How to quickly determine whether patients with chronic cough need corticosteroid treatment——construction of a predictive model for corticosteroid-responsive cough in chronic cough Authors: Baiyi Yi, Haodong Bai, Shujie Li, Tongyangzi Zhang, Yiqing Zhu, Wenxiu Luo, Xianghuai Xu, Li Yu Published: 2026 ABSTRACT Patients with chronic cough need to undergo a wide range of tests and rely on empirical medication to determine the underlying cause. Corticosteroid-responsive cough (CRC) accounts for the majority of the causes of chronic cough, and the diagnostic process is relatively complicated. To develop a predictive model for diagnosis of CRC, which is based on readily available clinical information. A retrospective dataset of 304 cases was used for training, and a prospective dataset of 131 cases was used for temporal single-center validation. Candidate predictors were screened via univariate analysis and refined using LASSO regression. A final multivariable model was constructed through stepwise logistic regression and presented as a clinical nomogram. The model’s performance was rigorously evaluated in terms of discrimination (AUC-ROC), calibration (calibration curve), and clinical utility (decision curve analysis). The robustness of the model predictors was further supported by their consistency with the key variables identified by an optimal machine learning model. Of 20 candidate variables based on patient basic information and test results, 7 variables were selected as optimal predictors to establish a CRC prediction model for chronic cough, including reflux symptoms, history of allergic diseases, blood eosinophils, IgE, CRP, FEV1/FVC, and FeNO50. The model demonstrated good discrimination and calibration in both the training set (AUC = 0.900 [0.865,0.935], MAE = 0.017) and the validation set (AUC = 0.840 [0.771,0.909], MAE = 0.046). The results of decision curve analysis revealed that using our models to predict CRC in chronic cough would add more benefit than either the treat-all scheme or the treat-none scheme. The variable selection for the nomogram was further validated by the high degree of agreement with the feature importance rankings derived from the top-performing machine learning model (Random Forest). Our study has developed a nomogram model, along with a simplified version, based on easily accessible clinical information and pulmonary function tests. The model demonstrates strong predictive performance and good calibration, providing clinicians with a rapid aid for initiating targeted treatment. Introduction Chronic cough is typically defined as a cough lasting more than 8 weeks, which is without abnormalities on chest radiographs1–3. The common etiologies of chronic cough include Cough Variant Asthma (CVA), Upper Airway Cough Syndrome (UACS), Eosinophilic Bronchitis (EB), Gastroesophageal Reflux-related Cough (GERC), and Atopic Cough (AC)2–6. Among them, CVA, EB and AC respond well to corticosteroid therapy7–9. Therefore, they are classified as Corticosteroid-Responsive Cough (CRC), and the remaining categories are classified as Non-Corticosteroid-Responsive Cough (NCRC)7,8. Studies have shown that CRC accounts for approximately 55.2–66% of chronic cough, making it the most prevalent etiological category even after empirical therapy4,7,10–12. The etiological diagnosis of chronic cough is a prerequisite for effective treatment. The definitive diagnosis of CVA, EB, and AC depends on tests such as bronchial provocation, induced sputum cytology, and allergen testing13–16. However, clinical limitations include a lack of access to these tests, patient reluctance, and the fact that positive results alone are insufficient for diagnosis. Empirical treatment—whereby clinicians prescribe a course of medication based on the patient’s predominant symptom presentation, such as inhaled corticosteroids, antihistamines or proton pump inhibitors—works on some patients, while certain agents carry risks of side effects. Treatment without clear indications can lead to unnecessary harm. In addition, if CRC can be quickly excluded in clinical practice, it can prompt doctors to further explore the causes of NCRC, thereby accelerating etiologic diagnosis. Yi F et al. have explored whether using FeNO alone or with sputum eosinophilia and skin prick tests aids CRC and NCRC diagnosis17. However, specialized tests such as induced sputum cell counting are not yet widely available in many clinical settings, and obtaining samples of adequate quality remains challenging. Furthermore, the spectrum of variables analyzed in the study was relatively narrow, which may not fully capture the breadth of clinical information typically accessible in practice, thereby limiting the comprehensive applicability of the findings. Given the current state of chronic cough, exploring easily obtainable clinical information in newly-diagnosed patients is crucial to determine initial treatments. We aim to build a CRC clinical prediction model using patients’ basic data, symptoms, results of common ancillary tests (e.g., blood eosinophil count, inflammatory markers, immunoglobulin levels) and pulmonary function tests. This model will help clinicians quickly decide if corticosteroids are suitable for chronic cough patients. Methods Data source We conducted a retrospective analysis of clinical data from patients with chronic cough who visited the respiratory outpatient department of Tongji Hospital of Tongji University (Shanghai, China) from January 2022 to December 2024. Additionally, a prospective analysis was performed on clinical data from patients with chronic cough who attended the same department from January to June 2025. All candidate predictive variables were derived exclusively from the information contained within the electronic medical record system, encompassing: age, gender, BMI (Body Mass Index, calculated based on the patient’s height and weight), the duration of the cough symptoms, whether the cough is worse at night or during the day, the presence or absence of acid reflux or heartburn symptoms, the history