%)					0.823
Irregular	617 (46.3)	367 (45.9)	91 (45.5)	159 (47.7)
Regular	716 (53.7)	433 (54.1)	109 (54.5)	174 (52.3)
Lobulation sign, n (%)					0.989
Absent	576 (43.2)	347 (43.4)	86 (43)	143 (42.9)
Present	757 (56.8)	453 (56.6)	114 (57)	190 (57.1)
Spiculation sign, n (%)					0.619
Absent	701 (52.6)	424 (53)	109 (54.5)	168 (50.5)
Present	632 (47.4)	376 (47)	91 (45.5)	165 (49.5)
Vascular convergence sign, n (%)					0.642
Absent	308 (23.1)	188 (23.5)	49 (24.5)	71 (21.3)
Present	1025 (76.9)	612 (76.5)	151 (75.5)	262 (78.7)
Vacuole sign, n (%)					0.741
Absent	1096 (82.2)	657 (82.1)	168 (84)	271 (81.4)
Present	237 (17.8)	143 (17.9)	32 (16)	62 (18.6)
Pleural indentation sign, n (%)					0.403
Absent	562 (42.2)	341 (42.6)	90 (45)	131 (39.3)
Present	771 (57.8)	459 (57.4)	110 (55)	202 (60.7)
Sex, n (%)					0.341
Female	892 (66.9)	545 (68.1)	135 (67.5)	212 (63.7)
Male	441 (33.1)	255 (31.9)	65 (32.5)	121 (36.3)
Age(y), Median (Q1,Q3)	57 (51, 66)	58 (51, 66)	56 (50, 65.2)	56 (49, 65)	0.029
Lesion size(mm), Median (Q1,Q3)	16.1 (11.7, 21.6)	16.1 (11.7, 21.4)	15.4 (11.4, 21.1)	16.5 (11.7, 22.1)	0.453

Abbreviation: pGGN, pure ground-glass nodule; SN, solid nodule; SSN, subsolid nodule; LLL, left lower lobe; LUL, left upper lobe; RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe.

Dimensionality Reduction

In this study, we sequentially combined the T-test, Pearson correlation, and LASSO regression for dimensionality reduction among features. Following the application of the T-test and Pearson's correlation test, we retained 989 and 984 features, respectively. To further refine our feature selection, we utilized LASSO regression with 10-fold cross-validation, as depicted in Figure 3. This procedure resulted in the exclusion of 10 features with coefficients below 0.01, leaving us with a final subset of only 7 selected features. The constructed radiomic signatures are illustrated in Figure 4. Both dimensionality reduction techniques were initially executed on the training set and were later applied to the validation and test sets.

Figure 3.

Feature selection for the LASSO logistic regression and the predictive accuracy of the radiomics signature. (a) LASSO coefficient profiles for all features. (b) Selection of tuning parameter (λ) using 10-fold cross-validation with minimum criteria.

Figure 4.

Distribution of the selected radiomics features by LASSO.

Evaluation of Diagnostic Performance of Radiomic Models

For the selected radiomics features after dimensionality reduction (T-test→Pearson→LASSO), we evaluated the diagnostic performance of widely used ML methods to perform subsequent feature selection and construct the final radiomic model. The methods employed included LR, RF, GBC, XGBoost, LGBM, and CatBoost. As illustrated in Figure 5, the diagnostic performance on the validation set closely resembles that of the test set, indicating that the feature dimension reduction pipeline employed in this study is both effective and stable. Furthermore, the performance of the LR classification method, which has been previously used to classify LUAD patients’ pathological types and IASLC grading,^10–13 is significantly lower compared to the other ensemble methods.

Figure 5.

The ROC curves demonstrating the predictive performance of various machine learning methods used in radiomics model construction are presented for the validation set (a) and the testing set (b).

The diagnostic performance metrics for various radiomics approaches, which include dimensionality reduction techniques (T-test→Pearson→LASSO) and ML for subsequent feature selection, are summarized in Table 2. Notably, in the test set, the AUC for the combined dimensionality reduction method of T-test→Pearson→LASSO, along with different ML feature selection methods, significantly increased from 0.818 to 0.869. Among these, the LGBM achieved the highest performance, reaching an AUC of 0.869, while the LR displayed the lowest performance, with an AUC of 0.818 . Therefore, we define this approach as the Radiomics (TPL- LGBM) model.

