Association of CT-assessed sarcopenia and the development of immune-related adverse events and overall survival: a multistate survival analysis.

Introduction

Background
Immune checkpoint inhibitors (ICIs) have improved survival in patients with non-small cell lung cancer (NSCLC) compared to chemotherapy (1). However, their effectiveness is limited by clinical challenges, including immune-related adverse events (irAEs) (2). While ICIs have a better toxicity profile compared to chemotherapy, irAEs can lead to treatment interruption (3). Moreover, despite their advantages, complete response rates and median overall survival (OS) remain modest, especially in advanced stages of the disease.

Rationale and knowledge gap
Identifying predictors of ICI response is relevant for optimizing outcomes. Skeletal muscle mass (SMM) has emerged as a predictor of ICI response, with higher SMM associated with better survival (4,5). Low-dose computed tomography (LDCT), part of routine lung cancer diagnostics, allows cost-effective assessment of radiological sarcopenia and its association with irAEs and mortality. Previous studies have explored the relationship between CT-assessed sarcopenia and irAEs, with inconclusive results. Limitations of these studies include focusing on one single type of ICI, small sample sizes, and not accounting for the competing risk of death (4,6,7). Although irAEs correlate with improved survival (8), our previous work found the presence of CT-assessed sarcopenia to be associated with reduced irAE risk (9).
The trajectories of NSCLC patients from the start of ICI treatment to treatment success or death, often involving irAEs, are dynamic and complex. To date, the association between predictors such as CT-measured sarcopenia and transitions between health states, including irAEs and death, has not been adequately investigated. Multistate time-to-event models offer valuable insights into these transitions and evolving risks over time.

Objective
This study aims to (I) analyze transition probabilities from the start of treatment to irAEs and death and (II) evaluate the association between CT-measured sarcopenia, irAEs, and mortality in NSCLC patients treated with ICIs. Using a multistate time-to-event model, we extend previous findings (9), and capture dynamic transitions between health states, providing a comprehensive understanding of these relationships. We present this article in accordance with the STROBE reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-aw-1140/rc).

Methods

Study population and data collection
We performed a retrospective cohort analysis. Patients were included if they were adults >18 years of age and diagnosed with NSCLC, treated with ICIs between October 2015 and April 2021 at the University Medical Center Groningen (UMCG), the Netherlands, and who had available LDCT scans prior to ICI the start of treatment. OncoLifeS is a data biobank that collects biological specimens, clinical and pathological data from patients with cancer treated in the UMCG (10). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The OncoLifeS study was approved by the medical ethics committee of the University Medical Center Groningen (UMCG) (No. 2010/109) and has been ISO certified (9001:2008 Healthcare). It has been also registered in the Dutch Trial Register under the number: NL7839. These patients signed informed consent to join the Oncological Life Study (OncoLifeS).

Exposure definition
The definition of the exposure has been described previously (9). Briefly, SMM was assessed using LDCT scans, digitally available through electronic medical records, and performed within two months prior to the first ICI cycle, including the lumbar vertebra L3 level. The L3 level was chosen because it is representative of overall SMM (11). SMM was measured using an in-house machine learning software (SarcoMeas AI v1.00, Groningen, the Netherlands) that automatically delineates the skeletal muscle area in the CT slices (12) applying attenuation thresholds from –29 to +150 Hounsfield units (HU). If necessary, the images where manually corrected by an experienced radiologist. The L3 Skeletal Muscle Index (L3SMI) was calculated by dividing the skeletal muscle area at L3 (cm2) by the patient’s height squared (m2). We employed sex-specific cut-off points to categorize patients by the presence or absence of CT-assessed sarcopenia, as proposed by Prado et al. in a similar cohort to ours (oncology patients of a similar age, from a predominantly Western European ancestry). CT-assessed sarcopenia was defined as L3SMI <38.5 cm2/m2 in women and <52.4 cm2/m2 in men (13). A sensitivity analysis was performed using L3SMI as a continuous variable.

