The trial registration number (on ClinicalTrials.gov) of PLCO is NCT00002540, and the trial registration number (on ClinicalTrials.gov) of NLST is NCT00047385. The designs of the PLCO and NLST cancer screening trials have been previously published. Briefly, from 1993 to 2001, the PLCO cancer screening trial randomized 154,887 participants aged 55–74 years to the intervention arm receiving multiple screening exams for prostate, lung, colorectal, and ovarian cancers or to the control arm receiving usual care in 1:1 ratio (12–15). For the LC screening, participants in the screening arm received four annual posterior–anterior CXR. There were several changes in the screening protocol, including that never smokers randomized after December 1995 were no longer offered a T3 CXR exam unless they insisted on it. Positive screening exam of CXR was defined as one or more nodules, mass, hilar or mediastinal lymph node enlargement, infiltrate, consolidation, or alveolar opacity. Participants with positive CXR were encouraged to receive further diagnostic evaluation with their primary care physicians. From 2002 to 2004, the NLST cancer screening trial randomized 53,452 smokers aged 55–74 years with at least 30 pack-years of smoking history and at most 15 years smoking cessation history to receive three annual LDCT screening (the intervention arm) or posterior–anterior CXR screening (the control arm) in 1:1 ratio (6). Radiologists were required to review the images in two ways: looking at the image without reference to historical images (isolation read) and then looking at the image again with reference to historical images (comparison read). Positive exam was defined as any non-calcified nodule or mass with diameter ≥ 4 mm, or any other abnormalities appeared suspicious for lung cancer (in the radiologist’s judgment). Participants with either a positive result or a negative result but with other clinically significant abnormalities were strongly encouraged to receive a diagnostic evaluation for lung cancer or other suspected condition. Each participating center’s institutional review board approved the protocols, and all participants provided written informed consent.

In the PLCO trial, after excluding the participants in the control arm, a total of 77,443 participants in the screening arm were initially included in this study. After further excluding 6,811 participants who did not receive any CXR examination or without inadequate screen, 3,819 participants with one round of CXR examination, 4,937 participants with two rounds of CXR examination, 2,620 participants with three rounds of CXR examination enrolled in 1993–1995, 3,175 smoking participants with three rounds of CXR examination enrolled in 1996–2001, and 3,595 participants who did not meet the definitions of four classic CXR trajectories (details referred to the following section) in CXR, a total of 52,486 participants were finally included in this study ( Supplementary e-Figure S1 ). In the NLST trial, after excluding the participants who received LDCT in the intervention arm, a total of 26,730 participants who received CXR in the control arm were initially included in this study. After further excluding 4,037 participants without any CXR examinations and 699 participants who did not meet the definitions of four classic CXR trajectories (details referred to the following section) in CXR, a total of 21,994 participants were finally included in this study ( Supplementary e-Figure S2 ).

Based on multiple rounds of CXR screening, to achieve comparable CXR trajectories between the PLCO and NLST trials, four classic CXR trajectories were defined in this study, including stable negative CXR (CXR_SN), gain of positive CXR after persistent negative CXR and no subsequent negative diagnosis (CXR_GP), stable positive CXR (CXR_SP), and loss of positive CXR after persistent positive CXR and no subsequent positive diagnosis (CXR_LP). To avoid the confounding effect of instability symptoms on four classic CXR trajectories, disordered fluctuations on CXR were not included in this study, including that negative CXR progressed to positive CXR and subsequently regressed to negative CXR again, and that positive CXR regressed to negative CXR and subsequently progressed to positive CXR again.

