Measuring performance trajectories in lung cancer surgery: a longitudinal study using the French national hospital database from 2020 to 2024 Authors: Alain Bernard, Jonathan Cottenet, Pascale Tubert-Bitter, Catherine Quantin Published: 2026 ABSTRACT ObjectivesLung cancer surgery is associated with significant postoperative morbidity (15%–40%) and mortality (2%–5%), with considerable interhospital variability. This study conducts the first national longitudinal analysis of performance trajectories in lung cancer surgery using advanced trajectory modelling approaches to understand temporal dynamics of quality evolution across hospitals.MethodsWe included all adult patients (French national hospital database, n=56 299; 148 hospitals) who underwent lung resection for primary lung cancer in 2020–2024. The primary outcome was severe complications including major postoperative complications and 30-day mortality. Annual logistic regression models incorporating clinical and procedural variables calculated risk-adjusted rates (RAR) for each hospital. Hospitals were classified using funnel plots corrected for overdispersion. Hidden Markov Models estimated transition probabilities between performance categories, while group-based multitrajectory models identified subgroups of hospitals sharing similar evolution profiles, stratified by activity volume (low, <43; medium, 43-100; high, >100 procedures/year).ResultsPredictive models demonstrated excellent stable discrimination (mean area under the ROC curve=0.890). Markov models revealed strong performance inertia (persistence probabilities ≥78%). Trajectory analysis identified a complex non-linear relationship between volume and quality. Medium-volume hospitals showed optimal improvement trajectories (45.4–45.8% RAR reduction). Conversely, some low-volume centres experienced catastrophic deterioration (843% RAR increase), while paradoxically, a subgroup of high-volume hospitals also showed concerning performance degradation (151% increase).DiscussionThis study demonstrates heterogeneous performance trajectories in French lung cancer surgery, challenging the assumption that high-volume alone guarantees quality and suggesting organisational complexity challenges.ConclusionThis study shows the value of implementing systematic outcome monitoring using existing administrative data to target quality improvement interventions and reduce inter-hospital performance variations. WHAT IS ALREADY KNOWN ON THIS TOPIC For lung cancer surgery, while several European countries have implemented national surgical outcome monitoring systems enabling continuous quality improvement through benchmarking and outcome tracking, other countries such as France lack such a systematic evaluation framework despite possessing comprehensive national hospital discharge data. For lung cancer surgery, while several European countries have implemented national surgical outcome monitoring systems enabling continuous quality improvement through benchmarking and outcome tracking, other countries such as France lack such a systematic evaluation framework despite possessing comprehensive national hospital discharge data. WHAT THIS STUDY ADDS First national trajectory analysis of lung cancer surgery performance across all French hospitals (56 299 patients, 2020–2024) using Hidden Markov Models and group-based trajectory modelling. First national trajectory analysis of lung cancer surgery performance across all French hospitals (56 299 patients, 2020–2024) using Hidden Markov Models and group-based trajectory modelling. Trajectory modelling with existing administrative data can detect hospitals entering decline before major quality degradation occurs, thus enabling targeted interventions rather than reactive crisis management, contrary to static cross-sectional analyses that fail to capture the temporal dynamics of hospital quality trajectories: some low-volume hospitals experienced catastrophic decline with complication rates increasing nearly ninefold (a warning signal invisible to traditional static evaluations), medium-volume hospitals achieved the best improvement trajectories (45% reduction in risk-adjusted complications), while a subgroup of high-volume centres paradoxically deteriorated by 151%. Trajectory modelling with existing administrative data can detect hospitals entering decline before major quality degradation occurs, thus enabling targeted interventions rather than reactive crisis management, contrary to static cross-sectional analyses that fail to capture the temporal dynamics of hospital quality trajectories: some low-volume hospitals experienced catastrophic decline with complication rates increasing nearly ninefold (a warning signal invisible to traditional static evaluations), medium-volume hospitals achieved the best improvement trajectories (45% reduction in risk-adjusted complications), while a subgroup of high-volume centres paradoxically deteriorated by 151%. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Our findings demonstrate a clearly non-linear relationship between surgical volume and quality evolution, challenging the long-standing international assumption that high volume inherently guarantees superior outcomes Our findings demonstrate a clearly non-linear relationship between surgical volume and quality evolution, challenging the long-standing international assumption that high volume inherently guarantees superior outcomes Volume-based centralisation policies alone are insufficient: organisational factors and continuous performance feedback matter as much as surgical volume for ensuring quality care. Volume-based centralisation policies alone are insufficient: organisational factors and continuous performance feedback matter as much as surgical volume for ensuring quality care. Introduction Surgical resection remains the gold standard for early-stage lung cancer, offering the highest likelihood of long-term survival.1 2 However, postoperative risks remain substantial, with major complications occurring in 15%–40% of cases and perioperative mortality reaching 2%–5%.3 4 Such variability raises fundamental questions about the determinants of care quality and the persistent heterogeneity observed across healthcare institutions. Over the last two decades, assessing quality of care in thoracic oncology has become a central concern. Interhospital disparities consistently exceed what would be expected by chance, underscoring the role of organisational and human factors on outcomes.5 6 While the relationship between surgical volume and clinical results has been widely investigated, it remains debated and cannot alone define hospital performance.7 Several European countries have addressed this challenge through national monitoring systems providing systematic feedback and public reporting, such as the UK National Lung Cancer Audit, the Dutch DICA and Scandinavian national registries.8–11 In contrast, France lacks routine nationwide evaluation of surgical outcomes, despite the availability of comprehensive hospital discharge data, limiting the monitoring of trends and targeted quality improvement initiatives.12 Existing research often relies on cross-sectional or multiyear averaged comparisons that obscure temporal dynamics and fail to capture13 14 institutions undergoing improvement or deterioration. Trajectory-based analytical frameworks offer promising advances. Hidden Markov Models (HMM) enable transitions between quality states over time,15 while group-based multitrajectory models (GBMTM) identify clusters of institutions sharing similar performance trajectories.16 France constitutes an especially relevant setting for such analyses. Its universal healthcare system, mixed public–private landscape and comprehensive PMSI database (Programme de Médicalisation des Systèmes d’Information, French national hospital database) allow exhaustive longitudinal follow-up of hospital activity. Using these data, this study proposes a national longitudinal analysis of performance trajectories among all institutions performing lung cancer surgery in France between 2020 and 2024, combining risk adjustment, funnel plot classification and advanced trajectory modelling to establish baseline performance metrics and demonstrate the feasibility of trajectory-based quality assessment. Material and methods Data source and study population From the French national hospital database (PMSI),17 we included all adult patients who underwent surgical lung resection for primary lung cancer between January 2020 and December 2024 (this period was selected for several reasons detailed in online supplemental file 1). A total of 148 hospitals participated in thoracic oncologic surgery activity during the study period. These patients were identified through a principal discharge diagnosis of lung cancer (International Classification of Diseases-10 code C34) associated with a lung cancer surgery procedure (Common Classification of Medical Procedures codes for thoracotomy, video-assisted thoracic surgery (VATS) or robot-assisted surgery) during the same hospital stay. Patient characteristics At baseline, we identified for all patients age and gender, as well as surgery-related variables such as the surgical approach (thoracotomy, VATS or robot-assisted surgery), the type of resection (limited resection, lobectomy, bilobectomy and pneumonectomy), bronchoplasty and the extent of the pulmonary resection. We also included comorbidities described in online supplemental table 1. Finally, we calculated a modified Charlson Comorbidity Index as a marker of comorbidity.18 Hospital characteristics In France, hospitals are classified as either non-academic public, academic public, private non-profit or private for-profit. For each hospital, we also calculated the hospital volume, defined as the mean number of thoracic procedures performed per year (2020–2024), allowing facilities classification into three categories according