Noninvasive prediction of ALK fusion status in non-small cell lung cancer by a machine learning model combining CT images and clinical information

Abstract

Background

As the primary contributor to global cancer mortality, lung cancer requires precise anaplastic lymphoma kinase (ALK) genotyping to implement personalized targeted treatment for non-small cell lung cancer (NSCLC) patients. Invasive biopsy-based ALK detection is clinically limited by sampling deviation, procedural complications, and insufficient tumor specimens. The objective of this study was to develop and validate a machine learning model that integrates computed tomography (CT) radiomics features with clinicopathological data to non-invasively predict ALK fusion status in patients with NSCLC.

Methods

This retrospective multi-center study enrolled 722 NSCLC patients (291 ALK-positive, 431 ALK-negative). From segmented tumor regions, 2,264 radiomics features were derived. After feature selection using the least absolute shrinkage and selection operator (LASSO) regression, three predictive models were constructed and compared: a clinical & region of interest (ROI) model, a radiomics model, and a combined model integrating both feature types. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity in training, test, and validation cohorts.

Results

The final models included 14, 67, and 53 features in the clinical & ROI model, radiomics model, and combined model, respectively. The combined model demonstrated superior predictive performance, achieving an AUC of 0.997 in the training cohort, 0.988 in the test cohort, and 0.973 in the validation cohort. It significantly outperformed the clinical & ROI model (AUC: 0.997 vs. 0.923 in the training cohort, P<0.001) and showed a superior performance compared to the radiomics model (AUC: 0.997 vs. 0.995 in the training cohort, P=0.18), though not statistically significant.

Conclusions

A machine learning model combining CT radiomics and clinical data exhibited robust performance in predicting ALK fusion status in NSCLC patients. This non-invasive approach shows significant potential as a clinical tool for pre-therapeutic selection of patients who may benefit from ALK-targeted therapies.

Highlight box.

Key findings

• This study developed and validated a machine learning model based on computed tomography (CT) and clinical data that can effectively predict anaplastic lymphoma kinase (ALK) fusion status in patients with non-small cell lung cancer (NSCLC).

What is known and what is new?

• CT radiomics features can non-invasively capture tumor phenotypic information related to molecular alterations in NSCLC, including ALK fusions.

• This study developed and rigorously validated a combined least absolute shrinkage and selection operator (LASSO)-based machine learning model that integrates CT radiomics features with clinical and region of interest (ROI) imaging characteristics. The combined model achieved uniquely high discrimination and outperformed the clinical & ROI model and radiomics model.

What is the implication, and what should change now?

• The study implies that CT radiomics can non-invasively predict ALK status with high accuracy, strengthening the evidence for a non-invasive, imaging-clinical tool to guide ALK-targeted therapy selection.

Introduction

Lung cancer remains the principal cause of cancer-associated mortality on a global scale, with non-small cell lung cancer (NSCLC) contributing to approximately 85% of all lung cancer diagnoses (1). The identification of actionable driver mutations has revolutionized the treatment landscape for NSCLC, with targeted therapies offering improved survival and quality of life compared to traditional chemotherapy (2,3). Among these mutations, the anaplastic lymphoma kinase (ALK) gene rearrangement is a significant predictor of response to ALK tyrosine kinase inhibitors (TKIs), highlighting the importance of accurate and non-invasive methods for mutation detection.

ALK rearrangements are found in approximately 3–13% of patients with NSCLC and are associated with a distinct clinical profile and response to TKIs (4,5). Accurate and early detection of ALK mutations is therefore crucial for guiding personalized treatment strategies. Traditional methods for detecting ALK mutations, such as fluorescence in situ hybridization (FISH) and immunohistochemistry (IHC), require invasive tissue biopsies and are subject to sampling errors and interobserver variability.

The advent of radiomics, the extraction of high-dimensional data from medical images, has introduced a non-invasive approach to characterize tumors and predict molecular profiles (6-8). Computed tomography (CT) scans, being widely available and routinely used in clinical practice, offer a rich source of data for radiomics analysis. Recent studies have demonstrated the potential of CT radiomics in predicting various molecular alterations in lung cancer, including epidermal growth factor receptor (EGFR) mutations and programmed cell death ligand 1 (PD-L1) expression levels (9,10). The ability to predict ALK mutation status from CT images would be a significant advancement, allowing for the stratification of patients who may benefit from ALK-targeted therapies without the need for invasive procedures. Several studies have begun to explore the relationship between CT radiomics features and ALK mutation status. For instance, one study retrospectively analyzed CT images from lung adenocarcinoma patients and identified a set of radiomics features that could predict ALK fusion status with considerable accuracy (11). Another investigation integrated radiomics features with clinical and traditional CT features to establish a predictive model for ALK mutation status, achieving high accuracy (12).

