Abstract Background Prediction of relapse in hormone receptor-positive patients treated with adjuvant hormonal therapy is an area of active research. Several tissue based genomic tests have been developed and are used in clinical practice [e.g. OncotypeDX (Genomic Health), MammaPrint (Agendia)] to evaluate the risk of recurrence in early stage breast cancer. However, a cost–effective serum based test, that would allow identification of patients at high risk of early relapse, is of clinical interest. Methods We used MALDI ToF Mass spectrometry to obtain mass spectra from pre-surgery serum samples from 499 patients treated with adjuvant hormonal therapy. Spectra were subjected to pre-processing and 84 peaks (features) were selected for the analysis. We used a novel proprietary approach, utilizing recent advances in learning theory, to create a diagnostic test to classify patients as Early Relapse or No Early Relapse. The method creates many multivariate classifiers that are filtered for performance and combined using logistic regression with dropout regularization into a single master classifier. To avoid bias introduced through a particular split of training and test, many realizations of the development set are created. The performance of each master classifier is examined and the classifiers are combined using a majority vote procedure to serve as the final test. The method allows the use of smaller training sets and minimizes overfitting. Results 22 out of 499 patients had an early relapse in < 5 years. Samples from these patients were matched, mindful of treatment and HER2 status, with 22 samples from patients without relapse with the longest duration of relapse free survival (RFS), to serve as the development set. The remaining samples, including those from patients with late relapses (> 5 years), were set aside for additional testing. The performance of the 200 master classifiers created from the 200 test/training splits of the development set was evaluated showing a median overall accuracy of 70%, specificity of 73%, sensitivity of 67%, and hazard ratio (HR) of 3.5. The final classifier was created using the 200 master classifiers from the realizations of Training and Test splits which were combined using the majority vote procedure. Classification of patients in the development set from the majority vote of master classifiers resulted in a significant separation in survival curves (log-rank p<0.0001, HR 6.2 Median RFS Early Relapse 2.6 years, No Early Relapse not reached) with overall accuracy of 79%, specificity of 82%, and sensitivity of 77%. In the combined population of all patients, the separation was also significant (log-rank p=0.028, HR 2.0, median RFS not reached in both classifications). In the multivariate analysis of the overall population, classification remained significant (p=0.007, adjusted HR 3.5) along with menopausal status, nodal status, and tumor size. Conclusions We created a classifier from pre-surgery serum samples that can identify patients at risk of early relapse (<5 years) when treated with adjuvant hormonal therapies. Such a test would have clinical utility in identifying patients who may need a revised adjuvant treatment strategy. Citation Format: Heidi Fiegl, Christian Marth, Krista Meyer, Julia Grigorieva, Heinrich Roder. Serum–based test to identify patients with early relapse treated with adjuvant hormonal therapy [abstract]. In: Proceedings of the Thirty-Seventh Annual CTRC-AACR San Antonio Breast Cancer Symposium: 2014 Dec 9-13; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2015;75(9 Suppl):Abstract nr P6-08-23.