Understanding heterogeneity in pulmonary embolism cases: identification of risk groups using cluster analysis

Objective: The aim of this study was to evaluate the heterogeneity of patients with pulmonary embolism (PE) based on their clinical, symptomatic, laboratory, and radiological characteristics, and to identify subgroups with similar profiles in order to define their risk stratification. In this context, Two-Step Cluster analysis was applied to determine distinct clinical phenotypes, with the goal of facilitating personalized diagnostic, follow-up, and treatment strategies. Furthermore, the identified clusters were assessed in relation to biomarker levels, prognostic scores, radiological findings, and survival outcomes to demonstrate their clinical significance, thereby contributing to the early identification of high-risk patients and the improvement of patient management. Materials and Methods: This retrospective study included 366 patients diagnosed with PE in our clinic between 2020 and 2024. The diagnosis was confirmed by clinical findings, laboratory tests, imaging modalities (CT angiography, echocardiography, Doppler ultrasonography), and clinical scoring systems (Wells, Geneva, PESI). Demographic data, symptoms, vital signs, laboratory parameters, radiological features, and risk factors were recorded. Statistical analyses were performed using ANOVA/Kruskal-Wallis and chi-square tests, with p<0,05 considered significant. Cluster analysis was performed with the Two-Step Cluster method, and the optimal number of clusters was determined using the Silhouette method. All analyses were conducted with SPSS version 22.0. Results: The mean age of the study population (n=366) was 62.4±16.2 years; 53.8% (n=197) were male and 46.2% (n=169) were female. The most common presenting symptoms were dyspnea 67.5% (n=247), chest pain 47.5% (n=174), and cough 20.8% (n=76). The most frequent physical and vital signs were oxygen desaturation 29.5% x (n=108), tachypnea 24.9% (n=91), and tachycardia 22.7% (n=83). Comorbidities were present in 36.1% (n=132) of patients. Based on 18 variables derived from symptoms and clinical findings, Two-Step Cluster analysis identified five distinct phenotypes (Silhouette=0.65). The analysis revealed significant intercluster differences in symptom profiles, hemodynamic parameters, biomarkers, and prognostic scores. Predictor Importance analysis demonstrated that "leg swelling" had the strongest discriminative power among clusters, followed by palpitations, tachycardia, oxygen desaturation, and leg tenderness (all p<0,05). Radiological findings, including right ventricular dilatation, pulmonary artery enlargement, and bilateral/segmental-subsegmental thrombi, differed significantly among clusters (p<0,05), being most prominent in Cluster 5, consistent with its high-risk clinical profile. In terms of symptom distribution, dyspnea, palpitations, tachycardia, tachypnea, oxygen desaturation, and hypotension were significantly more frequent in Cluster 5 (all p<0,05). Syncope and hemoptysis were also more common in this cluster (p<0,05). Leg swelling and tenderness were predominantly observed in Cluster 3, supporting its association with concomitant DVT (p<0,05). Biochemical and hemodynamic parameters including D-dimer, troponin T, NT-proBNP, CK-MB, and sPAP differed significantly across clusters (all p<0,05). Cluster 5 showed the highest levels of these markers, indicating cardiac strain and myocardial injury. D-dimer levels were elevated in both Cluster 3 and Cluster 5 (p<0,01), consistent with increased thrombotic activity. Hospitalization duration was also longer in Cluster 5 (p<0,05). Smoking history (pack-years) did not differ significantly among clusters (p>0,05). Multiple comparison analyses confirmed that Cluster 5 exhibited significantly higher levels of D-dimer, troponin T, and CK-MB (all p<0,01), whereas NT-proBNP was higher but not statistically significant (p=0,08). Length of hospital stay was significantly longer in Cluster 5 compared with Clusters 1 and 4 (p<0,05). The identified phenotypes were characterized as follows: "Symptomatic Stable PE Phenotype" (Cluster 1) represented lower-risk and stable cases; "Elderly PE xi Phenotype with Subclinical Cardiac Involvement" (Cluster 2) included older patients with high comorbidity burden; "Outpatient-Treatable PE Phenotype" (Cluster 3) reflected younger, low-risk cases with mild disease course; "Malignancy-Associated Silent PE Phenotype" (Cluster 4) consisted of cases with high malignancy prevalence, often incidentally diagnosed; and "Hemodynamically Unstable High-Risk PE Phenotype" (Cluster 5) represented the most severe group, with elevated biomarker levels, prolonged hospitalization, higher rates of massive PE (50%, p<0,001), and lower survival. Overall survival was significantly reduced in the unstable phenotype (p<0,001). Conclusion: The timely diagnosis of PE is critical due to its association with high mortality and morbidity, which can be mitigated with early treatment. This study demonstrates that PE is not a homogeneous disease but rather exhibits substantial heterogeneity in clinical course and prognosis. In conclusion, data-driven methods such as cluster analysis provide valuable insights into disease heterogeneity and may serve as useful tools for improving patient management and guiding individualized therapeutic strategies in PE. Keywords: Cluster analysis, pulmonary embolism, malignancy, comorbidity, heterogeneity, phenotype, prognosis, mortality