After informed consent, all participants in the PLCO trial were provided with a baseline questionnaire to collect participant-reported information on demographic and potential risk factors associated with PLCO cancers, such as demographics, smoking history, family history of cancer, and medical history. Similar baseline variables with different information were collected in the NLST trial. To achieve comparable data in both the PLCO and NLST trials, the following index variables were finally included in this study, including age (55–59, 60–64, 65–69, and 70–74 years), sex (female and male), race/ethnicity (non-Hispanic white and other), education level (<senior high school, senior high school, college, and above), marital status (married/living as married, widowed/divorced/separated, and never married), smoking status (never smoking [only for the PLCO trial], ever smoking, and current smoking), body mass index (BMI) (<18.5, 18.5–25, 25–30, and > 30 kg/m²), and family history of lung cancer (no and yes). BMI was calculated as weight in kilograms divided by the square of height in meters (kg/m²).

The primary endpoints events of this study were LC incidence and mortality. In the PLCO trial, the incidence and mortality of LC were ascertained primarily by mails of Annual Study Update (ASU) questionnaire after last-round CXR and supplemented by repeated mails or phone calls to participants who were not response to the ASU questionnaire. The mortality was further supplemented by periodic linkage to the National Death Index (NDI), and a more accurate assessment of LC deaths was adjudicated by an independent Death Review Process (DRP). The cancer data were collected until 31 December 2009, and mortality data were collected through 2018 in the PLCO trial.

The NLST confirmed diagnoses of LC through medical record abstraction (MRA), which was triggered by annual or semi-annual study update form, positive CT or CXR screening exam, direct report by relatives or physicians, and supplemented by NDI Plus searches. If the MRA process did not find records indicating a LC diagnosis, the LC was not considered confirmed, even if a source such as a death certificate indicated LC. Endpoint verification process was used to determine definitively whether LC was the cause of death. Active follow-up data were collected on cancer diagnoses and deaths that occurred through 31 December 2009. Extended follow-up data were collected for deaths through 31 December 2015.

In both the PLCO and NLST trials, if LC was diagnosed, further information was recorded about cancer characteristics (including histopathological type, grade, location, size of tumor, and TNM components of stage), initial treatment, and cancer progression. Furthermore, in this study, the primary outcomes in both the PLCO and NLST trials were censored at the date of the LC diagnosis (for LC incidence only), death, loss of follow-up, or end of the follow-up period, whichever comes first.

In this analysis for secondary data, analysis of variance or chi-square test was performed to compare the distribution of baseline variables and pathological characteristics of LC between different groups. Log-rank test was initially used to compare the LC incidence and mortality between four classic CXR trajectories. Multivariable Cox proportional hazard regression model was used to analyze the associations of CXR trajectories with LC incidence and mortality after adjusting all available baseline variables (including age, sex, race, education levels, marital status, smoking status, family history of lung cancer, and BMI). The associations were measured as hazard ratio (HR) and 95% confidence interval (CI). Due to the potential heterogeneity between the PLCO and NLST trials, meta-analyses with random-effect models were conducted to pool the study-specific associations from the PLCO and NLST trials.

Subgroup analyses were performed to evaluate the stratified associations of CXR trajectories with LC risks by smoking status in the PLCO trial but not the NLST trial due to lack of non-smokers in the NLST trial. Further analyses were conducted to interaction between CXR trajectories and smoking with LC risk in the PLCO trial. Additionally, multivariable logistic regression models were used to investigate the potential factors associated with progression and regression in CXR, and associations were measured as odd ratio (OR) and 95% confidence interval (CI).

All statistical analyses were conducted via R software (version 4.1.0). A p-value < 0.05 was considered statistically significant.

As presented in Supplementary e-Table S1 , in the PLCO trial, compared to participants with CXR_SN, those with CXR_GP seemed to have older age (70–74 years, 15.7% vs. 11.5%, p-value < 0.001), more men (53.1% vs. 50.7%, p-value = 0.015), lesser white Americans (88.7% vs. 89.9%, p-value = 0.047), lower education level (< senior high school, 7.7% vs. 6.2%, p-value = 0.001), lesser married status (75.2% vs. 78.7%, p-value < 0.001), and more current smoker (12.7% vs. 8.3%, p-value <0.001), while there was no significant difference in baseline characteristics between participants with CXR_SP and CXR_LP. In Supplementary e-Table S2 , in the NLST trial, only older age was observed in participants with CXR_GP than those with CXR_SN (70–74 years, 11.0% vs. 8.1%, p-value < 0.001), while younger age (70–74 years, 11.4% vs. 14.2%, p-value =0.011), more white Americans (92.2% vs. 87.6%, p-value = 0.047), and higher BMI (>30 kg/m², 25.3% vs. 14.8%, p-value = 0.003) was observed in participants with CXR_LP than those with CXR_SP.

