The Usefulness of Machine Learning Models to Predict Patient‐Reported Outcome Measures in Chronic Rhinosinusitis With Nasal Polyps Authors: Yang Shen, Pan‐Hui Xiong, Bo‐Wen Zheng, Chen‐Xi Li, Jun‐Liang Chen, Yue Gu, Yan‐Han Yang, Tao Lu, Yu‐Cheng Yang Published: 2026 ABSTRACT ABSTRACTBackground and ObjectiveThe treatment available for chronic rhinosinusitis with nasal polyps (CRSwNP) has remained unsatisfactory. Patient‐reported outcome measures (PROMs), capturing patient‐perceived health status and well‐being, are vehicles for measuring and improving the efficacy of care. This study aimed to establish machine learning (ML) models to predict PROMs in CRSwNP patients using minimally invasive and easily acquired clinical data.MethodsWe collected commonly available clinical predictive data from 437 patients and established four separate ML models in the training set: a least absolute shrinkage and selection operator (LASSO)‐based Logistic regression, a random forest (RF) regression, a gradient‐boosted decision tree (GBDT), and a deep neural network (DNN). In the test and independent external validation sets, the predictive performance of these models was measured by calculating C statistics, expected prediction results, and decision curves. A feature‐ranking analysis was performed using the ML algorithm. We then developed a predictive nomogram using LASSO‐based Logistic regression.ResultsThe models performed well on an independent external validation set, with no statistically significant differences in generalization ability metrics between groups. LASSO regression identified key features of the predictive PROMs. A nomogram was created based on multivariate Logistic regression with LASSO regularization.ConclusionsAll four ML models demonstrated similar performance in predicting PROMs in CRSwNP patients from which a clinical nomogram was developed. Early prediction of the subjective treatment response is crucial, as it influences clinician decisions and facilitates effective doctor‐patient communication preoperatively; this could lead to more precise and personalized treatment for CRSwNP patients. Introduction Chronic rhinosinusitis with nasal polyps (CRSwNP) is a common and complex inflammatory disorder of the upper airway affecting > 4% of the general population [1]. Nasal polyps (NP) are usually benign edematous masses in the nasal and paranasal cavities but may cause severe nasal congestion, discharge and postnasal drip, facial pain, and pressure, headache, a decreased or lost sense of smell, and symptoms derived from lower airway involvement [2]. These symptoms often last for many years and significantly affect the patients' quality of life, impair their work productivity and social life [3], and result in a large financial burden for the patient and the healthcare system [2]. Traditionally, the standard‐of‐care medical management of CRSwNP includes intranasal and oral corticosteroids, long‐term antibiotics, nasal saline irrigation, and endoscopic sinus surgery (ESS) [4]. Pharmacological therapy is used to control tissue inflammation, and surgery is performed to remove polyps and polypoid mucosal tissue. However, in most cases, the combination of pharmacological therapy and ESS cannot completely treat this disease. In patients with uncontrolled severe CRSwNP, the frequency of recurrence after surgery is as common as 40% after 18 months [5] and 62% after 10 years [6]. Thus, preoperative prediction methods could help identify high‐risk patients for uncontrolled prognosis, inform doctors' treatment options (e.g., surgery or emerging biologic therapy), and facilitate effective doctor‐patient communication preoperatively. Patient‐reported outcome measures (PROMs) are essential for tailoring the most appropriate intervention to the individual patient and providing complete post‐treatment efficacy evaluation. Compared with objective clinical measures (i.e., laboratory, computed tomography [CT] scan, and endoscopic outcomes), PROMs capture specific data on the clinical interventions and tangible improvements in patient health status, productivity, and overall well‐being. Improving shared decision‐making, symptom management, patient satisfaction, and quality of life [7]. Currently, several healthcare systems worldwide have focused on PROMs as vehicles for measuring and improving the value of care, and are increasingly adopted globally [8, 9, 10]. Given the importance of understanding PROMs during disease management, a commonly used method, the visual analog scale (VAS), was selected as a tool to predict prognosis in this retrospective study. Machine learning (ML) has been previously applied in CRS research. Ishikawa et al. [11] used ML to thoroughly investigate features of chronic eosinophilic pneumonia (CEP) as a comorbidity of CRS and to predict the likelihood of eosinophilic chronic rhinosinusitis (ECRS) and CEP in patients. The team also performed ML‐based classification according to the presence and severity of blood eosinophilia, which may provide novel insights into the pathophysiology of CRS [12]. Furthermore, Morse et al. [13] utilized ML to characterize patterns of olfactory impairment associated with CRS. Therefore, in this study, we initially employed four minimally invasive ML models that rely on easily obtainable clinical variables. We then compared their performance to quantify the validated prediction models further. We found good performance on all four ML models and determined predictive values of each clinical variable in determining PROM‐prompted prognosis. Finally, we established a nomogram using a least absolute shrinkage and selection