Table 2.

Comprehensive Comparison of the Diagnosis Performance of LR, RF, GBC, XGBoost, LGBM and CatBoost. The Best and Second Best Results are Highlighted in Bold Font and Underlined Ones, Respectively.

Method	AUC	Accuracy	Sensitivity	Specificity	Precision
LR	0.818	0.766	0.735	0.820	0.876
RF	0.855	0.790	0.825	0.730	0.841
GBC	0.828	0.763	0.768	0.754	0.844
XGBoost	0.863	0.814	0.829	0.787	0.871
LGBM	0.869	0.820	0.787	0.877	0.917
CatBoost	0.859	0.799	0.801	0.795	0.871

The Visualizations of Classification Effect

The t-distributed Stochastic Neighbor Embedding (t-SNE) technique was employed to visually display the effects of the classification task for the Radiomics (TPL-LGBM) model. The t-SNE visualizations of both the original features and the features after dimensionality reduction are shown in Figure 6. From these visualizations, it is evident that the data from the original features is not easily distinguishable, with the two categories appearing entangled. However, after applying dimensionality reduction, the data can be more clearly separated, which enhances the final classification performance. These results strongly indicate that the Radiomics (TPL-LGBM) model utilized in this study is effective.

Figure 6.

The t-Stochastic Neighbor Embedding (t-SNE) visualizations of original features (a) and features after dimensionality reduction (b).

Comparison of Alternative Radiomic Models

We conducted an alternative radiomics model that utilizes PCA for dimensionality reduction and ML for feature selection. This approach was compared with the Radiomics (TPL-LGBM) model.

The theoretical framework for dimensionality reduction using PCA is presented in Algorithm 2. The main goal of PCA^40–42 is to identify a new set of orthogonal bases that transforms the data from an (n)-dimensional space to a (k)-dimensional space, such that the distance between the data points is maximized. In other words, PCA seeks to maximize the variance among the data points after they are projected onto the plane formed by these orthogonal bases. When the variance is maximized for each individual orthogonal basis, it inherently ensures that the projected distances are also maximized on the corresponding plane.

PCA reduces the data from (n) dimensions to (k) dimensions, and the next crucial step is determining the appropriate value for (k). A common selection criterion involves evaluating the ratio of variance before and after projection, and for this study, we established a variance ratio threshold of 0.90. Ultimately, the data was successfully compressed to (k = 24) using PCA.

A comparison of the classification performance on the testing set between two mainstream radiomics approaches was conducted: the first utilizes PCA for dimensionality reduction and ML for feature selection, while the second employs the T-test→Pearson→LASSO for dimensionality reduction followed by ML for feature selection. The results are visually represented in a bar graph (Figure 7). The findings indicate that the dimensionality reduction method using the T-test→Pearson→LASSO outperforms PCA. This advantage can be attributed to PCA's potential inability to effectively eliminate redundancy among the original features, as PCA operates in an unsupervised manner, whereas LASSO is a supervised learning method.

Figure 7.

Comparison of classification performance between two mainstream radiomics approaches on the testing set.

Algorithm 2 Radiomics Model Framework: PCA for Dimensionality Reduction, and ML for Feature Selection.

Input: The training samples {(

y_{i}

x_{i}

) | i = 1,…, n}. Inference:

1: Tumor segmentation; // §

2: Radiomics feature extraction: features

x_{i}

; // §

3: Dimensionality reduction: PCA, features

x_{i}^{″}

; // §

4: Feature selection:

{\hat{y}}_{i} = f_{M L} ({x^{″'}}_{i})

; // §

Output: Diagnostic prediction

{\hat{y}}_{i}^{(K)}

Discussion

The IASLC grading system for lung adenocarcinoma provides critical prognostic information, making preoperative identification of tumor grade essential for tailoring treatment strategies. Our radiomics model offers a noninvasive approach to differentiate Grade 1 from Grade 2–3 tumors, thereby directly informing surgical decisions. Specifically, a prediction indicating Grade 2–3 pathology could justify anatomic lobectomy with lymph node dissection, which aligns with the aggressive biological behavior of these grades. Conversely, a Grade 1 result may support regular surveillance rather than immediate surgical intervention. Integrating this tool into multidisciplinary workflows—for instance, within radiology reporting systems—would thus provide a standardized, quantitative assessment to augment preoperative planning.