After a median follow-up of 16.95 and 10.30 years in the PLCO and NLST trials, a total of 889 (1.7%) LC cases and 1,186 (2.3%) LC deaths were documented in the PLCO trial, while a total of 532 (2.4%) LC cases and 823 (3.7%) LC deaths were recorded in the NLST trial. As shown in Figure 1 , participants with CXR_SP seemed to have the higher risks of LC incidence and mortality than participants with other CXP trajectories in the PLCO trial (all p-values < 0.01), and almost the same results were also observed in the NLST trial.

As shown in Table 1 , after adjusting all available baseline factors, compared to CXR_SN, CXR_GP was significantly associated with increased LC incidence [HR (95%CIs) of 2.82(2.32–3.42) for PLCO, 4.03(3.08–5.28) for NLST, and 3.33(2.35–4.71) for meta-analysis] and LC mortality [1.63(1.33–2.00) for PLCO, 1.45(1.05–2.01) for NLST, and 1.68(1.41–2.01) for meta-analysis]. CXR_LP was significantly associated with decreased LC incidence [0.43(0.26–0.71) for PLCO, 0.16(0.09–0.29) for NLST, and 0.27(0.10–0.70) for meta-analysis results] and LC mortality [0.55(0.34–0.88) for PLCO, 0.29(0.17–0.50) for NLST, and 0.40(0.22–0.76) for meta-analysis results] compared to CXR_SP.

Moreover, after incorporating the four CXR trajectories into a single variable, compared to CXR_SN, the adjusted HR (95%CIs) of LC incidence based on meta-analyses were 2.88(1.50–5.52), 3.86(2.03–7.35), and 1.08(0.80–1.46) for CXR_GP, CXR_SP, and CXR_LP, respectively, while adjusted HR (95%CIs) of LC mortality were 1.58(1.33–1.87), 2.56(1.53–4.29) and 1.05(0.89–1.25) ( Table 1 ).

In the PLCO trial, after stratifying by smoking status and adjusting other available baseline factors ( Table 2 ), compared to CXR_SN, CXR_GP was still significantly associated with increased LC incidence across different subgroups [HR (95%CIs): 4.94(2.79–8.73), 2.94(2.24–3.86) and 2.34(1.71–3.21) in never, sever, and current smokers] and associated with increased LC mortality in both ever and current smokers [1.88(1.43–2.47) and 1.43(1.03–1.97)] but not in never smokers [1.16(0.51–2.65)]. Compared to CXR_SP, CXR_LP was significantly associated with decreased risks of both LC incidence and mortality in both never and ever smokers [0.24(0.07–0.82) and 0.38(0.18–0.80) for LC incidence, 0.41(0.13–1.29) and 0.40(0.21–0.78) for LC mortality] but not in current smokers [0.60 (0.25–1.42) for LC incidence and 0.96(0.39–2.35) for LC mortality] ( Table 2 ).

After incorporating the four CXR trajectories into a single variable, among never smokers, compared to CXR_SN, higher risks of LC incidence and LC mortality were observed in CXR_SP than CXR_GP [HRs (95%CIs): 7.74(2.81–21.37) vs. 4.61(1.70–12.55)], and no significant differences in LC incidence and LC mortality were observed between CXR_SN and CXR_LP. Similar trends were observed among ever and current smokers. However, higher risks of LC incidence were observed in CXR_GP than CXR_SP among both ever smokers [HRs (95%CIs): 2.95(2.25–3.88) vs. 2.59(1.33–5.02)] and current smokers [2.33(1.70–3.18) vs. 2.26(1.06–4.83)]. Moreover, among current smokers, even regression to negative CXR after continuing positive CXR, an increased risk of LC was still observed, with HRs (95%CIs) of 1.53(1.08–2.17) ( Table 2 ).