of allergic diseases, blood eosinophils (hereinafter abbreviated as EOS), C-reactive protein (CRP), immunoglobulin E (IgE), and selected pulmonary function indicators. Data collection and the entry of etiological diagnosis were conducted separately by two researchers. After that, a third researcher checked the data and performed the statistical analysis. Participants Both the retrospective and prospective cohorts were required to conform to the following criteria. Inclusion Criteria: (1) aged over 18 years; (2) cough lasting no less than 8 weeks and with a normal chest radiograph; (3) a documented therapeutic response assessment (Cough Symptom Score reduction ≥ 50% from baseline) or confirmed etiological diagnosis, enabling classification of chronic cough. Exclusion Criteria: (1) Pregnant or lactating patients; (2) Patients with severe cardiac insufficiency, severe hepatic or renal insufficiency, mental or cognitive dysfunction, hearing or communication impairment, malignancy, or parasitic disease; (3) Patients with other respiratory conditions that may induce cough; (4) Cases with ≥ 10% of retrospective data missing or with inconsistent medical histories; (5) Participants lost to follow-up or who withdrew from the prospective cohort. The definition of chronic cough, the criteria for its etiological classification, and the definition of CRC were based on the Chinese National Guidelines on Diagnosis and Management of Cough (2021)1. CRC encompassed patients with a confirmed etiological diagnosis of CVA, EB, or AC, as well as patients without a definitive etiological diagnosis whose cough responded to empiric corticosteroid therapy. For patients in whom a specific etiology was confirmed according to the diagnostic criteria described later, the standard treatment duration was 8 weeks for CVA or EB and 4 weeks for AC. For patients who were unwilling or unable to undergo further etiological investigations, empiric corticosteroid therapy was administered for 8–12 weeks. If an initial response was achieved, the corticosteroid dose was tapered; if symptoms recurred during tapering, the treatment course was extended accordingly. Therapeutic response was assessed using the Cough Symptom Score (CSS). The CSS is a validated patient‑reported instrument that rates daytime cough and nighttime cough on a 0–3 scale each, yielding a total score ranging from 0 to 6. CSS was recorded at baseline and at the end of the treatment window. A ≥ 50% reduction in CSS from baseline was defined as a positive therapeutic response, and these patients were classified as CRC. Patients who did not meet this threshold were classified as NCRC. Patients without a documented therapeutic trial or with incomplete CSS assessments were excluded from the CRC/NCRC classification. CVA: (1) chronic cough, often accompanied by a significant night cough; (2) positive bronchial provocation test; (3) responsive to anti-asthma treatment. EB: (1) chronic cough, usually without wheezing or dyspnea; (2) ≥ 3% eosinophils in induced sputum differential cell count; (3) negative bronchial provocation test; (4) exclusion of other causes such as asthma, chronic obstructive pulmonary disease, or other known causes of eosinophilic airway inflammation. AC: (1) chronic cough predominantly dry; (2) positive allergen-specific IgE or skin prick test to common aeroallergens; (3) normal pulmonary function; (4) exclusion of other causes such as asthma, EB, GERC, and UACS. GERC: (1) chronic cough predominantly occurring during daytime; (2) esophageal reflux monitoring demonstrates an acid exposure time (AET) > 6% and a symptom-association probability (SAP) ≥ 95%; (3) markedly responsive to anti-reflux therapy. UACS: (1) cough occurring during the daytime or triggered by postural changes; (2) clinical manifestations indicative of nasal or pharyngeal disease, substantiated by corresponding ancillary investigations; (3) effective response to etiology-directed treatment of the underlying disorder. Patients diagnosed with CVA, EB, or AC, or for whom empiric ICS therapy proved effective, are all classified as CRC. Study design A total of 435 patients were included, of whom 304 constituted the training set (retrospective cohort) and 131 constituted a temporal single‑center validation set (prospective cohort). Candidate variables were initially screened by univariate logistic regression to exclude factors with negligible individual association. We then applied LASSO regression with 10‑fold cross‑validation to penalize model complexity and select a parsimonious subset of predictors, followed by multiple forward/backward stepwise logistic regressions to refine the variable set and identify the model with the lowest Akaike Information Criterion (AIC). To facilitate rapid clinical assessment at the point of care, the model was further simplified by dichotomizing continuous variables using cut-off values derived from established clinical thresholds or relevant published studies (Fig. 1).Fig. 1Research pathway diagram.*: Patients meeting any of the specific disease states outlined in exclusion criteria (1), (2), or (3); **: Patients who, upon re-evaluation of their initial clinical data, fail to strictly meet the diagnostic criteria for chronic cough according to Chinese national guideline on diagnosis and management of cough (2021); ***: Data missingness > 10% or lack of efficacy after medication. Research pathway diagram. *: Patients meeting any of the specific disease states outlined in exclusion criteria (1), (2), or (3); **: Patients who, upon re-evaluation of their initial clinical data, fail to strictly meet the diagnostic criteria for chronic cough according to Chinese national guideline on diagnosis and management of cough (2021); ***: Data missingness > 10% or lack of efficacy after medication. To enhance the robustness of the predictive model, we employed a multi-method validation strategy. In addition to the logistic regression nomogram, six machine learning models were constructed on the same training set: Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), K‑Nearest