The radiomics process encompasses several key stages, including image preprocessing, feature extraction, dimensionality reduction, and feature selection. In previous studies, researchers have developed CT-based radiomics approaches to classify the pathological types and IASLC grading of LUAD patients. These studies particularly focused on using a combination ofthe T-test, Pearson correlation, and LASSO for dimensionality reduction, followed by the radiomics score for classification based solely on LR. However, enhancing the robustness and accuracy of radiomics models continues to be a challenge, particularly during the stages of dimensionality reduction and feature selection.

In this study, we focused on systematically optimizing the radiomics workflow for IASLC grading by investigating both dimensionality reduction and machine learning strategies. The main findings of the study are as follows: i) We demonstrated, for the first time, the robustness and accuracy of radiomics models in differentiating between Grade 1 and Grade 2–3 tumors in IASLC grading for LUAD based on a large multicenter CT database; ii) the T-test, Pearson correlation, and LASSO approach prove to be more suitable for radiomics feature dimensionality reduction compared to PCA; iii) we enhanced the commonly used LR feature selection method by employing the more advanced LGBM for distinguishing between Grade 1 and Grade 2–3 tumors according to the IASLC grading system for LUAD patients.

The findings from the radiomics experiment reveal distinct advantages of certain methods over others in the realms of dimensionality reduction and feature selection. Firstly, the observation that the T-test, Pearson correlation, and LASSO approach are more suitable for radiomics feature dimensionality reduction compared to PCA stems from their fundamental methodological differences. LASSO, in particular, stands out as a supervised learning method that considers the outcome variable (distinguishing between Grade 1 and Grade 2-3 tumors) when selecting features. This process potentially retains the features that are most predictive of the outcome. In contrast, PCA is an unsupervised learning technique focused on reducing dimensionality by retaining components that explain the most variance without regard to the outcome. This methodological difference likely explains why supervised approaches like LASSO are more effective in identifying relevant features for predicting specific outcomes in radiomics studies.

Furthermore, enhancing the commonly used LR feature selection method by employing the more advanced LGBM to distinguish between Grade 1 and Grade 2–3 tumors underscores the superiority of advanced ML algorithms in capturing complex patterns in image data. LGBM is recognized for its efficient handling of large datasets, its robustness against overfitting through techniques such as gradient boosting and regularization, and its ability to model non-linear relationships and interactions between features. These strengths make LGBM particularly well-suited for tasks where the relationship between features and the outcome is intricate and not easily modeled by traditional linear approaches like LR. This advantage is particularly evident when compared with recent state-of-the-art studies published in 2024 and 2025. For instance, recent works by Chen et al,¹¹ Zuo et al,¹² and Koo et al¹³ have successfully validated the robustness of the TPL dimensionality reduction pipeline for IASLC grading; however, these studies predominantly relied on LR-based models for classification. While effective, linear models may fall short in capturing high-order non-linear interactions within radiomics data. In contrast, our study demonstrates that integrating the proven TPL pipeline with the non-linear LGBM classifier yields a superior AUC of 0.869 compared to LR (AUC = 0.818 in our cohort). This direct comparison highlights the novelty of our work: moving beyond standard linear modeling to unlock the full potential of radiomics features through advanced ensemble learning. The adoption of LGBM over LR for feature selection highlights the advantages of advanced machine learning algorithms in capturing the nuanced patterns and interactions within radiomics data, ultimately enhancing the accuracy and reliability of tumor grading in LUAD patients.

