Non-invasive differentiation of light chain amyloidosis and multiple myeloma based on Raman spectroscopy analysis using one-dimensional convolutional neural networks.

Light chain amyloidosis (AL) and multiple myeloma (MM) are interrelated plasma cell disorders characterized by malignant proliferation, yet they demonstrate distinct pathophysiological mechanisms and clinical progression patterns. The clinical differentiation between these conditions presents significant challenges, frequently resulting in delayed diagnosis, particularly for AL amyloidosis, which adversely affects patient prognosis. Current diagnostic methodologies predominantly depend on invasive tissue biopsies and extended serological testing, underscoring the urgent requirement for rapid, non-invasive auxiliary diagnostic approaches. In this investigation, we developed an innovative analytical framework integrating serum Raman spectroscopy with an advanced one-dimensional convolutional neural network (1D-CNN) to achieve precise discrimination between AL and MM. Serum specimens were collected from clinically diagnosed patients and analyzed using a 785 nm excitation Raman system spanning the spectral range of 200-2000 cm-1. Following comprehensive preprocessing procedures, our specially designed 1D-CNN architecture attained exceptional classification performance, demonstrating area under the curve (AUC) values of 0.94 for AL and 0.96 for MM, with overall accuracy reaching 92.5%, accompanied by 91.4% sensitivity and 93.1% specificity. The proposed model exhibited statistically superior performance (p < 0.01) compared to conventional machine learning algorithms, including support vector machines (AUC = 0.78), and other deep learning architectures. Critical spectral analysis identified prominent Raman band variations at 500 cm-1, 1150 cm-1, and 1750 cm-1, providing molecular-level insights into the discriminatory characteristics. This spectroscopy-based deep learning platform represents a substantial advancement in clinical diagnostics, offering a rapid, non-invasive methodology with significant potential for early disease screening and enhanced decision-support in differential diagnosis of plasma cell dyscrasias.