Background/objectivesMicroRNAs (miRNAs) are involved in chemosensitivity through their biological activities in various malignancies, including pancreatic cancer (PC). However, single-miRNA models offer limited predictability of treatment response. We investigated whether a multiple-miRNA prediction model optimized via machine learning could improve treatment response prediction. MethodsA total of 20 and 66 patients who underwent curative resection for PC after gemcitabine-based preoperative treatment were included in the discovery and validation cohorts, respectively. Patients were classified according to their response to preoperative treatment. In the discovery cohort, miRNA microarray and machine learning were used to identify candidate miRNAs (in peripheral plasma exosomes obtained before treatment) associated with treatment response. In the validation cohort, miRNA expression was analyzed using quantitative reverse transcription polymerase chain reaction to validate its ability to predict treatment response. ResultsIn the discovery cohort, six and three miRNAs were associated with good and poor responders, respectively. The combination of these miRNAs significantly improved predictive accuracy compared with using each single miRNA, with area under the curve (AUC) values increasing from 0.485 to 0.672 to 0.909 for good responders and from 0.475 to 0.606 to 0.788 for poor responders. In the validation cohort, improved predictive performance of the miRNA combination over single-miRNA prediction models was confirmed, with AUC values increasing from 0.461 to 0.669 to 0.777 for good responders and from 0.501 to 0.556 to 0.685 for poor responders. ConclusionsPeripheral blood miRNA profiles using an optimized combination of miRNAs may provide a more advanced prediction model for preoperative treatment response in PC.