SPECTROSCOPY DEVICE AND METHOD FOR DETERMINING BLOOD PARAMETERS IN NON-HEMOLYSIZED BLOOD SAMPLES

Description

[0001] The invention relates to a spectroscopy device and a method for determining blood parameters in non-hemolyzed blood samples.

[0002] Hemoglobin is an extremely important substance in the human body, contained in high concentrations in red blood cells, and responsible for transporting oxygen from the lungs to the cells and organs. Accurate knowledge of the total hemoglobin concentration (ctHb) is crucial in clinical settings, as a significant decrease can lead to oxygen deprivation (anemia) with pronounced symptoms ranging from shortness of breath and disorientation to death. Elevated levels can also occur in various diseases. However, not only the absolute amount of hemoglobin is important, but also its functionality. In a healthy person, approximately 98% of the hemoglobin in arterial blood is bound to oxygen; in this state, it is called oxyhemoglobin (O₂Hb). Without bound oxygen, it is called reduced hemoglobin or deoxyhemoglobin (HHb).Pathologically altered, dysfunctional forms of hemoglobin do not participate in oxygen transport and can therefore lead to a life-threatening situation within a very short time. They are called carboxyhemoglobin (COHb), methemoglobin (MetHb), and sulfhemoglobin (SHb). The cause can be carbon monoxide (CO) poisoning, but also poisoning by chemicals (nitrates, nitrites, nitrogen oxides, hydrogen peroxide, sulfur compounds), medications (sulfonamides, nitroglycerin, nitric oxide, and local anesthetics such as benzocaine and prilocaine), as well as hereditary, genetic factors.

[0003] The determination of ctHb and the fractional amounts of O₂Hb, HHb, COHb, MetHb, and SHb is usually performed spectroscopically using CO-oximeters. In whole blood, hemoglobin is concentrated in the red blood cells. The surrounding blood plasma has a lower refractive index, resulting in pronounced scattering effects at the interface. For this reason, the red blood cells are typically first destroyed by ultrasound or added reagents, so that the hemoglobin they contain is evenly distributed in the sample. This significantly simplifies optical analysis and allows the determination of ctHb and the fractional amounts of hemoglobin derivatives based on absorption at different wavelengths according to the Bouguer-Lambert-Beer law.This describes the relationship between the concentration c of a substance and the spectroscopically measurable attenuation of the transmitted intensity / compared to the incident intensity / 0 : . I = I 0 ⋅ e − ε λ ⋅ c ⋅ d

[0004] This is ε(λ) one of substance and wavelength λ dependent coefficient, which defines the optical attenuation, and d, the optical path length through the medium.

[0005] However, the necessity of hemolysis is seen as a disadvantage because it results in time losses compared to direct measurement in whole blood, and the blood sample is irreversibly altered, rendering it unavailable for further measurements. If hemolysis is incomplete, scattering effects can still occur, leading to errors in the determination of blood parameters. Another disadvantage is that additional consumables (e.g., reagents) or components (e.g., ultrasound generators) are required for hemolysis, resulting in increased space and maintenance requirements as well as additional costs.

[0006] A method for optically determining hemoglobin derivatives without prior hemolysis is described in US 6,262,798 B1. This method uses an empirically determined correction matrix to correct the absorbance of a whole blood sample measured at seven wavelengths, and subsequently to determine the concentrations using the Lambert-Beer method.

[0007] In US Patent 10,338,058 B2, the target wavelength accuracy of ±0.03 nm for a spectrometer used to analyze whole blood samples is achieved by incorporating a calibration lamp and employing complex temperature stabilization. Such a measurement method is not very robust and is not suitable for mobile use under fluctuating environmental conditions, such as those expected when using the instrument outside a controlled environment (like a clinic or laboratory).

[0008] Patent applications US 10,088,360 B2, US 10,088,468 B2, US 10,151,630 B2, US 2017 / 0227397 A1, US 2019 / 0017993 A1, and US 2019 / 0033287 A1 protect various aspects, such as the compensation for the thermal expansion of achromatic lenses and other features of an optical spectrometer for whole blood analysis. The chemometric method kernel orthogonal projection to latent structures (k-OPLS) is used to determine hemoglobin derivatives from whole blood based on the principle of diffuse transmission.

[0009] In [REDMER, B. et al.: Determination of hemoglobin derivatives in unaltered whole blood samples using Support Vector regression in the spectral range from 450 to 700nm. Proceedings of SPIE, Vol. 11247, 14.02.2020. DOI: 10.1117 / 12.2544806] and in DE 10 2020 104 266, the application of support vector machines (SVM) for the determination of blood parameters, such as ctHb, hemoglobin derivatives and hematocrit, in the spectroscopic analysis of whole blood samples is proposed.

