Logistic Regression Outperforms Random Forest for Predicting Malignancy in BI-RADS 4 Breast Lesions

In 216 BI-RADS 4 breast lesions, a logistic regression model with 6 imaging/elastography features best predicted malignancy — random forest appeared superior (AUC 1.00) but showed clear overfitting. LR model AUC in validation not reported in source. Single-center, n=212.

• Retrospective, single-center study; 212 patients (216 lesions) split 70/30 into training (n=151) and validation (n=65) sets; 8 ML algorithms compared via LASSO/logistic regression feature selection.
• Emax-2 shell stiffness cutoff of 104.71 kPa showed the highest diagnostic efficacy for the BI-RADS 4a subgroup; exact AUC for the final LR model in the validation cohort not reported in the source abstract.
• Key limitation: small single-center sample with a 70/30 random split — no external validation cohort — limiting generalizability; the random forest AUC of 1.00 on training data signals overfitting rather than true performance.

RadPigeon summaries are original and for information only. They are not clinical advice.