Abstract
Background:
Early identification of rapid progression in patients with preserved ratio but chronic cough remains challenging.
Objective:
To develop and validate an integrated multimodal prediction model that combines clinical phenotypes, small airway function indices, and chest CT quantitative features for stratifying progression risk in individuals with preserved ratio, chronic cough, and early COPD features.
Design:
Retrospective cohort study with external validation.
Methods:
In this retrospective cohort study of 408 participants with chronic cough (⩾3 months) and preserved spirometry (FEV1/FVC ⩾ 0.7), we integrated clinical profiles, pulmonary function (including small airway indices), and AI-quantified chest CT quantitative features. A three-stage modeling approach was used: variable screening with generalized estimating equations, predictor selection via LASSO regression, and construction of a Random Survival Forest–Cox hybrid model to predict a composite outcome of accelerated lung function decline (annual FEV1 decline >40 mL) or acute exacerbation. External validation was performed using an independent cohort (COPDGene, n = 312, 82 outcome events), with 3-year follow-up for prediction. The model underwent internal validation (10-fold cross-validation), with interpretability assessed via SHapley Additive exPlanations (SHAP).
Results:
Multivariable analysis confirmed small airway dysfunction and mucus plug burden as independent predictors of accelerated FEV1 decline (all p < 0.001), with a significant synergistic interaction (β = 2.34, p = 0.003). The final nine-feature model showed excellent discrimination, with a time-dependent AUC of 0.872 (95% CI: 0.843–0.901) on internal validation and 0.843 (95% CI: 0.812–0.874) on external validation. SHAP analysis identified small airway parameters and mucus plug burden as the top contributors, jointly accounting for 68.4% of predictive output. A nomogram and online calculator were developed for clinical use.
Conclusion:
We developed and validated a robust multimodal prediction model that accurately stratifies progression risk in individuals with preserved ratio, chronic cough, and early COPD features. This tool facilitates personalized risk assessment and holds promise for guiding precision management strategies to improve outcomes in this high-risk population.
Plain language summary
Chronic Obstructive Pulmonary Disease (COPD) is a serious lung condition. Catching it early is key to slowing it down. However, many patients have a persistent cough but their standard breathing tests appear nearly normal. Doctors currently lack a reliable way to identify which of these patients will get worse quickly.
We studied 408 patients with chronic cough (lasting 3+ months) whose initial breathing tests were borderline. We combined three types of information: clinical profile, detailed breathing tests (especially for small airways), and chest CT analysis (such as mucus plugs in airways). Using this data, we built a model to predict which patients would experience rapid lung function decline or serious flare ups.
• The final model uses just 9 pieces of information easily collected in clinics.
• It accurately identified high risk patients, and accuracy remained strong when tested on a separate patient group.
• Two factors mattered most: small airway function and mucus plugs on CT scans, together accounting for 68.4% of the model's predictive power.
• When both factors are present, risk increases more than the sum of each alone — a synergistic effect.
Keywords
Introduction
Chronic obstructive pulmonary disease (COPD) continues to pose a significant global health burden, characterized by high rates of disability and mortality. 1 While the clinical diagnosis of COPD is traditionally based on a fixed spirometric threshold (post-bronchodilator FEV1/FVC < 0.7), there is growing recognition of a prolonged “pre-COPD” phase that precedes the development of persistent airflow limitation.2–4 This phase is marked by early pathological changes, such as small airway disease and mucus hypersecretion, which often manifest clinically through symptoms such as chronic cough and represent a critical window for early intervention.5–7 Individuals with lung function in the “gray zone” (post-bronchodilator FEV1/FVC ⩾ 0.7) who exhibit such clinical features are considered to be at the highest risk of progressing to overt COPD.8,9 This high-risk profile often aligns with the preserved ratio impaired spirometry (PRISm) phenotype, defined by a reduced FEV1 in the context of a preserved FEV1/FVC ratio, which is increasingly recognized as a transitional state associated with increased symptoms, exacerbations, and mortality.10,11 However, in our cohort, the mean FEV1% predicted was 89.7% ± 9.8%, meaning many participants did not meet the strict PRISm definition. Therefore, we hereafter refer to the study population as “individuals with preserved ratio, chronic cough, and early COPD features” rather than “PRISm patients.” Despite its clinical significance, a significant gap remains in current practice: there are no effective tools that integrate multidimensional data to accurately identify individuals within this heterogeneous PRISm group who are likely to experience rapid disease progression. 12 The dynamic and heterogeneous nature of the PRISm phenotype further complicates accurate risk stratification. 13
To address this gap, we conducted a retrospective cohort study to develop and validate an individualized risk prediction model. Our approach leveraged a multimodal data fusion strategy, systematically quantifying the predictive value of chronic cough, small airway dysfunction, and CT quantitative features—with a particular focus on mucus plug burden—for disease progression. This research seeks to bridge the gap between evolving pathophysiological understanding and clinical application, providing both evidence-based support and a practical tool to facilitate a paradigm shift in COPD management toward earlier, prevention-focused, and precision medicine-based care.