Outcome definition
In this study we focused on two clinical outcomes: the development of severe and all immune related adverse event and death by any cause. The outcomes have been described previously (9). Briefly, the first outcome, irAE development, was collected from medical records. IrAEs were defined using the Naranjo algorithm, where a score of ≥2 was considered a possible association and included in the analysis (14). Medical records were reviewed for any reported irAEs from the start of treatment until death or the end of the study, set at 2 years after the start of ICI therapy for the last included patient (August 21, 2023). The Common Terminology Criteria for Adverse Events (CTCAE) version 5.0 was used to classify irAEs as mild (CTCAE score 1–2) or severe (CTCAE score ≥3) (15). The second outcome, OS, was defined as death from any cause, as recorded in the civil registry. Life status was followed from the start of ICI therapy until death or censored at the end of the study.

Adjustment variables definition
The variables used in this study to fit our multi state time-to event models have been described previously (9). Briefly, data regarding sex, age, body mass index (BMI), tumor stage, performance status, ICI employed, combination with other systemic therapies (other immunotherapy agents, chemotherapy or targeted therapy), line of treatment were retrieved from electronic medical records. Age was recorded at the start of the ICI therapy. Tumor stage was categorized into stages III and IV using the tumor-node-metastasis (TNM) classification (8th edition) (16). BMI was calculated as weight in kilograms divided by height squared in meters and was collected at the start of ICI treatment. Performance status was assessed at ICI initiation using the Eastern Cooperative Oncology Group (ECOG) criteria, where patients were classified as asymptomatic [0], symptomatic but fully ambulatory [1], symptomatic with less than 50% of the day in bed [2], symptomatic with more than 50% of the day in bed [3], bedbound [4], or deceased [5] (17). Combined treatment was defined as chemotherapy with one or more ICIs or two ICIs administered together. Treatment line was classified as first-line (no previous systemic treatment or recurrence after >24 months) or second line and beyond (one or more previous treatments).

Statistical analysis
Descriptive statistics were used to present the clinical and pathological characteristics, stratified by the presence of irAEs. These analyses were presented based on irAE severity (severe and any-grade irAEs), as well as for the group without irAEs.

Multistate time-to-event model
To explore the overall behavior of the transition between health states and the association of CT-assessed sarcopenia with these transitions, we developed a semi-Markov multistate time-to-event models with clock reset on entry of the irAE state for both any-grade irAEs and severe irAEs. The construction of the models was made as follows (Figure 1).

States
❖ State 0 (start of ICI treatment): initial state when patients start treatment with ICI therapy.

❖ State 1 [development of (severe) irAEs]: patients progress to this state if they experience (severe) irAEs during treatment. This was defined as a transient state, meaning that patients can transition to and from this state. A clock reset was performed when patients entered this state.

❖ State 2 (death): death was defined as the absorbing state in the model, meaning that patients can transition to it but cannot move away from it.

Transitions
Three transitions were possible:
❖ Transition 0–1: transition between “no event” (state 0) and the “development of irAEs” (state 1).

❖ Transition 0–2: transition from “no event” (state 0) to “death” (state 2), without experiencing irAEs.

❖ Transition 1–2: transition to “death” (state 2) after the “development of irAEs” (state 1).

Death without prior irAEs was treated as a competing risk for the transition from treatment initiation to irAE.

Transition probabilities
Transition probabilities were calculated from the transition hazards obtained from the semi-Markov models described above and were plotted for the overall population as well as stratified by CT-assessed sarcopenia status. The Aalen-Johansen method was used to estimate transition probabilities while accounting for the competing risk of death. These probabilities were visualized using stacked probability in state curves.

Association of transition probabilities with CT-assessed sarcopenia
The association of CT-assessed sarcopenia with each allowed transition was evaluated with univariate and multivariable competing risks Cox proportional hazards models. Each multivariable model was adjusted for confounders, namely age, sex, performance status, BMI, stage of disease, treatment line, and combination therapy. The same multistate time-to-event models were applied to the entire dataset as an overall but also stratified by both irAE groups: “Severe” irAEs and “any-grade” irAEs. All statistical analyses were conducted using R software, with the “mstate” package. A significance level of 0.05 was used for all hypothesis tests.

Sensitivity analysis: time-dependent Cox model with interaction
To support the findings from the primary multi-state model, we conducted a sensitivity analysis using a Cox proportional hazards model in which irAE occurrence was included as a time-dependent covariate. This approach allowed us to dynamically assess the impact of irAEs on OS. An interaction term between the time-dependent irAE variable and baseline CT-assessed sarcopenia status was included to evaluate whether the effect of experiencing an irAE on survival differed by sarcopenia status. All models were adjusted for potential confounders, including age, sex, cancer stage, ECOG performance status (PS), treatment combination, and BMI. Separate models were fitted for severe irAEs and any-grade irAEs.

