We included anonymized electronic medical records (EMRs) from inpatients aged 18–80 years hospitalized in The First Affiliated Hospital of Guangzhou Medical University (a tertiary hospital specialized in respiratory and critical care medicine) between July 2017 and December 2020 (training set), and between January 2021 and December 2022 (validation set) because of exacerbation onset. All hospitalized patients had a primary discharge physician-diagnosis of COPD (ICD-10-CM code: J44.), bronchiectasis (ICD-10-CM code: J47.) or both. Apart from extracting ICD-10 codes for diagnosis, we integrated chest high-resolution computed tomography (HRCT) and symptoms for confirming the diagnosis of bronchiectasis.1 Furthermore, we integrated spirometric findings and clinical symptoms to jointly diagnose COPD.2 We excluded patients with CBA without HRCT and spirometric results within 1 year (before and after hospitalization). The study was approved by ethics committee (ES-2023-K008-03) and additional informed consent was waived for the retrospective cohort.

We retrospectively extracted EMRs from hospital information system, including demographics [age, gender, body-mass index (BMI), residency], medical history (disease duration, smoking history, number of exacerbations in the past year, etc.), major comorbidities (asthma, cor pulmonale, hypertension, diabetes, etc.), symptoms and signs (cough, phlegm, shortness of breath, chest pain, hemoptysis), HRCT findings (modified Reiff score, emphysema, etc.), spirometry [forced vital capacity percentage predicted (FVC% pred), forced expiratory volume in one second percentage predicted (FEV1% pred) and FEV1/FVC%], laboratory findings (blood routine test, sputum culture, C-reactive protein, total IgE, etc.), medications, clinical outcomes [length of stay, non-invasive ventilation, invasive ventilation, intensive care unit (ICU) admissions, re-admission, etc.]. We adopted the Charlson Comorbidities Index (CCI) for assessment of comorbidities, and the modified Reiff scores to evaluate the radiologic severity of bronchiectasis (mild: 1–6; moderate: 7–12; severe: 13–18 points). The Bronchiectasis Severity Index (BSI) was computed for patients with CBA. Because the modified Medical Research Council dyspnea scale was not recorded from EMR,11 we estimated this scale based on patient's description of dyspnea in the present history as a proxy.

For the cross-sectional data analysis, we only included the findings of the first examination soon after admission among patients who underwent multiple examinations during the same hospitalization. For patients with missing data of HRCT and spirometry during hospitalization, we replaced these with the data from our study site within 1 year before and after hospitalization for subsequent analysis. We included only the single most complete hospital examination record for patients with multiple admissions within 1 year. We also conducted a longitudinal follow-up of CBA patients by capturing any of their subsequent electronic health records, extending until June 2023.

We collected 95 variables from the retrospective study (Fig. 1, Table S1, online supplement). We excluded variables with missing data among >15% of patients (N=29), no direct relevance to CBA classification (N=36) and medications and outcomes (N=13). For the remaining 17 variables, we performed multiple imputation for missing data with multi-level generalized linear regression model.12 Next, we performed an exploratory factor analysis13 for continuous variables (Table S2) or system hierarchical clustering analysis14 for categorical variables (Fig. S1) with the remaining 17 variables, thereby leaving six variables in the clustering analysis. These included the modified Reiff score category (mild vs. moderate-to-severe), hospitalizations in the past year (0 vs. 1–2 vs. ≥3 times), sputum culture (positive vs. negative), blood neutrophil count, blood eosinophil count, and the CCI (Fig. 2).

Fig. 1.

Flowchart of patient recruitment and data analysis. Bx, bronchiectasis; COPD, Chronic Obstructive Pulmonary Disease; CBA, COPD-bronchiectasis association.

(0.65MB).

Fig. 2.

Ranking of the contribution of variables for clustering analysis and the core variables. (A) Importance ranking of the clustering variables. Comparison of the grading of hospitalization in the past year (B), bronchiectasis severity (modified Reiff score category) (C), sputum culture findings (D), blood neutrophil (E), Charlson comorbidity index (F), blood eosinophil (G).