[0010] Common chemometric methods, such as k-OPLS or support vector regression (SVR), are unfortunately very susceptible to instrument drift. This can significantly affect the accuracy and reliability of the measurement method and is caused, for example, by faulty calibration and manufacturing tolerances, instabilities resulting from statistical processes such as noise, as well as by environmental influences such as temperature fluctuations, mechanical vibrations, and changes in position.

[0011] [Damon T. DePaoli, Prudencio Tossou, Martin Parent, Dominic Sauvageau & Daniel C. Côté, "Convolutional Neural Networks for Spectroscopic Analysis in Retinal Oximetry", Scientific Reports volume 9, 11387 (2019), doi: 10.1038 / s41598-019-47621-7] describes the use of convolutional neural networks (CNNs) in the spectroscopic determination of oxygen saturation in the retina to correct statistical noise and spectral shift due to insufficient calibration of the measuring device.

[0012] In [Ding X et al: "Measuring Oxygen Saturation with Smartphone Cameras using Convolutional Neural Networks", IEEE Journal of Biomedical and Health Informatics, IEEE, Piscataway, NJ, USA, Vol. 23, No. 6, November 2019 (2019-11), pages 2603-2610, XP011754103, ISSN: 2168-2194, DOI: 10.1109 / JBHI.2018.2887209], a method for evaluating measurement data generated by an optical setup based on image acquisition of a non-hemolyzed blood sample is described. In this case, the measurement data are evaluated using a predictive model implemented as a Convolutional Neural Network (CNN), whereby the model was trained using training data that includes at least partially augmented data.

[0013] Finally, in [Liu J et al.: "Deep Convolutional Neural Networks for Raman Spectrum Recognition: A Unified Solution", ARXIV.ORG, Cornel University Library, 201 Olin Library Cornell University Ithaca, NY 14853, August 18, 2017 (2017-08-18), XP080955100, DOI: 10.1039 / C7AN01371J] a Convolutional Neural Network (CNN) is described that was trained with augmented data and applied to optical Raman measurements of the RRUFF database.

[0014] In [M. Chatzidakis, GA Botton, "Towards calibration-invariant spectroscopy using deep learning", Scientific Reports volume 9, 2126 (2019), doi: 10.1038 / s41598-019-38482-1] the correction of design-related (when using different measuring instruments of the same type) or statistical deviations with respect to wavelength in electron energy loss spectroscopy (EELS) is described, using CNNs.

[0015] The corrections made in these cases do not take into account deviations of the measuring instrument caused by unpredictable environmental influences such as temperature fluctuations or changes in the position of the measuring instrument, but only those deviations that are causally foreseeable, recurring and therefore calculable.

[0016] However, deviations caused by unpredictable environmental influences such as temperature fluctuations or changes in the position of the measuring device play a crucial role in measurements with mobile measuring devices (in-situ measurements).

[0017] Therefore, the object of the invention is to provide a device and a method which enables the spectroscopic determination of blood parameters in non-hemolyzed whole blood samples.

[0018] In particular, the object of the invention is to provide a device and a method which enables the spectroscopic determination of blood parameters in non-hemolyzed whole blood samples in situ using mobile measuring devices.

[0019] The object of the invention is solved by a device and a method according to the main claims.

[0020] In particular, the problem underlying the invention is solved by a spectroscopy device for determining blood parameters in non-hemolyzed blood samples, which has at least one light source for emitting radiation in the wavelength range between 300 nm and 1000 nm, a measuring cuvette for receiving the blood sample during a measurement, which is optically transparent at least in a measuring area through which the radiation passes at least temporarily during the measurement, an integrator configured to receive the light forward scattered by the irradiated blood sample, a spectrometer to which the radiation received by the integrator is fed, and a data transmission unit comprising a microcontroller, wherein the spectrometer is configuredto generate a measurement signal based on at least one property of the radiation and to transmit it at least temporarily via an interface to the data processing unit, wherein the data processing unit and / or the microcontroller is configured to generate measurement data based on the measurement signal and to evaluate the measurement data with a predictive model based on at least one convolutional neural network (CNN) trained at least partially with training data containing at least partially augmented data, and to transmit a result of the evaluation to an output via an interface. According to the invention, the evaluation of the measurement data is therefore carried out with a specially designed predictive model comprising at least one convolutional neural network (CNN), since this model was trained with training data that has special properties, in particular a scope thatThis is a feature previously unknown for training predictive models in spectroscopic applications in medical technology. The data processing unit is therefore designed such that the evaluation of the measurement data generated by the optical setup, which enables transmission measurements on a non-hemolyzed blood sample, not only yields accurate results but also exhibits exceptional robustness. In the context of the invention's description and specific embodiments, robustness refers to the system's insensitivity to, for example, temperature fluctuations, vibrations, component and manufacturing tolerances, calibration errors, and similar factors. This characteristic is essential for application scenarios where such disturbances are prevalent, such as in pre-hospital emergency medical services or in many emerging and developing countries.in which known analytical methods and equipment cannot be used.