Results
We included 363 patients, predominantly males (N=221, 60.9%), with a mean age of 64.4 years [standard deviation (SD) =9.1] and mean BMI =25.6 kg/m2 (SD =4.4). Most patients were classified as having CT-assessed sarcopenia (82.4%), with a mean overall L3SMI of 42.0 cm2/m2 (SD =8.1). The mean sex-stratified L3SMI values were 45.2cm2/m2 (SD =7.6) and 37.1 cm2/m2 (SD =6.0) for males and females, respectively, with a normal distribution in both groups (Figure S1). Most patients had stage IV disease (92.8%) and usually received ICI monotherapy (82.9%) as a second and further line of treatment (66.4%). No missing data on predictors or outcomes was found. Details on demographic characteristics can be found in Table 1.

Multi state model

Transition frequencies
During the study period, a total of 166 (45.7%) patients developed any-grade irAEs and 76 (20.7%) developed severe irAEs. Details on the type of irAEs developed can be found on Table S1. At the end of the study period (96 months total follow-up), 301 (82.9%) patients had died. Of these, 117 (32.2%) patients had developed irAE and 60 (16.5%) had severe irAEs. Thirteen patients were censored at the end of the study. Any-grade and severe irAEs were more common in non-sarcopenic patients (any-grade irAEs =59%, severe irAEs =25%) compared to the sarcopenic patients (any-grade irAEs =42%, severe irAEs =20%) (Table 2).

Transition probabilities after treatment initiation, overall model description
In both the “severe” and “all” irAE models, similar transition patterns were observed. Initially, the probability of remaining event-free decreased sharply after the start of treatment. The probability of developing both severe and any-grade irAEs increased with time, reaching a peak around 6 months after starting treatment {6-month transition probability =28% [95% confidence interval (CI): 23–32%] for any-grade irAEs, 11% (95% CI: 8–14%) for severe irAEs}, and then gradually decreased as patients died. The probability of immediate death (without irAEs) increased steadily over time, reaching 50% after 10 months in both models. After the occurrence of all severe irAE, the probability of death gradually increased to 50% after 12 months in both severe and any-grade irAE models. In Figures 2,3, we present the time-dependent stacked survival curves showing the trajectories described above.

Association between CT-assessed sarcopenia and the transitions to irAEs and mortality

Severe irAEs
Both the sarcopenic and non-sarcopenic groups showed a decrease in the probability of remaining event-free shortly after the start of ICI treatment. No significant association was found between CT-assessed sarcopenia and the development of severe irAEs (HR =0.81, 95% CI: 0.42–1.58). The probability of death (without irAEs) increased consistently in both groups. The univariate model indicated a higher risk of death (without irAEs) after starting treatment in the sarcopenic group (HR =1.60, 95% CI: 1.11–2.29). However, after adjustment for confounding variables, this association was no longer significant (HR =1.19, 95% CI: 0.82–1.74). After the development of a severe irAE, survival decreased more rapidly in the non-sarcopenic group. The multivariable model showed a significantly lower risk of death after developing severe irAEs in the non-sarcopenic group compared to the sarcopenic group (HR =0.39, 95% CI: 0.23–0.65) (Table 3, Figure S2).

Any-grade irAEs
Regarding the analysis on any-grade irAEs, both the sarcopenic and non-sarcopenic groups showed a quick decline in the probability of remaining event-free after the start of ICI treatment. In the multivariable analysis, we found a significantly lower risk of developing irAEs in the non-sarcopenic group compared with the sarcopenic group (HR =0.63, 95% CI: 0.42–0.96). The probability of death (without irAEs) increased steadily over time and was similar in both the sarcopenic and non-sarcopenic groups, with no significant difference observed in the multivariable model (HR =1.08, 95% CI: 0.70–1.67). Similarly, the probability of death after the development or any-grade irAEs increased constantly over time and was similar in both sarcopenic and non-sarcopenic groups, with no significant difference observed in the multivariable model (HR =0.81, 95% CI: 0.52–1.26) (Table 3, Figure S3).

