Utilizing Artificial Intelligence to Reconcile Indeterminate Lupus Anticoagulant Panel Results

Abstract Lupus anticoagulant (LA) is one laboratory criterion for diagnosing antiphospholipid antibody syndrome (APS), a condition associated with hypercoagulability, pregnancy complications, and death. LA diagnosis requires a complex panel, often performed at a specialized laboratory. Ruling out false positive or transient LAs requires ≥ two positive tests ≥ 12 weeks apart. Most LA tests are negative, with small subsets resulting positive or indeterminate. Indeterminates leave clinicians with limited understanding of appropriate clinical follow-up. Thus, the International Society on Thrombosis and Hemostasis discourages weak resulting nomenclature such as “indeterminate.” Our study used artificial intelligence to predict LA indeterminate results as positive or negative, identifying patients requiring further testing. This IRB approved study reviewed nearly 4000 LA tests performed from 2019-2023. Parameters included CBC, thromboelastography, renal/hepatic function assays, coagulation tests, antiphospholipid antibodies and LA tests. LA test components included screening tests (APTT, Dilute Russel Viper Venom Test (DRVVT) screen), mixing studies (1:1 APTT mix, DRVVT mix), phospholipid-dependent confirmation (DRVVT, hexagonal phase, platelet neutralization), and factor VIII. A positive LA test fulfilled 4 criteria: at least one positive screening test, mixing study and phospholipid-dependent assay and factor VIII inhibitor exclusion. An LA panel meeting 2-3 criteria resulted indeterminate. Studies meeting <2 criteria were negative. This analysis was performed using Machine Intelligence Learning Optimizer (MILO) platform which displays the performance of multiple models. MILO was trained with two classifications: positive and negative cases. Data was split into training (100 training positives; 100 negatives) and testing (primary and secondary validation) sets (100 positives; 1388 negatives). LA panels affected by drug interactions (9.4%) or containing incomplete datasets (21%) were excluded. Sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) determined model efficacy. 2903 negative (84%), 318 positive (9.1%), and 246 indeterminate (7.1%) LA results were included from 3467 patients. Only 84 of the 246 indeterminate patients (34%) were retested. Of these, 28 remained indeterminate, 43 became negative and 13 resulted positive. The best performing MILO models were the K-nearest neighbor (KNN) and logistic regression. The ability of these models to determine which indeterminate cases would become positive or negative was evaluated. KNN displayed the highest specificity (0.88) and PPV (0.57). All models had a high NPV and specificity due to substantive exposure to negative cases for the training and testing. Exclusion of some parameters did not improve the performance of the models. However, even the best model (various KNN models) had low sensitivity and PPV due to a small pool of positive training/testing cases and a much smaller group of indeterminate LA with subsequent positive results. Creating models accurately predicting indeterminate cases that may become positive enables appropriate clinical follow-up. Thus, improving model accuracy will require a larger sample size of positive indeterminates with follow-up.