Materials and methods
Study design and participant selection
This single-center retrospective cohort study was approved by the Institutional Ethics Committee (Approval No: JHYK-EC2024-090). Informed consent was waived because of the use of anonymized retrospective data. We consecutively screened patients who visited our center between January 2019 and December 2020. Inclusion criteria were: (1) baseline chronic cough (duration ⩾ 3 months); (2) post-bronchodilator FEV1/FVC ⩾ 0.70; and (3) availability of at least one follow-up spirometry record obtained ⩾12 months after baseline. Exclusion criteria included a prior diagnosis of chronic respiratory diseases (e.g., COPD, asthma, bronchiectasis, or interstitial lung disease) or other explained causes of chronic cough (e.g., gastroesophageal reflux disease, upper airway cough syndrome, or use of angiotensin-converting-enzyme inhibitors). A total of 408 participants met the criteria and constituted the primary analysis cohort. The inclusion and exclusion criteria are summarized in Table 1.
Inclusion and exclusion criteria.
ACEI, angiotensin-converting enzyme inhibitor; COPD, chronic obstructive pulmonary disease; FEV1, forced expiratory volume in 1 second; FEV1/FVC, ratio of forced expiratory volume in 1 second to forced vital capacity; FVC, forced vital capacity; GERD, gastroesophageal reflux disease; ILD, interstitial lung disease; UACS, upper airway cough syndrome.
We employed a dynamic enrollment approach: the first visit meeting the inclusion criteria was defined as the baseline (index date), and follow-up started from that point. Follow-up continued until the last available spirometry date, occurrence of an endpoint event, or the study cutoff date (December 31, 2024), whichever came first. To ensure reliable estimation of annual lung function decline, each participant was required to have at least two spirometric measurements (including baseline).
A matched control group of 200 asymptomatic healthy individuals (matched for age, sex, and smoking history) was recruited during the same period. All controls had normal baseline lung function, defined as FEV1/FVC > 0.75 and FEV1% predicted > 80%.
Data collection and definitions
Baseline demographic characteristics, smoking history (pack-years), and body mass index (BMI) were extracted from the hospital’s electronic health record system. Key measurements and instruments are summarized in Table 2.
Key measurements and instruments.
FEF25–75%, forced expiratory flow at 25–75% of forced vital capacity; FEF50%, forced expiratory flow at 50% of forced vital capacity; FEV1, forced expiratory volume in 1 second; FEV1/FVC, ratio of forced expiratory volume in 1 second to forced vital capacity; FVC, forced vital capacity; LAA856%, low-attenuation area percentage below −856 Hounsfield units; MMEF, maximal mid-expiratory flow.
Clinical and pulmonary function data
Spirometry was performed according to American Thoracic Society/European Respiratory Society (ATS/ERS) guidelines 14 using a MasterScreen™ pneumotachograph (Vyaire, Hochberg, Germany). Recorded parameters included FEV1, FVC, FEV1/FVC ratio, and small airway indices: maximal mid-expiratory flow (MMEF, also known as forced expiratory flow between 25% and 75% of FVC, i.e., FEF25–75%), forced expiratory flow at 50% of FVC (FEF50%), and forced expiratory flow at 25–75% of FVC (FEF25–75%). All values are expressed as percentages of predicted values. For clarity, we use MMEF and FEF25–75% interchangeably in this manuscript, as they represent the same physiological parameter. 14
CT image acquisition and radiomic analysis
All participants underwent non-contrast chest CT on a 128-slice GE scanner (GE Healthcare, USA) with standardized parameters: tube voltage 120 kV, automatic tube-current modulation, slice thickness 0.625 mm, and reconstruction thickness 1.0 mm. Quantitative CT analysis was performed using an in-house deep-learning model based on a U-Net architecture (implemented in Python 3.8 with TensorFlow 2.4). The model achieved a Dice similarity coefficient of 0.85 for mucus plug segmentation on an independent test set and automatically quantified the following imaging biomarkers:
Because the features extracted were limited to these two predefined quantitative metrics rather than high-dimensional radiomic features (e.g., texture, shape, and wavelet), we use the term “quantitative CT features” throughout this manuscript rather than “radiomic features.”