Sensitivity analysis: SMI as continuous variable
When modeled as a continuous variable, SMI was not significantly associated with any transition in the multistate model. For the transition from no event to irAE, the adjusted HR per unit increase in SMI was 1.01 (95% CI: 0.98–1.03). For the transition from no event to death without prior irAE, the HR was 0.99 (95% CI: 0.96–1.01). For the transition from irAE to death, the HR was 0.99 (95% CI: 0.96–1.02) (Table S2).

Sensitivity analysis: time-dependent Cox model with interaction term
In our sensitivity analysis based on Cox regression with irAEs modeled as a time-varying covariate, the model that considered only severe irAEs showed that experiencing a severe irAE was associated with a significantly higher risk of death (HR =3.42, 95% CI: 1.84–6.36). The interaction term between irAE and sarcopenia was also statistically significant (HR =0.28, 95% CI: 0.14–0.57), indicating a lower risk of death among patients with baseline CT-assessed sarcopenia compared to their non-sarcopenic counterparts after experiencing a severe irAE.
In contrast, in the model analyzing any-grade irAEs, the main effect of irAE was not significantly associated with OS (HR =1.08, 95% CI: 0.60–1.92). Similarly, the interaction term between irAE and sarcopenia was not statistically significant (HR =0.64, 95% CI: 0.34–1.19), suggesting no strong evidence that the effect of experiencing any-grade irAEs on survival differed between sarcopenic and non-sarcopenic patients.
CT-assessed sarcopenia was not associated with OS in both the severe-irAEs (HR =1.13, 95% CI: 0.71–1.78), or in the any-grade irAEs models (HR =1.25, 95% CI: 0.84–1.86) (Table S3).

Discussion

Key findings
This study provides novel insights into the associations between CT-assessed sarcopenia, irAEs, and OS in NSCLC patients treated with ICIs using multistate survival analysis. By modeling the dynamic transitions between treatment initiation, irAEs, and death, we captured patient trajectories after start of ICI treatment. In our analysis, CT-assessed sarcopenia was not significantly associated with the development of severe irAEs. However, sarcopenic patients had a lower risk of death after the occurrence of severe irAEs. Furthermore, sarcopenic patients had a significantly lower likelihood of developing any-grade irAE compared to non-sarcopenic patients. CT-assessed sarcopenia was not significantly associated with mortality (without any-grade irAEs) or with mortality after the development of any-grade irAE. Our sensitivity analysis supported these findings. Patients who experienced severe irAEs had a higher overall risk of death, but among those with sarcopenia, the occurrence of a severe irAE was associated with a lower risk of death. In contrast, when considering any-grade irAEs, neither the presence of an irAE nor the interaction between sarcopenia and irAEs was significantly associated with survival.

Strengths and limitations
This study has several strengths. The use of a multistate model allowed for a detailed analysis of dynamic transitions, while adjusting for key confounders such as ECOG performance status, BMI, and treatment line. This approach provided valuable insights by simultaneously evaluating multiple outcomes, accounting for competing risks, and illustrating the impact of CT-assessed sarcopenia on transitions to censoring or death following the development of any or severe irAEs. Furthermore, we performed a sensitivity analysis using a more traditional statistical approach to expand and increase the robustness of our findings.
However, limitations should be acknowledged. First, the retrospective design restricted access to certain variables relevant to this analysis, such as PD-L1 expression and smoking status. Second, as a tertiary-level hospital, UMCG primarily manages patients with complex clinical conditions, reflected on a higher observed incidence of CT-assessed sarcopenia compared to other studies (18-20) which may explain the lack of association between sarcopenia and OS, as previously discussed above. This may result in the overestimation or underestimation of the effect of CT- assessed sarcopenia on survival and irAEs, potentially limiting the generalizability of our findings. Additionally, our definition of sarcopenia relied on CT-assessed SMM, while a robust measure, does not reflect the functional aspects of muscle health. The absence of data on fat infiltration within skeletal muscle, a metric that could offer deeper insights, combined with the lack of functional muscle measurements, such as handgrip strength or the Sit-to-Stand Test, and the absence of a clinical diagnosis of sarcopenia, further limits our analysis to the interpretation of SMM alone, without accounting for muscle quality. Moreover, our study did not incorporate information on the clinical management of irAEs, such as corticosteroid treatment, duration of immunosuppression, or decisions regarding ICI rechallenge, because these variables were not routinely captured in our real-world dataset. Differences in post-irAE management may influence outcomes after an irAE and could therefore affect the transition from irAE to death in the multistate framework. As such, our estimates reflect the natural course of irAEs under routine clinical care at our center but do not account for heterogeneity in downstream management. This represents a limitation and may introduce some degree of residual confounding. Furthermore, because patients with ECOG ≥2 are rarely treated with ICIs in Dutch practice, our cohort likely underrepresents frailer sarcopenic individuals. As a result, any adverse association between sarcopenia and outcomes may be underestimated, despite the adjustment for ECOG PS. Finally, several potential confounders could not be fully accounted for. Although treatment type (monotherapy vs. combination therapy) was included in all models, baseline steroid or proton pump inhibitor (PPI) use, autoimmune disease, diabetes, and detailed metastatic patterns were not systematically recorded and therefore could not be reliably adjusted for. These omissions may bias estimates in either direction—for example, steroids may suppress irAEs, PPIs may increase them, and conditions such as autoimmune disease, diabetes, or liver metastases could influence both irAE risk and mortality, potentially distorting the sarcopenia-outcome relationship. While ICI treatment in patients with autoimmune diseases and chronical use of steroids is relatively uncommon in Dutch practice, their incomplete documentation introduces residual confounding that should be considered when interpreting our findings.
Although real-world data introduce important limitations, they reflect the population actually treated in routine practice. This context helps ensure that, despite these constraints, the findings remain relevant to everyday clinical care.

