From raw clinical data to robust prediction: an AI framework for early lymphedema detection

Breast cancer related lymphedema (BCRL) is a frequent postoperative complication that can substantially impair physical function and quality of life. While objective detection tools such as bioimpedance spectroscopy and imaging are accurate, they are resource intensive and difficult to deploy at scale in routine clinical workflows. Early identification of patients at elevated risk using routinely available clinical data may enable timely surveillance and intervention. We developed a comparative machine learning framework for postoperative BCRL risk prediction using routinely collected clinical data from 1,328 breast cancer patients treated at a single tertiary care center. Twenty four demographic, clinical, pathological, and treatment related variables were analyzed in tabular form. Nine classifiers spanning linear, ensemble, boosting, and transformer based models were systematically evaluated across seven preprocessing pipelines addressing missing data, feature scaling, class imbalance, and decision threshold optimization. Model performance was assessed using minority class F1 (F1_m), accuracy, and the Brier score to jointly evaluate discrimination and probabilistic calibration. Predictive performance varied substantially across preprocessing strategies. When emphasis was placed on minority-class discrimination and probabilistic calibration, ensemble-based models including Random Forest, CatBoost, and LightGBM, as well as the transformer-based TabPFN, achieved the strongest results under the task-specific Custom preprocessing pipeline. Among these, TabPFN attained the highest overall performance, reaching F1_m values of approximately 0.86 and accuracy around 0.89 on the held-out test set. Random Forest, CatBoost, and LightGBM demonstrated robust and competitive performance, with F1_m values in the range of approximately 0.78 to 0.80 and accuracies around 0.83 to 0.86, together with stable Brier scores. In contrast, preprocessing pipelines that avoided targeted data exclusion yielded more conservative but consistent performance, highlighting trade-offs between discrimination, calibration, and data retention in real-world clinical settings. This study demonstrates that preprocessing choices play a critical role in BCRL risk modeling using real world clinical data. While task specific preprocessing can substantially improve minority class performance, such gains should be interpreted cautiously in light of data exclusion and potential selection bias. The proposed framework provides a transparent and reproducible foundation for retrospective BCRL risk stratification and offers a methodological template for future multi center validation studies in postoperative breast cancer care.