The present review was performed in accordance with AMSTAR 2 (17). The methods and protocol for the present study were pre-registered, in accordance with standard procedures, in the International Platform of Registered Systematic Review and Meta-analysis Protocols (registration no. 202390105; DOI: 10.37766/inplasy2023.9.0105).

A comprehensive literature search was performed using the following key terms: ‘Risk factor’, ‘predictive’, ‘spread through air spaces’, ‘lung adenocarcinoma’ and ‘nomograms’. This search used the PubMed (pubmed.ncbi.nlm.nih.gov/), Embase (embase.com/),Scopus(https://www.scopus.com/),Wiley(https://onlinelibrary.wiley.com/) and Web of Science(https://www.webofscience.com/wos/) databases, with a cut-off date of May 1, 2023. The references of included studies were also systematically reviewed to obtain potentially relevant publications (Table I).

The inclusion criteria included the following: i) Study focuses on patients who have been diagnosed with LUAD and who exhibit STAS; ii) objective of the study is to develop a predictive model to accurately identify the presence of STAS in patients with LUAD. Tumor STAS was defined as tumor cells (micropapillary structures, solid nests, or single cells-spreading within air spaces in the lung parenchyma beyond the edge of the main tumor (3). The present study selected studies that included patients who underwent segmental or lobar resection.

The exclusion criteria were as follows: i) Predictive models that were not constructed based on radiological features; ii) no clear inclusion and exclusion criteria; iii) reviews and lecture-type literature; iv) literature for which the full text could not be obtained and v) literature for which data could not be extracted.

The initial screening of titles and abstracts was independently conducted by two researchers (CL and PW) using Cochrane handbook's guidelines for systematic reviews of interventions (18), adhering to the predefined inclusion and exclusion criteria. Discrepancies or uncertainty about article inclusion were resolved through discussion or consultation with a third reviewer (XL).

The quality of articles was assessed by two independent reviewers using Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) (19), which is a tool for assessing quality of diagnostic studies, focusing on ‘risk-of-bias’ and ‘applicability concerns’. The risk-of-bias was assessed across four domains: Patient selection, index test, reference standard and flow and timing. The applicability was evaluated for the first three domains and rated as ‘yes’, ‘no’ or ‘unclear’, with ‘yes’ denoting low risk, ‘no’ indicating high risk and ‘unclear’ suggesting insufficient information. In cases of disagreement, a third reviewer was used for resolution. A Measurement Tool to Assess systematic Reviews was used for a stringent quality assessment (14).

Additionally, the methodological quality of studies was appraised using the Cochrane Handbook's risk-of-bias assessment tool (RevMan v.5.3.5, The Cochrane Collaboration) (20), covering six aspects (selection, performance bias, detection, attrition, reporting bias and other bias), which were categorized as ‘yes’, ‘no’ or ‘unclear’ to indicate the level of bias.

The data extracted primarily encompassed the following aspects: i) Characteristics of the included literature, such as author information, publication date, country of origin, predictive models and regression methods employed and predictive factors investigated; ii) details of the study subjects, such as sample size, sex distribution and tumor stage (according to the 8th edition of the AJCC staging standards) (21), with all participants having undergone surgery and iii) evaluation of effect indicators.

The effectiveness of various predictive models were evaluated based on their accuracy, sensitivity (SEN), and specificity (SPE). Predictive models were categorized according to their unique features for NMA to assess performance in predicting STAS. This NMA, conducted using Stata software (version 17.0; StataCorp LP) within a Bayesian framework, used the Markov Chain Monte Carlo Subset Simulation (22) in accordance with the PRISMA NMA guidelines (23). A nodal approach for quantifying and clarifying concordance between direct and indirect comparisons was adopted. The consistency criterion for the NMA was P>0.05. Network diagrams visually represented diagnostic methods, with nodes symbolizing each method and lines representing direct comparisons. The size of nodes and thickness of lines corresponded to the number of studies. To detect possible publication bias in selected studies, funnel plots were constructed for each measure of diagnostic efficiency, employing symmetry criteria as a key validation technique. Statistical heterogeneity was evaluated using I² statistic, a measure in meta-analytical methods. This quantifies the proportion of the total variation in study estimates due to heterogeneity rather than chance. An I² value of 0% indicates no observed heterogeneity, whilst higher values suggest increasing heterogeneity, with guidelines typically considering 25, 50 and 75% as low, moderate and high heterogeneity, respectively (18). Additionally, to ascertain the relative superiority of one method, the level of certainty for predictive models was quantified. This assessment was performed using surface under the cumulative ranking curve (SUCRA), forest plots and league tables.

Subgroup diagnostic meta-analyses were performed using Stata to assess relative predictive efficiency of composite models. The effectiveness of these models was evaluated using the area under the curve (AUC) derived from the summary receiver operating characteristics (sROC). Additionally, the Fagan plot was utilized to quantify the overall discriminatory power of a diagnostic test (24).

