Prognostic associations between patient-reported outcomes and overall survival among patients with advanced squamous non-small cell lung cancer: results from the RATIONALE-307 trial Authors: Jie Wang, Xinmin Yu, Shiangjiin Leaw, Frederick B Barnes, Timothy Victor, Gisoo Barnes Published: 2026 ABSTRACT AbstractPurposeIn the RATIONALE-307 trial, favorable overall survival (OS) outcomes with tislelizumab plus chemotherapy versus chemotherapy alone have been reported, along with improvements in patient-reported outcomes (PROs), including global health status/quality of life (GHS/QoL) and lung cancer–specific symptoms such as cough and dyspnea. Here, we aimed to evaluate the association between patient-reported symptoms and OS among patients with first-line advanced sq-NSCLC.MethodsPROs were assessed using the EORTC QLQ-C30 and QLQ-LC13 instruments. A novel three-component joint model (JM) was developed, linking linear mixed and Cox models to evaluate associations between PRO change from baseline scores, recurrent-symptomatic deterioration (RS-D) events, and OS. The association between PROs and OS was summarized via association-effect hazard ratios (HRs) from the JMs.ResultsOverall, 328 patients were included: 111 in the tislelizumab plus paclitaxel–carboplatin arm (Tisle + PC), 110 in the tislelizumab plus nab-paclitaxel–carboplatin arm (Tisle + nPC), and 107 in the paclitaxel–carboplatin (PC) alone arm. Compared to PC alone, Tisle + PC and Tisle + nPC arms significantly improved GHS/QoL (mean difference: 6.29 and 10.23, respectively; both p < 0.05). Dyspnea significantly improved in both Tisle arms; chest pain improved in the Tisle + nPC arm only. In the Tisle + PC arm, chest pain and cough were significantly associated with a 36.3% reduced risk of death (both HRs = 0.64; p = 0.02).ConclusionIn this analysis, compared to PC alone, Tisle + PC improved GHS/QoL and dyspnea in both arms, while chest pain improvement was observed only with Tisle + nPC.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s11136-026-04211-1.Keywords: Patient-reported outcomes, Overall survival, Tislelizumab, Squamous non-small cell lung cancer, Joint models Introduction Lung cancer (LC) remains the leading cause of cancer-related death in both men and women, with an estimated 733,000 deaths in China and 1.8 million deaths reported globally in 2022 [1]. Incidence and mortality vary across regions and between sexes, with disproportionately higher burdens observed in Asia and among male populations [2]. Non-small cell lung cancer (NSCLC) accounts for roughly 80% to 85% of all LCs, originating primarily from epithelial lung cells, with squamous cell carcinoma representing 20% to 30% of all NSCLC cases. Among patients with squamous NSCLC (sq-NSCLC), patient-reported sq-NSCLC-specific symptoms are prevalent and contribute substantially to poor health-related quality of life (HRQoL) [3–6]. A growing body of evidence supports the prognostic value of patient-reported outcomes (PROs) as independent predictors of overall survival (OS) across various tumor types, including advanced HER2-positive breast cancer and metastatic lung cancer [7–14]. LC–specific symptoms captured via PROs can provide complementary insights that extend beyond what is measurable through invasive procedures or imaging-based biomarkers traditionally used to assess treatment response. By providing a direct reflection of the patient experience, PROs can enhance clinical decision–making, guide treatment selection, and contribute to a more comprehensive evaluation of a therapy’s benefit–risk profile. As the oncology field continues to shift toward more patient-centered frameworks, proper consideration and analytical focus on the longitudinal trajectory of PROs is essential to fully capture the dynamic and evolving nature of the patient experience. Joint modeling (JM) has emerged in recent years as a powerful methodological solution to better characterize treatment effects in randomized clinical trials (RCTs) [15–17]. JM enables the simultaneous evaluation of longitudinal PRO data and time-to-event outcomes within a unified statistical framework, under the assumption that there is a shared process that is driving outcomes. Such an approach allows for more efficient and less biased estimates of treatment effects and supports individual-level risk prediction to better understand the associations between symptom trajectories and OS [15, 18, 19]. In contrast, traditional hypothesis testing used in RCTs often evaluates PROs and clinical outcomes independently, overlooking their inherent interdependence. This separation may result in less consistent and less efficient treatment effect estimates, increased bias, and suboptimal use of available data. By explicitly linking PRO trajectories with OS, JMs provide a more appropriate and informative analytic solution for evaluating patient-reported endpoints in oncology RCTs [20–22]. RATIONALE-307 (NCT03594747) was a randomized, open-label, multicenter, phase 3 trial that examined the efficacy and safety of tislelizumab combined with paclitaxel and carboplatin (Tisle + PC) or nab‑paclitaxel and carboplatin (Tisle + nPC), versus paclitaxel and carboplatin alone (PC arm), as first-line treatment for advanced sq-NSCLC [23]. After 4 years of follow-up (i.e., median OS follow-up of 44.8 months), median OS reached 26.1 (95% CI 19.0–33.8) and 23.3 months (95% CI 18.8–26.4) in the tislelizumab arms compared with 19.4 months (95% CI 