Diagnostic performance of machine learning-based radiomics models for predicting epidermal growth factor receptor mutation status in lung adenocarcinoma in Chinese patients: A systematic review and meta-analysis

Abstract

Objective

This systematic review and meta-analysis evaluates the diagnostic performance of machine learning-based radiomics models for predicting epidermal growth factor receptor mutation status in Chinese patients with lung adenocarcinoma.

Methods

Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines and prospectively registered in the International Prospective Register of Systematic Reviews (CRD420251273027), a systematic search of PubMed, Embase, Web of Science, the Cochrane Library, Scopus, China National Knowledge Infrastructure, Wanfang, VIP, and Chinese Biomedical Literature Database was conducted from inception to 31 October 2025. Two reviewers independently screened studies, extracted data, and assessed bias using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. A bivariate random-effects model was used to synthesize the data. Subgroup analyses were conducted for three factors: (a) imaging modality (computed tomography vs. positron emission tomography-computed tomography); (b) algorithm type (deep learning vs. conventional machine learning); and (3) validation strategy (external vs. internal).

Results

Thirteen studies encompassing 6628 patients were included. The pooled sensitivity was 71% (95% confidence interval: 68–74), the pooled specificity was 81% (95% confidence interval: 78–84), and the summary area under the curve was 0.85 (95% confidence interval: 0.82–0.88). Deep learning models significantly outperformed conventional machine learning models (area under the curve: 0.871 vs. 0.798; P = 0.012). Computed tomography-based models yielded higher accuracy than positron emission tomography-computed tomography-based models (area under the curve: 0.879 vs. 0.828; P = 0.038). Models validated on independent external cohorts demonstrated superior performance compared with those relying solely on internal validation (area under the curve: 0.922 vs. 0.841; P = 0.006). Imaging modality was a significant source of heterogeneity (P < 0.05). No threshold effect or publication bias was detected.

Conclusion

Machine learning-based radiomics models exhibit promising diagnostic accuracy for the noninvasive prediction of epidermal growth factor receptor mutations in Chinese patients with lung adenocarcinoma. Computed tomography-based deep learning models subjected to independent external validation represent the current optimal approach. However, the retrospective nature and substantial heterogeneity of the included studies necessitate large-scale, prospective, multicenter trials with standardized workflows before clinical translation.

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide. Non-small cell lung cancer (NSCLC) constitutes 80%–85% of all cases, with lung adenocarcinoma (LUAD) being the predominant histological subtype.^1,2 In Asian populations, epidermal growth factor receptor (EGFR) mutations occur in nearly 50% of patients with LUAD, a prevalence markedly higher than the 15% observed in Western cohorts. ³ Tumors harboring these sensitizing mutations respond favorably to tyrosine kinase inhibitors (TKIs), which significantly prolong progression-free and overall survival.^4,5

Currently, tissue biopsy coupled with molecular genetic testing remains the gold standard for determining EGFR status. ⁶ However, this invasive approach carries inherent risks of sampling error, procedural complications, and delays in obtaining results, fundamentally limiting its utility for real-time monitoring.

Radiomics offers a noninvasive alternative by extracting high-dimensional quantitative features from routine medical images. These features are used to develop predictive models based on machine learning (ML) algorithms, ranging from conventional methods (e.g. support vector machines and random forests) to deep learning (DL) architectures such as convolutional neural networks. ⁷ While numerous studies have explored computed tomography (CT)- or positron emission tomography-computed tomography (PET/CT)-based radiomics for predicting EGFR mutations in LUAD,^8,9 most have been small-scale, single-center, retrospective studies. Substantial methodological variations in algorithm selection, feature processing, and validation strategies have resulted in considerable heterogeneity in performance, underscoring the need for a systematic synthesis of the existing evidence.

Synthesizing these data specifically for Chinese populations is not merely beneficial but imperative. The approximately 50% prevalence of sensitizing EGFR mutations in this population—more than three times that observed in Western cohorts—dramatically alters the pre-test probability. This baseline difference fundamentally changes the clinical utility and performance characteristics of any predictive model. Furthermore, emerging evidence suggests that EGFR mutation subtypes and co-mutation patterns vary across ethnic groups, which may manifest as differences in imaging phenotypes. ¹⁰ A pan-ethnic meta-analysis risks obscuring these population-specific signals, potentially yielding summary estimates that are disconnected from the realities of Chinese clinical practice, where the demand for accurate, noninvasive prediction is greatest. Consequently, this review focuses specifically on sensitizing EGFR mutations (e.g. exon 19 deletions and exon 21 L858R mutations), excluding rare variants or primary resistance mutations.

