A novel multi-feature fusion technology of FTIR spectroscopy based on attention and adaptive gate for disease...

Fourier transform infrared (FTIR) spectroscopy has been extensively applied in the auxiliary diagnosis of diseases. However, in complex clinical environments, the information reflected by a single spectral feature is limited. Chaos theory focuses on studying complex behaviors in nonlinear dynamical systems, while statistics is dedicated to collecting, analyzing, interpreting, displaying, and organizing data. This study extracts chaotic features and statistical domain features from the time-domain form of FTIR spectroscopy and integrates them with the original FTIR spectroscopy through feature fusion model Cross Features Attention Multi-feature Adaptation Gate (CFAMAG), aiming to achieve better performance in disease auxiliary diagnosis. The CFAMAG addresses the issues of information misalignment, heterogeneity, and complementarity among different features by computing cross-feature attention between chaotic features, statistical domain features, and FTIR spectral features. Additionally, the design of gating vectors and displacement vectors allows CFAMAG to adjust the influence of non-FTIR spectral features on the feature space, facilitating better integration of chaotic features and statistical domain features. Experimental results demonstrate the outstanding and stable predictive performance of the CFAMAG model across four binary classification experiments and one multi-class classification experiment. In the experiments, the CFAMAG model shows an average improvement of 6.48 % in Accuracy (Acc), 7.73 % in Precision (Pre), 4.87 % in Sensitivity (Sen), 11.10 % in Specificity (Spe), 8.37 % in F1 score (F1), and 7.08 % in Area Under Curve (AUC). The proposed method enhances the accuracy of disease auxiliary diagnosis and holds significant reference value for disease auxiliary diagnosis.