16.0–23.4) in the PC arm, with corresponding 4-year survival rates of 32.2%, 26.0%, and 19.2%, respectively [52]. In addition, patients also reported improvements in PROs, particularly in global health status/quality of life (GHS/QoL) and LC–specific symptoms such as cough and dyspnea, warranting further exploration. While these findings highlight the potential added value of tislelizumab plus chemotherapy from a patient-centered perspective, the relationship between longitudinal PRO trajectories and OS remains poorly defined in the sq-NSCLC setting. In clinical practice, oncologists routinely consider changes in disease- and treatment-related symptoms when assessing treatment response and informing prognosis. Given the potential for PROs to capture meaningful changes in patient-reported functioning and symptoms, we conducted a secondary analysis to empirically evaluate the association between recurrent PRO-based deterioration and OS among patients with advanced sq-NSCLC enrolled in the RATIONALE-307 trial to help inform clinical decision-making. Methods Study design and patients The study design for RATIONALE-307 (ClinicalTrials.gov identifier: NCT03594747) has been previously published [23]. In brief, adult patients (aged 18–75 years) were randomized (1:1:1) to receive one of the following regimens: Tisle + PC, Tisle + nPC, or PC alone. Randomization was stratified by disease stage (IIIB versus IV) and tumor cell PD-L1 expression (< 1% versus 1%–49% versus ≥ 50%). Patients in the PC alone arm were permitted to cross over to tislelizumab 200-mg monotherapy upon disease progression, provided protocol-specified criteria were met. This trial was approved by the Institutional Review Board or Independent Ethics Committee at each study site and conducted in accordance with applicable regulatory requirements, Good Clinical Practice Guideline, the principles of the Declaration of Helsinki, and local laws and regulations. All patients were required to provide written informed consent prior to participation in the study. Study endpoints and assessments The primary endpoint in the RATIONALE-307 study was independent review committee (IRC)-assessed progression-free survival; key secondary endpoints included OS and IRC- and investigator-assessed objective response rate and duration of response. In the present analysis, we focused on the secondary PRO endpoint and evaluated its association with the key time-to-event outcome of OS, comparing the Tisle + PC arm or Tisle + nPC arm with the PC arm, as it is widely recognized as one of the most definitive and clinically meaningful outcomes in oncology, particularly for a highly lethal disease like advanced sq-NSCLC, where demonstrating an OS benefit may be preferred [24]. Two validated PROs for measuring HRQoL and LC–specific symptoms were included: the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire (EORTC QLQ-C30) [25] and the LC–specific EORTC Quality of Life Questionnaire-LC 13 (QLQ-LC13) [26]. PROs were assessed at baseline, and at every 6 weeks through the end of treatment. For the current analyses, data up to week 36 were utilized. This post-hoc analysis focuses on a subset of endpoints considered relevant to patients, physicians, and regulators (i.e., GHS/QoL, cough, dyspnea, and chest pain) [27–29], and for their clinical relevance in LC and sensitivity to treatment- and disease-related effects. JM results for the remaining PRO endpoints are found in Supplemental Tables S1/2. For the GHS/QoL scale, higher scores indicated better health or functioning; thus, a positive change from baseline reflected improvement, and a negative change reflected deterioration. In contrast, for symptom scales, higher scores indicated worse symptoms; therefore, a negative change from baseline reflected symptom improvement, while a positive change indicated symptom worsening. All PRO questionnaires were completed by patients at the clinical site prior to any other study procedures to ensure objective responses. Statistical analyses All post-baseline PRO assessments collected prior to the clinical event, including those from unscheduled visits, were included in the analysis. Only patients with at least one post-baseline PRO assessment and without crossover to subsequent immunotherapy during the PRO observation period were included in the JM. Patients who progressed and then crossed over to immunotherapy, and thus had no usable post-event PRO data, were excluded from the JM analysis set. Analyses were conducted by linking three components: (1) longitudinal PRO trajectories, (2) time to recurrent symptomatic deterioration (RS-D) events, and (3) OS. Longitudinal changes in PRO scores were modeled via a linear mixed-effects model (LMM), with subject and time specified as random effects. Fixed effects included treatment, time (visit), baseline PRO score, and treatment-by-time interaction to examine between-arm difference in change from baseline. Longitudinal changes in PRO scores were modeled via a linear mixed-effects model (LMM), with subject and time specified as random effects. Fixed effects included treatment, time (visit), baseline PRO score, and treatment-by-time interaction to examine between-arm difference in change from baseline. Two Cox proportional hazards submodels were fit: one for time to RS-D events and one for OS. The former estimated the hazard of RS-D events over time, accounting for treatment arm, while the latter evaluated OS, also adjusted