Logistic regression analysis for identifying the factors affecting development of non-invasive blood glucose calibration model by near-infrared spectroscopy

• Individual PLS models were successfully developed for 40% of diabetic patients.
• The disturbance factors for model development are statistically investigated.
• Medical examination data are analyzed by using logistic regression.
• Physical measuring conditions were likely to be more critical than blood elements.

This study was conducted to develop a statistical method for investigating the factors that hamper the development of individual non-invasive blood glucose calibration model using near-infrared (NIR) spectroscopy. First, the individual calibration models for the patients with diabetes mellitus were developed by applying partial least squares regression to relate data of NIR spectra measured at the hand with blood glucose concentrations. Second, items obtained via medical examinations were analyzed using logistic regression in order to specify the factors that hamper the development of successful models. Consequently, the individual calibration models for approximately 40% of patients were successfully developed with a mean standard error of cross-validation (SECV) of 25.0 mg/dL. For more than 60% of these patients, over 80% of the validation samples fell within zone A of the Clarke error grid representing clinically acceptable accuracy. According to the results of logistic regression analyses, body mass index (BMI) which may affect the variation in physical measurement conditions was considered to be an effective factor for developing successful calibration models besides the change in blood constituents. A statistical approach using logistic regression analysis represents a novel efficient method for investigating the factors contributing to develop successful individual calibration models.