[0021] In optical setups, instrument drift typically presents a significant problem during operation, one that has often remained unsatisfactorily resolved. The inventive technical solution enables particularly robust measurement and data evaluation by taking into account instrument drift and its underlying causes. This is achieved by considering device-specific characteristics and aging effects in the predictive model processed in the data processing unit, particularly the at least one Convolutional Neural Network (CNN) used.This is achieved by using specially acquired training data for training the predictive model, in particular the at least one convolutional neural network (CNN), which, according to the invention, do not necessarily have to be based exclusively on measurements, but contain at least partially augmented data. Alternatively, it is also conceivable that synthetic data could be used as training data, the difference between augmented and synthetic data being essentially that data augmentation requires existing data, in particular experimentally acquired data, while synthetic data are generated without such data.

[0022] It is therefore important that the data used to train the at least one Convolutional Neural Network (CNN) employed as a predictive model also reflect the expected interferences that can lead to instrument-related drifts in the measurement signals. By implementing the invention, it is therefore not strictly necessary, as is usually the case, to acquire large amounts of experimental data, since augmented and / or synthetic data, which reflect the expected interferences, are generated and used to train the Convolutional Neural Network(s) (CNN). Preferably, these augmented data are derived at least partially from experimentally obtained data. Alternatively or additionally, the augmented data are obtained using computational and / or simulation models, taking into account at least one property of the technical components used for measuring and generating the measured values.

[0023] According to a particular embodiment, the at least one convolutional neural network (CNN) implemented in the data processing unit and / or in the microcontroller is characterized by the fact that it was trained with training data augmented to take into account at least one, preferably at least two, significant disturbances, such as those that occur particularly in mobile measuring devices. Particularly preferably, the disturbances considered are a drift in the detected intensity of the radiation received by the integrator and / or the spectrometer and / or a shift in the frequency spectrum of this radiation.

[0024] A drift in the detected intensity of the received radiation can be caused by inaccurate control of the light source, wear and tear of the light source, contamination, component and manufacturing tolerances, or temperature influences, including at the detector. According to a particular embodiment, a drift of + / - 10% is simulated by multiplying the generated measured values ​​by a factor that takes on a value of 0.9 to 1.1. This drift is thus taken into account in the training data and therefore during the training of the at least one Convolutional Neural Network (CNN).

[0025] Furthermore, a shift in the frequency spectrum of the received radiation can be caused by mechanical influences on the optical bench of a spectrometer, as well as by temperature fluctuations, inaccurate calibration, and component and manufacturing tolerances. The instrument-related drift of the measured values, or the instrument drift of typical grating spectrometers, for example, is approximately 0.05 nm / K. According to a further embodiment of the invention, this drift is replicated by interpolation to a wavelength vector shifted by ±2 nm and is taken into account in the training data and thus during the training of the at least one convolutional neural network (CNN). This interpolation accounts for temperature fluctuations of approximately ±40 K.

[0026] The training data used for the at least one Convolutional Neural Network (CNN) implemented in the data processing unit and / or microcontroller is advantageously generated by not only experimentally varying clinically relevant blood parameters in the laboratory and using a large database of optical spectra and reference values ​​for modeling, i.e., for the predictive model, but also by using exclusively or at least partially augmented data. In a comparatively simple way, this increases the amount of training data available and thereby achieves a particularly effective increase in the robustness of the measurement data analysis.It should be noted that the collection of experimental data in the biomedical field is regularly associated with considerable effort. Therefore, the invention provides an efficient way to use a large amount of training data to train a convolutional neural network (CNN) implemented in the data processing unit and / or the microcontroller of a device for determining blood parameters in non-hemolyzed blood samples. The data processing unit and / or microcontroller used according to the invention is thus specifically designed through the implementation of the predictive model according to the invention.

[0027] According to a particularly suitable embodiment of the invention, at least partially augmented data exhibiting increased noise are used to train at least one Convolutional Neural Network (CNN). Such augmented data offers advantages for the training and stability of the CNNs trained with it.