for treatment. Two Cox proportional hazards submodels were fit: one for time to RS-D events and one for OS. The former estimated the hazard of RS-D events over time, accounting for treatment arm, while the latter evaluated OS, also adjusted for treatment. The three components were integrated into the JM using shared random effects to account for interdependencies among all outcomes [30]. The frailty and terminal event Cox proportional hazards models were based on risk intervals and right-censored at study conclusion and were adjusted for the following randomization factors: disease stage (IIIB versus IV) and level of tumor cell (TC) PD-L1 expression (< 1% TC expression versus 1%–49% versus ≥ 50%). Within the JM, missing not at random (MNAR) mechanisms were addressed by linking longitudinal PRO trajectories to time-to-event submodels. This JM allowed for modeling of informative dropout, thereby reducing bias in treatment effect estimation and increasing precision [16, 17, 31–33]. For each of the QLQ-C30 and QLQ-LC13 scales, an RS-D event was defined using the threshold of change from baseline established by Osoba et al. [34]: a decrease of ≥ 10 points in GHS/QoL scores and an increase of ≥ 10 points in symptom scores. Each unique deterioration event, defined as a deterioration separated by at least one non-deterioration change score (i.e., a return above the threshold boundary), was counted as an RS-D event. This requirement was applied to avoid counting persistent deterioration across adjacent visits or short-term fluctuations as multiple events. Patients were categorized as having zero (censored), one, or two or more RS-D events. Patients who did not experience RS-D events or disease progression by the end of the study were censored. The number of RS-D events from week 1 to week 36 by treatment arm and domain are found in Supplemental Table S3. Model convergence was assessed using Markov chain trace and density plots, as well as the Gelman-Rubin statistic [35]. An statistic with a value of 1.1 indicated acceptable convergence. Kaplan–Meier (KM) curves were generated to illustrate survival distributions for both OS and RS-D events. Treatment group comparisons of PROs were conducted using log-rank tests. Forest plots were generated to visualize the treatment effect on OS across different PRO domains. All statistical analyses were conducted using R statistical software (version 4.4.1) with the JMBayes2 (0.5-0), nlme (3.1–165), and survival (3.7-0) packages [36–39]. Conclusions regarding the significance of association were established at a threshold of p < 0.05. Results Patient characteristics At the data cutoff date of September 30, 2020, the intent-to-treat population consisted of 360 patients randomized to Tisle + PC (120 [33.3%]), Tisle + nPC (119 [33.1%]), or PC alone (121 [33.6%]), as described in detail elsewhere [23]. Overall, 328 patients contributed PRO data, with 218 participants included in the OS JM comparing Tisle + PC (n = 111 [50.92%]) to PC alone (n = 107 [49.08%]). Separately, 217 participants were included in the comparison between Tisle + nPC (n = 110 [50.69%]) and PC alone. A total of 23 patients in the Tisle + PC arm and 21 patients in the Tisle + nPC arm were censored because they did not have evaluable PRO data. Baseline study characteristics were similar between groups [23]. Linear mixed models: treatment efficacy for change from baseline in PRO scores Tables 1 and 2 present the LMM results for change-from-baseline scores and treatment-by-time interaction as well as the OS Cox model, after adjusting for RS-D, change from baseline in corresponding symptoms, and stratification factors. In the LMM derived from the JM, Tisle + PC demonstrated a clinically meaningful and statistically significant improvement in QLQ-C30 GHS/QoL compared to the PC arm (6.29 [95% CI 1.22–11.14]; p = 0.0169). This benefit was accompanied by a significant treatment-by-time interaction (− 2.20 [95% CI − 3.27, − 0.95]; p = 0.0101), indicating that the between-arm difference narrowed over time (Table 1). In the Tisle + nPC arm, a similar trend was observed; however, the numerical magnitude of improvement was slightly larger, with an overall mean change from baseline in GHS/QoL of 10.23 (95% CI 5.47–14.96; p < 0.001), compared to the PC arm (Table 2). The corresponding treatment-by-time interaction was also significant (− 2.53 [95% CI − 3.62, − 1.32]; p < 0.001). In both arms, the corresponding statistic were all approximately 1.1, indicating strong convergence and model stability for these estimates. All convergence density and trace plots for the JMs are provided in the Supplementary Materials. Table 1Joint models for OS: adjusted for Tisle + PC Efficacy, PRO ΔBL, and RS-D eventsPRO domainParameterβ (95% CI)P-value a HR (95% CI)QLQ-C30 GHS/QoLPRO ΔBLb – treatment effectc 6.290 (1.223, 11.140) 0.0169 1.002 –PRO ΔBLb – treatmentc × timed interaction − 2.202 (− 3.265, − 0.948) 0.0101 1.026 –RS-D event – treatment effectc–0.84651.0061.049 (0.645, 1.708)RS-D event – longitudinal effect–< 0.001 1.062 0.917 (0.891, 0.941) OS – longitudinal effect–0.18821.0011.007 (0.997, 1.017)OS – RS-D event (frailty)–0.84121.0051.473 (0.018, 87.957)eQLQ-LC13 coughingPRO ΔBLb – treatment effectc− 4.055 (− 9.752, 1.530)0.16211.249–PRO ΔBLb – treatmentc x timed interaction0.554 (-0.508, 1.771)0.58662.529–RS-D event – treatment effectc–0.12461.0551.804 (0.858, 3.924)RS-D event – longitudinal effect–< 0.001 1.025 1.084 (1.057, 1.118) OS – longitudinal effect–0.71251.0341.002 (0.992, 1.014)OS – RS-D event (frailty)–0.66921.0090.354 (0.006, 67.008)eQLQ-LC13 pain in chestPRO ΔBLb – treatment effectc− 3.483 (− 8.039, 0.990) 0.1251 1.043 –PRO ΔBLb – treatmentc × timed interaction1.296 (0.426, 2.373)0.00831.324–RS-D event – treatment effectc–0.72981.0350.886 (0.450, 1.691)RS-D event – longitudinal effect– 0.0000 1.080 1.103 (1.070, 1.144) OS – longitudinal effect–0.71911.0021.002 (0.992, 1.012)OS – RS-D event (frailty)–0.98541.0070.976 (0.010, 100.4)eQLQ-LC13 dyspneaPRO ΔBLb – treatment effectc − 6.915 (− 11.529, − 2.327) 0.0036 1.004 –PRO ΔBLb – treatmentc × timed interaction 1.543 (0.571, 2.468) < 0.001 1.016 –RS-D event – treatment effectc–0.05851.0011.808 (0.979, 3.387)RS-D event – longitudinal effect–< 0.001 1.011 1.103 (1.073, 1.139) OS – longitudinal effect–0.30381.0020.995 (0.985, 1.005)OS – RS-D event (frailty)–0.54791.0020.259 (0.005, 28.3)eEach model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1 expression in TCs (≥ 50% TCs versus 1%–49% TCs versus < 1% TCs) and baseline PRO measurement (in the linear mixed models). Significant effects are highlighted in boldCI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; RS-D, repeated symptom deterioration; TC, tumor cell; Tisle + PC, tislelizumab + paclitaxel and carboplatin chemotherapyaAn R ^ statistic with a value of 1.0 indicated acceptable convergence. bΔBLrepresents the change from baseline in patient-reported outcome scores. cTreatment effect was coded as Tisle + PC versus PC alone with the former as the effect group. dTime was defined as months since baseline. eAssociation parameter and not HR Joint models for OS: adjusted for Tisle + PC Efficacy, PRO ΔBL, and RS-D events Each model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1 expression in TCs (≥ 50% TCs versus 1%–49% TCs versus < 1% TCs) and baseline PRO measurement (in the linear mixed models). Significant effects are highlighted in bold CI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; RS-D, repeated symptom deterioration; TC, tumor cell; Tisle + PC, tislelizumab + paclitaxel and carboplatin chemotherapy aAn R ^ statistic with a value of 1.0 indicated acceptable convergence. bΔBLrepresents the change from baseline in patient-reported outcome scores. cTreatment effect was coded as Tisle + PC versus PC alone with the former as the effect group. dTime was defined as months since baseline. eAssociation parameter and not HR Table 2Joint models for OS: adjusted for Tisle + nPC Efficacy, PRO ΔBLb, and RS-D eventsPRO domainParameterβ (95% CI)P-value a HR (95% CI)QLQ-C30 GHS/QoLPRO ΔBLb – treatment effectc 10.234 (5.469, 14.957) < 0.001 1.002 –PRO ΔBLb – treatmentc × timed interaction− 2.534 (− 3.617, − 1.322)< 0.001 1.026 –RS-D event – treatment effectc–0.73011.0281.092 (0.645, 1.820)RS-D event – longitudinal effect–< 0.001 1.041 0.908 (0.882, 0.933) OS – longitudinal effect–0.24471.0181.005 (0.997, 1.013)OS – RS-D event (frailty)–0.90451.0100.733 (0.010, 52.373)eQLQ-LC13 coughingPRO ΔBLb – treatment effectc− 2.662 (− 8.736, 3.208)0.37921.062–PRO ΔBLb – treatmentc × timed interaction1.105 (− 0.300, 2.551)0.12891.296–RS-D event – treatment effectc–0.37481.0371.371 (0.680, 2.802)RS-D event – longitudinal effect–< 0.001 1.129 1.077 (1.052, 1.107) OS – longitudinal effect–0.36811.0701.004 (0.995, 1.015)OS – RS-D event (frailty)–0.56891.0650.244 (0.004, 36.796)eQLQ-LC13 pain in chestPRO ΔBLb – treatment effectc − 5.632 (− 9.843, − 1.291) 0.0122 1.103 –PRO ΔBLb – treatmentc × timed interaction 1.351 (0.042, 2.255) 0.0332 1.352 –RS-D event – treatment effectc–0.78231.0421.101 (0.561, 2.155)RS-D event – longitudinal effect–< 0.001 1.018 1.128 (1.088, 1.176) OS – longitudinal effect–0.20891.0080.993 (0.981, 1.004)OS – RS-D event (frailty)–0.80031.0460.497 (0.004, 116.9)eQLQ-LC13 dyspneaPRO ΔBLb – treatment effectc− 7.573 (− 12.21, − 3.004) 0.0007 1.008 –PRO ΔBLb – treatmentc × timed interaction 1.519 (0.576, 2.506) < 0.001 1.066 –RS-D event – treatment effectc–0.14991.0041.636 (0.839, 3.240)RS-D event – longitudinal effect–< 0.001 1.134 1.130 (1.087, 1.188) OS – longitudinal effect–0.62701.0280.997 (0.986, 1.008)OS – RS-D event (frailty)–0.56681.0150.238 (0.003, 52)eEach model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1 expression in TCs (≥ 50% TCs versus 1%–49% TCs versus < 1% TCs) and baseline PRO measurement (in the linear mixed models). Significant effects are highlighted in boldCI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; RS-D, repeated symptom deterioration; TC, tumor cell; Tisle + nPC, tislelizumab + nab‑paclitaxel and carboplatin chemotherapyaAn R ^ statistic with a value of 1.0 indicated acceptable convergence. bΔBLrepresents the change from baseline in patient-reported outcome scores. cTreatment effect was coded as Tisle + nPC versus PC alone, with the former as the effect group. dTime was defined as months since baseline. eAssociation parameter and not HR Joint models for OS: adjusted for Tisle + nPC Efficacy, PRO ΔBLb, and RS-D events Each model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1 expression in TCs (≥ 50% TCs versus 1%–49% TCs versus < 1% TCs) and baseline PRO measurement (in the linear mixed models). Significant effects are highlighted in bold CI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; RS-D, repeated symptom deterioration; TC, tumor cell; Tisle + nPC, tislelizumab + nab‑paclitaxel and carboplatin chemotherapy aAn R ^ statistic with a value of 1.0 indicated acceptable convergence. bΔBLrepresents the change from baseline in patient-reported outcome scores. cTreatment effect was coded as Tisle + nPC versus PC alone, with the former as the effect group. dTime was defined as months since baseline. eAssociation parameter and not HR For LC–specific symptoms assessed via the QLQ-LC13, statistical significance varied across domains. In the Tisle + PC arm, dyspnea-related symptom deterioration was significantly lower than in the PC arm (− 6.92 [95% CI − 11.53, − 2.33]; p = 0.0036), while no significant differences were observed for cough or chest pain (Table 1). Significant treatment-by-time interactions were observed for chest pain (p = 0.0083) and dyspnea (p < 0.001). In the Tisle + nPC arm, chest pain (− 5.63 [95% CI − 9.84, − 1.29]; p = 0.0122) and dyspnea (− 7.57 [95% CI − 12.21, − 3.00]; p = 0.0007) demonstrated significantly less deterioration from baseline compared to the PC arm (Table 2). The treatment-by-time interaction remained significant for both chest pain (1.35 [95% CI 0.04–2.26]; p = 0.0332) and dyspnea (1.52 [95% CI 0.58–2.51]; p < 0.001). The majority of parameters demonstrated strong convergence, with values near 1.1. Time-to-event models: adjusted associations between RS-D events and risks of OS In the Cox proportional hazards model, after adjusting for RS-D events, change from baseline in corresponding symptoms, and stratification factors, higher (better) longitudinal GHS/QoL scores were significantly associated with a reduced risk of future RS-D events (HR = 0.917 [95% CI 0.891–0.941]; p < 0.001), corresponding to an 8.3% relative reduction in risk per 10-point increase in GHS/QoL score (Table 1). Findings in the Tisle + nPC arm were nearly identical, with better longitudinal GHS/QoL scores significantly associated with a reduced risk of RS-D events (HR = 0.908 [95% CI 0.882–0.933]; p < 0.001), corresponding to a 9.2% relative reduction in risk per 10-point increase in GHS/QoL score, irrespective of treatment (Table 2). For sq-NSCLC-specific endpoints of cough, chest pain, and dyspnea, the longitudinal RS-D effects were all statistically significant (p < 0.001) (Table 1). Each 10-point increase in symptom severity over time was associated with an 8.4% increased risk of a future RS-D event due to cough (HR = 1.084 [95% CI 1.057–1.118]), a 10.3% increased risk for chest pain (HR = 1.103 [95% CI 1.070–1.144]), and a 10.3% increased risk for dyspnea (HR = 1.103 [95% CI 1.073–1.139]), irrespective of treatment. The treatment-by-time interaction was significant for chest pain (1.30; p = 0.0083) and dyspnea (1.54; p < 0.001). In the Tisle + PC arm, treatment effect OS HRs for both cough and chest pain were significantly associated with a (1 − 0.637) × 100 = 36.3% reduced risk of death events compared to Arm C (HR = 0.637; both p = 0.02) (Fig. 1). Although not statistically significant, HRs were lower across all PRO domains in the Tisle + nPC arm versus the PC arm, suggesting a numerically favorable effect of tislelizumab-based treatment on OS (Fig. 2). In the Tisle + nPC arm, for all three symptoms, the longitudinal RS-D effects were significant (p < 0.001). Each unit increase in symptom severity was associated with a 7.7% increased risk for cough (HR = 1.077 [95% CI 1.052–1.107]), a 12.8% increased risk for chest pain (HR = 1.128 [95% CI 1.088–1.176]), and a 13.0% increased risk for dyspnea (HR = 1.130 [95% CI 1.087–1.188]), irrespective of treatment (Table 2). Forest plot for tisle + PC vs. PC: OS – treatment effect. aAn R ^ statistic with a value of 1.0 indicated acceptable convergence. Notes Treatment effect was coded as Tisle + PC versus PC alone with the former as the effect group. Each model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1expression in TCs (≥50% TCs versus 1%–49% TCs versus <1% TCs) and baseline PRO measurement (in the linear mixed models). Abbreviations. CI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; PC, paclitaxel and carboplatin chemotherapy; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; TC, tumor cell; Tisle + PC, tislelizumab + paclitaxel and carboplatin chemotherapy Forest plot for tisle + nPC vs. PC: OS – treatment effect. aAn R ^ statistic with a value of 1.0 indicated acceptable convergence. Notes Treatment effect was coded as Tis+Nab-Pac+Carbo versus Pac+Carbo alone with the former as the effect group. Each model was adjusted for the following stratification factors: disease stage (IIIB versus IV) and PD-L1expression in TCs (≥50% TCs versus 1%–49% TCs versus <1% TCs) and baseline PRO measurement (in the linear mixed models). Abbreviations. CI, confidence interval; GHS/QoL, global health status/quality of life; HR, hazard ratio; PC, paclitaxel and carboplatin chemotherapy; PD-L1, programmed death-ligand 1; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; QLQ-LC13, quality of life questionnaire lung cancer 13 items; TC, tumor cell; Tisle + nPC, tislelizumab + nab‑paclitaxel and carboplatin chemotherapy Kaplan–Meier plots Figures 3, 4, 5 and 6 display KM curves for OS in the PRO-evaluable population, corresponding to the patients included in the PRO analyses of OS. KM plots for additional domains are available in the Supplementary Materials. In Figs. 3 and 5, comparing the Tisle + PC arm versus the PC arm, statistically significant improvements in OS were observed favoring Tisle + PC for both PRO domains. In contrast, Figs. 4 and 6, comparing