In untargeted lipidomics experiments, putative lipid identifications generated by automated analysis software require substantial manual filtering to arrive at usable high-confidence data. However, identification software tools do not make full use of the available data to assess the quality of lipid identifications. Here, we present a machine-learning-based model to provide coherent, holistic quality scores based on multiple lines of evidence. Underutilized metrics such as isotope ratios and chromatographic behavior allow for much higher accuracy of identification confidence. We find that approximately 50% of tandem mass spectrometry-based automated lipid identifications are incorrect but that multidimensional rescoring reduces false discoveries to only 7% while retaining 80% of true positives. Our method works with most chromatography methods and generalizes across a family of MS instruments. LipoCLEAN is available at https://github.com/stavis1/LipoCLEAN.