operator (LASSO)‐based Logistic regression model to provide a more intuitive and visual form for clinical application. Material and Methods Study Design and Population This retrospective, double‐center follow‐up study was conducted at the First Affiliated Hospital of Chongqing Medical University and Xishui People's Hospital. The protocols were approved by the ethics committee of the First Affiliated Hospital of Chongqing Medical University (approval number: 2022‐K301) for both sites. The requirement for informed consent was waived due to the retrospective and observational nature of the study. The diagnosis of NP was based on the clinical history, anterior rhinoscopy, nasal endoscopy, and CT scans, according to the current European Academy of Allergy and Clinical Immunology (EAACI) Position Paper on Rhinosinusitis and Nasal Polyps [14]. CRSwNP patients were enrolled based on the following inclusion criteria: age > 18 years, complete clinical data, and available patient records during follow‐up. Patients with anthrochoanal polyps, cystic fibrosis, fungal sinusitis, primary ciliary dyskinesia, or systemic diseases were excluded. This study followed the transparent reporting of a multivariable prediction model for individual prognoses or Diagnosis (TRIPOD) reporting guidelines for prognostic studies. Data from 437 CRSwNP patients who underwent Functional Endoscopic Sinus Surgery (FESS) between January 2015 and December 2022 were analyzed. The model development and internal validation involved a cohort of patients from the First Affiliated Hospital of Chongqing Medical University. A second cohort from Xishui People's Hospital was used for external validation. Clinical Measures Clinical data including patient age, sex, disease duration, history of asthma, recurrence, and smoking were collected. VAS assessing Chronic rhinosinusitis (CRS) symptoms, including nasal obstruction, rhinorrhea, facial pain/pressure, headache, loss of smell, and overall burden caused by CRS, were chosen as PROMs [15]. For the VAS scoring, patients rated the extent of their symptoms on six slider scales, one for total symptoms and five individual symptoms, including nasal obstruction, nasal discharge, smell impairment, facial pain/pressure, and sleep dysfunction/fatigue. The term “total symptoms” indicates the cumulative effect of all the above‐mentioned symptoms associated with CRS summarized as the impact on the patients' quality of life. The scoring was calibrated from 0 (no impact/discomfort) to 10 (maximum impact/discomfort) for each symptom based on the severity experienced in the previous month. The preoperative CT scans were scored according to the method described by Lund and Mackay [16]. The ratio of the ethmoid sinus score (E score) to the maxillary sinus score (M score), or E/M ratio, was calculated to confirm CRS. The diagnosis of allergic rhinitis (AR) was diagnosed based on typical symptoms and atopy test results. The atopic status was determined by detecting serum allergen‐specific IgE antibodies using ImmunoCAP (Phadia, Uppsala, Sweden). Asthma was diagnosed by a physician according to the Global Initiative for Asthma (GINA) [17], based on a history of recurrent dyspnea, wheezing, or coughing episodes and positive airway reversibility testing (forced expiratory volume 1 > 12%, reaching 200 mL after the inhalation of 400 mg of salbutamol) or positive provocation testing. The blood eosinophil count was determined using automated analysis preoperatively. Sample Collection During Surgery and Histological Eosinophil Analysis NP tissues were collected during surgery. The tissue sections were embedded in paraffin wax, sectioned at 4–5 μm thicknesses, and stained with hematoxylin and eosin (HE). Sections were observed under a low‐power magnification (×100) to obtain a general expression. At ×400, the absolute number of eosinophils was counted in five random non‐overlapping fields beneath the epithelial surface. The five counts were averaged to obtain the mean number of cells per high‐power field. Patients with tissue eosinophil count > 10/HPF were considered the ECRS subtype [18, 19]. All analyses were performed by two independent observers, neither of whom had access to the diagnosis or clinical data. Follow‐up and Outcome Assessment All operations were performed by the same team of rhinologists following accepted functional ESS principles. The scope of the procedure and the instruments used were decided on a case‐by‐case basis by the performing surgeon. Patients started to receive nasal glucocorticoid spray and hypertonic saline irrigations 1 week postoperatively, and the dressing was changed regularly until the surgical cavity became epithelialization. Patients were followed up at 1, 3, 6, 12, 18, and 24 months postoperatively. The outcome was assessed at 36 months post‐ESS using VAS scores. Controlled CRS is defined as the total symptom VAS ≤ 2, healthy or nearly healthy mucosa, and no requirement for systemic medicine to control the disease. Partly controlled CRS is defined as the total symptom VAS > 2 and ≤ 5, and the need for a course of antibiotics or corticosteroids in the last 3 months. Uncontrolled CRS is defined as the total symptom VAS > 5 and ≤ 10, coupled with the need for long‐term antibiotics or corticosteroids in the last month [20, 21, 22], factors known to impact the quality of life [23]. The outcome was the prognosis of NP (control or partly control, defined as VAS ≤ 5). Statistical and ML Analyses SPSS 27.0 (SPSS Inc., Chicago, Illinois, USA) and R (version 4.3.0; R Foundation, Vienna, Austria) were used for statistical analysis. The PROMs were