(0.21MB).

We employed a two-step clustering approach for patient categorization, beginning with the data compression into small subclusters to manage complex data sets effectively. This method utilized the log-likelihood distance for categorical data and the Euclidean distance for continuous variables to create dense subclusters. Following this step, an agglomerative hierarchical clustering method merged these subclusters based on the Ward's method. This has not only facilitated the efficient analysis of mixed data types but also automatically identified the most statistically significant cluster configuration, optimizing the handling of complex datasets.15 We then applied a decision-tree algorithm to all variables in clustering analysis to predict the specific cluster for individual patients, which allowed for calculation of the rates of misclassification.16

We further included an independent CBA validation cohort (hospitalized between January, 2021 and December, 2022 and 348 patients who had overlapped with the training set have been excluded). By comparing the clinical characteristics of CBA in the discovery cohort with those in validation cohort, we sought to confirm the robustness of clustering findings.

We summarized categorical variables as frequencies and proportions. We tested continuous variables for the Gaussian distribution (Kolmogorov–Smirnov test) and expressed as mean±standard deviation (SD) or median (interquartile range, IQR) as appropriate. We evaluated the differences between clusters for each variable using analysis-of-variance or Kruskal–Wallis test for continuous variables and Chi-square test or Fisher's exact test for categorical variables. Unless otherwise specified, we applied Bonferroni correction to two-group comparisons. We adopted multivariate Logistic regression models to analyze the risks of clinically significant events (e.g., ICU admission) during longitudinal follow-up, with the hazards ratio (HR) and 95% confidence interval (95%CI) being demonstrated. We defined statistical significance as two-sided P<0.05. We conducted statistical analyses using SPSS (Version 25.0; IBM Corp., Armonk, NY).

We initially identified 6635 patients with bronchiectasis alone, 15,356 with COPD alone, and 2109 with CBA. After exclusion of repeated inpatient records, 4511 and 9320 patients had a diagnosis of bronchiectasis and COPD, respectively. The corresponding proportion of CBA was 29.9% in bronchiectasis and 14.5% in COPD, respectively. Finally, after excluding patients with malignant tumor, we included 2928 patients with bronchiectasis, 5158 with COPD, and 1219 with CBA (Fig. 1).

Compared with COPD or bronchiectasis alone, CBA was consistently associated with a markedly greater age, longer disease duration, a higher proportion of patients with emphysema, higher blood neutrophil percentage, higher detection rate of Pseudomonas aeruginosa and more pronounced lung function impairment, evidenced by lower FEV1/FVC, FEV1%pred and FVC% pred (all P<0.001).

The proportion of intravenous antibiotic use was significantly higher in CBA (P<0.001). CBA yielded non-significantly higher proportion of patients with comorbidities (e.g. pulmonary aspergillosis, cor pulmonale), a longer length of hospital stay, and more frequent receipt of invasive ventilation and ICU admission compared with COPD or bronchiectasis alone (Table 1, Tables S3 and S4).

Of the 1219 hospitalized patients with CBA, 572 were excluded due to missing spirometry or HRCT records. The clinical characteristics of 647 patients who were included in and those excluded from clustering analysis were comparable in terms of gender distribution (males: 82.5% vs. 83.4%), age (median: 68 vs. 67 years), BMI (median: 21.3 vs. 21.0kg/m2), CCI (median: 1.0 vs. 1.0), and the proportion of patients admitted to ICU (8.2% vs. 9.3%) (all P>0.05, Table S5).

We identified five clusters of CBA (Table 2, Tables S6 and S7). Based on the stepwise discriminant analysis, the most influential variables were ordered as hospitalization in the past year (1.000), bronchiectasis severity (0.745), sputum culture (0.494), neutrophil count (0.023), eosinophils count (0.001) and CCI (0.001) (Fig. 2).