Outcome measures
Statistical analysis
All statistical analyses were performed using R software (version 4.2). The reporting of this study conforms to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement. 15 The statistical analysis approach was adapted from standard methods for high-dimensional prediction modeling in respiratory research, as described by Steyerberg et al. (Clinical Prediction Models, 2019) and implemented in the “glmnet” and “randomForestSRC” R packages. 16 The analytical strategy comprised three main stages: preliminary screening of associations, predictor selection, and predictive model building.
First, to preliminarily explore these associations, we employed generalized estimating equations (GEE) to assess the relationships between small airway function parameters and mucus plug volume with the annual rate of FEV1 decline, adjusting for age, sex, smoking history, and BMI. An interaction term was included to test for a synergistic effect. The GEE model used an exchangeable correlation structure and robust standard errors.
Subsequently, for variable selection in building the prediction model, all baseline candidate variables were included in a Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis (using the glmnet package in R) to identify predictors associated with the composite endpoint (defined as accelerated lung function decline [annual decline >40 mL/year] or occurrence of a COPD-related acute exacerbation). The optimal lambda (λ) was selected using the minimum cross-validated deviance.
Finally, based on the predictors selected by LASSO, a hybrid Random Survival Forest–Cox proportional hazards model (RSF–Cox) was constructed to predict an individual's risk of experiencing the aforementioned composite endpoint event. The RSF component used 500 trees with default node splitting, and the Cox component provided the linear predictor for the final hybrid model.
Model performance was evaluated through internal validation (10-fold cross-validation) and external validation using the independent COPDGene cohort (n = 312, 82 outcome events, 3-year follow-up). Discrimination was assessed using time-dependent AUC, and calibration was evaluated using calibration plots. Interpretability was enhanced using SHapley Additive exPlanations (SHAP) values. A clinical nomogram and an online risk assessment calculator were developed based on the coefficients from the Cox component of the hybrid model.
No formal sample size power calculation was performed before the study, as this was an exploratory retrospective analysis.
Results
Study population characteristics
A total of 608 participants were included in the final analysis, comprising 408 individuals with preserved ratio and chronic cough (67.9% male; mean age 65.2 ± 5.8 years) and 200 well-matched asymptomatic controls. The two groups were comparable in basic demographic characteristics including age, sex distribution, smoking history, and BMI (p > 0.05 for all), confirming successful matching. As expected, significant impairments in physiological and structural parameters were observed in the study group compared to controls. Specifically, they had lower FEV1% predicted values, reduced small airway function (MMEF, FEF50%, FEF25–75%), and higher mucus plug burden and air trapping indices (p < 0.001 for all). Although our cohort had a mean FEV1% predicted of 89.7%, which is higher than the conventional PRISm threshold (<80%), approximately 23% (94/408) of participants met the strict PRISm definition (FEV1% predicted < 80%). Sensitivity analyses (not shown) excluding these 94 participants yielded similar model performance (C-index 0.865), suggesting the model’s applicability across a broader spectrum of early disease. Comprehensive baseline characteristics are presented in Table 3.
Baseline characteristics of the study cohort and control group.
Data are presented as mean ± standard deviation or n (%), unless otherwise specified. Normality was checked using the Shapiro-Wilk test; all continuous variables met the normality assumption. Continuous variables were compared using independent t-tests (unequal variances adjusted with Welch’s correction where appropriate), and categorical variables were compared using the chi-square test. All p values are unadjusted.
BMI, body mass index; FEF25–75%, forced expiratory flow at 25%–75% of forced vital capacity; FEF50%, forced expiratory flow at 50% of forced vital capacity; FEV1, forced expiratory volume in 1 second; FEV1/FVC, ratio of forced expiratory volume in 1 second to forced vital capacity; FVC, forced vital capacity; LAA856%, low-attenuation area percentage below −856 Hounsfield units; LAA950%, low-attenuation area percentage below −950 Hounsfield units; MMEF, maximal mid-expiratory flow.
