Abstract Introduction The laboratory diagnostic criteria for plasma cell neoplasms such as multiple myeloma include the presence of a monoclonal paraprotein which may be identified by serum protein electrophoresis (SPEP). To date, there are no high-sensitivity screening methods for automated detection of serum paraproteins using machine learning. Methods Cases from a plasma cell neoplasm registry (n=501) were manually labeled for paraprotein status. Each case included categorical features for 31 laboratory results commonly reported in myeloma. We applied fast Fourier transform to the gamma region of the SPEP curve and constructed features corresponding to cumulative magnitudes in the Fourier domain. Data were split into a 75% training set and a 25% test set. Using all feature domains independently, as well as combinations of select feature domains, we then trained three models: logistic regression, support vector machine, and random forest. Via 5-fold cross-validation, we tuned all models to a 95% sensitivity target and evaluated them by AUC, specificity, positive predictive value (PPV), F1 score, accuracy and likelihood ratio. We evaluated performance in plasma cell neoplasm and clinically heterogeneous test sets at low, medium, and high paraprotein concentrations, performing error analysis on each set. Results The logistic regression model using only Fourier domain features performed with the highest AUC (0.944). This model achieved sensitivity 0.962, PPV 0.667, specificity 0.569, F1 score 0.787, and accuracy 0.755. The model with the highest likelihood ratio while maintaining sensitivity over 0.90 achieves AUC 0.944, sensitivity 0.904, PPV 0.860, specificity 0.862, F1 score 0.899, and accuracy 0.900. Stratification of sensitivity by paraprotein concentration showed the model achieves sensitivity 0.88 in samples with low paraprotein concentration (0.1 to 0.5 g/dL), and 1.0 in samples with medium (0.5 to 3.0 g/dL) and high (over 3.0 g/dL) concentrations. In the separate clinically heterogeneous test set, the overall sensitivity was 0.880 (0.870 for low, 0.885 for medium and 0.909 for high concentration samples). Error analysis on the misclassified samples with low paraprotein status showed that the algorithm may miss cases with low paraprotein concentration, as well as cases with paraproteins outside the gamma region. Discussion and Conclusion Our machine learning algorithm detects paraproteins with 94% sensitivity at over 0.85 specificity, PPV, and accuracy using the Fourier transformation of the SPEP gamma region. Future work to achieve higher sensitivity at similar specificity would allow clinical laboratories to use this as a screening method for paraprotein rule-out and drive efficiencies in laboratory workflows. Although we did not demonstrate it in this study, we believe that extending our model to include quantitative laboratory results that frequently accompany SPEP holds promise for increasing sensitivity within the scope of physicians standard clinical laboratory ordering practice. Additional strategies include deep learning models like convolutional and recurrent neural networks.