While these initial findings are promising, the development of a comprehensive radiomics model that can accurately predict ALK mutation status in NSCLC patients remains an area of active research. This study aims to establish and validate a radiomics model based on contrast-enhanced CT scans to predict ALK mutation status and differentiate ALK from EGFR mutations in NSCLC patients. We hypothesized that a radiomics signature, derived from quantitative imaging features, can serve as a robust predictor of ALK mutation status and potentially guide personalized treatment strategies. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1340/rc).

Methods

Patient population

This retrospective study reviewed patients with NSCLC pathologically confirmed by surgery or biopsy at three hospitals (Beijing Friendship Hospital, Capital Medical University; Beijing Chest Hospital, Capital Medical University; and China-Japan Friendship Hospital) from January 2013 to July 2024. ALK rearrangement status was determined according to the standard clinical diagnostic protocols at each participating center, using FISH, reverse transcription-polymerase chain reaction (RT-PCR), targeted next-generation sequencing (NGS), or IHC. All results were obtained from the medical records of the respective centers. The study enrolled patients over 18 years of age who satisfied the following inclusion criteria: (I) complete ALK mutation gene test results; (II) acquisition of preoperative CT data; and (III) documentation of clinicopathological information, including smoking history, age, sex, and histological subtype of lung cancer. Exclusion criteria included: (I) CT images with severe artifacts; (II) receipt of any treatment prior to CT examination; (III) an interval exceeding one month between CT examination and surgery or biopsy. A matched case-control design was employed for the selection of controls. Specifically, ALK-negative NSCLC patients were matched to ALK-positive cases at a ratio of approximately 1:1.5, with matching based on prespecified key confounders. All enrolled individuals fulfilled the identical set of inclusion and exclusion criteria.

Based on these criteria, 722 patients (431 ALK-negative and 291 ALK-positive) were deemed eligible for the investigation. Eighty percent of the cases were randomly selected from the ALK-positive and ALK-negative patients, respectively, to build an independent training cohort (577 cases, 233 ALK-positive and 344 ALK-negative), while the remainder were allocated into two test cohorts (test cohort: 72 cases, 29 ALK-positive and 43 ALK-negative; validation cohort: 73 cases, 29 ALK-positive and 44 ALK-negative). The flowchart of the eligibility and exclusion criteria is shown in Figure 1. The tumor lesions were all solitary. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (No. 2024-P2-357-01) and individual consent for this retrospective analysis was waived. All participating hospitals were informed of and agreed to the study.

Figure 1.

Flowchart illustrating the process of patient selection in this study. ALK, anaplastic lymphoma kinase; CT, computed tomography; NSCLC, non-small cell lung cancer.

Clinicopathological data

The clinical characteristics and pathological data of all patients were obtained from the electronic medical records system. Demographic data collected from patients included age, sex, and smoking history.

Pathological data included T stage, N stage, M stage, tumor TNM stage, tumor type and tumor location. The χ² test was used to determine whether there were significant differences between the two groups for categorical variables, while continuous variables were compared using the Mann-Whitney U test.

CT image acquisition

High-resolution CT images were obtained using scanners manufactured by Siemens, Philips, or GE. The scanning protocols utilized a tube voltage of 100–120 kVp and a tube current ranging from 60 to 250 mAs. Prior to image acquisition, patients received breath-hold training and were subsequently scanned in the supine position at full inspiration with suspended respiration. The scanning volume spanned from the thoracic inlet down to the costophrenic angle.