Literature review was performed using 565 articles. Subsequent to this, a meticulous screening process was undertaken to ensure the relevance and quality of sources. This entailed the removal of 249 articles due to duplication. Further scrutiny, focusing on titles and abstracts, led to the exclusion of an additional 288 articles that were not pertinent. The remaining pool of 28 articles was subjected to a more rigorous evaluation, which included accessibility of full-text versions and the feasibility of data extraction. This process led to the disqualification of 14 articles, leaving a final count of 14 articles (Fig. 1).

These 14 articles, collectively encompassing data from 3,734 participants, were exclusively focused on patients diagnosed with STAS in LUAD. The predictive models were categorized into the following four distinct types based on their methodological approaches and radiological characteristics: i) Models developed using logistic regression analysis to screen CT features (Features_CT); ii) models using machine learning (ML) techniques to screen tumor radiological characteristics (ML_Tumour); iii) models applying ML for the screening of both tumor and peritumor radiological features (ML_Peri_tumour) and iv) models that used logistic regression analysis for screening of PET/CT features (pet_CT).

In the process of tumor and peritumor segmentation, the open-source software 3D Slicer v4.8.1 (slicer.org) was used. Moreover, all studies used pathological findings as a benchmark, forming a control group against which predictive models were evaluated. The data enabling direct comparative analysis in outcomes were also assessed (Table II) (5,15,16,25–35).

In the present study involving 14 articles and 17 predictive models, NMA was performed using Stata and the QUADAS-2 tool was used to assess quality, risk of bias and applicability of the articles. The inter-rater reliability (κ-agreement) between the reviewers was 0.87. A high risk of bias was detected in a few articles (3/14) in terms of patient selection (1/14) and reference standards (2/14); however, the overall quality of the publications was satisfactory (Fig. 2).

NMA evaluated the relative risk (RR) values and 95% confidence intervals (CI) across different predictive models in terms of accuracy, SEN and SPE for STAS in LUAD.

NMA graph illustrates the comparative accuracy, SEN and SPE of predictive models (Fig. 3). Notably, CT_feature model group encompassed the largest sample size, followed by the ML_Peri_tumour model. Specifically, two studies directly compared the CT_feature and ML_Peri_tumour model, and one study contrasted the CT_feature model with ML_Tumour model (Fig. 3). Furthermore, the comprehensive evaluation of the included studies spanned all domains. Potential publication bias was assessed using funnel plots (Fig. 4). The roughly symmetric distribution suggested a negligible presence of publication bias or other forms of bias within the studies. This symmetry bolstered the reliability of findings.

Using the SUCRA, the accuracy of several predictive models for STAS was evaluated. Models ranked in descending order of accuracy were as follows: Control (100.0%); ML_Peri_tumour (56.5%); ML_Tumour (41.8%); pet_CT (41.4%) and Features_CT (10.3%; Fig. 5A). A detailed two-by-two comparative analysis is presented in Table IIIA, highlighting the predictive efficacy of these models. ML_Peri_tumour model demonstrated superior accuracy, particularly compared with Features_CT (RR=1.14; 95% CI, 0.99–1.32), ML_Tumour (RR=1.04; 95% CI, 0.83–1.30) and pet_CT (RR=1.04; 95% CI, 0.84–1.29). A heterogeneity test revealed I² value of 20.4%. Consistently, the forest plot demonstrated the highest predictive accuracy for ML_Peri_tumour model (Fig. 6A).

SEN for different predictive models for STAS, derived from the SUCRA was as follows: Control (99.9%); Features_CT (51.9%); ML_Peri_tumour (49.9%); ML_Tumour (42.8%); and pet_CT (5.5%; Fig. 5B). Table IIIB shows a comparative league table for a two-by-two analysis of these models. Features_CT model exhibited superior SEN, especially compared with ML_Peri_tumour (RR=1.00; 95% CI, 0.89–1.13), ML_Tumour (RR=1.02; 95% CI, 0.88–1.18) and pet_CT (RR=1.17; 95% CI, 0.97–1.40). The heterogeneity test indicated an I² of 17.8%. Additionally, forest plot highlighted the superior predictive SEN of the Features_CT model (Fig. 6B).

SPE of different predictive models for STAS, ascertained using the SUCRA, was as follows: Control (99.7%); pet_CT (53.9%); ML_Peri_tumour (48.0%); ML_Tumour (42.7%); and Features_CT (5.7%; Fig. 5C). A comprehensive league table in Table IIIC compares these models in a two-by-two format. The pet_CT model showed enhanced SPE, particularly against ML_Peri_tumour (RR=1.03; 95% CI, 0.77–1.39), ML_Tumour (RR=1.05; 95% CI, 0.74–1.50) and Features_CT (RR=1.24; 95% CI, 0.95–1.62). The heterogeneity test yielded I² of 9.1%. The forest plot indicated the superior predictive SPE of the pet_CT model (Fig. 6C).

Diagnostic MA scrutinized the predictive capabilities of the ML_Peri_tumour and Features_CT models. AUC of the sROC for the ML_Peri_tumour model was 0.86 (95% CI, 0.82–0.88), while for the Features_CT model it was 0.81 (95% CI, 0.77–0.84; Fig. 7A and B, respectively). Fagan plot analysis, which assessed the predictive potency of models, demonstrated the relative superiority of the ML_Peri_tumour model (Fig. 7C and D).