Background: Neoadjuvant therapy (NAT) is critical in the therapeutic strategy for locally advanced breast cancer (BC) patients. Biomarkers to predict the pathologic complete response (pCR) and to screen beneficial responders need to be urgently explored. Lipidomics is a high-throughput analysis technique, which has potential applications for peripheral biomarkers’ detection to overcome the difficulty in serial pathological biopsies and tumor heterogeneity. Machine learning is a method for mining high-dimensional information with high group and low intermolecular correlation. Here we construct multidimensional models with machine learning approach combining clinical clinicopathological and lipidomics data to predict NAT response. Methods: Plasma samples were collected from 119 BC patients before after two cycles of NAT. Peripheral lipidomic profiles at multiple levels of lipid composition, concentration, chain length, and saturation were dynamically monitored by using absolute quantitative lipidomics. Responders (pCR) and non-responders (non-pCR) groups were randomly sampled in a 1:1.5 ratio for subsequent analysis. Screening of candidate lipidomic biomarkers were performed according to the criteria “VIP>1, FC>2 or FC<0.5 and p<0.05”. Above feature screening were validated by using ensemble machine learning algorithms (Logistic regression, Random forest and Support vector machine). Pearson correlation coefficients were calculated for the expressions of candidate biomarkers. Multidimensional prediction models with the logistic regression algorithm were constructed by correlating candidate lipidomics features with clinicopathologic phenotypes prospectively collected with baseline and dynamic changes. The performance of the models constructed above were assessed by ROC analysis. Results: A total of 8 major classes, 39 subclasses, and 2292 molecules of lipid metabolites were identified in the 235 plasma samples. Most of the candidate biomarkers had lower correlations, indicating lower overlapping and more optimal combination of panels. The prediction models constructed by baseline correlating candidate lipidomics features with ensemble machine learning achieved an Area under the Curve (AUC) of 0.84 for the training set and 0.72 for the validation set with the accuracy 0.84, the specificity 0.92, and the sensitivity 0.71 when the cutoff value is 0.57.The prediction models constructed with lipidomics dynamic changes also achieved good performance. Conclusions: The study suggested a possibility that peripheral lipidomics provided a potential tool to develop multidimensional models with ensemble machine learning for predicting response to NAT for BC patients with good discrimination power, which might guide individual optimal NAT strategies and avoid unnecessary treatment.