Based on the reasoning outlined above, we consider this radiomics process, which incorporates dimensionality reduction techniques (T-test→Pearson→LASSO) followed by LGBM for subsequent feature selection, as detailed in Algorithm 1. This process is defined as the Radiomics (TPL-LGBM) model, aimed at accurately classifying IASLC grading for lung adenocarcinoma based on preoperative CT scans. Furthermore, we paid close attention to the potential risk of overfitting, a common challenge in high-dimensional radiomics studies. By employing LASSO regression with 10-fold cross-validation, we effectively penalized redundant features. Crucially, we monitored the generalization gap between the validation and testing datasets. The results showed a negligible difference in performance, with an AUC of 0.872 in the validation set and 0.869 in the testing set. This consistency confirms that our model has achieved a robust balance between complexity and generalizability, effectively mitigating the risk of overfitting. We also addressed the potential concern regarding the testing set size (n = 333) and class distribution. Although the prevalence of pure Grade 3 tumors was relatively low, our study design focused on the binary differentiation of Grade 1 versus Grade 2–3 to align with surgical management strategies. In this context, the testing set comprised 122 Grade 1 cases and 211 Grade 2–3 cases. This distribution provides a sufficient and balanced sample size to draw statistically confident conclusions regarding the model's generalization capability. To visually display the effects of the classification task, we employed the t-SNE technique.

t-SNE is a nonlinear dimensionality reduction method that explores the structure of high-dimensional data. It is particularly suitable for visualizing high-dimensional data because it effectively preserves the local structure of the original data in a lower-dimensional space. This characteristic has contributed to the growing popularity of t-SNE in the fields of ML and data analysis. The core idea behind t-SNE involves defining a probability distribution between data points in the high-dimensional space to represent the similarity between them. The t-SNE algorithm aims to ensure that the similarity matrix in the higher-dimensional space aligns as closely as possible with the similarity matrix in the lower-dimensional space. By minimizing the difference between these two distributions using Kullback-Leibler (KL) divergence, t-SNE effectively maps high-dimensional data into a low-dimensional space for visualization. In the gradient descent calculation, the input consists of the coordinates of each point in the low-dimensional space. These coordinates are the parameters we want to optimize, while the output represents the similarity between points in this lower-dimensional space, calculated from the current low-dimensional coordinates. The labels correspond to the similarity between points in the high-dimensional space, derived from the original raw high-dimensional data. Our goal is to adjust the coordinates (ie, inputs) of a point in the low-dimensional space so that the calculated similarity from these coordinates (ie, the output) closely matches the similarity of the known high-dimensional space (ie, the labels). To achieve this, we compute the gradient of the loss function (ie, the KL divergence) with respect to each low-dimensional coordinate and use this gradient to update the coordinates. This iterative process continues until a predetermined number of iterations is reached or until the change in the low-dimensional coordinates falls below a specified threshold. The t-SNE visualizations display both the original features and the features after dimensionality reduction (Figure 6). From these visualizations, it is evident that the data from the original features are not easily distinguishable, with the two categories appearing entangled. However, after applying dimensionality reduction, the data can be more clearly separated, which enhances the final classification performance. These results strongly indicate that the Radiomics (TPL-LGBM) model utilized in this study is effective.

This study is subject to several limitations. Firstly, being retrospective in nature, it is susceptible to selection bias concerning the retrospective pathologic diagnosis. Secondly, as the study involves a comprehensive analysis of a large multicenter CT database, the use of CT scanners from various manufacturers across multiple medical centers may contribute to variability in the results. Thirdly, although the data were collected from multiple centers, no independent external validation cohort was established, which limits the ability to fully assess the generalizability of the proposed radiomics model across different institutions. Fourthly, a follow-up study was not conducted in this report, which limits the ability of the radiomics model to further associate with surgical strategy selection and patient outcome prediction. Fifthly, regarding the multicenter nature of our data, we employed z-score normalization to standardize feature scales. However, we did not apply advanced harmonization techniques, such as ComBat, to explicitly correct for scanner-specific batch effects. While our model showed robust performance across centers, future studies could incorporate such harmonization methods to further minimize potential variations between different imaging protocols. Finally, although the proposed dimensionality reduction and classification framework demonstrated promising performance for IASLC histological grading, its generalizability and broader clinical applicability—such as for prognostic prediction or differentiation between benign and malignant lesions—have not yet been fully validated. Further studies with expanded datasets, independent external validation, and diverse clinical tasks are warranted to confirm its robustness and clinical utility.