Tisle + nPC to the PC arm, showed no significant difference in OS. Kaplan–Meier plot for OS by treatment arm (Tisle + PC vs. PC) in the QLQ-C30 GHS/QoL domain. Note The dashed lines represent the median time-to-event for each treatment arm. Abbreviations. CI, confidence interval; GHS/QoL, global health status/quality of life; OS, overall survival; PC, paclitaxel and carboplatin chemotherapy; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; Tisle + PC: tislelizumab + paclitaxel and carboplatin chemotherapy Kaplan–Meier plot for OS by treatment arm (Tisle + nPC vs. PC) in the QLQ-C30 GHS/QoL domain. Note The dashed lines represent the median time-to-event for each treatment arm. Abbreviations. CI, confidence interval; GHS/QoL, global health status/quality of life; OS, overall survival; PC, paclitaxel and carboplatin chemotherapy; PRO, patient-reported outcome; QLQ-C30, quality of life questionnaire-core 30; Tisle + nPC, tislelizumab + nab‑paclitaxel and carboplatin chemotherapy Kaplan–Meier plot for OS by treatment arm (Tisle + PC vs. PC) in QLQ-LC13 coughing domain. Note The dashed lines represent the median time-to-event for each treatment arm. Abbreviations. CI, confidence interval; OS, overall survival; PC, paclitaxel and carboplatin chemotherapy; PRO, patient-reported outcome; QLQ-LC13, quality of life questionnaire lung cancer 13 items; Tisle + PC, tislelizumab + paclitaxel and carboplatin chemotherapy Kaplan–Meier plot for OS by treatment arm (Tisle + nPC vs. PC) in QLQ-LC13 coughing. Note The dashed lines represent the median time-to-event for each treatment arm. Abbreviations. CI, confidence interval; OS, overall survival; PC, paclitaxel and carboplatin chemotherapy; PRO, patient-reported outcome; QLQ-LC13, quality of life questionnaire lung cancer 13 items; Tisle + nPC, tislelizumab + nab‑paclitaxel and carboplatin chemotherapy Discussion Using data from the RATIONALE-307 randomized trial, this post hoc analysis demonstrated that longitudinal PROs can provide independent prognostic information for OS among patients with sq-NSCLC. The novel JM-derived evidence demonstrated that Tisle + PC significantly improved GHS/QoL, dyspnea, and chest pain, while simultaneously demonstrating longer survival in the Tisle + PC arm compared to PC alone. The RS-D submodel showed that higher GHS/QoL scores were significantly associated with a reduced RS-D event risk, while worsening cough, chest pain, and dyspnea predicted increased RS-D event risk. To our knowledge, this study represents one of the first developments of a three-component JM framework in this setting, empirically evaluating the relationship between PROs and OS. In the linear mixed-effects submodel, both treatment arms (Tisle + PC and Tisle + nPC) demonstrated statistically significant and clinically meaningful improvements in GHS/QoL compared with PC alone. Dyspnea improved in both arms, irrespective of chemotherapy backbone, while improvements in chest pain were observed only in Tisle + nPC, suggesting a possible domain-specific effect of nab-paclitaxel. In the RS-D (recurrent event) submodel, worsening longitudinal PRO symptom burden was associated with a higher hazard of RS-D events, indicating a greater likelihood of subsequent symptomatic deterioration episodes over follow-up. Clinically, these trajectories may function as an early warning signal of decline and a prompt for closer monitoring and supportive intervention. The OS – treatment effect estimates from the JM demonstrated that the Tisle + PC arm experienced a lower risk of death relative to the PC arm for both cough and chest pain (both HRs = 0.64). This finding is directionally consistent with the OS HR observed in the primary analysis [23] for Tisle + PC versus PC alone (HR = 0.68; 95% CI 0.46–1.01), despite the additional adjustment for longitudinal symptom trajectories and RS-D events, reinforcing the observed OS benefit of tislelizumab obtained under such rigorous statistical scrutiny. Importantly, these findings also illustrate broader methodological implications for how PROs are evaluated. The European Society for Medical Oncology (ESMO) emphasizes strategies for treatment response and toxicity as critical research priorities. Within the framework of the ESMO Magnitude of Clinical Benefit Scale (MCBS), there is growing recognition that high-quality, clinically meaningful PRO data are essential for a balanced assessment of the benefit–risk profile of new therapies. The current MCBS framework primarily relies on GHS/QoL readouts, with less explicit emphasis on symptom-specific endpoints, and traditional hypothesis-testing approaches, such as comparing mean scores at fixed timepoints using mixed-effects model for repeated measures (MMRM), which often require large sample sizes to detect treatment differences. Additionally, single-event deterioration models, such as time-to-deterioration (TTD), assume unidirectional change (i.e., once deteriorated, always deteriorated) and fail to account for symptom recurrence. This assumption is often unrealistic in the context of PROs, where symptom trajectories are often transient and fluctuating, and, therefore, may oversimplify the complexity of patient experiences. As a result, clinically meaningful patterns may be obscured, and the impact of isolated events overstated. Rather than forcing the data to fit such constraints, we reframed TTD as a recurrent event process, thereby aligning the analytic approach with the