Enhancing Lupus Anticoagulant Testing with Machine Learning: Deep Neural Networks Match Expert Performance without the Need for Feature Selection.

BACKGROUND

The lupus anticoagulant (LAC) is an important laboratory criterion in the diagnosis of antiphospholipid antibody syndrome. LAC testing at the Mayo Clinic Special Coagulation Laboratory includes up to 13 tests, which are interpreted by trained physicians to identify the presence of LAC and rule out anticoagulant interferences. This feasibility study explored the use of two deep neural network (DNN) architectures for multilabel classification of LAC profiles as a first step toward automating interpretation.

METHODS

Seven thousand two hundred and two retrospective cases were randomly split (64:16:20) for training, validation, and test, respectively. LAC positivity by dilute Russell's viper venom time (LAC-DRVVT) and activated partial thromboplastin time (LAC-APTT) and the presence of warfarin (WAR) and heparin (HEP) were adjudicated by one expert. DNN architectures included: single-output DNNs using domain-knowledge input feature selection and a single-column multioutput DNN using all 13 inputs.

RESULTS

Domain-knowledge-naïve multioutput DNN achieved similar or improved performance for all 4 label prediction tasks compared with domain knowledge optimized DNNs: F1 scores of 0.977 vs 0.968 for LAC-DRVVT, 0.954 vs 0.945 for LAC-APTT, 0.961 vs 0.957 for HEP, and 0.995 vs 0.977 for WAR, respectively.

CONCLUSIONS

The comparable performance of the 4 domain knowledge optimized DNNs and the multioutput DNN using all 13 input features suggests that the DNN may learn feature importance or mapping to a task without explicit input selection. Given its relative simplicity and versatility, the multioutput DNN is the preferred choice for implementation in a clinical laboratory to standardize LAC diagnosis.