assessed through Logistic regression analysis 3 years postoperatively, with outcomes classified as uncontrolled versus partially controlled/controlled dichotomous variables. The relaxed LASSO was used to select the variables with the strongest relationship with uncontrolled prognosis outcomes. LASSO regression prevents multicollinearity and overfitting issues associated with excessive prediction variables. We employed a 10‐fold cross‐validation, centralizing and normalizing variables to select the optimal lambda value. A model with good performance and minimal independent variables was constructed using “lambda+1se.” Compared with other methods such as stepwise regression models, it has the advantages of variable selection and optimization of the number of predictive variables, resulting in a parsimonious model that utilizes the most important set of predictors [24]. Logistic regression models with LASSO regularization were used to predict the probability of an outcome. The predictive performance was assessed using receiver operating characteristic (ROC) curves and the corresponding area under the curve (AUC). The cutoff points for continuous variables were determined based on the maximal AUC. The nomogram was created by proportionally transforming each β coefficient in the multivariate Logistic regression analysis into a 0‐to‐100‐point scale. The points for all variables were then added to derive a total score, which was subsequently transformed into a predicted probability [25]. The predictive accuracy of the nomogram was evaluated using a concordance index (C‐index), ranging from 0.5 (no predictive value) to 1.0 (perfect accuracy). Next, we constructed three other ML prediction models: the random forest (RF) [26], the gradient‐boosted decision tree (GBDT) [27], and the deep neural network (DNN) [28] and compared the prediction performances of these models. A RF is a collection of decision trees created using bootstrapped samples of training data and random feature selection for correct fit tree induction. A GBDT is another ensemble‐approach decision tree addition model estimated by gradient descent. When implementing the RF and GBDT models, we used a grid search strategy to determine the best combination of hyperparameters by using the “ranger” and “caret” packages [29]. A DNN is a class of ML algorithms comprising multiple layers of nonlinear networks that process the values of the parameters learned by the unit to obtain the best outcome prediction. ML algorithms have been used to develop prediction models on large and small datasets [30, 31]. The value of these algorithms for predicting PROMs prognosis in CRSwNP remains unclear. Cross‐validation was applied to internally validate the stability of the model by randomly splitting patients in the training cohort into 10 equal samples. Seven of these samples were used to construct the models, and the model coefficients were applied to the remaining samples. This splitting process was repeated 10 times. To assess the external validity, the model was applied to an independent dataset from Xishui People's Hospital. In the test set, we measured the prediction performance of each model by computing (1) C‐statistics (the ROC), (2) prospective prediction results (sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio, and negative likelihood ratio), and (3) decision curve analysis. These results were compared across the models. ROC curves were compared using the DeLong test [32]. The value of p < 0.05 was considered statistically significant. Next, variable importance, a scaled measure with a maximum value of 100, was computed for all models. The development processes of the nomogram and ML are shown in Figure 1. The process of development of nomogram and machine learning methods. DNN, deep neural network; GBDT, gradient‐boosted decision tree; LASSO, least absolute shrinkage and selection operator; RF, random forest. Results Patient Characteristics The demographic and clinical characteristics of the study population are presented in Table 1. In the training and test sets, 307 patients underwent ESS to treat CRS, as shown in Table 1A; most baseline characteristics were not statistically different and were comparable between the two cohorts. Demographic and clinical characteristics of the independent external validation set (n = 130) are presented in Table 1B. Table 1ADemographic and clinical pathological characteristics of patients with CRSwNP.CohortCharacteristicsTraining set (n = 213)Test set (n = 94) p valueMale, n (%)146 (68.5)70 (74.5)0.343Female, n (%)67 (31.5)24 (25.5)0.157Age (y), mean ± SD43.5 ± 14.443.5 ± 13.90.984VAS (points), mean ± SD5.7 ± 1.75.6 ± 1.70.652Nasal obstruction (points), mean ± SD5.3 ± 2.55.5 ± 2.30.510Olfactory (points), mean ± SD3.8 ± 3.53.5 ± 3.30.451Purulent nasal discharge (points), mean ± SD3.7 ± 2.83.5 ± 2.70.645CT score (points), mean ± SD11.8 ± 4.911.5 ± 4.80.538E/M, mean ± SD1.8 ± 1.12.0 ± 1.10.183Polyp, n (%)149 (70.0)71 (75.5)0.388Asthma, n (%)29 (13.6)13 (13.8)0.995AR, n (%)92 (43.2)44 (46.8)0.643Allergy, n (%)54 (25.4)27 (28.7)0.633Lesion extent, n (%)179 (84.0)72 (76.6)0.163Blood eosinophil ratio (%), mean ± SD3.3 ± 2.23.4 ± 3.00.288Blood eosinophil number (×109), mean ± SD0.19 ± 0.130.22 ± 0.160.256Smoking, n (%)67 (31.5)30 (31.9)1.000Surgical history, n (%)57 (26.8)22 (23.4)0.574Tissue eosinophil ratio > 10% per HPF, n (%)123 (57.7)47 (50.0)0.257Tissue eosinophil number > 10 per HPF, n (%)77 (36.2)34 (36.2)1.000Abbreviations: AR, allergic rhinitis; CRSwNP, chronic rhinosinusitis with nasal polyps; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; HPF, high