Neighbors (KNN), and Decision Tree (DT). All candidate predictors described in Section 2.1 were entered into each model without prior feature selection, thereby allowing the machine learning algorithms to independently assess variable importance. Hyperparameters for each model were tuned via Bayesian optimization with 10‑fold cross‑validation on the training set, using the Brier score as the minimization target; a complete description of the search spaces and optimized hyperparameters is provided in Supplementary Table S2. Models were compared using AUC‑ROC (discrimination) and the Brier score (calibration). The best‑performing model was subsequently used for variable importance ranking and SHAP analysis, serving as cross‑method validation of the logistic regression‑derived predictors. The model’s discrimination was evaluated using the area under the receiver operating characteristic curve (AUC‑ROC). Calibration was assessed visually with a calibration curve and statistically via bootstrapping. Clinical utility was quantified using decision curve analysis (DCA). This study has been reviewed and approved by the Shanghai Tongji Hospital Ethics Committee and has been registered at the Chinese Clinical Trial Registry (https://www.chictr.org.cn/). Registration number: ChiCTR2500106049. Statistical analysis All statistical analyses and data visualizations were performed using R (version 4.4.2). R packages, namely “mice”, “rms”, “scales”, “glmnet”, “pROC”, “ggplot2”, “regplot”, “tidymodels”, “xgboost”, and “shapviz”, were used in the analysis. Patients with >10% missing data across candidate variables were excluded prior to analysis. For the remaining patients, missing values were imputed using multiple imputation by chained equations via the “mice” package, under the assumption of missing at random. The imputation model was generated on the training cohort and subsequently applied to the validation cohort to maintain consistency. Data are presented as numbers and percentages for categorical variables, while continuous data are expressed as mean ± standard deviation. Additionally, for some data with skewed distributions, the median (25th-75th percentile) was used to express the values. Continuous variables were compared using the t-test, while categorical variables were compared using the chi-square (χ2) test or Fisher’s exact test. All statistical tests were two-sided, and statistical significance was set at 0.05 unless otherwise specified. Reproducibility details All continuous predictors were standardized to z-scores using the mean and standard deviation derived exclusively from the training set; these same training-set statistics were then used to standardize the validation set, and no categorization was performed. Multi-category variables (Diurnal Difference, Reflux Symptoms, and History of allergic diseases) were each recoded into two binary indicator variables that separately capture the certainty and the direction of the clinical finding; detailed definitions and coding rules for all predictors are provided in Supplementary Table S1. If any indicator from a multi-category variable reached statistical significance (P < 0.05) in univariate screening, the entire pair of indicators for that variable was advanced to subsequent modelling. LASSO regression (glmnet package, α = 1) was applied with λ selected by 10-fold cross-validation and the one-standard-error rule, which additionally addresses potential collinearity by shrinking coefficients of redundant predictors to zero. Variables with non‑zero coefficients at the chosen λ were retained. Missing data were handled as described in Section 2.4. The retained variables were then evaluated by bidirectional stepwise logistic regression (stepAIC, MASS package, direction = “both”), where at each step a predictor was added or removed solely based on whether the change reduced the AIC; the model obtained upon convergence of this stepwise process served as the final multivariable model. The nomogram was built directly from this final model without manual inclusion or exclusion of variables. No shrinkage or optimism correction was applied, as the model was prospectively validated in an independent cohort and demonstrated excellent calibration (see Section 3.4 for details). Ethics approval This study was approved by the Institutional Review Board of [Shanghai Tongji Hospital Ethics Committee]. The trial was registered at the Chinese Clinical Trial Registry (https://www.chictr.org.cn/; Trial ID: [ChiCTR2500106049]). Results Characteristics of training and validation cohorts Table 1 presents the baseline characteristics of chronic cough patients in the training and validation cohorts. Collectively, CRC—comprising CVA, EB, AC, and cases classified as CRC based on treatment response alone—accounted for 144 patients in the training cohort and 64 in the validation cohort. NCRC—excluding those with concomitant CVA, EB or AC, as such patients may exhibit a partial response to corticosteroids—accounted for 160 patients in the training cohort and 67 in the validation cohort. No significant differences were observed between the two groups across most variables. Three variables showed statistically significant differences: reflux symptoms, diurnal difference in cough, and peak expiratory flow (PEF). Of these, PEF and diurnal difference were not included in the final prediction model. Although reflux symptoms were retained as a model predictor, the observed distributional difference stems from improved ascertainment precision—the prospective cohort’s detailed standardized inquiry markedly reduced “uncertain” classifications—rather than a true biological disparity, and therefore does not affect the validity of model construction or cross-cohort evaluation.Table 1Baseline characteristics of the study population.VariableAll (n = 435)Training cohort (n = 304)Validation cohort (n = 131)p valueSubtype CVA81 (18.62%)55 (18.09%)26 (19.85%)0.766 GERC98 (22.53%)61 (20.07%)37 (28.24%)0.081 EB26 (5.98%)19 (6.25%)7 (5.34%)0.884 UACS13 (2.99%)9 (2.96%)4 (3.05%)1.000 