clinical reality of how symptoms evolve in practice. This shift in analytic strategy enabled the detection of previously unidentified treatment effects that would likely have been attenuated or missed entirely using traditional single-event models. Collectively, these data provide a basis to engage ESMO and the oncology community in updating clinical benefit frameworks to explicitly incorporate longitudinal and event-based PRO dynamics, integrating both GHS/QoL and symptom-specific information, to improve the fidelity of the patient perspective in benefit–risk assessment. In application, there is growing precedent for the use of JMs within RCTs, particularly to integrate PROs into survival-based analyses [20, 40]. Beyond their analytic advantages, JMs offer strategic value in informing the selection of PRO endpoints that are treatment-sensitive, clinically meaningful, and prognostically informative. JMs also enable the generation of individual-level risk predictions, making it a valuable tool for both trial design and clinical decision-making. Findings from this study support broader adoption of JMs as a bridge between PRO data and traditional clinical endpoints, ensuring that future trials are more closely aligned with outcomes that matter to patients and stakeholders alike [41]. Clinically, the results highlight the importance of long-term symptom monitoring and risk stratification, as worsening cancer-specific symptom trajectories may serve as early indicators of treatment resistance or disease progression. Linking longitudinal symptom data with OS allows clinicians to better interpret fluctuations in symptom burden as potential signals of underlying disease change or therapeutic benefit, which may inform timely interventions to preserve HRQoL and potentially extend OS in patients with low or declining HRQoL. In practice, identifying patterns of RS-D events in key PRO domains such as dyspnea, cough, or chest pain could enable clinicians to adjust treatment strategies proactively or initiate supportive care earlier. This shifts care from reactive to anticipatory, leveraging PRO data not only to document patient experiences but also to inform monitoring protocols, guide personalized symptom management, and foster more effective clinician–patient dialogue aimed at optimizing therapeutic outcomes. Consistent with prior research, our findings reinforce that PROs can serve as independent prognostic indicators of OS in patients with advanced sq-NSCLC receiving immune checkpoint inhibitor therapy. However, despite the growing recognition of the prognostic value of PROs across multiple tumor types [7, 14, 42–45], there remains comparatively limited evidence supporting the routine use of JM frameworks within the RCT context, either prospectively or in post-hoc settings due in part to their methodological complexity and computational demand [46]. Consistent with our findings, in a notable study of metastatic pancreatic cancer, investigators compared a JM approach with an LMM to evaluate two chemotherapy regimens [21]. By incorporating the risk of death into the analysis, the JM generally produced larger estimated differences in PROs between treatment groups than the LMM. This study extends prior bodies of work by explicitly modeling patient-relevant RS-D events that are informed by the occurrence of a clinically meaningful worsening threshold (e.g., 10-point change) in relation to OS events. This study has several methodological strengths that enhance the interpretability of the findings. First, this framework provides a suitable method for missing data handling, reduces parameter estimation bias, and increases statistical efficiency through the incorporation of correlations between longitudinal, recurrent, and OS [30]. Second, because we moved from a marginal modeling approach, such as the MMRM, to a conditional random-effects model, we were able to observe a more robust assessment of how covariates influence individual patient outcomes over time. Third, by treating time as a continuous variable, we were able to align the mixed model with the survival model’s time scale, in contrast to the discrete time approach (e.g., treatment cycles) commonly used in RCT analyses. Fourth, model reliability was supported by rigorous Bayesian convergence diagnostics, including Markov chain trace and density plots, and Gelman-Rubin statistics, which reinforce confidence in the stability of posterior estimates. Despite these strengths, several limitations should be acknowledged. First, this study was limited by the open-label design, which may have influenced patient-reported questionnaire responses. However, reports and systematic reviews have found that PRO symptom scores often do not differ between arms, even when clinician-reported toxicities differ substantially [47–50]. Second, while the JM implicitly accounts for MNAR data, it does not explicitly model dropout mechanisms such as patient withdrawal due to adverse events or physician decision. Incorporating such information as competing risks or truncation indicators could further reduce bias and improve model validity. Third, the model’s reliability depends on the correct specification of the random effects structure and the functional form of the association parameters. Model misspecification could bias estimates or reduce generalizability. Fourth, one of the main barriers to implementation may be attributed to perceived difficulties in the interpretation