to activity volume terciles: <43 procedures per year, 43–100 procedures per year and >100 procedures per year. Primary outcome variable Our main outcome was the major complication rate based on the Clavien–Dindo classification (online supplemental table 2), transformed into a binary variable: the variable was equal to 1 if the Clavien–Dindo classification was higher than 2. This grade >2 rating included severe 30-day postoperative complications and 30-day in-hospital mortality. Severe postoperative complications were defined as the presence of one or more of the postoperative conditions (online supplemental table 3). Thirty-day in-hospital mortality was defined as all deaths occurring during the same hospital stay as the operation or within 30 days of the operation. Hereafter, we will refer to this outcome (Clavien–Dindo >2) as ‘severe complication’. These criteria have been validated in previous studies.19 Statistical analyses Model specification A logistic regression model was applied separately for each year from 2020 to 2024 to predict the individual probability of major complications and to calculate risk-adjusted rates (RAR) by hospital. The model incorporated for each year the following covariates: age (continuous variable), sex, comorbidities, Charlson score, extended resection, sleeve lobectomy, type of pulmonary resection and surgical approach. These variables were selected based on clinical relevance established in prior literature18 19 and data availability in the PMSI database. RAR calculation For each hospital and each year, the RAR was calculated according to the standardised indirect morbidity ratio method multiplied by the overall complication rate.20 Annual validation of predictive models Each annual model was subjected to a rigorous validation procedure evaluating several dimensions of predictive performance according to recent methodological recommendations for clinical prediction models. Model discrimination was estimated by the area under the ROC curve. Calibration was evaluated using several complementary metrics: Brier score, Brier scaled and root mean square error.21 Hospital classification Funnel plots were constructed annually according to the methodology recommended by Verburg et al22 for comparative evaluation of hospital quality. This approach distinguishes between performance variations attributable to chance from systematic differences reflecting actual quality of care. Using these funnel plots, hospitals were classified as worse, as expected or better. The overdispersion coefficient φ was calculated according to the method of Verburg et al22 to quantify inter-hospital heterogeneity beyond the expected binomial variability. HMM To estimate the trajectories of hospitals classified as worse, as expected or better during the period from 2020 to 2024, we used HMM.15 In this context, each hospital is classified annually according to three performance categories: worse (insufficient performance), as accepted (satisfactory performance) or better (superior performance). This methodology estimates transition probabilities between states, allowing us to quantify, for example, the probability that a hospital classified as accepted in a given year will improve its performance (better), maintain its status (as accepted) or regress (worse) the following year. GBMTM The GBMTM identifies groups of hospitals sharing similar RAR evolution trajectories over the 2020–2024 period.16 Each group is described by a polynomial curve of time (order 3), and models with 1 to 6 classes were estimated to document the solution landscape. The final number of classes was set at three based on a compromise between goodness of fit (Akaike’s Information Criterion (AIC)/Bayesian Information Criterion (BIC) and size-adjusted size Bayesian Information Criterion (sBIC)), parsimony principle and clinical readability. The model was applied to hospitals classified according to their activity volume (low, medium, high). All analyses described above are more detailed in online supplemental file 2. All statistical analyses were performed using R V.4.4.1 and Stata 19.0 (StataCorp, College Station, TX, USA). The statistical significance threshold was set at p<0.05. Ethics Patient consent was not required, as the French national hospital data are based on pseudonymised data, that is, they do not contain any identifying information. The patient’s identity is pseudonymised, which allows data from the same patient to be linked without knowing the patient’s identity. The study was approved by the National Committee for Data Protection: declaration of conformity to the methodology of reference 05 obtained on 7 August 2018 under the number 2204633 v0. Results Population characteristics During the 2020–2024 period, the number of hospitals performing pulmonary resections for cancer was 148, and 56 299 patients were included in the analysis. Table 1 presents baseline patient characteristics for the overall cohort, stratified by