Materials and methods

Literature search strategy

This systematic review strictly adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement and was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO) (CRD420251273027).

A systematic search of PubMed, Embase, Web of Science, the Cochrane Library, Scopus, China National Knowledge Infrastructure (CNKI), Wanfang, VIP Database, and Chinese Biomedical Literature (CBM) Database was conducted to retrieve all relevant records from inception to 31 October 2025. The search strategy used a combination of Medical Subject Headings (MeSH) terms and free-text keywords across five domains: disease (“lung adenocarcinoma,” “LUAD,” “non-small cell lung cancer”), methodology (“radiomics,” “radiogenomics”), technology (“machine learning,” “deep learning,” “artificial intelligence”); imaging (“computed tomography,” “CT,” “PET/CT”), and biomarker (“EGFR,” “epidermal growth factor receptor mutation”). Manual screening of reference lists was also performed to supplement the electronic search.

Inclusion and exclusion criteria

Retrieved records were managed using EndNote 2025 (Clarivate Analytics). Studies were selected based on predefined eligibility criteria.

The inclusion criteria are as follows:

• Population. Pathologically confirmed LUAD patients without prior anti-tumor therapy;

• Index test. Radiomics models using ML algorithms derived from preoperative CT or PET/CT images;

• Reference standard. EGFR mutation status confirmed by tissue-based molecular assays (e.g. next-generation sequencing, Sanger sequencing, or amplification refractory mutation system);

• Study design. Diagnostic accuracy studies providing sufficient data to construct a 2 × 2 contingency table (true positive (TP), false positive (FP), true negative (TN), false negative (FN));

• Language. English or Chinese.

The exclusion criteria are as follows:

• Conference abstracts, reviews, meta-analyses, case reports, animal studies, and editorials;

• Studies lacking extractable diagnostic accuracy data;

• Duplicate publications (only the most complete or most recent version retained);

• Studies without full-text availability or outside the scope of the review.

Study selection and data extraction

Two investigators independently screened titles, abstracts, and full texts. Any discrepancies were resolved through discussion or by consultation with a third reviewer (G.C.). A standardized data extraction form was used to collect the following information: (a) baseline study characteristics (first author, publication year, and country); (b) patient demographics (sample size, age, sex, and tumor stage); (c) technical parameters (imaging modality, region of interest (ROI) segmentation, ML algorithm, and validation strategy); and (d) diagnostic performance metrics (sensitivity, specificity, accuracy, area under the curve (AUC), TP/FP/TN/FN values). To reduce overfitting bias, data were preferentially extracted from independent test sets or external validation cohorts rather than training sets.

Risk of bias assessment

Two reviewers independently evaluated methodological quality using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 tool. Risk of bias and applicability concerns were evaluated across four domains: patient selection, index test, reference standard, and flow and timing. Each domain was classified as having a low, high, or unclear risk of bias.

Statistical analysis

Statistical analyses were performed using Meta-Disc 1.4 and Stata 18.0. The presence of a threshold effect was evaluated using Spearman's correlation between the logit of sensitivity and the logit of (1−specificity), with P <0.05 indicating statistical significance. Heterogeneity was assessed using Cochran's Q test (P < 0.10) and the I² statistic, with I² >50% indicating substantial heterogeneity.

A bivariate random-effects model was used to pool diagnostic accuracy estimates in the presence of significant heterogeneity; otherwise, a fixed-effects model was applied. Pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and corresponding 95% confidence intervals (CIs) were calculated. A summary receiver operating characteristic (SROC) curve was constructed, with the AUC used as the overall performance metric.

Predefined subgroup analyses were conducted stratified by three factors: (a) imaging modality (CT vs. PET/CT); (b) algorithm type (DL vs. conventional ML); and (c) validation strategy (external vs. internal). In this study, ML refers to the broad category, including conventional ML (e.g. support vector machines (SVMs) and random forests (RFs)) and DL(e.g. convolutional neural networks (CNN) and residual networks (ResNet)). Between-group differences were statistically compared. Univariate meta-regression was performed to explore imaging modality and algorithm type as potential sources of heterogeneity. Galbraith plot analysis was used to assess the robustness of pooled estimates. Publication bias was assessed using Deeks’ funnel plot asymmetry test, P <0.05 indicating significant bias.

