Background: RSClin™ combines clinicopathological (CP) and genomic data to refine predictions of distant recurrence (DR) risk and chemotherapy (CT) benefit in node-negative (N0) hormone receptor (HR)-positive/human epidermal growth factor receptor 2 (HER2)-negative early breast cancer (eBC). Available in the United States, RSClin™ is inaccessible in Europe due to regulatory constraints. We developed and validated RSC4All, a machine learning (ML) nomogram replicating RSClin™ predictions and evaluating clinical impact by comparing DR risk and CT benefit estimates across recurrence score (RS), RSClin™, and RSC4All. Materials and methods: We retrospectively identified 290 HR-positive/HER2-negative eBC with both 21-gene RS and RSClin™ estimation at three European centers (development cohort). ML models were trained on 70% and internally validated on 30% of this cohort to reproduce RSClin™ estimates for DR risk and CT benefit. External validation used an independent cohort of 513 patients. Performance was assessed using receiver operating characteristic area under the curve (ROC AUC) (classification) and R2 (regression). Clinical agreement between RSC4All and RSClin™ was evaluated using Cohen's kappa and McNemar's test, including replication of RS to RSClin™ reclassification patterns. RS and RSClin™ classifications were compared in the pooled population and CP predictors of reclassification were explored. Results: ML classification models achieved excellent performance with ROC AUC 0.99 for both DR risk and CT benefit, while regression model had R2 = 0.82 and 0.77, respectively. In external validation, RSC4All maintained high accuracy (ROC AUC = 0.99; R2 = 0.82 and 0.76) with an almost perfect clinical agreement with RSClin™ (Cohen's kappa = 0.87 and 0.84) and replicated RS to RSClin™ reclassification direction in 94.1% (DR) and 93.0% (CT) of cases. In the pooled cohort, RSClin™ reclassified 22.2% of DR risk (balanced up/downgrades) and 12.2% for CT benefit (mostly upgrades), with high grade, high Ki-67, low progesterone receptor, and larger tumor size predicting upward reclassification. A web-based tool was developed for public access (https://rsc4all.streamlit.app). Conclusions: RSC4All accurately reproduces RSClin™ predictions and reclassification patterns, providing a freely accessible alternative where RSClin™ is unavailable. By integrating CP and genomic data, it refines risk stratifications and supports adjuvant decisions in N0 HR-positive/HER2-negative eBC. External validation in an independent cohort strengthens its generalizability and clinical applicability.
RSC4All as a machine learning nomogram to predict RSClin™ results in HR+/HER2- node-negative early breast cancer
Martorana, F.;Vigneri, P.;
2025-01-01
Abstract
Background: RSClin™ combines clinicopathological (CP) and genomic data to refine predictions of distant recurrence (DR) risk and chemotherapy (CT) benefit in node-negative (N0) hormone receptor (HR)-positive/human epidermal growth factor receptor 2 (HER2)-negative early breast cancer (eBC). Available in the United States, RSClin™ is inaccessible in Europe due to regulatory constraints. We developed and validated RSC4All, a machine learning (ML) nomogram replicating RSClin™ predictions and evaluating clinical impact by comparing DR risk and CT benefit estimates across recurrence score (RS), RSClin™, and RSC4All. Materials and methods: We retrospectively identified 290 HR-positive/HER2-negative eBC with both 21-gene RS and RSClin™ estimation at three European centers (development cohort). ML models were trained on 70% and internally validated on 30% of this cohort to reproduce RSClin™ estimates for DR risk and CT benefit. External validation used an independent cohort of 513 patients. Performance was assessed using receiver operating characteristic area under the curve (ROC AUC) (classification) and R2 (regression). Clinical agreement between RSC4All and RSClin™ was evaluated using Cohen's kappa and McNemar's test, including replication of RS to RSClin™ reclassification patterns. RS and RSClin™ classifications were compared in the pooled population and CP predictors of reclassification were explored. Results: ML classification models achieved excellent performance with ROC AUC 0.99 for both DR risk and CT benefit, while regression model had R2 = 0.82 and 0.77, respectively. In external validation, RSC4All maintained high accuracy (ROC AUC = 0.99; R2 = 0.82 and 0.76) with an almost perfect clinical agreement with RSClin™ (Cohen's kappa = 0.87 and 0.84) and replicated RS to RSClin™ reclassification direction in 94.1% (DR) and 93.0% (CT) of cases. In the pooled cohort, RSClin™ reclassified 22.2% of DR risk (balanced up/downgrades) and 12.2% for CT benefit (mostly upgrades), with high grade, high Ki-67, low progesterone receptor, and larger tumor size predicting upward reclassification. A web-based tool was developed for public access (https://rsc4all.streamlit.app). Conclusions: RSC4All accurately reproduces RSClin™ predictions and reclassification patterns, providing a freely accessible alternative where RSClin™ is unavailable. By integrating CP and genomic data, it refines risk stratifications and supports adjuvant decisions in N0 HR-positive/HER2-negative eBC. External validation in an independent cohort strengthens its generalizability and clinical applicability.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
