Abstract Background and aims A Los Angeles Motor Scale (LAMS) score of 4-5 is employed in the pre-hospital setting to screen large vessel occlusion (LVO) with moderate sensitivity. We aim to evaluate the adjunctive role of functional near-infrared spectroscopy (fNIRS) to the LAMS score in classifying patients with LVO using a deep learning architecture. Methods Clinical Characteristics (LAMS score, age(years), Sex, Systolic and Diastolic BP) were abstracted from patients with stroke, Transient ischemic attack, and stroke mimics. All underwent fNIRS assessment of the sensorimotor cortex during resting state (3 minutes) and finger tapping. LVO presence was classified with either LAMS alone or with a Hybrid Deep Neural Network (HDNN). The HDNN processed: a deep, dense network (128-64-32-neuron topology) to extract nonlinear fNIRS features from 28 channels and regularized clinical characteristics. Features were fused via a concatenation layer for final binary classification (LVO vs. non-LVO). Interpretability was assessed using integrated gradient rank channels. Results A total of 143 patients were enrolled, with a median age of 64 (56, 83) years, 43% female, and 29.4% (n = 42/143) were LVO. LAMS score alone classified LVO with an accuracy of 70.6% and a precision of 72.5%. HDNN (fNIRS+Clinical characteristics) classified LVO with an accuracy of 96% and precision of 93%. The highest-ranked predictive features were LAMS score, diastolic BP, and specific fNIRS channels (FC3-FC1, C2-CP2, CP3-C3, C4-FC4, C3-C5). Conclusions This study validates a multimodal deep learning approach for LVO detection. The feasibility of deploying 4-channel fNIRS in a prehospital setting needs assessment. Conflict of interest Mahesh Kate: Nothing to Disclose.
Duba et al. (Fri,) studied this question.