year. Annual activity increased steadily from 10 977 patients (19.5%) in 2020 to 12 936 patients (23%) in 2024, with a decrease in 2021 to 9588 patients (17%). Men represented approximately 70% of the study population across all years. A detailed breakdown of patient characteristics by sex is provided in online supplemental table 4. Table 1Patients characteristics according to year20202021202220232024P valueNumber of patients10 9779588819214 60612 936Pulmonary disease3533 (32.2%)2778 (29.0%)2763 (33.7%)3994 (27.3%)3806 (29.4%)<0.001Heart disease1766 (16.1%)1486 (15.5%)1359 (16.6%)2109 (14.4%)1862 (14.4%)<0.001Vascular peripheral disease928 (8.5%)970 (10.1%)663 (8.1%)1153 (7.9%)985 (7.6%)<0.001Neurological disease411 (3.7%)359 (3.7%)228 (2.8%)609 (4.2%)447 (3.5%)<0.001Renal disease256 (2.3%)298 (3.1%)267 (3.3%)434 (3.0%)328 (2.5%)<0.001Metabolic disease1377 (12.5%)1223 (12.8%)1079 (13.2%)1532 (10.5%)1628 (12.6%)<0.001Anaemia1572 (14.3%)1656 (17.3%)898 (11.0%)1322 (9.1%)1462 (11.3%)<0.001Infectious disease71 (0.6%)42 (0.4%)38 (0.5%)72 (0.5%)67 (0.5%)0.250Haematologic disease311 (2.8%)397 (4.1%)426 (5.2%)345 (2.4%)541 (4.2%)<0.001Other disease3727 (34.0%)3588 (37.4%)3975 (48.5%)5955 (40.8%)5741 (44.4%)<0.001Other treatment346 (3.2%)1331 (13.9%)739 (9.0%)989 (6.8%)1095 (8.5%)<0.001Male7154 (65.2%)6210 (64.8%)4917 (60.0%)8896 (60.9%)8147 (63.0%)<0.001Female3823 (34.8%)3378 (35.2%)3275 (40.0%)5710 (39.1%)4789 (37.0%)Age (years)65±9.865±9.664.8±9.565±9.566±9.5<0.001Lobectomy8401 (76.5%)7225 (75.4%)6637 (81.0%)11 184 (76.6%)10 161 (78.5%)<0.001Bilobectomy441 (4.0%)340 (3.5%)232 (2.8%)427 (2.9%)348 (2.7%)<0.001Pneumonectomy770 (7.0%)562 (5.9%)530 (6.5%)793 (5.4%)710 (5.5%)<0.001Charlson=11685 (15.4%)1263 (13.2%)691 (8.4%)1565 (10.7%)1117 (8.6%)<0.001Charlson=21602 (14.6%)1115 (11.6%)1206 (14.7%)1469 (10.1%)1568 (12.1%)<0.001Charlson≥33803 (34.6%)3473 (36.2%)3634 (44.4%)5673 (38.8%)5690 (44.0%)<0.001VATS2285 (20.8%)1589 (16.6%)1565 (19.1%)3749 (25.7%)2871 (22.2%)<0.001VATS or robot1420 (12.9%)1447 (15.1%)1404 (17.1%)1900 (13.0%)1704 (13.2%)<0.001Robot382 (3.5%)921 (9.6%)592 (7.2%)608 (4.2%)1555 (12.0%)<0.001Extended resection769 (7.0%)1122 (11.7%)697 (8.5%)1506 (10.3%)1370 (10.6%)<0.001Sleeve lobectomy245 (2.2%)87 (0.9%)133 (1.6%)545 (3.7%)298 (2.3%)<0.001VATS, video-assisted thoracic surgery. Patients characteristics according to year VATS, video-assisted thoracic surgery. The major complication rate was 34.75% (n=3814) in 2020 and 31.25% (n=4075) in 2024. Within the composite severe complication outcome (Clavien-Dindo >2), 30-day in-hospital mortality rates were 3.2% (n=351) in 2020, 3.0% (n=288) in 2021, 2.8% (n=312) in 2022, 2.7% (n=343) in 2023 and 2.6% (n=336) in 2024, showing a consistent downward trend over the study period. Performance and validation of predictive models The discriminant performance of the logistic regression models remained excellent and stable for each year throughout the period (2020–2024), with a mean AUC of 0.890 (range: 0.884-0.896, table 2). This stability suggests that the predictive factors for major complications have not substantially evolved between 2020 and 2024. Calibration scores were also satisfactory, with Brier scaled values ranging from 0.446 to 0.469, indicating that the models correctly predicted absolute probabilities of complications. Table 2Validation performance of predictive models by year (2020–2024)YearAUCBrier scoreRMSEφ20200.8900.4520.2851.4520210.8840.4460.2871.5220220.8960.4690.2791.3820230.8910.4610.2831.4120240.8890.4580.2811.43All φ values were significantly greater than 1, justifying the correction of funnel plot control limits.AUC, area under the ROC curve; RMSE, root mean square error; φ, overdispersion coefficient. Validation performance of predictive models by year (2020–2024) All φ values were significantly greater than 1, justifying the correction of funnel plot control limits. AUC, area under the ROC curve; RMSE, root mean square error; φ, overdispersion coefficient. Hospital classification Online supplemental figure 1 presents the annual funnel plots of hospital performance in oncologic thoracic surgery from 2020 to 2024. Hospitals are classified into three categories: better than expected (below the lower red limit), as expected (between the red limits) and worse than expected (above the upper red limit). Table 3 presents the annual hospital ranking based on funnel plots from 2020 to 2024, showing the distribution of hospitals across the three performance categories. Table 3Annual hospital ranking based on funnel plots (2020–2024)YearWorse (n, %)As expected (n, %)Better (n, %)202010 (6.8%)126 (86.3%)10 (6.8%)20219 (6.3%)127 (89.4%)6 (4.2%)202211 (7.4%)130 (87.8%)7 (4.7%)20239 (6.4%)125 (88.7%)7 (5.0%)202410 (7.0%)125 (88.0%)7 (4.9%) Annual hospital ranking based on funnel plots (2020–2024) Transition analysis using markov models HMM was used to estimate annual transition probabilities between the three performance categories (better, as