Pleural effusion (PE) is a common condition, with an incidence of 400 cases/100,000 population/years in Spain1 and 1.5 million of patients each year in United States,2 with about 60 possible etiologies.1,3–5 PE is a response of the pleura to specific insults (local, pulmonary or systemic). The characteristics of the PE vary with the underlying physiopathological mechanism, the etiology (local or systemic) of the disease, and a variety of factors including, but not limited to, the type of PE (acute or chronic), and previous use of medicines (diuretics, antibiotics, among other). However, pleural response patterns have not yet been associated with the cause of PE.

The analysis of pleural fluid (PF) may have diagnostic potential and be useful to determine the etiology of pleural effusion.6 Diagnosis of PE is generally based on biochemical parameters in PF as measured individually.7–18 However, this approach does not account for the fact the pleura may respond differently to the same insult, and PEs of the same etiology may show a different biochemistry. Combined models of analytical parameters have been documented to have a high discrimination power for PEs.19–22 This model may help identify pleural response patterns associated with the etiology of the disease. Cluster analysis has been used to identify subtypes or phenotypes of patients diagnosed with a specific respiratory disease such as asthma, chronic obstructive pulmonary disease, or obstructive sleep apnea hypopnea syndrome.23–25 Therefore, clustering can also be useful in PE.

The objective of this study was to create homogeneous groups of patients according to the analytic characteristics of their PF by the use of principal component and cluster analysis. This study will enable us to identify pleural response patterns and/or the underlying etiopathogenic mechanism of the disease.

We prospectively studied all consecutive patients with PE seen in our center between January 1, 2011 and December 31, 2017. The protocol is an extension of the study evaluated and approved by the Clinical Research Ethics Committee of Galicia (Registry 2010/378).

Pleural fluid samples were taken by thoracentesis prior to any treatment. All patients signed an informed consent before any procedure was performed (chest CT-scan with contrast, thoracentesis, pleural biopsy or thoracoscopy).

PEs were studied according to our local standard protocol. PF samples were sent to microbiology (for Ziehl-Neelsen staining and cultures in aerobic, anaerobic and Löwenstein media), cytology and biochemistry. Aliquots of PF were centrifuged immediately for 15min at 1500×g, and the supernatants were stored at −80°C. All biochemistry determinations were performed using standard commercially available methods. A predefined algorithm (Fig. 1) was followed in all cases to establish the diagnosis. Video-assisted thoracic surgery was performed using a standard technique.26 All thoracenteses and biopsies were performed by experienced pulmonologists. In patients who underwent more than a thoracentesis, only the first PF chemistry panel was used for statistical analysis.

Fig. 1.

Diagnostic algorithm followed.

BF, bronchofibroscopy; CT, spiral computed tomography.

(0.13MB).

Definitions for malignant, paramalignant, tuberculous and infectious (including non-complicated, complicated and empyemas) PEs, as well as PE associated with heart failure and other diseases are provided in the Online data supplement.1,27–29 The clinical, radiological, and analytical parameters measured in PF and in blood are also described in the Online data supplement.

The flowchart of patients is shown in Fig. 2. The initial data set corresponded to 2515 subjects whose ages ranged from 15 to 93 years. After subjects with missing data were excluded, the final sample was composed of 1613 patients. The etiologies and origin of metastatic PEs of patients included in the final sample are displayed in Table 1. Demographic data for the cohort are shown in Table 2.

Fig. 2.

Flowchart of the patients studied with a pleural effusion.

(0.1MB).

PCA and cluster analysis

The percentages of variability explained by each dimension were 24.4% for the first axis, 17.7% for the second, 9.3% for the third, and 7.5% for the fourth one. The contribution of analytical parameters in PF to the first four dimensions is shown in Fig. 3.

Fig. 3.

Contribution of analytical analytical parameters of pleural fluid to each of the dimensions obtained in the analysis of major components.

(0.39MB).

The optimal number of clusters identified was six. Fig. 4(A–F) shows a biplot that displays the relationships between the variables (arrows) used for building clusters and patients (dots) based on their individual analytical PF characteristics. Results are projected onto the first four dimensions yielded by PCA. Colors correspond to the six groups obtained from clustering, showing the classification of patients based on their analytical PF characteristics.

Fig. 4.

Biplot representation displaying relationships between the variables (arrows) used for clustering. Distribution of patients (dots) based on the individual analytical characteristics of their pleural fluid. Results are projected onto the first four dimensions yielded by principal component analysis. Colors for observations correspond to the six clusters obtained from cluster analysis. Figures A–F. Combinations of the four dimensions taken two by two.

1, cluster 1; 2, cluster 2; 3, cluster 3; 4, cluster 4; 5, cluster 5; 6, cluster 6. Dim, dimension.