Results

Study selection and characteristics

The initial search yielded 1684 records. After removing 412 duplicates, 1272 records were screened based on titles and abstracts. Full-text review of 98 articles resulted in the inclusion of 13 studies in the meta-analysis. Figure 1 illustrates the PRISMA flow diagram.

Figure 1.

PRISMA 2020 flow.

PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses.

All included studies were conducted in China and encompassed 6628 patients (4763 in training cohorts and 1865 in testing cohorts). Seven studies used CT, while six used PET/CT. Three studies employed conventional ML, whereas 10 adopted DL approaches, predominantly CNN. Table 1 summarizes the baseline characteristics.^11–23

Table 1.

Characteristics and diagnostic performance of included studies.

Study (year)	Modality	Algorithm	Validation	TP	FP	FN	TN	Sensitivity (%)	Specificity (%)	AUC (95% CI)
Shao et al.11	PET/CT	DL (CNN)	Internal	48	14	23	27	67.6	65.9	0.73 (0.629–0.830)
Zhang et al.12	CT	DL (SE-CNN)	External	51	13	33	108	60.7	89.3	0.841
Ge et al.13	PET/CT	ML (SVM)	Internal	23	24	23	70	50.0	74.5	0.672 (0.586–0.741)
Huang et al.14	CT	DL (ResNet)	Internal	41	2	6	19	87.9	91.4	0.916
Zhang et al.15	CT	DL (3D-CNN)	External	49	7	16	60	75.4	89.6	0.874 (0.820–0.924)
Gao et al.16	PET/CT	ML (RF)	Internal	56	18	14	23	80.0	56.1	0.704
Huang et al.17	PET/CT	DL (Inc-v3)	Internal	52	13	6	57	89.4	81.4	0.921 (0.904–0.937)
Wang et al.18	CT	DL (3D-CNN)	Internal	38	11	10	41	78.6	78.6	0.781 (0.638–0.923)
Deng et al.19	PET/CT	ML (SVM)	Internal	14	1	3	16	82.0	94.0	0.94
Weng et al.20	CT	DL (ViT)	External	9	4	1	16	90.0	80.0	0.885
Yin et al.21	PET/CT	DL (SE-ResNet)	Internal	41	10	10	42	80.4	80.8	0.84 (0.75–0.90)
Song et al.22	CT	DL (EML)	Internal	100	23	51	93	66.2	80.2	0.813 (0.763–0.864)
Xiong et al.23	CT	DL (CNN)	Internal	53	6	40	62	57.1	91.0	0.776 (0.702–0.849)

TP: true positive; FP: false positive; FN: false negative; TN: true negative; AUC: area under the curve; CT: computed tomography; PET/CT: positron emission tomography-computed tomography; DL: deep learning; ML: machine learning; CNN: convolutional neural network; SE-CNN: squeeze-and-excitation convolutional neural network; ResNet: residual network; Inc-v3: Inception-v3; 3D-CNN: three-dimensional convolutional neural network; ViT: vision transformer; SVM: support vector machine; RF: random forest; EML: ensemble machine learning.

Risk of bias assessment

Figure 2 presents the QUADAS-2 assessment. Most studies showed a high risk of bias in patient selection, primarily due to retrospective study designs and non-consecutive patient enrollment. Applicability concerns related to the index test were generally low.

Figure 2.

QUADAS-2.

QUADAS-2: Quality Assessment of Diagnostic Accuracy Studies-2.

Meta-analysis of diagnostic performance

No threshold effect was detected (Spearman ρ = −0.129, P = 0.670). However, significant heterogeneity remained (Cochran's Q = 24.29, P < 0.001; I² = 92%), necessitating the use of a bivariate random-effects model.

The pooled sensitivity was 71% (95% CI: 68%–74%), and the pooled specificity was 81% (95% CI: 78%–84%). The PLR was 4.1 (95% CI: 3.0–5.5), the NLR was 0.31 (95% CI: 0.24–0.41), and the DOR was 12.5 (95% CI: 7.7–20.4). The SROC curve indicated favorable overall diagnostic performance, with a summary AUC of 0.85 (95% CI: 0.82–0.88) (Figure 3).

Figure 3.