Comparison with similar research and explanations of findings

Sarcopenia and reduced incidence of irAEs
Our findings align with prior research suggesting that body composition may reflect differences in underlying immunometabolic or inflammatory status during ICI therapy (4,5,16). The reduced incidence of irAEs in sarcopenic patients may stem from diminished immune activity or cytokine production associated with low SMM. Sarcopenia’s link to systemic inflammation and immune dysfunction might explain this outcome (21,22). However, this apparent protective role against irAEs contrasts with its association with poorer outcomes in other contexts, such as cito-toxic chemotherapy (23,24). Furthermore, previous studies have found higher incidence of irAEs in sarcopenic patients. Xue et al. reported a higher risk of irAEs among sarcopenic patients treated with anti-programmed cell death protein-1 (PD-1)/programmed death ligand-1 (PD-L1) therapy (25). However, their analysis used binary logistic regression, which does not account for event timing, competing risks, or early death—factors explicitly handled in our multistate framework and sources of potential immortal-time and misclassification bias when ignored. Xue et al. also applied Asian-specific SMI cut-offs in a younger, leaner, exclusively Asian population and combined all irAEs into a single binary endpoint without distinguishing severity or post-irAE mortality. These methodological and population differences likely contribute to the contrasting associations observed in our study. This reflects the complex interplay of immune activation in sarcopenic patients.

Sarcopenia and direct mortality (without irAEs)
Our multi-state model did not identify a significant association between baseline CT-assessed sarcopenia and direct mortality (without irAEs), likely due to the high incidence of sarcopenia in this tertiary-level hospital population compared to previous studies (18-20). This imbalance resulted in a small non-sarcopenic comparison group and wide CIs. A sensitivity analysis using a more traditional statistical method also showed no significant relationship between sarcopenia and OS. Although hazard ratios indicated a potential link between sarcopenia and direct mortality (without irAEs), statistical significance emerged only in the multi–state univariable severe irAEs model and diminished after adjustment for confounders. A larger sample size might clarify this association, consistent with prior research showing increased mortality risk in sarcopenic patients (26).

Sarcopenia and survival after severe irAEs
Our study shows a complex interplay between sarcopenia, severe irAEs, and survival in patients receiving immunotherapy. Sarcopenic patients who develop severe irAEs exhibit better survival than non-sarcopenic patients, who showed worse outcomes and increased mortality following severe irAE onset. Notably, this finding is supported by a large effect size (HR =0.45), highlighting the substantial survival differences associated with severe irAEs in sarcopenic patients. We hypothesize that, due to differing immune responses, non-sarcopenic individuals, with more robust immunity, can experience hyperactivation leading to excessive CD4+ effector memory T-cell activity, implicated in severe irAEs and increased mortality (27,28), this being reflected in a higher incidence and an earlier onset of severe irAEs in the non-sarcopenic group, compared to the sarcopenic group. Conversely, sarcopenic patients, often characterized by chronic inflammation and immunosenescence (28,29), may experience a delayed yet effective immune response in the form of severe irAEs, enhancing anti-tumor efficacy rather than causing lethal toxicity. These findings align with earlier evidence that severe irAEs can be both beneficial and harmful (30); a process that can depend on patient-specific immune and inflammatory status. The higher mortality in non-sarcopenic patients underscores how excessive immune responses can disrupt the balance between tumor control and immune-mediated toxicity.