ADAPF, adenosine deaminase pleural fluid; AlbPF, albumin pleural fluid; AlbS-PF, albumin serum-pleural fluid gradient; CA125PF, cancer antigen 125 pleural fluid; CEAPF, carcinoembryonic antigen pleural fluid; CholPF, cholesterol pleural fluid; CRPPF, C-reactive protein pleural fluid; IL-6PF, interleukin 6 pleural fluid; IL-8PF, interleukin 8 pleural fluid; LDHPF, lactate dehydrogenase pleural fluid; Lym, lymphocytes; PBNPPF, N-terminal pro-brain natriuretic peptide; ProtPF, proteins pleural fluid; Seg, segmented.

(0.6MB).

Significant differences were observed across the six clusters in terms of sex, age and analytical, radiological and clinical characteristics. Table 2 reports the values obtained for all parameters by cluster, with significant differences in all parameters (p=0.016 for gender and p<0.001 for the remainder of variables). Table 3 displays PEs in accordance with their etiology and the cluster in which the patient was included.

Table 3.

Classification of pleural effusions according to their etiology and the cluster they belong to.

Diagnosis	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Cluster 5	Cluster 6	Total
Metastatic	168	52	26	2	20	171	439
Primary of pleura	1	8	2	0	1	7	19
Myeloma	0	1	0	0	0	3	4
Leukemia	0	1	0	0	1	1	3
Lymphoma	0	9	1	1	4	34	49
Paramalignant	0	9	7	1	7	36	60
Tuberculosis	0	121	5	0	1	14	141
Simple parapneumonic	0	33	57	0	17	60	167
Complicated parapneumonic	0	10	75	2	0	2	89
Empyema	0	2	44	34	1	4	85
PE	0	3	5	0	0	8	16
Chylothorax	0	2	0	0	0	12	14
Pseudochylothorax	0	1	0	0	0	0	1
SLE	0	0	4	0	0	1	5
Rheumatoid arthritis	0	2	0	0	0	0	2
Post-CABG	0	2	0	0	0	1	3
Pancreatitis	0	0	2	0	2	0	4
Heart failure	0	2	2	0	279	86	369
Hepatic hydrothorax	0	0	0	0	23	11	34
Abscess/abdominal infection	0	1	1	0	2	4	8
After surgery	0	7	9	0	1	9	26
Thoracic trauma	0	5	4	0	2	7	18
Hemothorax	1	3	11	0	1	5	21
Viral pleuritis	0	1	4	0	1	4	10
Dressler syndrome	0	0	0	0	0	1	1
Hypoalbuminemia	0	0	0	0	6	1	7
Sarcoidosis	0	2	0	0	0	0	2
Uremia	0	0	0	0	7	1	8
Drugs	0	3	0	0	2	2	7
TOTAL	170	280	259	40	378	486	1613

PE, Pulmonary embolism; SLE, systemic lupus erythematosus; Post-CABG, after aortocoronary bypass revascularization surgery.

To the best of our knowledge, this is the first study to identify homogeneous groups of patients showing similar pleural response patterns based on the analytical characteristics of PF. The results obtained suggest that pleural response to aggressions is heterogeneous. On the one hand, pleural response to insults of the same etiology may vary (different clusters). On the other hand, pleural response to insults of different etiologies may be the same (same cluster).

In the last years, cluster-based analysis has been used to identify clusters of patients with other respiratory diseases based on shared clinical symptoms and phenotypes. Hence, five phenotypes of asthma were identified in a study based on level severity, with relevant differences among groups.23 In relation to COPD, five subgroups were established based on the clinical characteristics of patients (especially, severity of respiratory disease and presence of cardiovascular comorbidities and diabetes).24 A study in patients with obstructive sleep apnoea identified three groups of patients with different symptoms and comorbidities (sleep disorders, minimally symptomatic and daytime somnolence).25

PF accumulates in the pleural space by the action of different mechanisms.32 Pleural response varies according to the lung involved and the stimulus activated. Modification factors include time to thoracentesis, tumor load33 or previous use of drugs. We used biomarkers in PF to identify pleural response patterns, as they are the result of a process or disease, infection, cancer, or heart failure, and reflect the underlying physiopathological mechanism leading to PE. Combinations of six biomarkers were used to consider all factors that may be involved in a response without etiologies being identified. Diagnoses are generally based on critical and specific clinical and analytical findings. Clinically relevant data were obtained from clustering. Thus, a correlation was established between analytical PF parameters and the clinical characteristics of the patients who were more likely to be assigned to a specific cluster. Given a data set, each clustering algorithm can always generate a division, no matter whether the structure exists or not. To provide confidence for the clustering results derived from the algorithms used, we carried out an external evaluation based on patient's diagnosis and clinical data.