expected, worse). The results revealed strong inertia in performance states, with persistence probabilities in the initial state exceeding 78% for all states (table 4). Table 4Transition probability matrix between performance categories (Markov model)Initial state→ better→ as expected→ worseBetter34.0%65.9%0.1%As expected16.5%78.2%5.3%Worse2.2%64.2%33.6% Transition probability matrix between performance categories (Markov model) High-performing hospitals had a 34% probability of maintaining this performance the following year or a 65.9% probability of having as expected performance. Conversely, underperforming hospitals had a 33.6% probability of remaining in this category, with a low chance (2.2%) of improvement to the ‘better’ category. The most probable transitions were toward the intermediate ‘as expected’ category, which acts as a buffer state. Direct transitions between extreme categories (better↔worse) were rare (<2%), suggesting that performance changes follow a gradual process rather than abrupt jumps. Performance trajectory modelling using group-based trajectory model This analysis revealed an inverted U-shaped curve of performance according to volume (online supplemental figure 2). The detailed analysis by group of hospital volume, allowing the classification of institutions sharing similar evolution profiles, revealed interesting results. Medium-volume hospitals displayed the best improvement trajectories. Conversely, the extremes (low and high) presented subgroups with strong deterioration. Paradoxically, high-volume hospitals were not immune to difficulties, such as presented below. Low-volume hospitals The 34 low-volume hospitals (<43 procedures/year) exhibited very high homogeneity (entropy=0.001), with two diametrically opposed trajectories revealing sector polarisation. Small facilities showed an alarming bimodal trajectory. While Class 1 maintained slight improvement, Class 2 experienced catastrophic deterioration with RAR multiplied by 9.4 and a relative change of 843% (table 5). Table 5Evolution trajectories over the 2020–2024 period according to the volume of activity (low, medium, high) using the group-based multi-trajectory modelNumber of hospitalsRAR 2020RAR 2024Evolution% ChangeTrajectoryLow volume (<43 procedures/ year) Class 1170.3750.349−0.026−6.9%Improvement Class 2170.5485.170+4.622+843%DeteriorationMedium volume (43–100 procedures/year) Class 1180.6490.352−0.297−45.8%Improvement Class 2180.3790.207−0.172−45.4%Improvement Class 3170.3060.288−0.018−6.0%StabilityHigh volume (>100 procedures/year) Class 1180.4770.273−0.204−42.8%Improvement Class 2180.3020.285−0.017−5.5%Stability Class 3180.2110.531+0.320+151%Deterioration Evolution trajectories over the 2020–2024 period according to the volume of activity (low, medium, high) using the group-based multi-trajectory model Medium-volume hospitals The 53 medium-volume hospitals (43–100 procedures/year) represented the largest group and displayed moderate heterogeneity (entropy=0.119), with three distinct predominantly positive trajectories. Medium-volume hospitals demonstrated excellent adaptive capacity. Classes 1 and 2 showed remarkable improvements (45.8% and 45.4%), suggesting that this critical size offers an optimal balance between critical mass and organisational agility. Class 3 remained stable, consolidating an already satisfactory performance level (table 5). High-volume hospitals The 54 high-volume hospitals (>100procedures/year) presented notable homogeneity (entropy=0.016), with three trajectories revealing contrasting dynamics between operational excellence and structural difficulties. High-volume hospitals revealed a paradoxical situation. While Class 1 demonstrated strong improvement (42.8% reduction in RAR), Class 3 underwent concerning deterioration of 151% (table 5). Discussion This national study demonstrates for the first time the existence of fairly heterogeneous performance trajectories in lung cancer surgery in France over the period 2020–2024. The longitudinal analysis of 56 299 patients treated in 148 facilities reveals three major findings. First, annual logistic regression models show excellent stable discriminant performance (average AUC=0.890), confirming the robustness of predictive factors for major complications throughout the period. Second, the Markov models used identify strong inertia in performance states with persistence probabilities exceeding 78% for all categories, suggesting that performance changes follow gradual processes rather than abrupt transitions. Third, group trajectory modelling analysis reveals a complex non-linear relationship between activity volume and quality evolution, partially challenging the classical hypothesis that high volume systematically guarantees better outcomes, particularly for high-volume centres. Performance trajectories by activity volume Taken together, the trajectory analysis reveals an inverted U-shaped relationship between hospital volume and performance