Predictive models in near-infrared spectroscopic analysis for blood hemoglobin prediction

• Main Author: Mohd Idrus, Mohd Nazrul Effendy
• Format: Thesis
• Language: English; English; English
• Published: 2018
• Subjects: QD241-441 Organic chemistry
• Online Access: http://eprints.uthm.edu.my/430/1/24p%20MOHD%20NAZRUL%20EFFENDY%20MOHD%20IDRUS.pdf; http://eprints.uthm.edu.my/430/2/MOHD%20NAZRUL%20EFFENDY%20MOHD%20IDRUS%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/430/3/MOHD%20NAZRUL%20EFFENDY%20MOHD%20IDRUS%20WATERMARK.pdf; http://eprints.uthm.edu.my/430/

Determination of blood hemoglobin concentration is important to diagnose anaemia. Common clinical practice used to measure blood hemoglobin is by drawn some blood from the patient to be mixed with reagent chemicals for analysis. Alternatively, near-infrared spectroscopy (NIRS) technology can be used to measure blood hemoglobin level. NIRS is based on molecular overtone and combination vibrations produce absorption bands typically very broad, leading to complex spectral data makes the NIRS useful for analysis. Different types of predictive models such as linear, nonlinear, and hybrid predictive models were commonly used to predict component of interest from NIRS spectral data. However, different predictive model approached may achieve different accuracy of performance in predicting component of interest from NIRS spectral data. The aims of this study is to investigate the accuracy of linear partial least square (PLS), nonlinear artificial neural network (ANN), and hybrid partial least square - artificial neural network (PLS-ANN) predictive modelling in NIRS analysis. These predictive models were coupled with Savitzky-Golay (SG) preprocessing to remove unwanted signal from spectral data. The optimal numbers of frame length, latent variables, and hidden neurons used in SG preprocessing and predictive models were investigated. Results show ANN coupled with first order SG derivatives achieved the best prediction of performance with root mean square error of prediction (RMSEP) and the coefficient of determination of prediction (