[0028] To demonstrate the effectiveness of using augmented data for training, measurement data subjected to the aforementioned disturbances were examined for comparison. Surprisingly, it was found that even when the augmented data only reflects one or, in particular, two of the most severe deviations, the robustness of the measurement data analysis is also ensured for the intermediate levels of disturbance. This is a truly unexpected finding. The metrics coefficient of determination (R²), mean squared deviation (RMSE), and mean absolute deviation (MAE) were advantageously used to assess accuracy and robustness, including for the intermediate states.

[0029] According to another embodiment, it would be conceivable in principle to use augmented data when employing other machine learning methods, such as SVR or k-OPLS. However, studies regarding the use of k-OPLS have shown that this significantly increases the complexity of a predictive model and necessitates re-optimization of the model's hyperparameters. This is noteworthy because it has been found that traditional methods, such as Random Forest, Decision Trees, Elastic Net, and Lasso, are not capable of accurately modeling the optical behavior or optical properties of whole blood in a beam path, particularly of non-hemolyzed blot samples.

[0030] Furthermore, many traditional methods scale poorly with large datasets, such as those that can be generated relatively quickly through the use of augmented data as training data as provided for in the invention.

[0031] In contrast, the Convolutional Neural Networks (CNNs) used according to the invention scale very well and, as investigations have shown, do not require re-optimization of the hyperparameters. Furthermore, unlike many traditional methods, Convolutional Neural Networks (CNNs) can be trained quickly on a graphics processing unit (GPU), which offers significant advantages for practical applications.

[0032] The inventors have now determined and experimentally proven that the blood parameters ctHb, O2Hb, HHb, COHb, MetHb and hematocrit in non-hemolyzed whole blood samples can be determined extremely reliably and with high accuracy, taking into account deviations caused by unpredictable environmental influences such as temperature fluctuations or changes in the position of the measuring device, when the spectroscopic device and the method described below are used.

[0033] The whole blood spectroscopy device according to the invention comprises the following components: At least one light source for emitting radiation in the wavelength range between 300 nm and 1000 nm, preferably in the wavelength range 450 nm to 700 nm; measuring cuvette; integrator; spectrometer, preferably with dispersive optical element and detector; microcontroller; output; memory; predictive models (preferably CNNs trained with at least partially augmented data)

[0034] In a preferred embodiment, the device according to the invention has a lens or a reflector.

[0035] In another preferred embodiment, the device according to the invention has a pinhole aperture.

[0036] In another preferred embodiment, the device according to the invention has an optical waveguide.

[0037] The at least one light source for emitting radiation in the wavelength range between 300 nm and 1000 nm, preferably in the wavelength range of 450 nm to 700 nm, can be, for example, a tungsten-halogen lamp, a xenon lamp, or a white LED. Alternatively, a single light source can be used in which the emission of several LEDs is achieved by coupling them into the same beam path using mirrors, lenses, or special optical waveguides, or by mounting them on a common substrate.

[0038] Suitable lenses or reflectors are generally known to those skilled in the art. The function of both is to collimate (i.e., focus) the light in order to obtain a higher signal component at the detector.

[0039] The measuring cuvette can also be designed as a flow cell. During the measurement, it serves as a sample vessel and is made of a material that is optically transparent in the spectral range between 300 nm and 1000 nm, preferably between 450 nm and 700 nm. The measuring cuvette produces a sample layer thickness of between 60 µm and 150 µm, preferably between 80 µm and 120 µm, and particularly preferably between 90 µm and 110 µm.

[0040] The integrator consists of a diffuser plate, optical lenses, mirrors, or an integrating sphere. It serves to collect the light scattered forward by the sample so that it can be detected regardless of the scattering angle, which can vary, especially in blood samples. The use of an integrating sphere for the spatial integration of the radiation components transmitted by the sample is preferred because it avoids any polarization effects that might occur and could lead to deviations in the detected light intensity. Furthermore, it eliminates the need for other optical elements, such as lenses or mirrors, which are sensitive to mechanical vibrations.

[0041] Pinhole apertures are well known to those skilled in the art; they limit the beam diameter. The use of a pinhole aperture is particularly advantageous when using an integrating sphere. The ratio of beam diameter to port size of the integrating sphere determines the maximum angle of incidence into the sphere, thus ensuring that a certain percentage of forward-scattered light reaches the detector.

[0042] The optical waveguide directs the light from the integrator to the spectrometer. Alternatively, the spectrometer could also be mounted directly on the integrating sphere.

[0043] A key requirement when selecting a suitable spectrometer is that it is sensitive to the spectral range between 300 nm and 1000 nm and has an interface that can transmit the data to a microcontroller.