AC22 (5.06%)16 (5.26%)6 (4.58%)0.952 CRC*79 (18.16%)54 (17.76%)25 (19.08%)0.848 NCRC*140 (32.18%)104 (34.21%)36 (27.48%)0.205Age(year)50.09 ± 15.4949.97 ± 15.3550.38 ± 15.860.800Gender0.661 Female294 (67.59%)203 (66.78%)91 (69.47%) Male141 (32.41%)101 (33.22%)40 (30.53%)BMI(kg/m2)23.92 ± 3.5123.82 ± 3.4724.15 ± 3.600.372Duration(months)12.00 [5.00;54.50]12.00 [4.00;51.00]15.00 [6.00;54.50]0.405Atopy0.337 Absent145 (33.33%)97 (31.91%)48 (36.64%) Uncertain135 (31.03%)92 (30.26%)43 (32.82%) Present155 (35.63%)115 (37.83%)40 (30.53%)Reflux Symptoms0.005 Absent215 (49.43%)150 (49.34%)65 (49.62%) Uncertain100 (22.99%)81 (26.64%)19 (14.50%) Present120 (27.59%)73 (24.01%)47 (35.88%)Diurnal ifference0.006 Daytime Heavier256 (58.85%)185 (60.86%)71 (54.20%) Same Weight101 (23.22%)76 (25.00%)25 (19.08%) Nighttime Heavier78 (17.93%)43 (14.14%)35 (26.72%)EOS(×109/L)0.12 [0.06;0.22]0.12 [0.06;0.22]0.10 [0.07;0.22]0.856CRP(mg/L)1.10 [0.50;2.35]1.10 [0.50;2.52]1.10 [0.50;2.10]0.535IgE(IU/mL)29.90 [10.25;95.15]33.00 [11.30;106.75]26.10 [9.30;85.00]0.242FEV1(%pre)99.28 ± 15.3198.51 ± 15.49101.05 ± 14.790.114FEV1/FVC(%)80.61 ± 7.8780.82 ± 8.280.14 ± 7.040.385MMEF(%pre)81.26 ± 28.9381.03 ± 28.4181.8 ± 30.210.801PEF(%pre)90.65 ± 18.6489.47 ± 19.3393.39 ± 16.680.044MEF75(%pre)91.19 ± 22.8490.59 ± 22.892.58 ± 22.960.404MEF50(%pre)86.33 ± 28.8786.03 ± 28.6387.02 ± 29.510.745MEF25(%pre)80.95 ± 55.8678.6 ± 35.0486.4 ± 86.670.321FeNO50(ppb)16.00 [10.00;26.00]15.50 [11.00;26.00]16.00 [10.00;24.00]0.491FeNO200(ppb)8.00 [6.00;12.0]8.00 [6.00;12.00]7.00 [5.00;11.50]0.220CaNO(ppb)3.70 [2.40;5.60]3.70 [2.40;5.73]3.60 [2.30;5.35]0.682BMI body mass index, CRP C - reactive protein, IgE immunoglobulin E, FEV1% forced expiratory volume in 1 s percentage, FEV1/FVC forced expiratory volume in 1 s/forced vital capacity, MMEF maximal mid - expiratory flow, PEF peak expiratory flow, MEF75/50/25 maximal expiratory flow at 75/50/25% of vital capacity, FeNO50 Fractional Exhaled Nitric Oxide at 50 mL/s, FeNO200 fractional exhaled nitric oxide at 200 mL/s, CaNO concentration of alveolar nitric oxide.Categorical variables are presented as: number (percentage). Numerical variables are presented as: mean ± SD or median [25th percentile; 75th percentile]. *: The “CRC” and “NCRC” in the table are defined solely on the basis of treatment response, excluding cases in which a specific etiologic subtype has already been established; those cases are reported in their respective subtype columns. Table 1Baseline characteristics of the study population.VariableAll (n = 435)Training cohort (n = 304)Validation cohort (n = 131)p valueSubtype CVA81 (18.62%)55 (18.09%)26 (19.85%)0.766 GERC98 (22.53%)61 (20.07%)37 (28.24%)0.081 EB26 (5.98%)19 (6.25%)7 (5.34%)0.884 UACS13 (2.99%)9 (2.96%)4 (3.05%)1.000 AC22 (5.06%)16 (5.26%)6 (4.58%)0.952 CRC*79 (18.16%)54 (17.76%)25 (19.08%)0.848 NCRC*140 (32.18%)104 (34.21%)36 (27.48%)0.205Age(year)50.09 ± 15.4949.97 ± 15.3550.38 ± 15.860.800Gender0.661 Female294 (67.59%)203 (66.78%)91 (69.47%) Male141 (32.41%)101 (33.22%)40 (30.53%)BMI(kg/m2)23.92 ± 3.5123.82 ± 3.4724.15 ± 3.600.372Duration(months)12.00 [5.00;54.50]12.00 [4.00;51.00]15.00 [6.00;54.50]0.405Atopy0.337 Absent145 (33.33%)97 (31.91%)48 (36.64%) Uncertain135 (31.03%)92 (30.26%)43 (32.82%) Present155 (35.63%)115 (37.83%)40 (30.53%)Reflux Symptoms0.005 Absent215 (49.43%)150 (49.34%)65 (49.62%) Uncertain100 (22.99%)81 (26.64%)19 (14.50%) Present120 (27.59%)73 (24.01%)47 (35.88%)Diurnal ifference0.006 Daytime Heavier256 (58.85%)185 (60.86%)71 (54.20%) Same Weight101 (23.22%)76 (25.00%)25 (19.08%) Nighttime Heavier78 (17.93%)43 (14.14%)35 (26.72%)EOS(×109/L)0.12 [0.06;0.22]0.12 [0.06;0.22]0.10 [0.07;0.22]0.856CRP(mg/L)1.10 [0.50;2.35]1.10 [0.50;2.52]1.10 [0.50;2.10]0.535IgE(IU/mL)29.90 [10.25;95.15]33.00 [11.30;106.75]26.10 [9.30;85.00]0.242FEV1(%pre)99.28 ± 15.3198.51 ± 15.49101.05 ± 14.790.114FEV1/FVC(%)80.61 ± 7.8780.82 ± 8.280.14 ± 7.040.385MMEF(%pre)81.26 ± 28.9381.03 ± 28.4181.8 ± 30.210.801PEF(%pre)90.65 ± 18.6489.47 ± 19.3393.39 ± 16.680.044MEF75(%pre)91.19 ± 22.8490.59 ± 22.892.58 ± 22.960.404MEF50(%pre)86.33 ± 28.8786.03 ± 28.6387.02 ± 29.510.745MEF25(%pre)80.95 ± 55.8678.6 ± 35.0486.4 ± 86.670.321FeNO50(ppb)16.00 [10.00;26.00]15.50 [11.00;26.00]16.00 [10.00;24.00]0.491FeNO200(ppb)8.00 [6.00;12.0]8.00 [6.00;12.00]7.00 [5.00;11.50]0.220CaNO(ppb)3.70 [2.40;5.60]3.70 [2.40;5.73]3.60 [2.30;5.35]0.682BMI body mass index, CRP C - reactive protein, IgE immunoglobulin E, FEV1% forced expiratory volume in 1 s percentage, FEV1/FVC forced expiratory volume in 1 s/forced vital capacity, MMEF maximal mid - expiratory flow, PEF peak expiratory flow, MEF75/50/25 maximal expiratory flow at 75/50/25% of vital capacity, FeNO50 Fractional Exhaled Nitric Oxide at 50 mL/s, FeNO200 fractional exhaled nitric oxide at 200 mL/s, CaNO concentration of alveolar nitric oxide.Categorical variables are presented as: number (percentage). Numerical variables are presented as: mean ± SD or median [25th percentile; 75th percentile]. *: The “CRC” and “NCRC” in the table are defined solely on the basis of treatment response, excluding cases in which a specific etiologic subtype has already been established; those cases are reported in their respective subtype columns. Baseline characteristics of the study population. BMI body mass index, CRP C - reactive protein, IgE immunoglobulin E, FEV1% forced expiratory volume in 1 s percentage, FEV1/FVC forced expiratory volume in 1 s/forced vital capacity, MMEF maximal mid - expiratory flow, PEF peak expiratory flow, MEF75/50/25 maximal expiratory flow at 75/50/25% of vital capacity, FeNO50 Fractional Exhaled Nitric Oxide at 50 mL/s, FeNO200 fractional exhaled nitric oxide at 200 mL/s, CaNO concentration of alveolar nitric oxide. Categorical variables are presented as: number (percentage). Numerical variables are presented as: mean ± SD or median [25th percentile; 75th percentile]. *: The “CRC” and “NCRC” in the table are defined solely on the basis of treatment response, excluding cases in which a specific etiologic subtype has already