Forest plots of diagnostic accuracy by subgroup. (a) Subgroup by algorithm (deep learning vs. conventional machine learning); (b) subgroup by imaging modality (PET/CT vs. CT). Squares represent point estimates for each study with size proportional to study weight. Horizontal lines represent 95% confidence intervals. Diamonds represent pooled estimates. Heterogeneity: I² = 73.2%, P < 0.001 (random-effects model).

ET-CT: positron emission tomography-computed tomography; CT: computed tomography.

Subgroup analysis and meta-regression

Figures 3 and 4 display the forest plots and hierarchical SROC (HSROC) curves. Algorithm type. DL models achieved a significantly higher summary AUC (0.871, 95% CI: 0.832–0.918) than conventional ML models (0.798, 95% CI: 0.517–1.000; P = 0.012). However, the wide CI for conventional ML should be interpreted with caution due to the limited number of studies (n = 3).

Figure 4.

Pooled diagnostic performance: forest plots and SROC curve. (a) Forest plots of pooled sensitivity and specificity; (b) SROC curve. Squares denote point estimates with 95% CIs (diamonds: pooled estimates). In (b), circle size reflects sample size; the dashed line indicates random chance.

CI: confidence interval; SROC: summary receiver operating characteristic.

Imaging modality. CT-based models demonstrated higher diagnostic accuracy than PET/CT-based models (AUC: 0.879, 95% CI: 0.820–0.939 vs. 0.828, 95% CI: 0.721–0.935; P = 0.038).

Validation strategy. Studies incorporating independent external validation exhibited superior performance (AUC: 0.922, 95% CI: 0.850–0.993) compared with those restricted to internal validation (AUC: 0.841, 95% CI: 0.776–0.907; P = 0.006).

Univariable meta-regression was performed to explore potential sources of heterogeneity. Imaging modality emerged as a significant moderator of the DOR, with a coefficient of 0.45 (95% CI: 0.12–0.78; P < 0.05), explaining 37% of the between-study variance (R² = 0.37). Consistent with this finding, subgroup analysis showed that CT-based studies yielded a higher pooled OR than PET/CT-based studies (1.86 vs. 1.03). In contrast, ML algorithm type (DL vs. conventional ML) was not significantly associated with the DOR (coefficient = −0.08, 95% CI: −0.35–0.19; P = 0.56), as reflected by overlapping 95% CIs between subgroups. Similarly, subgroup analysis stratified by validation type (internal vs. external) showed no significant difference in pooled OR (1.54 vs. 1.09), consistent with the meta-regression results indicating no significant moderating effect of validation approach.

Sensitivity analysis and publication bias

Galbraith plot revealed no significant outliers. Deeks’ funnel plot asymmetry test indicated no significant publication bias (P = 0.46) (Figure 5).

Figure 5.

Deeks’ funnel plot asymmetry test for publication bias. The x-axis represents the natural logarithm of the diagnostic odds ratio, and the y-axis represents the inverse square root of the effective sample size (1/√ESS).

ESS: effective sample size.

Discussion

To our knowledge, this is the first meta-analysis specifically evaluating ML-based radiomics models for predicting EGFR mutations in Chinese patients with LUAD. The findings demonstrate the potential of this noninvasive approach, with a summary AUC of 0.85 (95% CI: 0.82–0.88), alongside balanced sensitivity (71%) and specificity (81%).

The exclusive focus on the Chinese population, which carries the highest global burden of EGFR-mutated lung cancer, substantially enhances the clinical relevance of these findings.^24,25 The high pre-test probability in this cohort suggests that radiomics models validated in this setting provide a more robust and clinically actionable benchmark than those derived from heterogeneous populations.

Furthermore, although EGFR mutations commonly occur across exons 18–21 in this population, the lack of disaggregated data prevented subgroup analyses based on specific mutation subtypes. This represents an important knowledge gap that should be addressed in future primary studies.

Imaging modality: rethinking the value proposition

A key finding of this study is the superior accuracy of CT-based models compared with PET/CT-based models, a difference that contributes significantly to observed heterogeneity. This challenges the assumption that multimodal metabolic imaging is inherently necessary for EGFR mutation prediction. For this task, the high-dimensional morphologic and textural features derived from routine CT appear not only sufficient but potentially more generalizable in clinical practice. ²⁶ This observation is consistent with the principle that model performance is constrained by the intrinsic “information density” of the input data relative to the clinical prediction task. ²⁷ Although PET/CT provides additional metabolic information, its incremental predictive value may be limited by higher cost and lower inter-center standardization. However, definitive conclusions require prospective intra-patient comparative studies.