The six clusters identified represent pleural response patterns and classify them as neoplastic, inflammatory, infectious, superinfectious, paucicellular and miscellaneous (cluster 1–6, respectively). The results obtained are clinically relevant. PEs of the same etiology were not included in the same cluster. Thus, PEs of etiologies as disparate as pleural metastasis, non-complicated infection, or heart failure were included in six, four, and four different clusters, respectively (Table 3). Table 4 shows significant differences in analytical values for metastatic PEs across clusters 1, 2 and 6. The same occurred in non-complicated parapneumonic PEs across clusters 2, 3 and 6; and PEs caused for heart failure (cluster 5 and 6) (Table 4). These differences suggest that either, the underlying mechanism causing PE varies (i.e. in neoplastic PEs, the mechanism involved is not always the same), or the pattern of response (which are modulated by the modification factors described above) differs, which translates into a different modulation of PE biomarkers. The reason why pleural responds differently to insults of the same etiology is unknown. A reason might be that PE changes as the disease progresses, as it occurs with infectious PE. Another explanation may be that modification factors modulate biomarkers. The fact that PEs of the same etiology were categorized into different clusters indicates that an etiology cannot be ruled out based on a single parameter. Thus, a metastatic PE may show an inflammatory, infectious or paucicellular response pattern. Simple and complicated parapneumonic PEs may be lymphocyte-predominant. Heart-failure-associated PE may show high cholesterol, protein and albumin levels and lower NT-proBNP levels (cluster 6).

All clusters are composed of PEs of different etiologies (Table 3). Otherwise said, aggressions of different etiologies may induce a similar pleural response. No studies have been performed so far to explain this phenomenon. Finally, cluster 6 – a group composed of 486 patients (30% of the series) – included PEs of all etiologies and does not show any particular characteristic that may lead to a conclusive diagnosis (Table 2). This cluster includes patients of an advanced age (75 years), and it consists of patients with metastatic, paramalignant, and tuberculous PEs, lymphoma, heart failure or non-complicated infectious diseases that do not feature the characteristics commonly associated with these etiologies (Table 4). The non-specific pleural response exhibited by these patients suggests that several mechanisms may be involved in the production of PE, which would generate different pleural stimuli. Another explanation is that modification factors may modulate biomarkers. Whatever the cause is, this cluster illustrates the difficulties that physicians encounter in identifying the etiology of a PE.

Clinical or radiological characteristics were not considered for clustering, as the level of dyspnea, chest pain or PE size do not necessarily reflect a specific pleural response pattern. In contrast, the results obtained in this study support the identification of patient clusters based on biomarker analysis. Hence, time to thoracentesis was longer (39% after more than 30 days), the percentage of malignant lesions on imaging scans was higher (87%), and the proportion of patients with general syndrome was higher (54%) in cluster 1 (mostly, metastatic PEs). On the other hand, the percentage of patients with pulmonary lesions (71% and 75%, respectively) and high temperature (58% and 65%, respectively) was higher in clusters 3 and 4, which were primarily composed of infectious PEs. As many as 65% of PEs in cluster 4 were purulent. Finally, patients in cluster 5 – who prevailingly had heart failure – were the ones with a higher incidence of bilateral PE (38%) and a lower incidence of left-side PE (23%).

A limitation of this study is that patients were recruited from a single site. Therefore, the validity of this methodology in other sites needs to be confirmed. Although 312 patients were excluded for a potential double diagnosis, we cannot guarantee that all PEs were associated with a single etiology.

In conclusion, according to the analytical characteristics of the PF, six homogeneous groups of patients were identified that demonstrate that the pleural response to aggressions is heterogeneous: PEs of the same etiology show different response patterns, while PEs of different etiologies may show the same response pattern. This phenomenon illustrates the challenges faced when establishing a diagnosis in the presence of a PE. Further studies are required to better understand the modulating role of the cause of pleural effusion, identify the mechanisms that lead to the production of PE, and determine modification factors. Finally, further research should be conducted to elucidate the implications of the results of these studies in clinical practice.

This study was supported by an unrestricted grant by Laboratorios Menarini S.A.

Lucía Ferreiro. Corresponding author. Author. Conception and design. Analysis and interpretation of data. Drafted the submitted article. Approval of the final version.

Óscar Lado-Baleato. Co-Author. Conception and design. Analysis and interpretation of data. Revised the article critically. Approval of the final version.

María E. Toubes. Co-Author. Acquisition of data. Analysis and interpretation of data. Revised the article critically. Approval of the final version.

Juan Suárez-Antelo. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Antonio Pose-Reino. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Esther San José. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Adriana Lama. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Nuria Rodríguez-Núñez. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

José Manuel Álvarez Dobaño. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Francisco Javier González Barcala. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Jorge Ricoy. Co-Author. Acquisition of data. Revised the article critically. Approval of the final version.

Francisco Gude. Co-Author. Conception and design. Analysis and interpretation of data. Revised the article critically. Approval of the final version.

Luis Valdés. Author. Guarantor. Conception and design. Analysis and interpretation of data. Drafted the submitted article. Approval of the final version.

The authors declare that they have no conflicts of interest.