improvement. This inverted U-shaped curve has been previously hypothesised in the surgical literature23 but rarely demonstrated longitudinally at the national level. Rather than a monotonically positive volume–outcome association, our findings demonstrate that medium-volume hospitals (43–100 procedures/year) occupy the apex of this curve, achieving the most favourable quality trajectories. At both extremes of the volume spectrum, subgroups of hospitals exhibited deteriorating performance. This non-linear pattern suggests the existence of two distinct thresholds that are explained in online supplemental file 3. The detailed analysis by group of hospital volume, allowing the classification of institutions sharing similar evolution profiles, revealed interesting results. Facilities with medium volume (43–100 procedures/year) show the best improvement trajectories, with RAR reductions of 45.8% and 45.4% in Classes 1 and 2, while Class 3 maintains stable and satisfactory performance. This finding suggests that these facilities benefit from an optimal balance between sufficient critical mass to maintain surgical expertise and organisational agility enabling rapid adaptation to evolving practices. Conversely, low-volume facilities (<43 procedures/year) reveal an alarming situation. While Class 1 maintains slight improvement, Class 2 undergoes catastrophic deterioration with a 9.4-fold increase in RAR (843% increase). This situation suggests that some low-volume centres face a major structural crisis, likely related to insufficient critical mass to maintain the necessary technical and organisational expertise. However, these results should be interpreted with caution as it partly reflects the small baseline event counts inherent to low-volume centres. Paradoxically, high-volume facilities (>100 procedures/year) are not immune to difficulties. While Class 1 demonstrates strong improvement (42.8% RAR reduction), Class 3 experiences concerning deterioration of 151%. Importantly, since the RAR is derived from logistic regression models that incorporate all available patient-level and procedural variables, a shift in case-mix complexity captured by these variables cannot explain this deterioration. The persistent increase in RAR after risk-adjustment therefore points to factors not captured by the model. Three mechanisms may be involved which are detailed in online supplemental file 4. This methodology applied to hospital trajectories in the USA13 for myocardial infarction and heart failure management showed that hospitals identified as highly performing were frequently similar across targeted pathologies and over time, suggesting that performance may be determined by care systems influencing different pathological conditions comparably. Performance trajectory evaluation could offer incremental information compared with traditional static single-year assessments, and value-based health programmes might benefit from a more comprehensive hospital-level approach in addition to a pathology-targeted approach.13 Temporal dynamics and performance inertia Markov models reveal strong inertia, with persistence probabilities ranging from 78.2% for the ‘as expected’ category to 34% for extreme categories (‘better’ and ‘worse’ performance status). This inertia has important practical implications: it suggests that quality improvement interventions must be early and sustained because once a facility enters a deterioration trajectory, it is then difficult to reverse the trend if the intervention is late. The ‘as expected’ category functions as a central buffer state in the Markov transition structure, absorbing the majority of transitions from both extreme categories: 65.9% of hospitals classified as ‘better’ and 64.2% of those classified as ‘worse’ transition to this intermediate state in the following year, while direct transitions between extreme categories remain exceptionally rare (<2%). This finding has several important implications detailed in online supplemental file 5. This method is ultimately rarely used to estimate hospital trajectories; in Italy, it enabled estimation of quality of care trajectories for elderly patients, but to our knowledge this method has never been used for surgery.24 Unlike several European countries that have implemented national systems for systematic evaluation of surgical outcomes (UK National Lung Cancer Audit, Dutch Institute for Clinical Auditing, Swedish national registries, Danish clinical databases),8–11 France does not have a systematic evaluation framework despite the existence of a national medico-administrative database, including all hospitals (public and private). This absence of routine monitoring represents a significant gap in quality surveillance and limits the ability to identify underperforming facilities, track temporal trends or guide quality improvement initiatives at the national level. Several factors explain why France has lagged behind in implementing routine surgical