Algorithm comparison: a nuanced perspective

As anticipated, DL models outperformed conventional ML models. This finding is consistent with the theoretical advantage of DL in automated feature representation learning. However, the observed difference (AUC: 0.871 vs. 0.798) should be interpreted with caution, as the small number of conventional ML studies resulted in wide CIs, limiting the strength of any definitive conclusion. More importantly, this comparison may underestimate the potential of both approaches. The future of predictive oncology is likely to involve multimodal data integration, combining imaging, clinical, and genomic information. Within such high-dimensional frameworks, the ability of DL methods to learn abstract representations and integrate heterogeneous data types is likely to provide a comparative advantage over conventional ML approaches.^28,29 Nevertheless, conventional ML remains valuable due to its interpretability, whereas DL is expected to play a central role in the development of more complex, integrated predictive models.

The imperative of rigorous validation

The notable finding of this study is the substantial performance difference between models using independent external validation and those relying solely on internal validation (AUC: 0.922 vs. 0.841). This difference highlights a critical principle in medical artificial intelligence: robust validation strategies are more indicative of real-world clinical utility than high performance metrics derived from internal datasets alone. A model achieving an AUC of 0.95 in internal validation but failing in external validation has limited clinical applicability, a limitation that has been frequently observed in lung cancer radiomics research.^30,31 Therefore, independent external validation should not be considered an optional enhancement but rather a fundamental requirement for assessing the clinical readiness of predictive models.^32,33

Limitations and future directions

This study has several limitations that should be considered when interpreting the findings. First, although a bivariate random-effects model was used, the high heterogeneity (I² = 92%) and wide CIs in some subgroups indicate that the pooled estimates reflect a highly variable evidence base rather than a precise or universal effect estimate. Differences in radiomics workflows—including manual versus semi-automated segmentation and the use of different feature extraction platforms (e.g. PyRadiomics versus in-house software)—limited direct comparability across studies and hindered identification of an optimal technical pipeline. Second, the exclusive reliance on retrospective studies conducted in a single country (China) introduces potential selection and spectrum bias, which may have inflated diagnostic performance estimates. ³⁴ Although focusing on Chinese patients enhances the clinical relevance for this high-prevalence population, it limits the generalizability of the findings to other ethnic groups with different EGFR mutation profiles. Third, the relatively small number of included studies, particularly within the conventional ML subgroup, reduced the statistical power of meta-regression analyses, making these findings exploratory in nature. Furthermore, insufficient primary data prevented subgroup analyses based on specific EGFR mutation subtypes (e.g. exon 19 deletions versus L858R mutations) or co-mutation patterns. The predictive value of radiomics may vary significantly across these molecularly distinct entities, representing a substantial knowledge gap. Finally, from a clinical translation perspective, the “black-box” nature of many advanced ML and DL models, together with the absence of cost-effectiveness analyses and integration studies within real-world clinical decision-making pathways, highlights substantial practical barriers that cannot be addressed by diagnostic accuracy metrics alone.^35,36

To bridge these gaps, we propose a targeted agenda for future research: (a) large-scale, prospective, multicenter trials using pre-registered, standardized protocols; (b) head-to-head, intra-patient comparisons of CT and PET/CT radiomics to definitively quantify their comparative value; (c) mandatory independent external validation as a core reporting requirement for all predictive model studies; and (d) expansion beyond binary EGFR prediction toward models capable of distinguishing specific sensitizing mutations, resistance mechanisms, and alterations in other driver genes (e.g. anaplastic lymphoma kinase (ALK) and ROS proto-oncogene 1 (ROS1)). Only through such rigorous methodological advancement can radiomics be fully integrated into precision oncology.

Conclusion

ML-based radiomics models demonstrate robust diagnostic accuracy for predicting EGFR mutation status in Chinese patients with LUAD, achieving a summary AUC of 0.85. CT-based DL models subjected to independent external validation represent the most promising technical configuration currently available. However, the retrospective design and substantial methodological heterogeneity across the included studies preclude immediate clinical deployment. High-quality prospective multicenter trials are essential to validate these findings and standardize radiomics workflows before routine clinical implementation.