Conclusion

In this study, we demonstrated that the T- test→Pearson→LASSO approach is more suitable for radiomics feature dimensionality reduction compared to PCA. Furthermore, we enhanced the commonly used LR feature selection method in medical research by employing the more advanced LGBM technique to distinguish between Grade 1 and Grade 2–3 tumors, aligning with the preoperative IASLC grading system for LUAD. The Radiomics (TPL-LGBM) model shows promising performance and may assist physicians in formulating comprehensive treatment strategies.

Supplemental Digital Content 1

CT Scan Parameters

Preoperative chest CT examinations were performed at several institutions using multidetector CT scanners: Center 1 (uCT550 or CT760, Shanghai United Imaging Healthcare, Shanghai, China), Center 2 (Brilliance iCT, Philips Healthcare, Best, the Netherlands), and Center 3 (Revolution CT 256-slice, GE Healthcare, Chicago, IL, USA).

Patients were scanned in the supine position using helical acquisition and were asked to hold their breath during imaging. Coverage extended from the lung apices to below the costophrenic angles. Volumetric acquisition parameters were: tube voltage 120 kV, tube current–time product 180–280 mA, and matrix size 512×512. Images underwent standard preprocessing and were reconstructed with a standard algorithm; slice thickness and reconstruction interval ranged from 0.55 to 0.625 mm.

Footnotes

Abbreviations

Acknowledgments

We would like to express our sincere appreciation to Editage () for their language editing services, which have significantly improved the quality and readability of this manuscript.

ORCID iDs

Xiaohong Fan

Ethical Approval

This study was approved by the local ethics committees: the Medical Ethics Committee of Xiangtan Central Hospital (Approval No. 2021-07-009, July 19, 2021), the Clinical Trial Ethics Committee of the Affiliated Hospital of Southwest Medical University (Approval No. KY2020147, August 12, 2020), and the Medical Ethics Committee of the Affiliated Hospital of Guilin Medical University (Approval No. 2023YJSLL-121, December 22, 2023).

Consent to Participate

Given the retrospective nature of our investigation, informed consent was obtained from all participants.

Consent for Publication

Not applicable.

Author Contributors

(I) Conception and design: Xiaohong Fan, Zhichao Zuo; (II) Administrative support: Zhichao Zuo, Jinqiu Deng; (III) Provision of study materials or patients: Yujie Feng; (IV) Collection and assembly of data: Wanyin Qi, Shanyue Lin; (V) Data analysis and interpretation: Ying Zeng, Xiaohong Fan; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability

The data for this study are available by contacting the corresponding author upon reasonable request.

Guarantor

Zhichao Zuo(zuozhichao1991@sina.com).

References

Zuo

Zhang

Peng

, et al. Survival nomogram for stage IB non-small-cell lung cancer patients, based on the seer database and an external validation cohort. Ann Surg Oncol. 2021;28(7):3941-3950. DOI:10.1245/s10434-020-09362-0

Moreira

Ocampo

Xia

, et al. A grading system for invasive pulmonary adenocarcinoma: a proposal from the international association for the study of lung cancer pathology committee. J Thorac Oncol. 2020;15(10):1599-1610. DOI:10.1016/j.jtho.2020.06.001

Rokutan-Kurata

Yoshizawa

Ueno

, et al. Validation study of the international association for the study of lung cancer histologic grading system of invasive lung adenocarcinoma. J Thorac Oncol. 2021;16(10):1753-1758. DOI:10.1016/j.jtho.2021.04.008

Asamura

Nishimura

Giroux

, et al. IASLC lung cancer staging project: the new database to inform revisions in the ninth edition of the Tnm classification of lung cancer. J Thorac Oncol. 2023;18(5):564-575. DOI:10.1016/j.jtho.2023.01.088