outcome monitoring, unlike the UK,25 the Netherlands26 or Scandinavian countries,27 as explained in online supplemental file 6. Our study demonstrates the feasibility of continuous surveillance using already available administrative data. The trajectory modelling approach offers a relevant alternative to static cross-sectional evaluations practiced in other countries, enabling early identification of facilities at risk of deterioration before major degradation occurs. This prospective approach represents a significant methodological advance over traditional benchmarking systems that are limited to point-in-time comparisons. Study strengths The methodological approach is innovative by jointly applying HMM and GBMTM in thoracic oncological surgery, offering a dynamic and prospective vision of quality evolution. The exhaustiveness and representativeness of the national PMSI database including 56 299 patients treated in 148 facilities over 5 years, ensuring complete coverage of the French healthcare system. The robustness of risk adjustment: predictive models with excellent discrimination (AUC≈0.89) and satisfactory calibration, integrating 20 relevant clinical and procedural variables. The longitudinal perspective adds insights beyond cross-sectional analyses by capturing temporal dynamics and population-level trends. The feasibility and reproducibility through the use of already available administrative data, enabling immediate implementation without additional collection. Limitations Clinical data such as respiratory function and the complete TNM stage (we only have the metastatic stage) and histological data are not directly available in the PMSI database. However, a study showed that this point ultimately has limited impact on predictive models.28 Moreover, the 30-day outcome captured in-hospital events (including transfers) but deaths occurring outside the hospital within 30 days are not captured, which may underestimate the true complication and mortality rates. Furthermore, our hospital classification relies solely on activity volume, without considering structural (technical platform, intensive care unit availability, staff/bed ratio), organisational (enhanced recovery protocols, quality programmes) or human factors (surgical expertise, training). In particular, the paradoxical deterioration observed in some high-volume centres likely reflects organisational mechanisms that cannot be measured through administrative data. Further studies integrating clinical registries with organisational data are thus needed to disentangle these underlying mechanisms. In addition, we acknowledge that the GBMTM framework as implemented does not directly produce CIs for RAR values within each trajectory class. Furthermore, the study period (2020–2024) encompasses the COVID-19 pandemic, which may have affected hospital performance through multiple mechanisms including restricted ICU bed availability, staff burnout, hospital reorganisation and delayed patient presentations. While the decrease in surgical volume observed in 2021 likely reflects pandemic-related disruption, several elements suggest that the deterioration trajectories identified are not primarily attributable to COVID-19 effects, as explained in online supplemental file 7. Assessing the association between pandemic-related workload and performance trajectories would be a valuable complementary analysis. Further research is needed to carry out such a study, which would require a specific design to identify COVID-19-related admission burden by hospital, taking into account the evolution of this impact over time due to geographical and temporal variations in the epidemic. Conclusion This nationwide study provides the first evidence of heterogeneous performance trajectories in lung cancer surgery in France. Our findings reveal a non-linear relationship between surgical volume and quality evolution, challenging the widely held international assumption that high volume consistently ensures superior outcomes. Notably, several high-volume centres exhibited concerning patterns of deterioration, underscoring the need for vigilant monitoring and targeted interventions. In contrast, medium-volume hospitals demonstrated the most favourable improvement trajectories. Beyond these substantive results, in countries that do not yet have a national, systematic evaluation system, such as France, our study provides clear evidence that trajectory-based quality assessment can be implemented effectively using routinely available data sources. In countries where systematic evaluation systems already exist, this dynamic approach may represent a decisive improvement. As it enables the early and reliable identification of facilities entering a decline, well before major deterioration becomes detectable, it could therefore ensure more accurate targeting of quality-improvement interventions and more rigorous monitoring of their effectiveness over time, thus improving patient outcomes.