Radiomics feature extraction and selection

Figure 2 illustrates the methodological workflow implemented via the uAI Research Portal (Shanghai United Imaging Intelligent Medical Technology Co., Ltd., Shanghai, China), an online artificial intelligence-driven research website specifically designed for medical image analysis and radiomics. This platform provides a comprehensive suite of machine learning algorithms and supports end-to-end workflows from data preprocessing to model evaluation. Initially, CT datasets were uploaded to the platform. Utilizing a semiautomatic segmentation tool, we delineated the region of interest (ROI) layer-by-layer through interactive manual correction, thereby constructing a three-dimensional volumetric ROI. Following segmentation, radiomics features were extracted from within this volume. The feature set encompasses various categories, including first-order statistics, gray-level co-occurrence matrices (GLCM), and gray-level dependence matrices (GLDM). Feature extraction was then conducted with 24 imaging filters in the uAI research portal, resulting in 2,264 features. Finally, Z-score normalization was performed on the entire feature set to standardize the distribution (mean = 0, standard deviation = 1). The K-best method was first used to remove features with low correlation with classification labels in order to reduce computational complexity and prevent overfitting. Then, least absolute shrinkage and selection operator (LASSO) was used to remove redundant features.

Figure 2.

The workflow of data analysis. (A) CT images and clinical features of NSCLC patients were retrospectively collected. (B) Auto-detection, segmentation, and manual confirmation of the tumor lesion. The initial ROIs are generated in this step. (C) Collection of radiomics, ROI, and clinical features. (D-F) Illustrations of feature selection, model construction, and model evaluation, respectively. ALK, anaplastic lymphoma kinase; CT, computed tomography; Lasso, least absolute shrinkage and selection operator; NSCLC, non-small cell lung cancer; ROI, region of interest.

Model construction and evaluation

To establish robust predictive algorithms, we developed three distinct models based on varying combinations of input variables: clinical features, ROI features, and radiomics features. Specifically, the models were constructed using radiomics features alone, a combination of clinical and ROI features, and a combined model simultaneously leveraging all three categories of features. To optimize the model architecture, LASSO logistic regression was employed for dimensionality reduction, ensuring the retention of only the most discriminative features. Furthermore, the radiomics score (Rad-score) for each individual was derived through a weighted linear summation of the selected features, utilizing their respective LASSO regression coefficients.

Various machine learning algorithms, including random forest (RF), support vector machine (SVM), Gaussian process, logistic regression (LR), K-nearest neighbor (KNN), decision tree, and stochastic gradient descent (SGD) algorithms, were employed for the model training. To rigorously evaluate model performance and predictive efficacy, we employed a comprehensive set of metrics, including the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, accuracy, and F1-score.

Statistical analysis

All machine learning procedures, including algorithm implementation, parameter tuning, model validation, and the DeLong’s test, were performed using the uAI Research Portal, with all underlying computations conducted in Python. Statistical analyses of clinical characteristics for the enrolled subjects were performed using IBM SPSS Statistics software (version 26.0; IBM Corp., Armonk, NY, USA) (13). Continuous variables were reported as mean ± standard deviation for normally distributed data and analyzed using the independent Student’s t-test. In contrast, non-normally distributed data were expressed as medians and compared utilizing the Mann-Whitney U test. Categorical variables were presented as frequencies and evaluated using the χ² test or Fisher’s exact test, as appropriate. To assess the predictive efficacy of the models, we calculated the AUC alongside sensitivity, specificity, accuracy, precision, and the F1-score. The DeLong’s test was employed to statistically compare the differences between ROC curves. Furthermore, decision curve analysis was performed to determine the clinical net benefit of each model across various threshold probabilities. A two-sided P value of less than 0.05 was considered indicative of statistical significance.

Results

Clinicopathological characteristics of patients

A total of 722 patients were included in the analysis, comprising 431 ALK-negative and 291 ALK-positive cases. These 722 patients were divided into three cohorts: training cohort, test cohort, and validation cohort. In each cohort, the ratio of ALK-positive to ALK-negative patients was approximately 1:1.5. The baseline clinical and pathological characteristics are summarized in Table 1. Among the entire cohort, 562 cases (77.8%) underwent surgical resection and 160 cases (22.2%) underwent diagnostic biopsy. There were significant differences in age, sex, T stage, N stage, M stage, and TNM stage between ALK-positive and ALK-negative patients (P<0.001). Compared with the ALK-negative group, ALK-positive patients were younger and had a higher proportion of female patients. Contrary to previous studies, the proportion of distant metastasis in ALK-positive patients was lower in this study. The difference in tumor location between the two groups was statistically significant. ALK-positive tumors were less frequently located in the right upper lobe (17.9% vs. 31.1%) and more frequently in the left upper lobe (35.4% vs. 25.3%) compared with ALK-negative tumors. Additionally, the proportion of patients without smoking history was higher in the ALK-positive group compared with the ALK-negative group (P=0.005).