power field; SD, standard deviation; VAS, Visual Analogue Scale.John Wiley & Sons, Ltd. Demographic and clinical pathological characteristics of patients with CRSwNP. Abbreviations: AR, allergic rhinitis; CRSwNP, chronic rhinosinusitis with nasal polyps; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; HPF, high power field; SD, standard deviation; VAS, Visual Analogue Scale. Table 1BDemographic and clinical pathological characteristics of patients with CRSwNP in external validation set (n = 130).CharacteristicsValuesMale, n (%)70 (53.8)Female, n (%)60 (46.2)Age (y), mean ± SD40.8 ± 13.9VAS (points), mean ± SD6.1 ± 1.2Nasal obstruction (points), mean ± SD5.3 ± 1.2Olfactory (points), mean ± SD1.0 ± 1.9purulent nasal discharge (points), mean ± SD3.8 ± 1.3CT score (points), mean ± SD14.3 ± 5.4E/M, mean ± SD1.7 ± 0.6Polyp, n (%)31 (23.8)Asthma, n (%)10 (7.7)AR, n (%)16 (12.3)Allergy, n (%)19 (14.6)Lesion extent, n (%)17 (13.1)Blood eosinophil ratio (%), mean ± SD3.1 ± 3.3Blood eosinophil number (×109), mean ± SD0.27 ± 0.37Smoking, n (%)35 (26.9)Surgical history, n (%)6 (4.6)Tissue eosinophil ratio > 10% per HPF, n (%)35 (26.9)Tissue eosinophil number > 10 per HPF, n (%)19 (14.6)Abbreviations: AR, allergic rhinitis; CRSwNP, chronic rhinosinusitis with nasal polyps; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; HPF, high power field; SD, standard deviation; VAS, Visual Analogue Scale.John Wiley & Sons, Ltd. Demographic and clinical pathological characteristics of patients with CRSwNP in external validation set (n = 130). Abbreviations: AR, allergic rhinitis; CRSwNP, chronic rhinosinusitis with nasal polyps; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; HPF, high power field; SD, standard deviation; VAS, Visual Analogue Scale. LASSO‐Based Logistic Regression Model for PROMs The results of the univariate and multivariate analyses of the clinical predictors of PROMs are presented in Table 2. Table 2Univariate and multivariate analysis of predictors of the prognosis after FESS (n = 213).UnivariateMultivariateVariable β OR (95%CI) p value β OR (95%CI) p valueVAS−0.3370.714 (0.608–0.838)< 0.001−0.3110.733 (0.621–0.865)< 0.001E/M−0.3830.682 (0.546–0.851)< 0.001−0.2570.774 (0.605–0.988)0.040Asthma1.3603.896 (1.981–7.663)< 0.0011.1593.187 (1.556–6.528)0.002Tissue eosinophil number0.7912.206 (1.359–3.582)0.0010.6281.874 (1.087–3.234)0.024Olfactory−0.1010.904 (0.845–0.968)0.004−0.0260.974 (0.900–1.055)0.909Nasal obstruction−0.1650.848 (0.763–0.942)0.002−0.0080.992 (0.868–1.134)0.909CT−0.0530.948 (0.902–0.996)0.0340.0081.008 (0.947–1.072)0.810AR−0.7030.495 (0.306–0.799)0.004−0.4150.660 (0.355–1.227)0.189Allergy−0.6910.501 (0.297–0.845)0.009−0.0390.962 (0.474–1.954)0.915Male−0.0330.968 (0.523–1.790)0.917Age−0.0040.996 (0.976–1.016)0.678Purulent nasal discharge−0.0300.971 (0.877–1.075)0.569Polyp−0.4520.636 (0.333–1.216)0.171Lesion extent−0.4130.642 (0.417–1.284)0.276Blood eosinophil ratio (%)−0.0540.947 (0.865–1.037)0.240Blood eosinophil number (×109)−0.9410.390 (0.092–1.660)0.203Smoking−0.1620.851 (0.463–1.563)0.602Tissue eosinophil ratio > 10%−0.0560.923 (0.832–1.021)0.265Surgical history−0.5230.593 (0.317–1.110)0.102Abbreviations: AR, allergic rhinitis; CI, confidence interval; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; EOS, eosinophil; FESS, functional endoscopic sinus surgery; OR, odds ratio; VAS, Visual Analogue Scale.John Wiley & Sons, Ltd. Univariate and multivariate analysis of predictors of the prognosis after FESS (n = 213). Abbreviations: AR, allergic rhinitis; CI, confidence interval; CT, computed tomography; E/M, ethmoidal/maxillary sinus density ratio; EOS, eosinophil; FESS, functional endoscopic sinus surgery; OR, odds ratio; VAS, Visual Analogue Scale. To identify candidate prognostic predictors, we performed LASSO regression on the 11 clinical variables, which aided in the selection and optimization of predictors. With the optimal λ and γ, five variables were selected by the algorithm: pre‐FESS VAS, E/M, tissue eosinophil blood count (EOS), asthma comorbidity, and AR comorbidity. Multivariate Logistic regression models were used to confirm the regression coefficients of these LASSO‐selected variables and investigate their predictive values of these risk variables (Table 3), resulting in an AUC of 0.724 (95% CI 0.652–0.795) and 0.748 (95% CI, 0.636–0.860) in training and test cohorts, respectively. These values were significantly higher than the AUC obtained for each variable in the model (Figure 2A,B), implying a better predictive capability. Table 3Estimated values and parameters of regression coefficients of nomogram (n = 213).Parameter β SEWald chi‐square testOR95%CI p valueVAS−0.2560.1026.2530.774(0.634–0.946)0.012E/M−0.2950.1553.6020.745(0.549–1.010)0.058Asthma−1.0950.4446.0870.335(0.140–0.798)0.014AR−0.8000.3285.9640.449(0.236–0.854)0.015Tissue eosinophil number−0.2250.3590.3920.799(0.395–1.615)0.531Abbreviations: AR, allergic rhinitis; CI, confidence interval; E/M, ethmoidal/maxillary sinus density ratio; EOS, eosinophil; OR, odds ratio; SE, standard error; VAS, Visual Analogue Scale.John Wiley & Sons, Ltd. Estimated values and parameters of regression coefficients of nomogram (n = 213). Abbreviations: AR, allergic rhinitis; CI, confidence interval; E/M, ethmoidal/maxillary sinus density ratio; EOS, eosinophil; OR, odds ratio; SE, standard error; VAS, Visual Analogue Scale. Receiver operating characteristic (ROC) curves and nomogram calibration curves. (A) ROC curve of the nomogram in the training set. (B) Area under the curve (AUC) comparison between nomogram and univariate prediction. (C) Calibration curve of