Association between predictors and FEV1 decline
GEE were used to evaluate the relationships between small airway dysfunction, mucus plug burden, and the annual rate of FEV1 decline. In univariable analysis, all small airway function parameters and mucus plug volume showed significant associations with accelerated FEV1 decline (p < 0.001). After adjusting for age, sex, smoking history, and BMI in the multivariable GEE model, both small airway dysfunction and mucus plug burden remained independent risk factors for faster annual FEV1 decline (Table 4). Critically, a statistically significant synergistic interaction was observed between MMEF %predicted and mucus plug volume (interaction term β = 2.34, 95% CI 0.78–3.90, p = 0.003), indicating that their combined presence confers a greater risk than the sum of their individual effects. The correlation patterns between these key predictors and FEV1 decline are shown in Figure 1.
Association between small airway function, mucus plug burden, and annual FEV1 decline (GEE model).
CI, confidence interval; FEF25–75%, forced expiratory flow at 25%–75% of forced vital capacity; FEF50%, forced expiratory flow at 50% of forced vital capacity; FEV1, forced expiratory volume in 1 second; GEE, generalized estimating equations; MMEF, maximal mid-expiratory flow.
Correlation between key predictors and annual FEV1 decline rate. (a) MMEF %predicted showed a significant negative correlation with FEV1 decline rate (r = −0.52, p < 0.001). (b) Mucus plug volume demonstrated a positive correlation with FEV1 decline rate (r = 0.46, p < 0.001). The shaded areas represent 95% confidence intervals.
Prediction model development and validation
LASSO regression with 10-fold cross-validation selected nine predictors for the final model: MMEF %predicted, FEF50% %predicted, FEF25–75% %predicted, mucus plug volume, age, smoking history (pack-years), BMI, air trapping index (LAA856%), and long-acting bronchodilator use. Note that FEF25–75% was included as the ninth predictor; it is highly correlated with MMEF but provided additional predictive information in the LASSO selection.
The RSF–Cox hybrid model showed excellent predictive performance. During internal validation, it achieved a time-dependent AUC of 0.872 (95% CI: 0.843–0.901) for predicting the composite outcome of accelerated lung function decline or acute exacerbation. Performance remained strong in the external validation cohort, with a time-dependent AUC of 0.843 (95% CI: 0.812–0.874) (Figure 2(a)).
Model performance and calibration. (a) Time-dependent ROC curves demonstrating discriminatory power in derivation (Time-dependent AUC: 0.872) and external validation (Time-dependent AUC: 0.843) cohorts at 3-year follow-up. (b) Calibration plot for external validation cohort showing excellent agreement between predicted and observed risk (intercept = 0.08, slope = 0.94).
Calibration analysis indicated good agreement between predicted and observed outcomes in both cohorts (Figure 2(b)). In the external validation set, the calibration intercept was 0.08 and slope 0.94, demonstrating minimal overfitting and good generalizability.
Subgroup analysis by smoking status: In participants who continued smoking during follow-up (n = 221, 54.2%), the model achieved a C-index of 0.861 (95% CI: 0.822–0.900); in those who quit smoking (n = 187, 45.8%), the C-index was 0.848 (95% CI: 0.801–0.895), indicating similar predictive performance irrespective of smoking cessation.
Use of inhaled corticosteroids (ICS): Data on ICS use at baseline were available for 389 participants (95.3%). Overall, 102 (26.2%) were treated with ICS (with or without long-acting bronchodilators). The model’s performance did not differ significantly between ICS users and non-users (C-index 0.865 vs 0.870, p = 0.54).
Model interpretation and feature importance
SHapley Additive exPlanations (SHAP) analysis elucidated the model's decision-making process. As illustrated in Figure 3(a), small airway function parameters (collectively contributing 42.1%) and mucus plug volume (26.3%) emerged as the most influential predictors, together accounting for 68.4% of the model’s predictive power. Age (12.1%) and smoking history (9.5%) constituted the next most important features. The SHAP plot shows the top 8 predictors (the remaining predictors beyond the top 8 had negligible contributions)
SHAP analysis for model interpretation. (a) Feature importance plot showing the mean absolute SHAP values for the top 8 predictors. Small airway parameters and mucus plug volume collectively accounted for 68.4% of the model's predictive power. (b) SHAP summary plot displaying the impact and value distribution of each feature on model output.