Table 1. Clinical characteristics of ALK⁻ and ALK⁺ NSCLC patients in this study.

Characteristics	ALK− (n=431)	ALK+ (n=291)	P value
Age at diagnosis, years, median (range)	66 (59–70)	55 (46–63)	<0.001
Sex, n (%)			<0.001
Male	195 (45.2)	172 (59.1)
Female	236 (54.8)	119 (40.9)
Smoking history, n (%)			0.005
No	295 (68.4)	227 (78.0)
Yes	136 (31.6)	64 (22.0)
Tumor pathology, n (%)			0.002
LUAD	393 (91.2)	278 (95.6)
LUSC	21 (4.9)	7 (2.4)
LCNEC	1 (0.2)	3 (1.0)
ASC	1 (0.2)	3 (1.0)
Others	15 (3.5)	0 (0.0)
Lobe, n (%)			0.001
LUL	109 (25.3)	103 (35.4)
LLL	76 (17.6)	61 (21.0)
RUL	134 (31.1)	52 (17.9)
RML	32 (7.4)	21 (7.2)
RLL	80 (18.6)	54 (18.5)
T stage, n (%)			<0.001
T1a	44 (10.2)	83 (28.5)
T1b	105 (24.4)	81 (27.8)
T1c	83 (19.3)	41 (14.1)
T2a	69 (16.0)	44 (15.1)
T2b	48 (11.1)	16 (5.5)
T3	45 (10.4)	18 (6.2)
T4	37 (8.6)	8 (2.8)
N stage, n (%)			<0.001
N0	249 (57.8)	200 (68.7)
N1	29 (6.7)	36 (12.4)
N2	84 (19.5)	54 (18.6)
N3	69 (16.0)	1 (0.3)
M stage, n (%)			<0.001
M0	327 (75.9)	280 (96.2)
M1	104 (24.1)	11 (3.8)
Clinical stage, n (%)			<0.001
IA1	38 (8.8)	82 (28.2)
IA2	94 (21.8)	57 (19.6)
IA3	53 (12.3)	23 (7.9)
IB	37 (8.6)	40 (13.7)
IIA	10 (2.3)	4 (1.4)
IIB	18 (4.2)	13 (4.5)
IIIA	45 (10.4)	25 (8.6)
IIIB	24 (5.6)	24 (8.2)
IIIC	9 (2.1)	11 (3.8)
IV	103 (23.9)	12 (4.1)

ALK, anaplastic lymphoma kinase; ASC, adenosquamous carcinoma of the lung; LCNEC, large cell neuroendocrine carcinoma; LLL, left lower lobe; LUAD, lung adenocarcinoma; LUL, left upper lobe; LUSC, lung squamous cell carcinoma; NSCLC, non-small cell lung cancer; RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe.

Feature selection and model construction

Three predictive models were developed based on varying combinations of clinical, ROI, and radiomics features. Specifically, models were built using radiomics features alone, a combination of clinical and ROI features and finally, a combined model incorporating all three types of features. The process of the feature selection sequence in different models was depicted in Figure 3.

Figure 3.

Illustration of the feature selection procedure in the three models. Each vertical panel exhibits the selection process for each of the three predictive models. Each symbol indicates a different type of feature. The number of selected features along with the optimal AUC obtained at each selection step is shown at the top of each sub-panel. In the radiomics model, 2,264 extracted radiomics features were used to begin the selection. In the clinical & ROI model, the initial features included 9 clinical features and 22 ROI features. In the combined model, the initial features included 2,264 radiomics features and 31 clinical and tumor ROI imaging features. The features were selected to maximize the AUC of the predictive model at the final step. AUC, area under receiver operating characteristic curve; Lasso, least absolute shrinkage and selection operator; ROI, region of interest.

From 2,264 radiomics features, we performed feature selection and 67 features were selected by the LASSO regression algorithm, including 7 first-order statistical features, and 60 texture features, which were subsequently used to establish a radiomics model. Subsequently, LASSO regression was performed, and 14 features—comprising both clinical and ROI features—were selected from the initial 31 clinical and tumor ROI imaging features: minimum CT value of tumor, maximum intensity of tumor, minimum intensity of tumor, standard deviation of tumor intensity, malignancy probability, location of nodule center on Y axis, location of nodule center on Z axis, nodule width (mm), M stage, N stage, T stage, age, sex, TNM stage. These 14 features were used to construct a clinical & ROI model. Finally, we selected 53 features from 2,264 radiomics features in conjunction with 31 clinical and ROI features using LASSO regression. These 53 features were used to construct a combined model.