been established; those cases are reported in their respective subtype columns. The development of prediction model Of 20 items based on patient basic information and test results, 10 variables showed statistical significance in univariate logistic regression (Table 2). Lasso regression, validated by 10-fold cross-validation and using Lambda.1se as the critical value, further screened the predictors and removed the indicators FeNO200 and CaNO (Fig. 2 & 3).Fig. 2The coefficient path plot of Lasso regression.Fig. 3The sparsity plot of Lasso regression.Table 2Univariate and multivariate logistic regression based on training cohort.VariableUnivariateMultivariateCRC (n = 143)NCRC (n = 161)p valueOR (95%CI)p valueOR (95%CI)Age48.66 ± 15.5451.14 ± 15.130.1600.848 (0.672–1.067)Gender0.5441.16 (0.719–1.872) Female93 (65.0%)110 (68.3%) Male50 (35.0%)51 (31.7%)BMI24.27 ± 3.7223.42 ± 3.19<0.051.288 (1.022–1.634)Duration(months)11.44 ± 26.5714.01 ± 24.510.3700.905 (0.714–1.117)Atopy*1<0.0016.868 (4.114–11.73)<0.00111.74 (5.263–27.95) Absent21 (14.7%)76 (47.2%) Uncertain36 (25.2%)56 (34.8%) Present86 (60.1%)29 (18.0%)Reflux Symptoms0.0020.361 (0.201–0.629)<0.010.302 (0.128–0.668) Absent78 (54.5%)72 (44.7%) Uncertain44 (30.8%)37 (23.0%) Present21 (14.7%)52 (32.3%)Diurnal Difference0.5330.854 (0.507–1.437) Daytime Heavier88 (61.5%)97 (60.2%) Same Weight38 (26.6%)38 (23.6%) NighttimeHeavier17 (11.9%)26 (16.1%)EOS0.30 ± 0.320.11 ± 0.08<0.0015.354 (3.221–9.634)<0.0013.656 (1.980–7.547)CRP3.15 ± 5.642.02 ± 2.84<0.051.301 (1.036–1.718)<0.051.487 (1.103–2.120)IgE165.82 ± 314.6845.51 ± 68.79<0.0015.181 (2.768–10.74)<0.012.749 (1.510–5.619)FEV1%96.33 ± 16.67100.36 ± 14.17<0.050.767 (0.604–0.964)FEV1/FVC79.76 ± 9.8381.85 ± 6.48<0.050.779 (0.617–0.972)0.1440.769 (0.537–1.088)MMEF78.17 ± 29.8383.58 ± 26.930.0990.822 (0.648–1.036)PEF87.89 ± 20.6790.88 ± 18.000.1830.86 (0.687–1.07)MEF7587.91 ± 25.1292.97 ± 20.300.0550.797 (0.63–1.002)MEF5083.25 ± 31.0788.51 ± 26.130.1110.829 (0.655–1.042)MEF2574.48 ± 35.5382.27 ± 34.300.0550.694 (0.475–1.001)FeNO5034.92 ± 33.7014.81 ± 8.72<0.0016.313 (3.564–12.24)<0.0016.153 (2.927–14.28)FeNO20014.64 ± 12.977.45 ± 3.68<0.0014.518 (2.725–8.14)CaNO5.72 ± 6.214.21 ± 3.01<0.051.39 (1.099–1.824) The coefficient path plot of Lasso regression. The sparsity plot of Lasso regression. Table 2Univariate and multivariate logistic regression based on training cohort.VariableUnivariateMultivariateCRC (n = 143)NCRC (n = 161)p valueOR (95%CI)p valueOR (95%CI)Age48.66 ± 15.5451.14 ± 15.130.1600.848 (0.672–1.067)Gender0.5441.16 (0.719–1.872) Female93 (65.0%)110 (68.3%) Male50 (35.0%)51 (31.7%)BMI24.27 ± 3.7223.42 ± 3.19<0.051.288 (1.022–1.634)Duration(months)11.44 ± 26.5714.01 ± 24.510.3700.905 (0.714–1.117)Atopy*1<0.0016.868 (4.114–11.73)<0.00111.74 (5.263–27.95) Absent21 (14.7%)76 (47.2%) Uncertain36 (25.2%)56 (34.8%) Present86 (60.1%)29 (18.0%)Reflux Symptoms0.0020.361 (0.201–0.629)<0.010.302 (0.128–0.668) Absent78 (54.5%)72 (44.7%) Uncertain44 (30.8%)37 (23.0%) Present21 (14.7%)52 (32.3%)Diurnal Difference0.5330.854 (0.507–1.437) Daytime Heavier88 (61.5%)97 (60.2%) Same Weight38 (26.6%)38 (23.6%) NighttimeHeavier17 (11.9%)26 (16.1%)EOS0.30 ± 0.320.11 ± 0.08<0.0015.354 (3.221–9.634)<0.0013.656 (1.980–7.547)CRP3.15 ± 5.642.02 ± 2.84<0.051.301 (1.036–1.718)<0.051.487 (1.103–2.120)IgE165.82 ± 314.6845.51 ± 68.79<0.0015.181 (2.768–10.74)<0.012.749 (1.510–5.619)FEV1%96.33 ± 16.67100.36 ± 14.17<0.050.767 (0.604–0.964)FEV1/FVC79.76 ± 9.8381.85 ± 6.48<0.050.779 (0.617–0.972)0.1440.769 (0.537–1.088)MMEF78.17 ± 29.8383.58 ± 26.930.0990.822 (0.648–1.036)PEF87.89 ± 20.6790.88 ± 18.000.1830.86 (0.687–1.07)MEF7587.91 ± 25.1292.97 ± 20.300.0550.797 (0.63–1.002)MEF5083.25 ± 31.0788.51 ± 26.130.1110.829 (0.655–1.042)MEF2574.48 ± 35.5382.27 ± 34.300.0550.694 (0.475–1.001)FeNO5034.92 ± 33.7014.81 ± 8.72<0.0016.313 (3.564–12.24)<0.0016.153 (2.927–14.28)FeNO20014.64 ± 12.977.45 ± 3.68<0.0014.518 (2.725–8.14)CaNO5.72 ± 6.214.21 ± 3.01<0.051.39 (1.099–1.824) Univariate and multivariate logistic regression based on training cohort. Subsequently, multivariate logistic regression was performed again, with both forward and backward stepwise selection, retaining only 7 valuable predictors (Fig. 4), including: reflux symptoms (OR = 0.302, 95% CI = 0.128–0.668, P < 0.01), history of allergic diseases (OR = 11.74, 95% CI = 5.263–27.95, P < 0.001), blood eosinophils (OR = 3.656, 95% CI = 1.980–7.547, P < 0.001), IgE (OR = 2.749, 95% CI = 1.510–5.619, P < 0.01), CRP (OR = 1.487, 95% CI = 1.103–2.120, P < 0.05), FEV1/FVC (OR = 0.769, 95% CI = 0.537–1.088, P = 0.144) and FeNO (OR = 6.153, 95% CI = 2.927–14.28, P < 0.001). Additionally, although the variable FEV1/FVC had a p-value > 0.05 in the multivariate logistic regression, it was retained in the model because it satisfied the minimum Akaike Information Criterion (AIC).Fig. 4Forest Plot of Multivariate Logistic Regression Analysis. Forest Plot of Multivariate Logistic Regression Analysis. Using the regression coefficients of multivariate logistic regression models to weight each feature in our models, we developed a risk score formula to predict CRC in chronic cough: Risk Score = n₀ + (n₁ × a) + (n₂ × b) + (n₃ × c) + (n₄ × d) + (n₅ × e) + (n₆ × f) + (n₇ × g) + (n₈ × h) + (n₉ × i). Predicted risk = 1/(1 + e−risk score). The corresponding numbers in the formula are presented in Table 3. The model that incorporated the above predictors was developed and presented as the nomogram (Fig. 5). Clinicians can determine the corresponding points for each variable by referencing the answers to the aforementioned questions and the specific numerical values of the associated metrics on the nomogram. By summing the points of all variables, clinicians can derive an estimated probability to assist in predicting corticosteroid responsiveness in patients with chronic cough.Fig. 5Nomogram of the prediction model for corticosteroid responsive cough in chronic cough.Table 3Coefficient Comparison Table of Forecast Formula.ConstantVariablen₀−0.3620CodeNameExplanationn₁−0.6478aAllergies_00if the history