Cluster 1 (CBA-MS, N=120) consisted of patients with predominantly moderate-to-severe bronchiectasis. Overall, CBA-MS exhibited more pronounced clinical symptoms and severe radiologic features of bronchiectasis. Apart from the highest Reiff score (median: 10.0), CBA-MS was characterized by the highest rate of cor pulmonale (30.8%), history of hemoptysis (22.5%) and detection of P. aeruginosa (39.1%). Regarding clinical outcomes, patients in CBA-MS yielded the highest rate of ICU admissions during hospitalization (14.2%).

Patients in cluster 2 (CBA-FH, N=108) had frequent hospitalization (more than 2 times) due to exacerbations of bronchiectasis in the past year. All patients had mild bronchiectasis and poorer lung function (83.8% had FEV1% predicted<50%: adjusted P<0.001). Patients with CBA-AE had a highest proportion of corticosteroids use (oral corticosteroids accounting for 40.7%, adjusted P=0.042), and the re-admission rate was highest (67.6%, adjusted P<0.001) among the five clusters.

Patients in cluster 3 (CBA-BI, N=163) consistently had bacterial infection as determined with sputum bacterial culture. By contrast, patients in cluster 4 (CBA-NB, N=143) had on average 1–2 hospitalizations per year but no bacterial infection The clinical characteristics of CBA-BI were overall similar with those of CBA-NB, except that CBA-BI had a higher proportion of patients with C-reactive protein >0.6mg/dl (86.5% vs. 64.2%) and antibiotics use (35.0% vs. 28.0%) (both adjusted P<0.05).

Cluster 5 (CBA-NHB, N=113) comprised patients with no bacterial infection or hospitalization in the past year. CBA-NHB yielded the lowest proportion of a history of cor pulmonale (5.3%) and procalcitonin>0.05ng/mL (29.9%) and consistently had mild bronchiectasis (100%). Although the rate of ICU admission (1.8%) and readmission within the past year (1.2%) was lowest in CBA-NB, the use of non-invasive (25.7%) and invasive ventilation (4.4%) was nominally lower than CBA-MS only.

We employed a decision-tree model to predict the cluster assignment (Fig. 3). Based on the three cardinal variables, bronchiectasis severity in HRCT, hospitalization in the past year and sputum culture, 91.8% (95% CI, 89.70–93.92%) of patients could be assigned to the correct cluster. Our model demonstrated a strong predictive accuracy, effectively identifying the true positives, as indicated by the high sensitivity (97.5% [95% CI, 96.26–98.68%]) and positive predictive values (96.0% [95% CI, 94.50–97.52%]). However, the specificity (55.6% [95% CI, 51.73–59.39%]) and negative predictive values (66.7% [95% CI, 63.04–70.30%]) revealed some further room for improvement in classifying the true negatives accurately. The moderate Matthews correlation coefficient (0.58 [95% CI, 0.45–0.69]) has highlighted the model's overall balanced performance in classification tasks (Table S8).

Fig. 3.

Scheme of predicting clinical cluster assignment and the major clinical characteristics of the clusters. Patients were assigned to the five clusters by using three core variables: bronchiectasis severity, hospitalization in the past year, and sputum culture (upper panel). The lower panel demonstrates the gradings for the five major variables or dimensions – lung function impairment, blood neutrophil count, detection of P. aeruginosa from sputum, ICU admission, readmission within 1 year of follow-up. The downward arrows indicate the decreased trend as compared with the normal levels, the plus signs denote the magnitude of increase or the likelihood of positivity as compared with the absence, the normal levels, and the minus signs signal the absence of the findings. ICU; intensive care unit.

(0.32MB).

Next, we included 218 inpatients with CBA (validation set) for validation of the clustering findings of the decision-tree model. Compared with the training set, there were no significant differences in the baseline distribution in gender, age, BMI, CCI and duration of symptoms (all P>0.05) (Fig. 1, Table S9). Five clusters were classified by applying the discriminant analysis findings of the decision-tree model. Comparison of clusters of the training set and validation set revealed no significant differences in the CCI, proportion of emphysema, FEV1% pred<50%, neutrophil ratio, length of hospital stay and the use of invasive ventilation (all P>0.05, Table S10).

