The utility of PORT for patients with NSCLC is disputed [1–4]. While retrospective studies suggest PORT's benefits [5, 6], recent RCTs [7, 8] demonstrated non-significant improvement in DFS or OS from PORT for unselected patients with NSCLC, albeit with a significant reduction in locoregional recurrence. Consequently, the focus has shifted to identifying high-risk individuals who might benefit from PORT. Prior studies revealed that PORT could enhance OS in carefully selected patients; however, these studies were based solely on clinical features and lacked external validation [13, 14]. Radiomics, which use quantitative imaging characteristics, has emerged as a promising tool for disease diagnosis and prognosis [22]. Our study showed that the RPI was a promising marker, and LRPI could identify patients who would benefit from PORT.
The LRPI is the first radiomics-based index to predict the benefit from PORT in patients with NSCLC. Previous studies have explored the prognostic value of radiomic signatures in early-stage NSCLC and the benefits of chemotherapy. One study established a radiomic signature by applying a LASSO-Cox regression model to 13 radiomic features of 329 patients with stage I or II NSCLC to predict DFS and benefit from chemotherapy [23]. A comprehensive model combining radiomics and major clinicopathological characteristics improved the DFS estimate (C-index: 0.74) compared with the clinicopathological model alone (C-index: 0.71). Another study used LASSO-Cox with radiomics to categorise the OS risk of pathologic stage IA pure-solid NSCLC in 800 patients [24]. In both the internal and external validation sets, the radiomics signature for predicting the 5-year OS demonstrated an area under the curve of 0.78 and 0.75, respectively. In our study, five independent validation cohorts were used to evaluate the prognostic value of the RPI, and we verified the ability of the LRPI to predict the benefits of PORT.
This study calculated the LRPI to identify patients who could benefit from PORT. The moderate-risk LRPI group exhibited a significant improvement in OS. By contrast, PORT did not improve OS in the low-risk LRPI groups, indicating that administering potentially toxic PORT provided no additional benefit for the low-risk groups [25, 26]. For patients with high risk, the prognosis is often determined by controlling distant metastatic, rather than locoregional, disease management. In this context, a study revealed that PORT was beneficial for a specific patient group, characterised by a low risk of distant metastasis and high risk of locoregional recurrence [27]. This finding aligns with our observations, further underscoring the critical role of tailored therapeutic strategies in optimising patient outcomes.
In the multivariate analysis, RPI and PLN remained significant predictors for OS, whereas pT did not. The RPI may embody the characteristics of pT and histological information. Previous studies have shown that radiomic features can predict pT [28], histology [29], and grade [30] in NSCLC. PLN remained a significant variable in the multivariate analysis. The radiomic features were extracted from the tumour volume; thus, they could not represent information from the mediastinal lymph nodes. PLN has been reported to correlate with OS [31–33], and patients with high PLN have been reported to benefit from PORT [33, 34]. In addition, studies have shown that the lymph nodes' phenotypic information enhances the primary tumour’s performance for predicting the pathological response [35] and progression-free survival [36] in NSCLC.
To include the RPI in a clinical model, we developed an LRPI that combined the RPI with PLN based on the results that PLN and RPIs remained significant predictors of OS in the multivariate analysis. In predicting OS, our model outperformed the pT stage and clinical model, including the pT and PLN. Previous studies have shown that radiomic features can predict pT stage [28], histology [29], pathological grade [30], and lymph vesicular invasion [37] in NSCLC. The RPI could represent more information than the clinicopathological variables; thus, it would be much more convenient to use the RPI as it can be fully automatically calculated from the CT image. In contrast, clinicopathological variables require experienced doctors to identify [38]. Additionally, we verified the potential of an LRPI to predict the benefits of PORT.
We identified a relationship between GLNN and TSR. GLNN measures the similarity in intensity values across the image, with a lower GLNN value indicating a greater homogeneity or uniformity of intensities. CT images provide density information of the tumour, which may vary along with the TSR. A higher TSR indicated a more hostile tumour microenvironment, often associated with poorer outcomes [39]. Besides its relationship with TSR, GLNN may also reflect intratumor heterogeneity. The structural diversity is linked to a varying blood supply within the tumour [40], leading to a hypoxic environmessociated with a less effective response to radiotherapy in NSCLC [41, 42].
The Reactome pathways associated with RPI involved PI3K/AKT, KEAP1/NFE1L2, and antigen presentation. The PI3K/AKT and KEAP1/NFE1L2 pathways are associated with resistance to radiotherapy [43]. The PI3K/AKT pathway has been strongly linked to both initiation and progression in NSCLC [44, 45]. In contrast, the KEAP1/NFE2L2 pathway is involved in managing cellular stress and can shield cancer cells from the impact of radiation therapy [46, 47]. Moreover, KEGG GSEA analysis consistently demonstrated that RPI was related to immune system processes, including antigen presentation. The prominence of immune system-related functions suggests that cancer cells might escape host immune surveillance, underscoring the potential of immunotherapies. With the release of encouraging results from clinical trials involving the administration of immunotherapy to patients with peri-operative NSCLC [48, 49], a novel adjuvant treatment approach has emerged. Immunotherapy demonstrates substantial potential to enhance long-term survival rates for patients with operable lung cancer. Future research should therefore further explore these biological processes as potential therapeutic targets for preventing NSCLC recurrence.
Our study had several limitations. First, our study follows a retrospective design, inherently prone to selection bias that could not be completely mitigated. Despite the utilisation of participants from the PORT-C RCT, it is crucial to acknowledge that preoperative CT scans were limited to a subset of the overall cohort. Consequently, this inclusion constraint may introduce inherent biases that warrant consideration when interpreting the study results. Second, we restricted the reconstruction kernels and slice thicknesses to maintain the reproducibility of radiomic features, thereby likely limiting the expansion of this model. However, our model showed satisfactory performance in the external cohorts, which had a variety of scanners, kernels, and slice thicknesses. Last, external cohort 4 — comprising patients at varying disease stages who underwent definitive radiotherapy — may not precisely mirror the population composition observed in cohort 1. Nevertheless, validation of the RPI — initially developed from cohort 1 — was possible within this external cohort. This outcome underscores the robustness of the RPI as a stable prognostic variable to OS.