the training set nomogram. (D) Calibration curves comparing the ideal model, nomogram, and bias‐corrected model in the test set. Machine Learning Models for Prognosis Prediction To further quantify the validated prediction models with good performance and the relative predictive values of each minimally invasive clinical variable in determining prognosis, we used three other ML models and compared their performance with the multivariate Logistic regression model. The discriminative powers of the different models, represented by the ROC curves, are shown in Figure 3A. Table 4 lists the generalization ability metrics of the four models in the test set, including the AUC, sensitivity, and specificity. No statistically significant differences were observed between the groups. The discrimination power of the DNN model was the lowest (C‐statistic, 0.717; 95% CI, 0.596–0.838), with the highest positive predictive value (0.875; 95% CI, 0.788–0.962) and positive likelihood ratio (2.908; 95% CI, 1.514–5.584) (Table 4). Prediction ability of the Logistic regression with LASSO regularization model and machine learning models for the prognosis after FESS in the test set. (A) ROC curve of the machine learning models in the test set. (B) Calibration curve in the test set. AUC, area under the curve; DNN, deep neural network; GBDT, gradient‐boosted decision tree; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic. Table 4Prediction ability of the Logistic regression with LASSO regularization model and three machine learning models for the prognosis after FESS in the test set (n = 94).Outcome and modelC statistic (95% CI) p valueSensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)PLR (95% CI)NLR (95% CI)Logistic regression with LASSO regularization0.748 (0.636–0.860)Reference0.767 (0.660–0.874)0.688 (0.527–0.848)0.821 (0.721–0.922)0.611 (0.452–0.770)2.453 (1.440–4.179)0.339 (0.203–0.568)Random forest0.764 (0.655–0.872)0.8440.754 (0.649–0.859)0.704 (0.531–0.876)0.860 (0.769–0.950)0.543 (0.378–0.708)2.544 (1.400–4.625)0.350 (0.214–0.571)Gradient‐boosted decision tree0.751 (0.645–0.856)0.9730.867 (0.781–0.953)0.516 (0.340–0.692)0.776 (0.676–0.876)0.667 (0.478–0.855)1.791 (1.229–2.611)0.258 (0.125–0.536)Deep neural network0.717 (0.596–0.838)0.7170.754 (0.649–0.859)0.741 (0.575–0.906)0.875 (0.788–0.962)0.556 (0.393–0.718)2.908 (1.514–5.584)0.332 (0.206–0.537) Note: Comparison of area under the curve of the reference model (conventional triage approaches) with those of each machine learning approach using a DeLong test.Abbreviations: CI, confidence interval; FESS, functional endoscopic sinus surgery; LASSO, least absolute shrinkage and selection operator; NLR, negative likelihood ratio; NPV, negative predictive value; PLR, positive likelihood ratio; PPV, positive predictive value.John Wiley & Sons, Ltd. Prediction ability of the Logistic regression with LASSO regularization model and three machine learning models for the prognosis after FESS in the test set (n = 94). Note: Comparison of area under the curve of the reference model (conventional triage approaches) with those of each machine learning approach using a DeLong test. Abbreviations: CI, confidence interval; FESS, functional endoscopic sinus surgery; LASSO, least absolute shrinkage and selection operator; NLR, negative likelihood ratio; NPV, negative predictive value; PLR, positive likelihood ratio; PPV, positive predictive value. The RF model presented the highest discriminatory power (C‐statistic, 0.764; 95% CI, 0.655–0.872) and negative likelihood ratio (0.350; 95% CI, 0.214–0.571). Although the GBDT model showed the highest sensitivity (0.867; 95%CI, 0.781–0.953), it displayed the lowest specificity (0.516; 95% CI, 0.340–0.692), which can be selected and used according to different scenarios in clinical work. In the external validation, the four ML models demonstrated satisfactory predictive proficiency, generalization ability, and AUC (Figure 4A and Table 5). Prediction ability of the Logistic regression with LASSO regularization model and machine learning models for the prognosis after FESS in the external validation set. (A) ROC curve of the machine learning models in the external validation set. (B) Calibration curve in the external validation set. AUC, area under the curve; DNN, deep neural network; FESS, functional endoscopic sinus surgery; GBDT, gradient‐boosted decision tree; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic. Table 5Prediction ability of the Logistic regression with LASSO regularization model and three machine learning models for the prognosis after FESS in the external validation set (n = 130).Outcome and modelC statistic (95% CI) p valueSensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)PLR (95% CI)NLR (95% CI)Logistic regression with LASSO regularization0.713 (0.594–0.832)Reference0.600 (0.481–0.791)0.815 (0.668–0.961)0.886 (0.793–0.980)0.458 (0.317–0.599)3.240 (1.433–7.325)0.491 (0.347–0.695)Random forest0.735 (0.614–0.856)0.80050.765 (0.664–0.866)0.750 (0.577–0.923)0.897 (0.818–0.975)0.529 (0.362–0.697)3.059 (1.511–6.193)0.314 (0.193–0.510)Gradient‐boosted decision tree0.712 (0.598–0.827)0.99030.733 (0.621–0.845)0.710 (0.550–0.869)0.830 (0.729–0.931)0.579 (0.422–0.736)2.526 (1.427–4.472)0.376 (0.233–0.605)Deep neural network0.663 (0.563–0.763)0.48880.756 (0.667–0.844)0.525 (0.370–0.680)0.782 (0.695–0.868)0.488 (0.339–0.638)1.591 (1.125–2.249)0.466 (0.292–0.743)Abbreviations: CI, confidence interval; FESS, functional endoscopic sinus surgery; LASSO, least absolute shrinkage and selection operator; NLR, negative likelihood ratio; NPV, negative predictive