The SHAP summary plot (Figure 3(b)) further delineates the directional impact of each variable: higher mucus plug volumes (represented in red) consistently elevate predicted risk, whereas better small airway function (indicated by higher MMEF values, shown in blue) lowers the risk score. This pattern reinforces the clinical coherence of the model and enhances its interpretability for potential translation into practice.
Discussion
This study successfully developed and validated a multimodal model that integrates clinical phenotype, small airway function indices, and chest CT quantitative features to predict the risk of disease progression in individuals with preserved ratio, chronic cough, and early COPD features. The final model demonstrated excellent discriminative ability in both internal and external validation (time-dependent AUCs of 0.872 and 0.843, respectively). Its primary value lies in transcending the limitations of conventional single-dimensional indicators (e.g., relying solely on FEV1% predicted or symptoms alone) by quantifying and integrating early functional and structural pathological alterations, thereby providing a more precise and robust tool for identifying high-risk individuals. This approach addresses the need highlighted in recent literature for more sophisticated phenotyping and dynamic risk assessment in PRISm and other pre-COPD states, which are often heterogeneous and transient. 17
Model innovation and pathophysiological insights
A key innovation of this study is the identification of a significant synergistic effect between small airway dysfunction and mucus plug burden in driving early COPD progression (interaction term β = 2.34, p = 0.003). This finding supports a theoretical framework of a self-perpetuating cycle involving “symptom-function-structure” interactions: chronic cough serves as the clinical phenotype, indicating persistent airway inflammation and mucus hypersecretion; decreased small airway function indices (e.g., MMEF% predicted) reflect functional airflow limitation; and AI-quantified mucus plug volume directly represents physical airway obstruction. Our multivariable GEE analysis confirmed both as independent risk factors for accelerated FEV1 decline (all p < 0.001), and their coexistence produces an effect greater than the sum of their individual impacts.
This observation aligns with recent research trends. A recent systematic review estimated the global prevalence of PRISm at 12% and confirmed its association with adverse outcomes, underscoring the importance of identifying progression drivers within this group. 17 Our study provides more direct and objective quantitative CT evidence for this theory by precisely quantifying mucus plugs—a structural alteration recently highlighted as a potential therapeutic target in COPD. 18
Pathophysiologically, this synergy can be explained through interconnected mechanisms: mucus plugs cause physical obstruction, directly increasing small airway resistance and worsening functional impairment. Conversely, impaired mucociliary clearance resulting from small airway dysfunction promotes the formation and persistence of mucus plugs. 19 Together, they create a localized microenvironment that perpetuates inflammation and tissue remodeling, forming a self-reinforcing cycle that accelerates disease progression. This implies that interventions for early COPD should consider combined strategies targeting both small airway function and mucus plug burden, which may yield synergistic therapeutic effects.
Clinical translation and application prospects
The developed model, along with the resulting nomogram and online risk calculator, holds clear potential for clinical translation.
First, the model incorporates predictors (e.g., age, smoking history, pulmonary function indices, and routine CT parameters) that are all derived from standard clinical data, requiring no complex biomarker assays. This ensures high generalizability across healthcare settings. Clinicians can use this tool to rapidly identify high-risk individuals (e.g., those with MMEF% predicted < 40% and mucus plug volume > 0.2 mL) for enrollment into more intensive monitoring programs (e.g., spirometry every 6 months), thereby optimizing resource allocation. This is particularly relevant given the known instability of the PRISm phenotype, where repeated measurements are crucial for accurate classification and prognosis.20,21
Second, SHAP analysis clearly identified small airway function parameters and mucus plug burden as the most important predictors, collectively contributing 68.4% to the model’s output. This finding provides clear targets for precision interventions. Based on these results, we hypothesize that for identified high-risk individuals, early intervention with extra-fine particle long-acting bronchodilators (for better small airway deposition) combined with mucolytic agents (to reduce mucus plug burden) might attenuate the accelerated FEV1 decline associated with these factors. This hypothesis should be tested in future randomized trials.