of effect estimates and the apparent complexity of the JM framework (e.g., association structures and frailty terms) that may not be immediately intuitive to clinical audiences. Fifth, because PRO trajectories and OS may be driven by shared underlying disease processes, the longitudinal and frailty effect parameter associations should not be interpreted causally, nor as additional ‘confirmatory’ evidence of treatment efficacy on survival. Finally, future research should seek to position the JM within the estimand framework [51] to better define the research objective, guide parameter selection, and inform optimal prespecification of PRO-related analyses in RCTs, as well as evaluate the robustness of these findings using alternative RS-D operational definitions (e.g., sustained deterioration requirements and/or alternative thresholds). Conclusions This JM analysis provides quantitative evidence that PROs offer independent prognostic value for OS in patients with sq-NSCLC. The addition of tislelizumab to platinum-based chemotherapy was associated with statistically significant improvements in GHS/QoL and dyspnea in the Tisle + PC arm, and in GHS/QoL, dyspnea, and chest pain in the Tisle + nPC arm, compared with PC alone. Notably, in the Tisle + PC arm, improvements in chest pain and cough were significantly associated with a 36.3% reduction in the risk of death compared to PC alone. Overall, these results, together with previously reported clinical efficacy and safety data, reinforce the use of tislelizumab plus chemotherapy as a first-line treatment option for patients with sq-NSCLC. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This study was sponsored by BeOne Medicines, Ltd. The authors thank the participants and their families, investigators, and all study site staff for their contributions to this study. We also thank Jason Allaire from Generativity Solution Group for providing medical writing support. This study was sponsored by BeOne Medicines, Ltd. The authors thank the participants and their families, investigators, and all study site staff for their contributions to this study. Abbreviations Confidence interval European Organization for Research and Treatment of Cancer European Society for Medical Oncology-Magnitude of Clinical Benefit Scale Global health status/quality of life Hazard ratio Joint model Kaplan–Meier Linear mixed model Missing not at random Mixed-effects model for repeated measures Nanoparticle albumin-bound paclitaxel Overall survival Paclitaxel and carboplatin chemotherapy Programmed death-ligand 1 Patient-reported outcome Quality of life questionnaire-core 30 Quality of life questionnaire-lung cancer 13-item module Randomized controlled trial Recurrent symptom deterioration Squamous non-small cell lung cancer Tumor cell Tislelizumab + paclitaxel and carboplatin chemotherapy Tislelizumab + nab‑paclitaxel and carboplatin chemotherapy Time to deterioration Author contributions Jie Wang, Xinmin Yu, Shiangjiin Leaw, Frederick B. Barnes, Gisoo Barnes, and Timothy Victor contributed to the concept and design of the study, as well as the acquisition, analysis, or interpretation of data. Shiangjiin Leaw, Frederick B. Barnes, Gisoo Barnes, and Timothy Victor were involved in the initial development of the manuscript. Jie Wang, Xinmin Yu, Shiangjiin Leaw, Frederick B. Barnes, Gisoo Barnes, and Timothy Victor critically revised the manuscript for important intellectual content. Statistical analysis was performed by Timothy Victor. Funding This study was funded by BeOne Medicines, Ltd, who obtained the regulatory and ethics approvals, developed the trial protocol, provided study drugs, collected and analyzed the data, and funded medical writing. The authors assisted with data collection, analysis, and interpretation. BeOne and the authors approved the final version of the manuscript for submission. Data availability On request, and subject to certain criteria, conditions, and exceptions, BeOne Medicines, Ltd. will provide access to individual deidentified participant data from applicable BeOne Medicines-sponsored studies. BeOne Medicines shares data only when permitted by applicable data privacy and security laws and regulations, shares when it is feasible to do so without compromising the privacy of the study participants, and other considerations. Data requests may be submitted to ClinicalTrials@beonemed.com. Declarations Competing interests GB, FBB, and SL are employees of and own stock in BeOne Medicines, Ltd. No other disclosures were reported. Ethical approval and consent to participate The trial was conducted in accordance with the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use guidelines, the principles of the Declaration of Helsinki, and country-specific regulatory requirements. The protocol was approved by the relevant Institutional Review Board/Independent Ethics Committee for each study site. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement On request, and subject to certain criteria, conditions, and exceptions, BeOne Medicines, Ltd. will provide access to individual deidentified participant data from applicable BeOne Medicines-sponsored studies. BeOne Medicines shares data only when permitted by applicable data privacy and security laws and regulations, shares when it is feasible to do so without compromising the privacy of the study participants, and other considerations. Data requests may be submitted to ClinicalTrials@beonemed.com.