Machine learning enables clinically feasible implementation of clot waveform analysis for differentiating causes of prolonged activated partial thromboplastin time

Activated partial thromboplastin time (APTT) is widely used as a coagulation screening test; nonetheless, its prolongation can result from diverse causes, complicating the diagnosis. Clot waveform analysis (CWA) has been proposed as a complementary method, although the conventional parameters have only moderate diagnostic value. Previously, we applied deep learning (DL) to multiwavelength CWA and achieved high classification accuracy; nevertheless, DL approaches are complex and challenging to interpret. In this follow-up study, we investigated whether machine learning (ML) applied to single-wavelength CWA could yield simpler, more interpretable models for clinical use. Overall, 683 prolonged APTT samples were classified into five groups: heparin ( n = 99), direct oral anticoagulants ( n = 249), warfarin ( n = 105), lupus anticoagulant ( n = 95), and factor VIII/IX deficiencies or inhibitors (FVIIIdef/FVIIIi/FIXdef, n = 135). Using 22 waveform-derived parameters at 660 nm, ML achieved a sensitivity and specificity of 82.8–99.0% and > 95%, respectively. Importantly, these parameters were associated with characteristic waveform patterns observed in different causes of prolonged APTT and may reflect differences in coagulation function. Additionally, validation using an independent set of 53 patient samples demonstrated high sensitivity and specificity. Therefore, ML using single-wavelength CWA achieves accuracy comparable to that of multiwavelength DL while enhancing interpretability and facilitating implementation in automated coagulation analyzers for broad clinical applicability.