of allergic diseases is clearly defined, a = 1; or a = 0n₂2.4176bAllergies_01if the history of allergic diseases is clearly present, b = 1; or b = 0n₃0.5659cHeartburn_00if the reflux symptom is clearly defined, c = 1; or c = 0n₄−1.5648dHeartburn_01if the reflux symptom is clearly present, d = 1; or d = 0n₅5.4196eEOSspecific values for blood eosinophils (×109/L)n₆0.0943fCRPspecific value of blood C-reactive protein(mg/L)n₇0.0027gIgEspecific values of IgE (IU/mL)n₈−0.0296hFEV1/FVCthe ratio of FEV1 to FVC (%)n₉0.0412iFeNO50specific value of FeNO measured at 50 mL/s (ppb) Nomogram of the prediction model for corticosteroid responsive cough in chronic cough. Table 3Coefficient Comparison Table of Forecast Formula.ConstantVariablen₀−0.3620CodeNameExplanationn₁−0.6478aAllergies_00if the history of allergic diseases is clearly defined, a = 1; or a = 0n₂2.4176bAllergies_01if the history of allergic diseases is clearly present, b = 1; or b = 0n₃0.5659cHeartburn_00if the reflux symptom is clearly defined, c = 1; or c = 0n₄−1.5648dHeartburn_01if the reflux symptom is clearly present, d = 1; or d = 0n₅5.4196eEOSspecific values for blood eosinophils (×109/L)n₆0.0943fCRPspecific value of blood C-reactive protein(mg/L)n₇0.0027gIgEspecific values of IgE (IU/mL)n₈−0.0296hFEV1/FVCthe ratio of FEV1 to FVC (%)n₉0.0412iFeNO50specific value of FeNO measured at 50 mL/s (ppb) Coefficient Comparison Table of Forecast Formula. The simplification of prediction model To simplify the model, continuous predictors were dichotomized using clinically relevant cut‑offs, and multi‑level categorical predictors were treated as ordinal variables. The updated risk score formula to predict CRC is as follows: Risk Score = m₀ + (m₁ × a’) + (m₂ × b’) + (m₃ × c’) + (m₄ × d’) + (m₅ × e’) + (m₆ × f’) + (m₇ × g’). Predicted risk = 1/(1 + e−risk score). The corresponding numbers in the formula are presented in Table 4 and the simplified model is presented as a nomogram in Fig. 6.Fig. 6Nomogram of the simplified prediction model for corticosteroid responsive cough in chronic cough.Table 4Coefficient Comparison Table of Simplified Forecast Formula.ConstantVariablem₀−2.3287CodeNameExplanationm₁2.5526a’Allergiespatient’s history of allergic disease (0 = none, 0.5 = uncertain, 1 = confirmed)m₂−1.0520b’Heartburnpresence of reflux/heartburn symptoms (0 = none, 0.5 = uncertain, 1 = confirmed)m₃2.1005c’if_EOSwhether blood eosinophil count ≥ 0.25 × 109/L (1 = yes, 0 = no)m₄0.6854d’if_CRPwhether C-reactive protein≥3 mg/L (1 = yes, 0 = no)m₅1.2491e’if_IgEwhether immunoglobulin E ≥ 150 IU/mL (1 = yes, 0 = no)m₆1.7996f’if_FEV1_FVCwhether FEV1/FVC ≥ 0.70 (1 = yes, 0 = no)m₇1.5587g’if_FeNOwhether FeNO50 ≥ 25 ppb (1 = yes, 0 = no) Nomogram of the simplified prediction model for corticosteroid responsive cough in chronic cough. Table 4Coefficient Comparison Table of Simplified Forecast Formula.ConstantVariablem₀−2.3287CodeNameExplanationm₁2.5526a’Allergiespatient’s history of allergic disease (0 = none, 0.5 = uncertain, 1 = confirmed)m₂−1.0520b’Heartburnpresence of reflux/heartburn symptoms (0 = none, 0.5 = uncertain, 1 = confirmed)m₃2.1005c’if_EOSwhether blood eosinophil count ≥ 0.25 × 109/L (1 = yes, 0 = no)m₄0.6854d’if_CRPwhether C-reactive protein≥3 mg/L (1 = yes, 0 = no)m₅1.2491e’if_IgEwhether immunoglobulin E ≥ 150 IU/mL (1 = yes, 0 = no)m₆1.7996f’if_FEV1_FVCwhether FEV1/FVC ≥ 0.70 (1 = yes, 0 = no)m₇1.5587g’if_FeNOwhether FeNO50 ≥ 25 ppb (1 = yes, 0 = no) Coefficient Comparison Table of Simplified Forecast Formula. The validation of prediction model The accuracy and discrimination of the nomogram were validated using ROC and calibration curves. The results showed that AUC was 0.900 (95% CI = 0.865–0.935) for the training cohort and 0.840 (95% CI = 0.771–0.909) for the validation cohort, suggesting an enhanced discrimination ability of the prediction model. For the simplified model, the AUC was 0.876 (95% CI = 0.836–0.916) and 0.824 (95% CI = 0.748–0.899) in the training and validation cohorts, respectively (Fig. 7). This indicates that the simplified model also has good discriminatory ability.Fig. 7ROC curves of the nomogram model and its simplified version for predicting CRC.A, B ROC curves of the original nomogram model in the training and validation sets, respectively. C, D ROC curves of the simplified nomogram model in the training and validation sets, respectively. ROC curves of the nomogram model and its simplified version for predicting CRC. A, B ROC curves of the original nomogram model in the training and validation sets, respectively. C, D ROC curves of the simplified nomogram model in the training and validation sets, respectively. The calibration of the nomogram was assessed using calibration slope, intercept, Brier score, and mean absolute error (MAE) in both the training and temporal validation cohorts, and further validated internally via 500 bootstrap resamples with bias correction (Fig. 8). In the training set, the original model showed perfect calibration (slope = 1.000, intercept ≈ 0, Brier = 0.122, MAE = 0.018); however, calibration declined in the validation set (slope = 0.573, intercept = −0.118, Brier = 0.170, MAE = 0.045), suggesting mild overfitting. By contrast, the simplified model exhibited substantially more consistent calibration across cohorts (training: slope = 1.000, intercept ≈ 0, Brier = 0.137, MAE = 0.027; validation: slope = 0.753, intercept = −0.015, Brier = 0.163, MAE = 0.025). Bootstrap internal validation confirmed minimal optimism for the simplified model, with the bias-corrected calibration curve closely tracking the ideal 45°line. These findings validate the calibration of the simplified nomogram, supporting its use for accurate CRC risk estimation in chronic cough.Fig. 8Calibration curves of the nomogram model and its simplified version for predicting CRC.A, B Calibration curves of the original nomogram model in the training and validation sets, respectively. C, D Calibration curves of the simplified nomogram model in the training and validation sets, respectively. Calibration curves of the nomogram