Model evaluation and comparison of predictive model performance

The performance of the three models in the training, test, and validation cohorts was evaluated using the AUC, sensitivity, specificity, accuracy, precision, and F1-score (Figure 4). A detailed overview of the diagnostic performance of the three models is presented in Table 2. The optimal cut-off value of the combined model was determined to be 0.45. At the optimal probability threshold determined by the Youden index, our model demonstrated excellent discriminative ability, with AUC values of 0.997, 0.988, and 0.973 in the training, test, and validation sets, respectively. At this threshold, the sensitivity and specificity were 0.961 and 0.988 for the training set, 0.966 and 0.954 for the test set, and 0.828 and 0.977 for the validation set, indicating satisfactory generalizability and robust classification performance. The performance of the three models was compared by DeLong’s test and the results of DeLong’s test were listed in Table 3. In the training cohort, DeLong’s test showed that the difference in AUC between the combined model and the clinical & ROI model was statistically significant (P<0.001), and the AUC between the radiomics model and the clinical & ROI model was also statistically significant (P<0.001). However, no statistically significant difference in AUC was observed between the combined model and the radiomics model (P=0.18). Despite the combined model showing the highest AUC among the three predictive models, the DeLong’s test indicated no statistically significant differences in AUC across the three models in either the test cohort (combined vs. clinical & ROI P=0.39; combined vs. radiomics P=0.10; radiomics vs. clinical & ROI P=0.66) or validation cohort (combined vs. clinical & ROI P=0.08; combined vs. radiomics P=0.21; radiomics vs. clinical & ROI P=0.27).

Figure 4.

The performance of the radiomics model, clinical & ROI model and combined model was evaluated in the training (A-C), test (D-F), and validation cohorts (G-I). AUC, area under receiver operating characteristic curve; BS, Brier score; ROI, region of interest.

Table 2. Diagnostic performance of each model in the training, test, and validation cohorts.

Models	AUC (95% CI)	Sensitivity	Specificity	Accuracy	Precision	F1-score
Clinical & ROI model
Training	0.923 (0.901–0.945)	0.811	0.878	0.851	0.818	0.815
Test	0.979 (0.953–1.000)	0.931	0.930	0.900	0.958	0.915
Validation	0.928 (0.868–0.988)	0.793	0.886	0.849	0.821	0.807
Radiomics model
Training	0.995 (0.991–0.999)	0.951	0.966	0.962	0.931	0.954
Test	0.973 (0.944–1.000)	0.861	0.954	0.903	0.824	0.889
Validation	0.958 (0.916–1.000)	0.724	0.955	0.863	0.913	0.808
Combined model
Training	0.997 (0.994–1.000)	0.961	0.988	0.978	0.983	0.972
Test	0.988 (0.972–1.000)	0.966	0.954	0.958	0.933	0.949
Validation	0.973 (0.940–1.000)	0.828	0.977	0.918	0.960	0.889

AUC, area under receiver operating characteristic curve; CI, confidence interval; ROI, region of interest.

Table 3. Comparison of models by DeLong test, NRI, IDI, and DCA in the training, test, and validation cohorts.

Cohorts	P value
	AUC†	NRI	IDI	DCA
Training
Combined model vs. clinical & ROI model	<0.001	<0.001	<0.001	<0.001 (0.055 to 0.083)
Combined model vs. radiomics model	0.18	<0.001	<0.001	<0.001 (0.008 to 0.016)
Clinical & ROI model vs. radiomics model	<0.001	<0.001	<0.001	<0.001 (−0.073 to −0.042)
Test
Combined model vs. clinical & ROI model	0.39	>0.99	0.23	0.24 (−0.011 to 0.050)
Combined model vs. radiomics model	0.10	0.37	0.001	0.002 (0.009 to 0.050)
Clinical & ROI model vs. radiomics model	0.66	0.51	0.35	0.64 (−0.029 to 0.052)
Validation
Combined model vs. clinical & ROI model	0.08	0.049	0.03	0.12 (−0.008 to 0.063)
Combined model vs. radiomics model	0.21	0.11	0.006	0.004 (0.008 to 0.057)
Clinical & ROI model vs. radiomics model	0.27	0.99	0.98	0.92 (−0.042 to 0.052)

^†, DeLong test. AUC, area under receiver operating characteristic curve; DCA, decision curve analysis; IDI, integrated discrimination improvement; NRI, net reclassification improvement; ROI, region of interest.