[0044] In the visible spectrum, grating spectrometers are typically used, which split the light into its different wavelengths using a dispersive element. A photosensitive element, such as a CCD- or CMOS-based sensor, serves as the detector, converting the incoming light into an electrical signal proportional to the light intensity. The photosensitive element must be sensitive to the spectral range between 300 nm and 1000 nm, preferably between 450 nm and 700 nm.

[0045] Alternatively or additionally, at least one optochemical sensor is used for determining blood parameters in non-hemolyzed blood samples. Optode This type of sensor is used to determine the blood parameters pH, pO2, and pCO2. It is also advantageously conceivable to use such a sensor to measure other blood parameters, such as glucose, lactate, and / or the electrolytes potassium, sodium, calcium, and magnesium.

[0046] In optodes, an indicator dye, sensitive only to a specific analyte, is embedded in a highly specific matrix. This matrix can be applied to various substrates, enabling different sensor designs. The indicator acts as an energy converter, activated by irradiation with light of a substance-specific wavelength, and its fluorescence is reduced or even quenched by the analyte, which is why it is also called a quencher.

[0047] Optodes are preferably manufactured in series or mass production, are comparatively small and inexpensive, especially when other required electronic and optical components are reused. Such optochemical sensors do not require a separate reference and are largely insensitive to electromagnetic fields. They have excellent sensitivity and specificity. Optodes are in thermodynamic equilibrium, not a steady state. Therefore, the analyte is not consumed during measurement, and the signal is independent of the flow velocity.

[0048] According to a particularly suitable embodiment, at least one optochemical sensor is provided for detecting the pH value of a non-hemolyzed blood sample. The pH measurement principle is preferably based on a ratiometric measurement of the pH-dependent luminescence, particularly preferably of the dye 8-hydroxypyrene-1,3,6-trisulfonic acid trisodium salt (HPTS, pyranine), which is immobilized in an organically modified silicon dioxide hydrogel matrix (ORMOSIL). The optochemical sensor or optode is advantageously coated with a second hydrogel layer containing carbon black to minimize interference from the fluorescence of sample components. Due to the preferably employed ratiometric measurement principle, signal changes caused by manufacturing tolerances, leeching, photobleaching, or changes in detector sensitivity are eliminated.

[0049] According to a specific further development, excitation is achieved using LEDs, particularly preferably with radiation having wavelengths of 405 nm and 456 nm, the LEDs preferably being arranged side by side. Alternatively, an emitter that alternately or sequentially emits corresponding excitation radiation is also conceivable.

[0050] Preferably, a photodiode serves as the detector. In a specific refinement, the luminescence maximum occurs at approximately 520 nm. Glass dye filters are preferably used in the excitation and detection paths to better separate them. The pH value is temperature-dependent and is therefore compensated for in a further step according to a specific refinement.

[0051] In a further particularly suitable embodiment, at least one optochemical sensor is provided for detecting the pCO₂ value of a non-hemolyzed blood sample. The corresponding optochemical sensor, or pCO₂ optode, is essentially identical in construction to the pH optode, but preferably has a thin silicone layer on its upper surface instead of the second hydrogel layer. This silicone layer, particularly in the form of a silicone membrane, is impermeable to ions but especially permeable to gases. The passage of CO₂ from the sample generates a pH shift, which is detected. This is comparable to the construction of a Severinghaus pCO₂ electrode.

[0052] The partial pressure of carbon dioxide depends on the local air pressure (Dalton's law) and is therefore preferably compensated in a final step by a pressure sensor that is also provided.

[0053] InIn another particularly suitable embodiment, at least one optochemical sensor is provided for detecting the pO₂ value of a non-hemolyzed blood sample. Preferably, a luminescence measurement based on phase fluorimetry is performed, which has the advantage that it is not affected by changes in the light source intensity and the detector sensitivity. This measurement is based on the quenching of luminescence by molecular oxygen, which leads to a reduction in intensity and decay time. The intensity of the light source is modulated at a specific frequency. The emitted luminescence is also modulated, but exhibits a phase shift. θ due to the finite lifetime of the excited state.

[0054] InIn a special embodiment, the optochemical sensor for pO2 measurement has the indicator dye Tris(4,7-diphenyl-1,10-phenanthroline)ruthenium(II) dichloride, which is preferably embedded in a transparent silicone matrix and covered with a black silicone membrane to reduce backscattering from the blood sample.

[0055] The relationship between luminescence and oxygen concentration is normally quantified by the Stern-Volmer equation (SV): I 0 I = τ 0 τ = 1 + K SV ⋅ pO 2 where I 0 and I The luminescence intensities in the absence and presence of oxygen are τ0 and τ0 are the decay times in the absence and presence of oxygen, respectively. K SV The Stern-Volmer constant is pO2, and pO2 indicates the partial pressure of oxygen.