model and its simplified version for predicting CRC. A, B Calibration curves of the original nomogram model in the training and validation sets, respectively. C, D Calibration curves of the simplified nomogram model in the training and validation sets, respectively. The clinical efficacy of the model was analyzed through clinical decision curve analysis. The outcomes revealed that both the training cohort and validation cohort achieved maximum benefits at threshold probability values ranging from 0.09–1.00 and 0.07–0.89 in the primary prediction model, while ranging from 0.10–1.00 and 0.13–0.88, respectively, in the simplified model. Within these threshold probability ranges, the number of predicted CRC patients exceeded the number of NCRC patients. Additionally, the loss-to-gain ratio remained consistently less than 1, indicating good clinical efficacy of the two models (Fig. 9).Fig. 9Decision curve analysis of the nomogram model and its simplified version for predicting CRC.A, B Decision curves of the original nomogram model in the training and validation sets, respectively. C, D Decision curves of the simplified nomogram model in the training and validation sets, respectively. Decision curve analysis of the nomogram model and its simplified version for predicting CRC. A, B Decision curves of the original nomogram model in the training and validation sets, respectively. C, D Decision curves of the simplified nomogram model in the training and validation sets, respectively. Machine‑learning analysis and SHAP interpretability Six machine learning models were trained and evaluated on the temporal validation set (Table 5). Among them, random forest achieved the highest Brier score (0.164) with an AUC of 0.834, followed by XGBoost (AUC = 0.800, Brier = 0.179). SHAP analysis of the random forest model (Fig. 10A) identified the seven features with the greatest contribution to prediction: allergic disease history (mean |SHAP | = 0.0960), blood eosinophils (0.0830), FeNO50(0.0812), IgE (0.0469), FeNO200 (0.0381), FEV1/FVC (0.0203), and MEF25(0.0184), five of which overlapped with the logistic regression predictors. Representative individual SHAP explanations are shown in Fig. 10B-D.Fig. 10SHAP visualization of the random forest model.A Contribution of predictors to the model output. B–D SHAP individual case explanation plots for three representative patients.Table 5AUC and Brier score of different models on the training and validation sets.NameTraining setValidation setAUCBrier scoreAUCBrier scoreMultivariable Logistic Regression, LR0.900 (0.865–0.936)0.122 (0.100–0.147)0.840 (0.771–0.909)0.170 (0.125–0.218)Random Forest, RF1.000 (1.000–1.000)0.022 (0.020–0.025)0.834 (0.761–0.906)0.164 (0.137–0.197)Extreme Gradient Boosting, XGB0.942 (0.917–0.967)0.104 (0.090–0.121)0.800 (0.722–0.878)0.179 (0.145–0.219)Light Gradient Boosting Machine, LGBM0.759 (0.706–0.813)0.230 (0.224–0.236)0.739 (0.652–0.826)0.226 (0.216–0.237)Support Vector Machine, SVM0.905 (0.868–0.941)0.562 (0.532–0.592)0.845 (0.776–0.914)0.544 (0.495–0.594)K-Nearest Neighbors, KNN1.000 (1.000–1.000)0.028 (0.024–0.033)0.719 (0.632–0.805)0.283 (0.223–0.358)Decision Tree, DT0.877 (0.837–0.916)0.117 (0.093–0.149)0.683 (0.586–0.779)0.214 (0.167–0.275) SHAP visualization of the random forest model. A Contribution of predictors to the model output. B–D SHAP individual case explanation plots for three representative patients. Table 5AUC and Brier score of different models on the training and validation sets.NameTraining setValidation setAUCBrier scoreAUCBrier scoreMultivariable Logistic Regression, LR0.900 (0.865–0.936)0.122 (0.100–0.147)0.840 (0.771–0.909)0.170 (0.125–0.218)Random Forest, RF1.000 (1.000–1.000)0.022 (0.020–0.025)0.834 (0.761–0.906)0.164 (0.137–0.197)Extreme Gradient Boosting, XGB0.942 (0.917–0.967)0.104 (0.090–0.121)0.800 (0.722–0.878)0.179 (0.145–0.219)Light Gradient Boosting Machine, LGBM0.759 (0.706–0.813)0.230 (0.224–0.236)0.739 (0.652–0.826)0.226 (0.216–0.237)Support Vector Machine, SVM0.905 (0.868–0.941)0.562 (0.532–0.592)0.845 (0.776–0.914)0.544 (0.495–0.594)K-Nearest Neighbors, KNN1.000 (1.000–1.000)0.028 (0.024–0.033)0.719 (0.632–0.805)0.283 (0.223–0.358)Decision Tree, DT0.877 (0.837–0.916)0.117 (0.093–0.149)0.683 (0.586–0.779)0.214 (0.167–0.275) AUC and Brier score of different models on the training and validation sets. Discussion In this study, a predictive diagnostic model was developed for CRC by integrating multiple clinical variables. In the prediction model, we selected reflux symptoms, history of allergic diseases, blood eosinophil count, CRP, IgE, FEV1/FVC, and FeNO as predictors to forecast the probability of CRC. Notably, five of these predictors overlap with the top seven variables identified by the optimal machine learning model (Random Forest). This high methodological convergence further supports that the association between these predictors and the disease is unlikely to be coincidental. EOS, FeNO and IgE are currently recognized biomarkers of type-2 inflammation18. Elevations in these biomarkers are closely associated with multiple type 2 inflammatory cytokines19–23. Many clinical studies have investigated the relationship between these three biomarkers and responsiveness to corticosteroid therapy. Price DB et al. demonstrated that EOS can be used to predict the response to inhaled corticosteroid (ICS) treatment in patients with asthma24. A study involving 142 non-smoking patients with chronic cough demonstrated that individuals with elevated blood eosinophil counts (≥237/μL) exhibited a significantly higher response rate to corticosteroid therapy than those with normal counts25. Moreover, a double-blind, randomized, placebo-controlled trial indicated that FeNO can predict responsiveness to corticosteroid therapy in patients with nonspecific respiratory symptoms, including cough26. A meta-analysis by Ambrosino P et al. concluded that an elevated pretreatment FeNO level helps identify patients with chronic cough who are likely to respond to ICS27. In contrast, serum IgE appears to be a less