Quantitative detection of serum biochemical indexes via UV-Vis-NIRS combined with deep neural networks

To achieve rapid, cost-efficient, convenient and accurate detection of five clinical serum biochemical indexes, namely glucose (GLU), triglycerides (TG), total cholesterol (TC), total protein (TP) and albumin (ALB), ultraviolet-visible-near infrared spectroscopy (UV-Vis-NIRS) technology combined with deep neural network (DNN) is firstly proposed in this study. The absorption spectra of 992 human serum are collected in 200-2500 nm. Different spectra preprocessing methods are studied and compared to eliminate interference, baseline offset, and highlight specificity information of biochemical indexes in the raw spectra. Moreover, the competitive adaptive reweighted sampling (CARS) algorithm is utilized to optimally select characteristic wavelengths related to biochemical indexes. A DNN, i.e., 1DCNN-LSTM model is established to quantitatively predict five biochemical indexes using stratified sampling with the training set and testing set divided in 7:3. Compared with the traditional machine learning (ML) and artificial neural network (ANN) algorithms, the results show that the quantitative prediction performances of 1DCNN-LSTM model are significant superior. Root mean square error of prediction (RMSEP) and determination coefficient (R²) of GLU, TG, TC, TP and ALB are 0.39 mmol/L, 0.36 mmol/L, 0.31 mmol/L, 1.26 g/L and 1.28 g/L, 0.97, 0.90, 0.93, 0.96 and 0.93, respectively. Finally, the advantage of UV-Vis-NIRS are verified by comparing with NIRS and UV-Vis alone. Results show that UV-Vis-NIRS combined with DNN can provide new idea and strong technical support in the clinical application of serum biochemical indexes detection.