value; PLR, positive likelihood ratio; PPV, positive predictive value.John Wiley & Sons, Ltd. Prediction ability of the Logistic regression with LASSO regularization model and three machine learning models for the prognosis after FESS in the external validation set (n = 130). Abbreviations: CI, confidence interval; FESS, functional endoscopic sinus surgery; LASSO, least absolute shrinkage and selection operator; NLR, negative likelihood ratio; NPV, negative predictive value; PLR, positive likelihood ratio; PPV, positive predictive value. In the decision curve analysis (Figure 3B), the net benefits of all models were better within the threshold probability range, with the RF model demonstrating the largest net benefit. However, the net benefits of the GBDT model are the lowest. In the external validation (Figure 4B), the net gains of all models were within the threshold probability range, with the RF model having the largest net gains and the GBDT model showing a low net gain in most cases. Variable Importance Figures 5 and 6 illustrate feature‐importance plots for the GBDT and RF models, respectively. Figure 5 ranks predictors by gain in the GBDT model, while Figure 6 ranks the top 10 predictors by MeanDecreaseGini in the RF model, with higher values indicating greater importance. In both models, VAS, CT, olfactory function, nasal obstruction, E/M, asthma comorbidity, and eosinophil‐related indicators were identified as important predictors. Furthermore, pre‐FESS VAS, E/M, tissue EOS, asthma comorbidity, and AR comorbidity were also included in the top 10 predictors of prognosis, which was consistent with the variables selected in the Logistic regression with LASSO regularization. Furthermore, we found that some variables (nasal obstruction, olfactory function, and CT E/M density ratio) were also important predictors of uncontrolled or recurring CRSwNP, which were not included in the logistic regression model. Importance of each predictor in the gradient‐boosted decision tree models. AR, allergic rhinitis; CT, computed tomography; EM, eosinophil‐to‐monocyte; VAS, visual analog scale. Importance of each predictor in the random forest models. AR, allergic rhinitis; CT, computed tomography; EM, eosinophil‐to‐monocyte; VAS, visual analog scale. Predictive Nomogram With LASSO‐Based Multivariate Logistic Regression A nomogram was created based on multivariate Logistic regression with LASSO regularization, to provide intuitive guidance to clinicians in deciding on treatment options and promoting effective doctor‐patient communication preoperatively. The b‐coefficients were used as weights using the “rms” package of the R software. Each variable had a specific score on a 0–100‐point scale to accurately predict the prognosis. Using these combined indicators, the nomogram provides a more accurate prediction of recurrence of CRSwNP than that achieved using a single peripheral factor alone. The scores for all variables were added to obtain the total points, and a vertical line was drawn from the corresponding position on the total point scale to the lowest point scale to predict prognosis (Figure 7). Calibration plots (apparent and bias‐corrected) overlapped with the ideal line created by the training and validation cohorts, showing adequate agreement between the predictive nomogram and the ML models (Figure 2D). Nomogram to predict prognosis after FESS. AR, allergic rhinitis; EM, eosinophil‐to‐monocyte; FESS, functional endoscopic sinus surgery; VAS, visual analog scale. Discussion In this study, we tested four ML models (LASSO‐based Logistic regression, RF, GBDT, and CNN) with clinical variables from PROMs to predict the prognosis and/or recurrence of CRSwNP in patients for 36 months post‐FESS and compared the performance of these models. Our results demonstrated that, for CRSwNP, the ML models demonstrated similar superiority using PROMs data points to predict post‐FESS clinical experience when compared to multivariate regression analysis. Using LASSO regression, the pre‐FESS VAS score, tissue eosinophil ratio, presence of asthma, and CT E/M density ratio were identified as important features for predicting the post‐FESS clinical experience of CRSwNP patients. A nomogram was created based on the multivariate Logistic regression using relaxed LASSO regularization. When directly compared with established prognostic factors reported in the literature, our LASSO‐derived model both corroborates and diverges from previous findings. Consistent with multiple studies, eosinophilic inflammation—represented by tissue EOS—and radiographic severity (E/M ratio) were reaffirmed as robust predictors of poor outcomes [33, 34, 35]. Similarly, comorbid asthma and AR have been consistently linked to worse postoperative courses and higher recurrence rates [35, 36]. However, a notable point of divergence lies in the predictive value of individual symptom domains. While several studies have identified olfactory‐specific VAS as a strong independent prognostic factor [33, 36], it did not retain independent significance in our final multivariate model. This discrepancy may be attributed to our study's focus on the comprehensive symptom burden (total VAS), which may encapsulate the overall disease impact more holistically than any single symptom. Alternatively, differences in patient populations, follow‐up durations, or the specific definition of ‘uncontrolled’ disease (PROMs‐based vs. endoscopic recurrence) could account for this variation. Patients with uncontrolled CRSwNP may face challenges of polyp recurrence, reoperation, and long‐term pharmacotherapy. Thus, pre‐ESS prediction of