reliable predictor of corticosteroid efficacy. Razi E et al. observed no significant difference in treatment response between groups with high and low total serum IgE levels28. Similarly, Joseph J et al. found that regular ICS use in patients with mild-to-moderate persistent asthma did not significantly alter total serum IgE concentrations29. Nevertheless, extant investigations have predominantly relied on isolated biomarkers to predict corticosteroid responsiveness—a strategy that remains inherently limited. Individual indices are susceptible to diurnal variation30,31, physical exertion32,33, dietary intake34,35, and other confounders36–38, while inter-individual heterogeneity further undermines both the specificity and the precision of such predictions39. This study integrates multiple established markers of type-2 inflammation, thereby enhancing both the sensitivity and specificity of treatment-response prediction. In addition, it incorporates several variables—such as CRP, a history of allergic disease, and reflux-related symptoms—for which current evidence does not yet support or refute their utility in predicting corticosteroid responsiveness. However, in this study, these variables were found to possess a certain diagnostic value for CRC within the overall model. CRP, an acute-phase protein synthesized by the liver, is elevated not only during infection but also in a variety of chronic inflammatory conditions, and it might therefore rise in non-infectious airway inflammation40–43. Reflux-related symptoms aid in identifying most NCRC, particularly GERC1,44,45, suggesting that their absence could serve as a protective indicator against CRC. A history of allergic disease is already integral to the diagnosis of AC and also contributes to predicting CRC46,47. Compared with the aforementioned type-2 inflammatory markers, these indicators, although offering only modest gains in model sensitivity, show considerable potential for enhancing specificity, as they rarely become abnormal in patients lacking the corresponding etiology. Collectively, by integrating the three established type-2 airway-inflammation biomarkers with readily obtainable clinical data and routine ancillary tests, this study developed a prediction model for corticosteroid-responsive cough within the chronic-cough population, achieving both high sensitivity and specificity. Additionally, other parameters—most notably small-airway indices such as MMEF, MEF25 and MEF75—have been linked to corticosteroid-responsive cough48,49. In a retrospective analysis by Yi F et al., more than half of patients with CVA exhibited small-airway dysfunction50. The study by Bai H et al. further demonstrated that MMEF, FEF75%, and FEF50% are diagnostically meaningful for CVA when specific cutoff values are applied51. In the present study, these indices were evaluated as candidate predictors but ultimately failed to enter the final model. We attribute this finding to the limited small-airway involvement in the patients enrolled in the training cohort. In the univariable logistic-regression screening, small-airway parameters—such as MEF25—yielded a p-value close to 0.05, indicating their potential as predictors of CRC. Furthermore, in the machine learning models used for validation, we also observed that small-airway‑related indicators ranked high in contribution; thus, we maintain confidence in their ability to predict CRC. Nevertheless, to remain consistent with the predefined study protocol, we refrained from forcibly including these metrics in the final prediction model. Finally, it should be noted that, in the simplified model, continuous variables were dichotomised using evidence-based cutoff values. Because the literature reports inconsistent thresholds for each variable, we systematically evaluated every documented candidate in the derivation cohort and retained the value that yielded the highest discriminatory performance in the validation cohort. We acknowledge that this cutoff selection process was partly informed by validation data, which may modestly influence the performance estimates of the simplified model. This was a pragmatic modeling decision intended to identify the most clinically suitable thresholds, and the simplified model’s metrics should therefore be interpreted with appropriate caution. Detailed criteria for cutoff selection are provided in Supplementary Material 1. For example, the selection of the optimal IgE cutoff was guided by previously reported threshold values52–55. Limitations This study has several limitations. First, as a single‑center study, the sample size satisfied the established 10‑events‑per‑variable (10 EPV) criterion for model development, although larger sample sizes are increasingly advocated to further enhance model stability and reduce overfitting56. Second, a substantial number of patients were excluded due to incomplete data, largely reflecting poor adherence. Such patients may have had less favorable initial treatment responses, potentially introducing selection bias by enriching the analytic cohort with more treatment‑responsive or common etiologies. Third, the CRC outcome combined patients with confirmed etiological diagnoses (CVA, EB, AC) and those identified solely by empirical corticosteroid response, which may introduce some outcome heterogeneity. Larger multicenter cohorts with improved data completeness and standardized outcome definitions are warranted to validate and refine the model. Conclusion In conclusion, we have developed a clinically applicable nomogram that estimates the likelihood of corticosteroid response in patients with chronic cough. The model demonstrated strong discrimination and calibration in single‑center temporal validation, providing a quantitative probability that may serve as an aid for treatment decision-making. Its clinical utility awaits further evaluation through external validation across multiple centers and prospective implementation studies. Supplementary information Supplementary Information TRIPODAI_checklist Supplementary Information TRIPODAI_checklist