An In vitro Laboratory Investigation on Layer Thickness-Independent Prediction of the Hemoglobin Concentration

Abstract

In the noninvasive determination of the hemoglobin concentration a main challenge is the "optical path". With sensors - fixed on human skin - the optical path cannot be exactly determined, as it is defined as the layer thickness in the Lambert Beer principle. The layer thickness is significantly involved in the optical interactions in the tissue. To circumvent this problem self-learning algorithms were evaluated which provide the hemoglobin concentration from reflection and transmission data without knowledge of the layer thickness. First various regression models were trained based on an high rate data set. To evaluate the six most promising models, a prediction dataset was measured in a prospective randomized and blinded study to guarantee integrity of the results. For both data sets, the transmission and reflection of diluted heparinized erythrocyte concentrate was determined with a double integrating sphere setup (laser diodes with 780 to 1310 nm). The evaluated hemoglobin concentrations ranged from 4 to 16 g/dl at a constant oxygen saturation above 97 %. Optical flow through cuvettes (1, 2, 3 mm) simulated different layer thicknesses of the blood. The evaluation of the predictions yielded that the layer thickness independent prediction of the hemoglobin concentration is feasible with the selected approaches. The mean absolute error (MAE) of the best regression model (GPRM - Matern 5/2) is 0.79 g/dl. In the clinically relevant tHb range of less than 8 g/dl the MAE was as low as 0.52 g/dl.

Current Status of Measurement Accuracy for Total Hemoglobin Concentration in the Clinical Context

“…The data and results presented in this investigation implicate two more valuable characteristics with a potential input on decision making: First, they may be embedded in a system with artificial learning for decision making [ 62 ]. Second, the data served as reference values with respect to the training of machine learning systems, i.e., regression models [ 63 , 64 ].…”