uncontrolled CRSwNP could help surgeons appropriately determine the need for additional surgery or other treatments. Previous studies have reported that clinical variables, such as a smell VAS score, previous NP surgeries, serum IgE, blood EOS, pre‐ESS ML Models, AR, and asthma are independent risk factors for CRSwNP prognosis [33, 35, 36]. However, few studies have focused on the prognosis using PROMs to lead their clinical data points, placing more emphasis on the subjective data collected from patient satisfaction and quality of life. To our knowledge, this is the first study to test four ML prediction models for predicting the prognosis of uncontrolled or recurring CRSwNP, and to compare their performance. In this study, we developed and validated a Logistic regression model and a nomogram to predict uncontrolled cases of CRSwNP. To simplify the prediction model and make it practical in a clinical setting, we focused only on risk factors that were easily accessible in routine preoperative examinations. The nomogram, based on pre‐FESS VAS, E/M, tissue EOS, asthma comorbidity, and AR comorbidity, had a discriminatory ability (C‐statistic) of 0.724 (95% CI 0.652–0.795) and 0.713 (0.594–0.832) in training and external validation cohorts, respectively. The predictive accuracy of pre‐ESS VAS alone in our study was represented by a C‐statistic of 0.635 (95% CI 0.555–0.715) and 0.691 (0.581–0.801) in training and validation cohorts, respectively. This indicates that using the nomogram based on PROMs data is slightly more effective than using a subjective pre‐operative VAS. Moreover, pre‐FESS VAS severity, which is measured by the combined impact of the E/M score, tissue EOS, and asthma/AR status, may be predictive of postoperative patient satisfaction. Consistent with previous research on predicting postoperative prognosis in CRSwNP patients [34], E/M and tissue EOS were independent risk factors that were predictive of uncontrolled or recurrent CRSwNP prognosis, but with low discriminative power and limited accuracy. Comorbidities have an important effect on the postoperative outcomes in CRSwNP. We showed that the comorbidities of asthma and AR were independently associated with an uncontrolled prognosis. Recently, ML models have been used to predict CRSwNP prognosis [35]. Some exhibit excellent accuracy. In this study, based on data from 437 patients with CRSwNP, we applied four modern ML models (Logistic regression‐based LASSO, RF, GBDT, and DNN) to routinely available preoperative information. These ML models demonstrated good predictive performance using PROMs to indicate uncontrolled or recurring CRSwNP. Which we confirmed by assessing AUCs, sensitivity, and specificity. In addition, decision curve analysis revealed that all ML models yielded a satisfactory net benefit. Our findings regarding the performance of ML models can be contextualized by direct comparison with prior applications in CRSwNP. For instance, Bai et al. [31] reported a higher AUC (0.79) using an RF model to predict polyp recurrence based on a set of clinical variables. The slightly lower discriminatory power of our best‐performing RF model (AUC 0.764) might stem from fundamental differences in the prediction target; we aimed to predict a PROMs‐based composite outcome of ‘uncontrolled’ disease at 36 months, which is a broader, more patient‐centric endpoint than radiographic or endoscopic recurrence alone. Furthermore, Guo et al. [33] utilized RF to identify biomarkers (e.g., CCL17 and MIP‐1β) for difficult‐to‐treat CRS. While our study did not incorporate such specific inflammatory cytokines, our model prioritized readily available clinical and PROMs data, enhancing its immediate practicality in routine preoperative assessment. This contrast highlights a trade‐off between biological precision and clinical feasibility. A deeper analysis of the performance discrepancies reveals insights into the models' characteristics. The Random Forest model, which aggregates numerous decorrelated decision trees, achieved the highest AUC, likely due to its robustness against overfitting with our moderate sample size and ability to model subtle interactions. Conversely, the Gradient‐Boosted Decision Tree model, while highly sensitive, showed lower specificity and net benefit. This pattern suggests that its sequential error‐correction approach might have overemphasized certain patterns in our training set, leading to slightly less calibrated probability estimates in validation. Importantly, the overall performance convergence underscores that for this clinical prediction task using readily available variables, multiple modeling approaches arrive at similarly useful predictions, with the choice of model being less critical than the inclusion of key predictors themselves. A key strength of our study is the external validation of our models using an independent cohort from a geographically distinct center, a step that strengthens generalizability compared to single‐center model developments [36]. To provide a more visual prediction, we developed a nomogram using a LASSO‐based regression. Partially consistent with a previous study [35], the nomogram based on the pre‐FESS VAS, E/M, tissue eosinophil count test (EOS), asthma comorbidity, and AR comorbidity had a satisfactory discriminatory ability of PROMs to predict recurrence in CRSwNP. This nomogram shares conceptual similarity with the tool developed by Du et al. [35], which also combines peripheral parameters for recurrence estimation. However, our model is distinct in its foundation on preoperative PROMs (VAS) as a core component, directly integrating