[0056] Because the indicator is embedded in a silicon matrix, the curve resulting from the SV deviates from its linear behavior. This can be described by the presence of at least two environments, a so-called two-sided model, in which the indicator is quenched at different rates: I 0 I = τ 0 τ = f 1 1 + K SV 1 ⋅ pO 2 + f 2 1 + K SV 2 ⋅ pO 2 − 1 where f 1 and f 2 = 1 - f 1 the proportions of luminescence for each component under unquenched conditions. K SV 1 and K SV 2 the corresponding Stern-Volmer constants for each component.

[0057] In a simple exponential decay, the following relationship applies to the phase shift: θ and the finite lifetime of the excited state τ tan θ = ωτ where ωThe angular modulation frequency of the light source is the relevant factor. The range of suitable modulation frequencies depends on the lifetime and must be large enough so that the phase shift is frequency-dependent, but lower than the frequencies at which the modulation is no longer measurable.

[0058] It is common practice to relate the phase shift measured at a single frequency to its apparent lifetime: tan θ 0 tan θ = f 1 + K SV 1 ⋅ pO 2 + 1 − f 1 + K SV 2 ⋅ pO 2 − 1 where θ 0 and θ the phase shifts in the absence or presence of oxygen. f and 1 - f are the proportions of the luminescence of the individual components under unquenched conditions (i.e., without oxygen). The quantities f, K SV 1, and K SV Two variables are frequency-dependent. Furthermore, all quantities are temperature-dependent, so the equation can be rewritten as follows. tan θ 0 ω T tan θ w T pO 2 = f ω T 1 + K SV 1 ω T ⋅ pO 2 + 1 − f ω T 1 + K SV 2 ω T ⋅ pO 2 − 1

[0059] The measurement data generated by the pO2 sensor are advantageously processed using an artificial neural network that analyzes a series of key figures K. K w T pO 2 ≡ tan θ 0 ω T tan θ w T pO 2 for the simultaneous prediction of temperature T and oxygen partial pressure pO2 from a series of measurements at different modulation frequencies w.

[0060] Measurements are performed at various modulation frequencies between 200 Hz and 100 kHz. The photodiode signal, along with the reference signal driving the LED current driver, is digitized using a fast ADC, and the phase is determined on the microcontroller using a software lock-in amplifier. A set of phase angles serves as input for the neural network, which ultimately outputs the oxygen partial pressure and temperature. In a final step, the phase angle is compensated for by a pressure sensor, which is also preferably included.

[0061] According to the invention, the data evaluation unit and / or the microcontroller receives the data from the spectrometer and / or a sensor, such as an optochemical sensor, processes this data and executes the predictive model(s) stored in memory, which consist of at least one Convolutional Neural Network (CNN) trained with training data that contains at least partially augmented data.

[0062] The microcontroller must be powerful enough to apply the predictive CNN models to the measurement data and must have the necessary interfaces. For example, it could be an STM32MP157. Ideally, it should be able to run an embedded operating system, such as Linux, with a suitable software framework like TensorFlow / Keras or PyTorch.

[0063] The output can be a standard interface, a display, a printer, or a radio module. Typical examples of interfaces are SPI, I²C, UART, or USB.

[0064] The predictive models used according to the invention are stored as a file in memory. They were previously trained using a CNN. The CNN is characterized by having at least one convolutional layer.

[0065] In the convolutional layer, the incoming signals are convolved using so-called filter kernels. The weights of the filter kernels were previously trained and stored in the predictive model as a file. The characteristic and distinguishable procedure is described in more detail below.

[0066] The input vector XThe neural network consists of n data points / values ​​xi (i=1... n). Each data point / value corresponds to an intensity value recorded by the spectrometer at a specific wavelength λ. The neural network is characterized by the fact that, in the first step, a convolution operation is performed between the input vector and one or more so-called filter kernels. The filter kernels are also vectors with m data points / values ​​yj (j=1...m). The number of filter kernels k, their length m, and the values ​​yj they contain were previously determined using training datasets and differ depending on the blood parameter to be determined from the input vector. The length m can be different for each filter kernel. A mathematical convolution operation is then performed between each filter kernel and the input vector. This operation is described in Figure 16This process is illustrated clearly. First, the data points with the same index are arranged opposite each other, i.e., i = j. The values ​​x and y are multiplied together, and the individual product results are summed. In the next step, the filter kernel is shifted by one value relative to the input vector, i.e., j = i + 1, and the same calculation is performed. This is repeated until the filter kernel reaches the end of the input vector. In this way, k new vectors of length n - m + 1 are generated. These vectors then serve as input vectors for the deeper layers of the neural network. In these layers, further convolution operations can be performed, or alternatively, the individual vector elements, weighted by specific factors, can serve as input signals for the neurons located in the deeper layers.

