Algorithm-Based Risk Identification in Patients with Breast Cancer-Related Lymphedema: A Cross-Sectional Study

SIMPLE SUMMARY: The current study employed a cohort of 294 patients from two hospitals in northern Italy initially assembled to highlight factors leading to one consequence of breast cancer (BC): upper limb unilateral lymphedema (BCRL). BCRL occurrence is a multi-factorial pathological condition that is not widespread, with a medium-long-term onset affecting not only physical function but also the quality of life of BC survivors. In the current study, we employed the data to stratify the risk of BCRL using unsupervised low-dimensional data embeddings and clustering. In the proposed approach, the ordinal and the binary patients’ clinical variables were modeled separately in two distinct embeddings. Afterward, they were merged; thus, the final representation was a single prognostic map displaying three clusters of patients with peculiar features. The characteristics of each group were extracted and evaluated, identifying the factors associated with the high-risk cluster. Our findings might provide future insight into a precise risk stratification to target high-risk patients with tailored therapeutic intervention and focus resources on patients who deserve more attention. ABSTRACT: Background: Breast cancer-related lymphedema (BCRL) could be one consequence of breast cancer (BC). Although several risk factors have been identified, a predictive algorithm still needs to be made available to determine the patient’s risk from an ensemble of clinical variables. Therefore, this study aimed to characterize the risk of BCRL by investigating the characteristics of autogenerated clusters of patients. Methods: The dataset under analysis was a multi-centric data collection of twenty-three clinical features from patients undergoing axillary dissection for BC and presenting BCRL or not. The patients’ variables were initially analyzed separately in two low-dimensional embeddings. Afterward, the two models were merged in a bi-dimensional prognostic map, with patients categorized into three clusters using a Gaussian mixture model. Results: The prognostic map represented the medical records of 294 women (mean age: [Formula: see text] years) grouped into three clusters with a different proportion of subjects affected by BCRL (probability that a patient with BCRL belonged to Cluster A: [Formula: see text]; Cluster B: [Formula: see text]; Cluster C: [Formula: see text]). The investigation evaluated intra- and inter-cluster factors and identified a subset of clinical variables meaningful in determining cluster membership and significantly associated with BCRL biological hazard. Conclusions: The results of this study provide potential insight for precise risk assessment of patients affected by BCRL, with implications in prevention strategies, for instance, focusing the resources on identifying patients at higher risk.