Among participants with first-round negative CXR, compared to CXR_SN among never smokers, even no progression in CXR, the adjusted HRs (95%CIs) of LC incidence were 7.39(5.60–9.75) for ever smokers and 31.45(23.58–41.95) for current smokers, while HRs (95%CIs) of LC mortality were 6.30(5.07–7.81) and 27.17(21.65–34.11). If participants gained positive CXR, LC incidence risk significantly climbed to 22.04(15.37–31.60) for ever smokers and 71.97(48.82–106.09) for current smokers, while LC mortality risk climbed to 11.90(8.58–16.50) and 38.92(27.04–56.02) ( Table 3 ). Among participants with first-round positive CXR, compared to CXR_SP among never smokers, even no regression in CXR, the adjusted HRs (95%CIs) of LC incidence were 2.93(0.89–9.63) for ever smokers and 12.07(3.43–42.45) for current smokers, while HRs (95%CIs) of LC mortality were 3.72(1.18–11.75) and 9.22(2.55–33.32). If participants lost their positive CXR, LC incidence risk significantly decreased to 1.08(0.38–3.07) for ever smokers and 7.73(2.72–22.02) for current smokers, while LC mortality risk decreased to 1.51(0.54–4.22) and 8.71(3.08–24.67) ( Table 3 ).

After combing all the participants, compared to CXR_SN among never smokers, the adjusted HRs (95%CIs) of LC incidence and mortality for CXR_SP among never smokers were 7.39(2.68–20.36) and 4.31(1.59–11.71). If never smokers lost their positive CXR, both the increased risks of LC incidence [1.74(0.83–3.64)] and mortality [1.71(0.96–3.06)] decreased to non-significant level. Similar trends were observed among both ever and current smokers. However, significant higher risks of LC incidence and mortality for CXR_SP were also observed for both ever and current smokers, especially for current smokers [68.24(31.00–150.18)]. Ever these smokers lose their positive CXR, the increased risks of LC incidence and mortality did not decrease to non-significant level ( Table 3 ).

As shown in Table 4 and Supplementary e-Tables S3,S4 , the proportion of advanced-stage LC in patients with CXR_GP was lower than those with CXR_SN [55.1% vs. 65.6% (p-value = 0.039) for PLCO, 39.2% vs. 66.3% (p-value < 0.001) for NLST, and 48.6% vs. 65.8% (p-value < 0.001) for meta-analysis], while there was no significant difference in cancer stages between CXR_SP and CXR_LP. Similar lower proportion of LC with poor-grade differentiation were observed in participants with CXR_GP compared to those with CXR_SN [49.6% vs. 60.9% (p-value = 0.020) for meta-analysis], while no significant difference in cancer grades was observed between CXR_SP and CXR_LP. Moreover, in the PLCO trial, non-significant differences in the histological types of LC were observed between CXR_SN and CXR_GP, or between CXR_SP and CXR_LP (both p-value > 0.05, Supplementary e-Table S3 ).

As shown in Table 5 , in the PLCO trial, based on the multivariable logistic regression, progression in CXR (namely, CXR_GP) was significantly associated with elder age [OR (95%CIs) for 70–74 years vs. 55–59 years: 1.58(1.39–1.79), p-value < 0.001], men [1.10(1.00–1.19), p-value = 0.042], widowed/divorced/separated status [1.19(1.07–1.31), p-value=0.001], and current smoking [1.79(1.59–2.04), p-value < 0.001]. However, regression in CXR (namely, CXR_LP) was not significantly associated with any currently available baseline variables ( Table 5 ). Similar associations were observed in the NLST trial ( Supplementary e-Table S5 ).