We analyzed the risks of oxygen therapy, non-invasive ventilation, mechanical ventilation and ICU admission with logistic regression after adjusting for the age, gender, BMI and CCI. Compared with CBA-NHB, CBA-MS (OR: 8.87, 95%CI: 1.98, 39.62, adjusted P=0.004), CBA-BI (OR: 6.87, 95%CI: 1.56, 30.38, adjusted P=0.011) and CBA-FH (OR: 4.81, 95%CI: 1.01, 22.91, adjusted P=0.048) consistently had a markedly higher risk of ICU admission (Fig. S2). We next employed Cox proportional regression analysis to investigate the risk of readmissions during longitudinal follow-up, which differed considerably among the clusters. Compared with CBA-NHB, CBA-FH (HR: 49.86, 95%CI: 18.24, 136.26, adjusted P<0.001), CBA-NB (HR: 19.18, 95%CI: 7.02, 52.42, adjusted P<0.001), CBA-MS (HR: 15.04, 95%CI: 5.43, 41.69, adjusted P<0.001) and CBA-BI (HR: 10.18, 95%CI: 3.70, 28.06, adjusted P<0.001) consistently exhibited a significantly higher risk of readmission of follow-up (Fig. 4A, Table S11). We also conducted receiver operation characteristics analysis to assess the risk of readmission within one years and two years of follow-up, the clustering grading (1=NHB, 2=BI, 3=MS, 4=NB, 5=FH) demonstrated a significant predictive capability (area under curve [95%CI]: 0.757 [0.716, 0.799] at 1 year; 0.764 [0.726, 0.802] at 2 years, both P<0.001, Fig. S3).

Fig. 4.

Longitudinal risk assessment and cluster membership dynamics of CBA. Panel A: Cox proportional hazards regression analysis demonstrating the risk for readmission over a longitudinal follow-up of 2.5–6 years, using CBA-NHB as the reference cluster, with adjustments for the gender, age, BMI and CCI (Training set). Panel B: Sankey plot illustrating the changes in the cluster membership over time during longitudinal follow-up (Training and validation set). The left side of the diagram represents the former cluster of patients with clusters labeled as CBA-BI (bacterial infection), CBA-FH (frequent hospitalizations), CBA-MS (moderate to severe bronchiectasis), CBA-NB (non-bacterial), and CBA-NHB (no hospitalizations or bacterial infection). The right side represents the re-evaluated clustering as per the rules of the decision-tree analysis, with the addition of ‘new’ to the original cluster labels—CBA-BI-new, CBA-FH-new, CBA-MS-new, CBA-NB-new, and CBA-NHB-new—the reassessment and potential shift in each patient's cluster membership. BMI, body mass index; CCI, Charlson Comorbidity Index; CI, confidence interval; HR adj, hazard ratio after adjusting for gender, age, BMI and CCI.

(0.16MB).

We identified 170 patients with CBA who had repeated hospitalization records within our study site (mean interval: 623.0 days). Overall, the clinical characteristics remained stable – no significant differences in blood routine test items, lung function metrics, bacterial culture findings, blood inflammatory markers, use of corticosteroids, and hospitalization outcomes (including invasive ventilation, length of stay, re-admission within 1 year, etc.). (all P>0.05, Table S12).

The decision-tree model for patients with re-admissions revealed that 31.8% of patients changed their cluster membership. Using the former hospitalization as the control event, Logistic regression analysis did not identify any significant differences in the risk of re-admissions between the former and the subsequent hospitalization. However, the Sankey diagram showed certain cluster membership changes which centered on CBA-FH, CBA-BI and CBA-NB (Fig. 4B). Regression analysis on these clusters revealed that CBA-FH exhibited a significant increase in the risk of hospitalization among patients whose cluster membership remained unchanged (HR: 10.57, 95%CI: 2.94, 38.09, P<0.001) (Table S13).