SMI as a continuous variable
The discrepancy between the significant effect of dichotomized sarcopenia and the non-significant associations observed when using continuous SMI likely reflects a nonlinear, threshold-dependent relationship between muscle mass and clinical outcomes. Prior work has shown that muscle depletion predominantly affects prognosis at the extreme lower end of the SMI distribution, whereas moderate differences within the normal or mildly reduced range have limited impact. This pattern means that risk increases sharply only once SMI falls below a clinically meaningful threshold. In our cohort, this nonlinear behavior likely diluted the per-unit effect estimated by the continuous analysis, resulting in non-significant HRs despite the clear prognostic signal captured by the validated sarcopenia cut-off. These findings emphasize that the prognostic value of muscle mass is not linear across its range but concentrated at the edges, reinforcing the clinical relevance of threshold-based sarcopenia definitions.

Implications and actions needed
These findings emphasize the importance of personalized management strategies in immunotherapy-treated patients. While our observational design does not allow causal inference, sarcopenic individuals may benefit from interventions to enhance immune function and muscle mass, while non-sarcopenic patients might require approaches that control immune overactivation and reduce toxicity. Baseline SMM stratification, readily assessed through routine LDCT scans, offers a cost-effective means to identify patients who need closer monitoring for severe irAEs or tailored therapeutic approaches. Moreover, the detrimental impact of severe irAEs, particularly among non-sarcopenic patients, highlights the need for proactive monitoring and timely intervention to improve survival outcomes. Preserving muscle mass and function through exercise and nutritional support could further optimize treatment responses. Future trials should investigate these strategies and evaluate the integration of SMM assessment into clinical workflows.
Furthermore, the use of multi-state models in the analysis of time-to-event data provides an added insight by capturing the dynamic and sequential nature of irAEs and mortality. These models allow for the examination of intermediate clinical events, such as irAE development, and their impact on subsequent transitions, offering a more nuanced understanding of patient trajectories. By accounting for different clinical states and transitions over time, multi-state models support more precise risk stratification and inform targeted interventions at each stage of the treatment course.
Further research should validate these findings in larger and more diverse cohorts, particularly those with favorable clinical characteristics and varied healthcare settings. Randomized controlled trials are needed to evaluate the impact of improving SMM on irAE development and survival, offering unbiased estimates. Real-world studies could further elucidate the mechanisms linking sarcopenia, irAEs, and survival. Prospective studies should explore whether targeting sarcopenia enhances ICI therapy outcomes, while the role of other body composition metrics, such as visceral fat and sarcopenic obesity, in modulating immune responses also requires further investigation. Moreover, inflammatory and nutritional biomarkers such as CRP and albumin were not available in this cohort. Because these markers largely represent downstream processes through which sarcopenia may influence irAE risk and outcomes, adjusting for them would likely attenuate or distort the association rather than clarify it. Future studies with systematic biomarker collection are needed to jointly evaluate sarcopenia with CRP, albumin, and related markers to better characterize underlying biological profiles.

Conclusions
CT-assessed sarcopenia significantly influences the trajectory of NSCLC patients treated with ICIs, with distinct patterns in irAE development and survival. Compared to non-sarcopenic patients, sarcopenic ones who develop severe irAEs experience significantly lower mortality. Additionally, they exhibit a lower incidence of any-grade irAEs. Sarcopenia is not associated with mortality without irAEs in our population. These findings emphasize the need to incorporate body composition analysis into routine oncology practice and the close monitoring of (severe) adverse events to optimize treatment outcomes and enhance personalized care strategies. Accordingly, we emphasize that these findings are associative, not causal, and further investigation, particularly with external validation and mechanistic analyses, is needed.

Supplementary
The article’s supplementary files as