Previous Abstract | Next Abstract
Printable Version
A3166
October 15, 2018
10/15/2018 3:30:00 PM - 10/15/2018 5:30:00 PM
Room North, Hall D, Area C
Validation of A Laboratory Setup for Generating A High Rate of Data Sets as Training Unit of Machine Learning Concerning the Optical Properties of Hemoglobin Under Clinical Conditions: Preliminary Results
Hartmut Gehring, M.D., Philipp Wegerich, M.Sc.
University Medical Center S-H, Campus Lübeck, Lübeck, Germany
Disclosures: H. Gehring: None. P. Wegerich: None.
Introduction: The non-invasive measurement of hemoglobin concentration is based on near-infrared optical methods. For the development of such sensors, knowledge of the optical properties of the main absorber hemoglobin - adapted to the clinical conditions - is essential and must be recorded according to the appropriate standard. A particular focus here is on the layer thickness to be penetrated. First, the knowledge of this layer thickness is an essential prerequisite for the application of the measuring principle - at least in the basics according to Lambert & Beer. Secondly, the penetration of the layer thickness with light is critically limited by the strong absorption of hemoglobin. Third, the layer thickness should also reflect a dimension, as can be found again in human tissue. The primary standard method for detecting these properties is based on the implementation of two integrating spheres. To generate a high rate of data sets, as they are necessary for the training of mathematical models, however, a setup is required, which allows a flexible and rapid implementation in the data generation. The primary goal of this in vitro laboratory investigation was to measure transmission and reflection under the preselected conditions. The second goal was to validate the data in terms of quality. The third goal was to create such a high rate of data sets that a training unit for machine learning techniques could be set up.

Methods: The test based on a transportable integrating sphere setup for standardised measurements on blood [1]. The optical system comprises laser diodes of the wavelengths 780 nm, 808 nm, 850 nm, 980 nm and 1310 nm. Data of transmission and reflection were continuously registered during flow through optical cuvettes with a layer thickness of 3.0 mm, 2.0 mm, and 1.0 mm. In order to control the optical setup based on a discontinuously measuring laser diode system, a spectrophotometer (Epp2000-UVN-100 (StellarNet Inc, USA) with the record of light intensity in the range of 690 nm to 1000 nm serves as first internal standard. To draw conclusions about the plausibility of the measured data, the Inverse Adding Doubling Algorithm [1] serves as a second internal standard. This determines the absorption coefficients (μa) as well as the reduced scattering coefficients (μs') of the measurement samples independent of the layer thickness. As reference for tHb in g/dL serves the mean data of 3 Hc 201+ (HemoCue, Sweden) devices with repeated measurements.

Results: Overall a number 396 of data sets were generated within this in vitro laboratory setup. Tab. 1 integrated basic information about data sampling. Fig. 1 demonstrates mean data of transmission on the left site distributed to the layer thicknesses. Right is an example of the spectrophotometric internal standard for a data set on layer thickness of 1.0 mm. The spectrum in counts/ms results in a light intensity comes to the sensor.

Conclusions: Based on this in vitro laboratory setup a high number rate of data sets can be provided for further training units in artificial intelligence methods in layer independent prediction of tHb.

[1] Wegerich P, et al. Current Directions in Biomedical Engineering 2017 DOI: 10.1515/cdbme-2017-0061

Figure 1
Figure 2

Copyright © 2018 American Society of Anesthesiologists