Kamigaichi

Hamada

Tsutani

. Segmentectomy for patients with early-stage pure-solid non-small cell lung cancer. Front Oncol. 2023;13:1287088. DOI:10.3389/fonc.2023.1287088

Jeon

Lee

Shin

, et al. Prognostic impact of micropapillary and solid histological subtype on patients undergoing curative resection for stage I lung adenocarcinoma according to the extent of pulmonary resection and lymph node assessment. Lung Cancer. 2022;168:21-29. DOI:10.1016/j.lungcan.2022.04.005

Chen

Copley

Viola

, et al. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol. 2023;93:97-113. DOI:10.1016/j.semcancer.2023.05.004

Tunali

Gillies

Schabath

. Application of radiomics and artificial intelligence for lung cancer precision medicine. Cold Spring Harbor Perspect Med. 2021;11(8):a039537. DOI:10.1101/cshperspect.a039537

Zuo

Zeng

Deng

, et al. Intratumoral heterogeneity score enhances invasiveness prediction in pulmonary ground-glass nodules via stacking ensemble machine learning. Insights Imaging. 2025;16(1):209.

10.

Yang

Cai

Chen

, et al. A CT-based radiomics nomogram combined with clinic-radiological characteristics for preoperative prediction of the novel IASLC grading of invasive pulmonary adenocarcinoma. Acad Radiol. 2023;30(9):1946-1961. DOI:10.1016/j.acra.2022.12.006

11.

Chen

You

, et al. Integrating IASLC grading and radiomics for predicting postoperative outcomes in stage IA invasive lung adenocarcinoma. Med Phys. 2024;51(9):6513-6524. DOI:10.1002/mp.17177

12.

Zuo

Zhang

Lin

, et al. Radiomics nomogram based on optimal volume of interest derived from high-resolution CT for preoperative prediction of iaslc grading in clinical IA lung adenocarcinomas: a multi-center, large-population study. Technol Cancer Res Treat. 2024;23:15330338241300734. DOI:10.1177/15330338241300734

13.

Koo

Kim

Lee

, et al. Deciphering the intratumoral histologic heterogeneity of lung adenocarcinoma using radiomics. Eur Radiol. 2025;35(8):4861-4872. DOI:10.1007/s00330-025-11397-4

14.

Zuo

Fan

Zeng

. Multiperspective tumor heterogeneity metrics for preoperative prediction of IASLC grading in clinical stage IA lung adenocarcinomas: a multicenter study. Comput Methods Programs Biomed. 2025;274:109137.

15.

Zuo

Deng

, et al. Quantifying intratumoral heterogeneity within sub-regions to predict high-grade patterns in clinical stage I solid lung adenocarcinoma. BMC Cancer. 2025;25(1):51. DOI:10.1186/s12885-025-13445-0

16.

Demircioglu

. Benchmarking feature selection methods in radiomics. Invest Radiol. 2022;57(7):433-443. DOI:10.1097/RLI.0000000000000855

17.

Parmar

Grossmann

Bussink

, et al. Machine learning methods for quantitative radiomic biomarkers. Sci Rep. 2015;5:13087. DOI:10.1038/srep13087

18.

NQK

Kha

Nguyen

, et al. Machine learning-based radiomics signatures for EGFR and KRAS mutations prediction in non-small-cell lung cancer. Int J Mol Sci. 2021;22(17):9254. DOI:10.3390/ijms22179254

19.

Huang

Lin

Xie

, et al. Development and validation of a preoperative CT-based radiomic nomogram to predict pathology invasiveness in patients with a solitary pulmonary nodule: a machine learning approach, multicenter, diagnostic study. Eur Radiol. 2021;32(3):1983-1996. DOI:10.1007/s00330-021-08268-z

20.

Kirienko

Sollini

Corbetta

, et al. Radiomics and gene expression profile to characterise the disease and predict outcome in patients with lung cancer. Eur J Nucl Med Mol Imaging. 2021;48(11):3643-3655. DOI:10.1007/s00259-021-05371-7

21.

