Despite the growing popularity of liquid chromatography-mass spectrometry (LC-MS)-based metabolomics, no study has yet to systematically compare the performance of different data acquisition modes in the discovery of significantly altered metabolic features, which is an important task of untargeted metabolomics for identifying clinical biomarkers and elucidating disease mechanism in comparative samples. In this work, we performed a comprehensive comparison of three most commonly used data acquisition modes, including full-scan, data-dependent acquisition (DDA), and data-independent acquisition (DIA), using a metabolomics study of human plasma samples from leukemia patients before and after one-month chemotherapy. After optimization of data processing parameters, we extracted and compared statistically significant metabolic features from the results of each data acquisition mode. We found that most significant features can be consistently found in all three data acquisition modes with similar statistical performance as evaluated by Pearson correlation and receiver operating characteristic (ROC) analysis. Upon comparison, DDA mode consistently generated fewer uniquely found significant features than full-scan and DIA modes. We then manually inspected over 2000 uniquely discovered significant features in each data acquisition mode and showed that these features can be generally categorized into four major types. Many significant features were missed in DDA mode, primarily due to its low capability of detecting or extracting these features from raw LC-MS data. We thus proposed a bioinformatic solution to rescue these missing significant features from the raw DDA data with good reproducibility and accuracy. Overall, our work asserts that data acquisition modes can influence metabolomics results, suggesting room for improvement of data acquisition modes for untargeted metabolomics.