Deep learning model for screening causes of activated partial thromboplastin time prolongation using clot waveform analysis at multiple wavelengths.

Activated partial thromboplastin time (APTT) prolongation occurs due to coagulation factor deficiencies/inhibitors, lupus anticoagulant (LA), and anticoagulant-taking, necessitating discrimination through further testing. Clot waveform analysis (CWA) can discriminate causes while measuring APTT, but conventional CWA exhibits moderate accuracy due to visual judgement and limited parameter use. We applied deep learning (DL) techniques to huge numerical data constituting clot waveforms and their first- and second-derivative curves (CWA curves) to leverage hidden features for developing an accurate classification model. We utilized a multi-wavelength detection system embedded in modern coagulometers to obtain multi-wavelength CWA curves. A convolutional neural network-based DL model was trained on 683 samples (135 hemophilic, 95 LA-positive, 99 heparin-treated, 105 warfarin-treated, and 249 direct oral anticoagulant-treated) and evaluated using 10-fold cross-validation. Conventional CWA parameters showed limited discrimination abilities (area under the curve [AUC] 0.532-0.858). DL models using single-wavelength CWA curves achieved higher performance (AUC 0.943-0.988), and multi-wavelength CWA curves further improved it (AUC 0.961-0.993) with high sensitivity (≥ 88.0%), specificity (> 92.0%), and overall accuracy (88.4%), although the performance may depend on reagents and/or analyzers. DL models using multi-wavelength CWA curves show promise as high-performance screening tools for classifying APTT prolongation causes and are best built in each laboratory.