Discussion

In this study, we developed and validated a machine learning-based predictive model integrating CT radiomics features and clinical information to noninvasively predict ALK fusion status in patients with NSCLC. Our results demonstrated that the combined model, which incorporates radiomics features along with clinical and ROI characteristics, achieved superior predictive performance compared to models using either radiomics or clinical & ROI features alone. The combined model exhibited excellent discrimination in the training, test, and validation cohorts, with AUC values of 0.997, 0.988, and 0.973, respectively. Our findings support the potential of combining radiomics features and clinical features as a noninvasive tool for molecular profiling, which could aid in stratifying patients for targeted therapy and reducing reliance on invasive biopsies. With further validation and refinement, such an approach may contribute to more personalized and efficient management of NSCLC.

The high predictive accuracy of the combined model underscores the complementary value of integrating multimodal data for molecular profiling. While radiomics features can capture intratumoral heterogeneity and phenotypic characteristics that are not discernible to the human eye, clinical variables such as age, sex, smoking history, and TNM stage provide essential contextual information that reflects the biological behavior of the tumor. The synergy between these data types enhances the model’s ability to distinguish ALK-positive from ALK-negative tumors, which is consistent with previous studies emphasizing the benefit of combining radiomics with clinical parameters (14,15).

Notably, our radiomics model alone also showed strong predictive capability, suggesting that quantitative radiomics features extracted from routine CT scans contain meaningful biological information related to ALK fusion. This is in line with emerging evidence that radiomics signatures can serve as noninvasive biomarkers for genetic alterations in NSCLC (11,16,17). However, the further improvement in performance seen in the combined model highlights the importance of a holistic approach that leverages both imaging and clinical data.

Several clinical characteristics were significantly associated with ALK positivity in our cohort, including younger age, female sex, and non-smoking history—findings that align with established epidemiological profiles of ALK-rearranged NSCLC (4,18). Interestingly, we also observed a lower rate of distant metastasis in ALK-positive patients compared to ALK-negative ones, which contrasts with some previous reports. This discrepancy may be attributable to differences in cohort composition or early detection rates in our study population.

Despite the favorable results, certain limitations should be acknowledged. First, the retrospective nature of the study, which carries a potential risk of selection bias, although we employed rigorous inclusion criteria and multi-center data to enhance generalizability. Second, the use of CT scans from different manufacturers and protocols may affect feature reproducibility, despite our efforts to standardize preprocessing and feature extraction. Future prospective studies with standardized imaging protocols are warranted to validate our findings. Third, while we included a relatively large sample size, external validation in more diverse populations is needed to ensure model robustness and clinical applicability.

In addition, the survival data of ALK-positive patients hold considerable clinical relevance. In our future research, we plan to collect longitudinal follow-up data from patients with ALK-positive NSCLC who have received ALK-TKI therapy and to extend our predictive framework to incorporate treatment response and survival benefit. This would allow stratification of patients likely to derive benefit from targeted therapy based on progression-free survival or overall survival. Such efforts will not only further validate the clinical utility of our radiomics biomarker, but may also provide valuable guidance for personalized treatment decisions.

Furthermore, although the combined model showed the highest AUC in the training cohort, the differences between models were not always statistically significant in the test and validation sets, as indicated by the DeLong’s test. This suggests that while the combined model is numerically superior, its incremental benefit over the radiomics-only model may be modest in certain settings. Future work should focus on optimizing feature selection and model interpretability to facilitate clinical translation.

Conclusions

In summary, we developed a machine learning model that effectively integrates CT radiomics features with clinical information to predict ALK fusion status in NSCLC patients. The combined model demonstrated high predictive accuracy and robustness across training, test, and validation cohorts, outperforming models based solely on clinical or radiomics features.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (No. 2024-P2-357-01) and individual consent for this retrospective analysis was waived. All participating hospitals were informed of and agreed to the study.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1340/dss

jtd-18-08-885-dss.pdf ^{(76.2KB, pdf)}

DOI: 10.21037/jtd-2026-1340