The flow diagram of the systematic review is shown in Fig. 2. After applying the inclusion and exclusion criteria, a total of ten initially retrieved articles were included in the study.

The process of our search strategy

The ten included articles were published from 2016 to 2023 [26,27,28,29,30,31,32,33,34,35]. The patient characteristics of the ten studies are summarized in Table 1. Most studies (9/10, 90%) [26,27,28, 30,31,32,33,34,35] were retrospectively designed, except one which was prospective [29]. All articles had multiple cohorts, such as training, testing, and validation groups. There were various prognostic outcomes, including disease-free survival (DFS), objective response rate (ORR), OS, progression-free survival (PFS), recurrence, and three studies contained more than one outcome [26, 29, 35]. Genetic data from three articles (3/10, 30%) were available on Gene Expression Omnibus [27, 31, 33]. In addition, one immunotherapy-related article contained patients receiving treatment according to their treatment plans [33].

The study details of the radiomics workflow, including acquisition parameters of the images, the region of interest (ROI) segmentation, feature extraction, selection, and a number of features are summarized in Table 2. One study [32] did not mention the slice thickness, and the other nine studies’ slice thickness ranged from 0.625 to 5 mm. The ROI segmentation method was not mentioned in one article [29], manual in two studies [27, 33], semiautomatic in five studies [26, 27, 30, 32, 34], and the remaining three were uncertain [28, 31, 35]. The number of extracted radiomic features ranged from 35 to 2996. The number of radiomic features in the final radiomic model ranged from 2 to 210. A total of eight articles constructed an independent radiomic model [26,27,28, 30, 31, 33,34,35]. The model constructed in the final radiomic model was usually a Cox proportional hazard model. More information is shown in Table 3.

Table 4 details the gene analyses. Nearly every article applied unique methods to selecting genotic features. Five articles constructed an independent genomic model [26, 28, 30, 31, 35]. However, the genomic approach used in one of the articles may also elucidate certain biological biases [33].

The details of the combination models and the corresponding performance metrics in the included studies were summarized in Table 5.

For the radiomic models in eight articles, the area under the receiver operating characteristic curve (AUC) was used to evaluate the performance in four studies [26, 30, 33, 35]. And the C-index was used to evaluate the performance in four studies [27, 28, 31, 35]. One of the articles did not show the performance [34].

For the genomic models in five articles, the AUC was used to evaluate the performance in three studies [26, 30, 35]. The C-index was used to evaluate the performance in three studies [28, 31, 35].

The AUC was used for combination models to evaluate the performance in four studies [26, 30, 33, 35], ranging from 0.72 to 0.99. Notably, all AUC values from combined models were higher than their respective independent models. The C-index was used to evaluate the performance in seven studies [27,28,29, 31, 32, 34, 35].

As shown in Supplementary Table 1, the mean RQS of the study is 12.2 (range 2–23). All the papers contain discriminant statistics, but only 20% (2/10) [29, 34] of the studies included calibration statistics. In 90% of the manuscripts [26, 27, 29,30,31,32,33,34,35], feature reduction was carried out to explain the possibility of overfitting. Importantly, all studies have the verification, 80% of which use internal testing [26, 28,29,30,31,32, 34, 35]. The ROB results and the assessment of the applicability of these studies are shown in Supplementary Table 2. Overall, all included articles were regarded as having a high ROB.