Recent evidence supports targeting the mucus phenotype. Beyond clinical trials, real-world data also indicate that treatment with mucoactive drugs is associated with reduced COPD exacerbations over extended follow-up. 22 This reinforces the potential of mucoregulatory strategies in disease management. Building on this, our multimodal prediction model—integrating small airway dysfunction and CT quantitative features—can accurately identify high-risk patient subgroups in early or pre-disease stages. 23 Identifying such individuals aligns with the “treatable traits” paradigm, increasingly advocated for application in pre-COPD to enable early intervention. 23 Therefore, our model provides a crucial stratification tool and rationale for designing future precision intervention trials targeting similar phenotypes in incipient or early COPD populations.
Finally, integrating this risk assessment model into electronic medical record systems could enable automated risk alerts, facilitating a shift from reactive treatment to proactive prevention in chronic disease management. This aligns not only with precision medicine principles but may also yield significant health economic benefits by reducing future costs associated with acute exacerbations through early intervention. 24
Study limitations and future directions
This study has several limitations. First, the retrospective, single-center design may introduce selection bias. Although we enhanced generalizability through external validation using an independent cohort (COPDGene), future prospective, multicenter studies are warranted to further validate the model’s performance and evaluate its real-world impact on clinical decision-making and patient outcomes. 25 Second, we did not perform a formal sample size power calculation before the study; the sample size was determined by available eligible patients. This is a limitation, and future prospective studies should include power calculations. Third, the current model does not include inflammatory biomarkers such as blood eosinophil counts. These indicators are gaining increasing attention for guiding COPD inflammatory endotyping and treatment selection 26 ; incorporating them in future iterations could potentially enhance predictive accuracy and phenotyping value.
Fourth, the quantification of mucus plugs relies on a specific AI algorithm, which is not yet widely available. To improve generalizability, we have made the core segmentation code available upon request, and we encourage external validation using alternative AI tools. Additionally, our study focused on functional and imaging markers. Incorporating static lung volume measurements (e.g., Total Lung Capacity), which have been shown to provide complementary physiological information and help differentiate underlying pathophysiology in PRISm, could further refine the model. 27 Fifth, long-acting bronchodilator use was included as a baseline predictor; however, we did not have detailed information on whether this treatment was initiated before or after baseline. To address potential bias, we performed a sensitivity analysis excluding participants with bronchodilator use (n=54, 13.2%), and the model performance remained similar (C-index 0.861). We have clarified in the Methods that this variable reflects baseline medication status.
The most critical next step involves conducting intervention studies based on model-derived risk stratification. Categorizing participants into different risk strata according to our model and applying management or treatment strategies of varying intensity would directly test the ultimate value of this model in guiding clinical practice and improving long-term patient outcomes. 28 This work holds significant importance for advancing COPD secondary prevention into the era of precision medicine.29,30
Conclusion
In conclusion, this study successfully developed and validated an efficient and interpretable multimodal prediction model. By integrating clinical, functional, and imaging data, the model accurately identifies individuals with preserved ratio, chronic cough, and early COPD features who are at high risk for rapid disease progression. It not only provides a practical tool for early risk stratification and precision management of COPD but also offers novel pathophysiological insights into the interaction between functional and structural changes in early disease stages.
Supplemental Material
sj-docx-1-tar-10.1177_17534666261466906 – Supplemental material for Development and validation of a multimodal risk prediction model for early COPD progression: a retrospective study integrating clinical, functional, and CT radiomic features
Supplemental material, sj-docx-1-tar-10.1177_17534666261466906 for Development and validation of a multimodal risk prediction model for early COPD progression: a retrospective study integrating clinical, functional, and CT radiomic features by Jia Zhang, Qin Yang, Shuangme Dai, Jiaxin Li, Zhe Zhang, Zihan Jia, Weihua Zhu, Rui Yang, Xuelian Bai, Wenya Li, Xin Zhang, Xinjun Zhang, Jichao Chen and Wailong Zou in Therapeutic Advances in Respiratory Disease
Footnotes
Acknowledgements
The authors gratefully acknowledge all patients who participated in this study. We also thank the clinical and technical staff at our institution for their assistance in data collection and management. We are grateful to our colleagues for their valuable suggestions on statistical analysis and algorithm implementation. This study received no external funding. All authors contributed to the study design, data analysis, manuscript preparation, and final approval.
Declarations
Supplemental material
Supplemental material for this article is available online.
Artificial intelligence statement
No generative AI (e.g., ChatGPT) was used in the creation of this manuscript’s text, references, figures, or other content.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