[0067] The learned weights of the filter kernel, in the form of a file, are read into the predictive model by a software framework, such as TensorFlow / Keras or PyTorch, thereby enabling a prediction on new, previously unknown data.

[0068] A common format is the Hierarchical Data Format (HDF). If the CNN is exported from TensorFlow / Keras to an HDFv5 (.h5) file, for example, the structure of the CNN is described at the beginning of the file header. A 1D convolutional layer, such as that found in spectroscopic data, then appears as "Conv1D" visible to the microcontroller.

[0069] Predictive models trained using CNNs are available in formats such as HDF5. They are generated using machine learning frameworks like TensorFlow / Keras and require components of these frameworks to be present on the microcontroller. They receive as input the transmissions measured by the spectrometer across the recorded spectral range and output the values ​​for the blood parameters. Multiple parameters can also be predicted by a single predictive model.

[0070] In addition to a device, the invention also relates to a method for the spectroscopic determination of parameters in non-hemolyzed blood samples comprising the following steps: Emission of radiation with a light source in the wavelength range between 300 nm and 1000 nm in the direction of a measurement area; arrangement of a measuring cuvette containing a blood sample, at least partially optically transparent in the wavelength range of 300 nm to 1000 nm, in a measurement area through which the radiation is passed at least temporarily during the measurement; reception of the light scattered forward by the blood sample with an integrator; feeding the radiation received by the integrator to a spectrometer, by which at least one measurement signal is generated based on at least one property of the radiation, and the measurement signal is transmitted at least temporarily via an interface to a data processing unit with a microcontroller, wherein the data processing unit and / or the microcontroller generates at least one measurement datum or measurement data based on the measurement signal.which are evaluated using a predictive model based on at least one Convolutional Neural Network (CNN) trained using training data that contains at least partially augmented data, and the transmission of a result of the evaluation of the measurement data to an output via an interface (10).

[0071] In In a particular embodiment, it is provided that a reference spectrum I R is determined in the context of a transmission measurement, with a detector, in a wavelength range of 300 nm to 1000 nm without an absorbing and / or scattering medium, for example a blood sample (5), in a beam path in the measurement area, at least partially the blood sample is introduced into the beam path in the measurement area, a sample spectrum IP is determined within the framework of a transmission measurement carried out in the measuring area with the detector, the transmission through the sample is measured at different integration times of the detector to generate a measurement signal based on the incident radiation, while the light source used (1) is switched off to obtain dark spectra I RD and I PD to obtain, I 0 by subtracting I RD from I Around I by subtracting I PD from IP to be determined, normalization of I 0 and I , so that both have an identical timescale, I through I 0 is divided to obtain an input vector X to determine at least one predictive model, including a mean and a variance of X adjusted by stored coefficients, Xon the at least one stored predictive model (12) based on at least one Convolutional Neural Network (CNN) trained with training data containing at least partially augmented data, and that information about at least one blood parameter of the blood sample (5) is output based on at least one result value obtained by applying the trained predictive model (12).

[0072] Furthermore, it is advantageous if at least one pooling layer is used in the Convolutional Neural Network (CNN).

[0073] According to a further embodiment of the method according to the invention, it is provided that during the transmission measurement for determining the reference spectrum, distilled water, air and / or an optically transparent rinsing solution in the wavelength range of 300 nm to 1000 nm, in particular in a wavelength range of 450 nm to 700 nm, is arranged in the beam path.

[0074] Furthermore, the invention also relates to a method for evaluating measurement data generated with an optical setup based on a transmission measurement of a non-hemolyzed blood sample, wherein the measurement data are evaluated using a predictive model implemented as a Convolutional Neural Network (CNN), which was trained using training data that contains at least partially augmented data.

[0075] The underlying predictive models were previously determined using at least one CNN, where the CNN has at least one pooling layer.

[0076] Common chemometric methods, such as k-OPLS or SVR, are highly susceptible to instrument drift. This has a significant impact on the accuracy and reliability of a spectrophotometer, such as those used in CO oximetry.

[0077] Basically, two effects can be distinguished: Fluctuations in the detected light intensity, for example due to heating, contamination, or wear of the light source, but also changes in detector sensitivity resulting from temperature fluctuations or induced mechanical stresses within the optical bench. Spectral shifts in the measured transmission spectra, such as those that occur due to temperature fluctuations in grating spectrometers typically used in the visible spectral range or that can be caused by inaccurate, faulty, or missing factory calibration.