Zuo

Zhang

Chen

, et al. Ct radiomic nomogram using optimal volume of interest for preoperatively predicting invasive mucinous adenocarcinomas in patients with INCI- dental pulmonary nodules: a multicenter, large-scale study. Technol Cancer Res Treat. 2024;23:15330338241308307. DOI:10.1177/15330338241308307

22.

van Griethuysen

Fedorov

Parmar

, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104-e107. DOI:10.1158/0008-5472.CAN-17-0339

23.

Box

. Guinness, gosset, fisher, and small samples. Stat Sci. 1987;2(1):45-52. DOI:10.1214/SS/1177013437

24.

Pearson

. Mathematical contributions to the theory of evolution. III. regression, heredity, and panmixia. Philos Trans R Soc A. 1896;187:253-318.

25.

Ling

Zuo

, et al. Predicting neoadjuvant chemotherapy response in locally advanced gastric cancer using a machine learning model combining radiomics and clinical biomarkers. Digital Health. 2025;11:20552076251341740. DOI:10.1177/20552076251341740

26.

Yang

Fan

Lin

, et al. Assessment of lymphovascular invasion in breast cancer using a combined MRI morphological features, radiomics, and deep learning approach based on dynamic contrast-enhanced MRI. J Magn Reson Imaging. 2023;59(6):2238-2249. DOI:10.1002/jmri.29060

27.

Zuo

Fan

Tang

, et al. Deep learning-powered 3D segmentation derives factors associated with lymphovascular invasion and prognosis in clinical t1 stage non-small cell lung cancer. Heliyon. 2023;9(4):e15147. DOI:10.1016/j.heliyon.2023.e15147

28.

Jiang

Zeng

Zuo

, et al. Leveraging multimodal MRI-based radiomics analysis with diverse machine learning models to evaluate lymphovascular invasion in clinically node-negative breast cancer. Heliyon. 2024;10(1):e23916. DOI:10.1016/j.heliyon.2023.e23916

29.

Efron

Hastie

Johnstone

, et al. Least angle regression. Ann Stat. 2004;32(2):407-499. DOI:10.1214/009053604000000067

30.

Yang

, et al. Machine learning for lung cancer diagnosis, treatment, and prognosis. Genom Proteom Bioinform. 2022;20(5):850-866. DOI:10.1016/j.gpb.2022.11.003

31.

Huang

Yang

Shen

, et al. Artificial intelligence in lung cancer diagnosis and prognosis: current application and future perspective. Semin Cancer Biol. 2023;89:30-37. DOI:10.1016/j.semcancer.2023.01.006

32.

Gortmaker

. Theory and methods–applied logistic regression by David W. Hosmer Jr and Stanley Lemeshow. Contemp Sociol. 1994;23(1):159. DOI:10.1007/978-3-030-85292-45

33.

Cutler

Stevens

. Random forests. In Ensemble machine learning: methods and applications, 2012: 157-175.

34.

Friedman

. Greedy function approximation: a gradient boosting machine. Ann Stat. 2001;29(5):1189-1232. DOI:10.1214/aos/1013203451

35.

Chen

Guestrin

. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. ACM, pp. 785-794. DOI:10.1145/2939672.2939785.

36.

Meng

Finley

, et al. Lightgbm: A highly efficient gradient boosting decision tree. In Advances in neural information processing systems (NIPS), pp. 3149-3157.

37.

Prokhorenkova

Gusev

Vorobev

, et al. Catboost: unbiased boosting with categorical features. In Advances in neural information processing systems (NIPS), pp. 6639- 6649.

38.

Pedregosa

Varoquaux

Gramfort

, et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825-2830.

39.

Buitinck

Louppe

Blondel

, et al. API design for machine learning software: experiences from the scikit-learn project. In ECML PKDD Workshop: languages for data mining and machine learning, pp. 108-122.

40.

Tipping

Bishop

. Probabilistic principal component analysis. J R Stat Soc Ser B: Stat Methodol. 1999;61(3):611-622.

41.

Minka

. Automatic choice of dimensionality for PCA. In Advances in neural information processing systems (NIPS), pp. 598-604.

42.

Halko

Martinsson

Tropp

. Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions. SIAM Rev. 2011;53(2):217-288.