the patient's subjective disease burden into the prognostic calculus. This aligns with a growing paradigm shift towards value‐based care that prioritizes patient‐reported experiences [8, 9, 10]. We believe that this nomogram holds significant clinical value, as it leverages patient‐reported data to provide a prognostic tool that is both straightforward and suitable for rapid implementation in routine clinical practice. This directly enhances clinical management by: first, enabling physicians to proactively formulate intensified therapeutic regimens for high‐risk patients, such as considering early initiation of biologic agents or more extensive surgical intervention, thereby facilitating personalized care; second, improving doctor‐patient communication and treatment adherence through visual risk stratification via the nomogram; third, guiding intensified postoperative surveillance for high‐risk individuals to achieve early detection and intervention, thereby reducing the rate of revision surgery; and fourth, optimizing the allocation of healthcare resources by targeting more aggressive therapies to those most in need. Collectively, these contributions can improve overall therapeutic efficacy, cost‐effectiveness, and long‐term patient outcomes. The integration of these predictive models, particularly the LASSO‐based nomogram, into routine clinical practice is both feasible and desirable. From a practical standpoint, the nomogram can be implemented as a simple, web‐based calculator or integrated directly into the electronic health record (EHR) system. During a preoperative consultation, the clinician would input five readily available variables: the pre‐FESS VAS score, the blood eosinophil‐to‐monocyte ratio (E/M), the tissue eosinophil count, and the presence of asthma and allergic rhinitis comorbidities. The tool would then instantly provide a visual risk stratification, classifying the patient's likelihood of an unsatisfactory patient‐reported outcome. This would serve as a structured aid for shared decision‐making, allowing the clinician to visually explain the rationale for considering more intensive treatment strategies—such as earlier biologic therapy or extended postoperative monitoring—for high‐risk patients. This process not only personalizes the treatment plan but also enhances treatment adherence by making the prognostic assessment transparent. The model's reliance on pre‐existing, minimally invasive clinical data ensures it can be adopted without imposing additional diagnostic burdens or costs, making it a practical tool for improving value‐based care in real‐world otolaryngology practice. This study has some limitations. The sample size limited the predictive data points included in this study. However, owing to this limitation, objective biomarkers and clinical features, such as polyp status and the presence of AR, are preferred, which are easily assessed by otolaryngologists. In contrast, the limited sample size resulted in individual epidemiological characteristics such as asthma and AR percentages being higher in the study population than in the general population. This is a well‐recognized selection bias in research and may limit the generalizability of pilot surveys. In addition, owing to the limitations of the pathological samples, the prediction of inflammatory factor addition was not included in this study. In the future, various inflammatory factors could be included to add to the data points used to predict the postoperative prognosis of patients with CRSwNP more comprehensively and accurately. Therefore, prospective studies with larger sample sizes should include additional potential predictors. Additionally, the inclusion of more sensitive biomarkers in model development would enhance future investigation. Conclusions In summary, we constructed four ML models for predicting recurrence or poor prognosis in CRSwNP after FESS and validated them externally. These models were developed using data routinely collected in clinical practice. Based on the Logistic regression model, a clinical nomogram was prepared using the five best factors for predicting prognosis, including the pre‐FESS VAS score, E/M ratio, tissue EOS, asthma comorbidity, and AR comorbidity. The four ML models exhibited similar superiority in predicting recurrence or poor prognosis of CRSwNP. Early prediction of treatment response is important and may greatly influence clinician options, providing insights for more precise personalized treatment of these patients. Author Contributions Yang Shen and Pan‐Hui Xiong contributed to the conception and design of the study. Yang Shen and Pan‐Hui Xiong collected the clinical data, organized the database, and wrote the first draft of the manuscript. Pan‐Hui Xiong performed the statistical analysis. Bo‐Wen Zheng, Chen‐Xi Li, Jun‐Liang Chen, Yue Gu, Yan‐Han Yang, and Tao Lu revised the manuscript. Yu‐Cheng Yang supervised the study and revised the manuscript. All authors have contributed to the article and approved the submitted version. Ethics Statement This retrospective, double‐center follow‐up study was conducted at the First Affiliated Hospital of Chongqing Medical University and Xishui People's Hospital. The study protocol was approved by the Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (Approval No. 2022‐K301) for both participating sites. The requirement for informed consent was waived by the committee due to the retrospective and observational nature of the study. All procedures were performed in accordance with the Declaration of Helsinki. Conflicts of Interest The authors declare no conflicts of interest.