[0078] Using the device and method according to the invention, the spectroscopic determination of blood parameters in non-hemolyzed whole blood samples is possible in situ with mobile measuring devices.

[0079] The device and method according to the invention enable the operation of extremely robust, lightweight, and compact mobile measuring devices for the spectroscopic determination of blood oxygen saturation and blood parameters in non-hemolyzed whole blood samples. A wavelength reference and temperature stabilization of the sensor are even unnecessary.

[0080] The CNNs used to determine the predictive models were identified through experimental work on a laboratory setup, which is described below.

[0081] Without limiting the generality of the teaching, transmission spectra of 358 blood samples from 7 different individuals were recorded in a laboratory setup to demonstrate the performance of the device and method according to the invention, and these were transferred to a database together with the reference values ​​for ctHb, O2Hb, COHb and MetHb determined on a radiometer OSM 3 CO-oximeter.

[0082] The HHb percentage was also recorded in the database. It was calculated by subtracting the sum of the other hemoglobin derivatives O₂Hb, COHb, and MetHb from 100%. The hematocrit of the samples was also determined by centrifugation of capillary tubes and recorded in the database.

[0083] Prior to measurement, the blood parameters relevant to the optical behavior were experimentally adjusted over a wide range to simulate clinical disease patterns. All blood samples were washed three times with phosphate-buffered saline (PBS, pH 7.4) to eliminate undefined variations due to plasma proteins, remove free hemoglobin, and precisely control the osmotic concentration.

[0084] The hematocrit of the samples was varied by centrifugation and subsequent mixing of the cellular and liquid components. Since hematocrit correlates excellently with ctHb, the amount of total hemoglobin could be adjusted in this way. The samples were then gassed in a tonometer with varying proportions of moistened O₂, CO₂, and CO in N₂. The samples were contained in a hollow glass sphere, which was continuously swirled in a water bath at 37.2 °C. Depending on the composition and duration of the gassing, the fractional proportions of O₂Hb, HHb, and COHb were adjusted. The proportion of MetHb was varied in some samples by adding sodium nitrite.

[0085] This database was then used to train predictive models that can predict the parameters ctHb, O2Hb, HHb, COHb, MetHb and hematocrit based on transmission spectra of new, previously unknown blood samples.

[0086] A separate predictive model was created for each parameter. This allows for individual optimization of the hyperparameters, but is not technically necessary. The predictive models were trained using the machine learning frameworks Scikit-learn and TensorFlow / Keras in the Python programming language.

[0087] The background noise was first subtracted from the transmission spectra, the result was normalized to 1 / s, and divided by the transmission without a blood sample in the beam path. Negative values ​​were set to zero. The data were then standardized so that the mean was zero and the variance was one. Fivefold cross-validation was performed to avoid under- or overfitting of the predictive models.

[0088] For testing the models, 45 additional blood samples from 12 different individuals were optically measured in the laboratory setup. The procedure was identical to the previous one; however, the samples were not washed with PBS and therefore still contained plasma proteins. The group of individuals used for testing the models was separate from the group of individuals used for training.

[0089] To demonstrate the performance of the algorithm for CO-oximetry, various algorithms were compared: linear regression, ridge regression, SVR, random forest, decision trees, Lasso, elastic net, gradient boosting, extreme gradient boosting, k-OPLS, and CNN. The mean squared deviation (MSE) was used as the loss function during training, with the goal of maximizing the negative MSE.

[0090] The prediction error was quantified using the coefficient of determination (R²), the root mean square error (RMSE), and the mean absolute error (MAE). The main focus was on instrument drift.

[0091] In Figure 1Without limiting the general concept of the invention, an advantageous embodiment of a device designed according to the invention is shown. The light emitted by the light source 1 is collimated by means of a lens 2 and limited to a defined measuring window by a pinhole aperture 3. The light then shines through a cuvette 4 in which the blood sample 5 is located. The light emerging from the blood sample 5 is collected in an integrating sphere 6 and guided by means of an optical fiber 7 to a spectrally resolving sensor 8. The spectrally resolved light intensities measured by the sensor 8 are acquired by means of a microprocessor 9 of a data processing unit, stored in a memory 11, and processed by a predictive model 12, also stored there and trained by a convolutional neural network (CNN).The blood parameters extracted by the CNN, for example ctHb, O2Hb, HHb, COHb, MetHb and hematocrit, are then displayed on a display unit 10.

[0092] Figures 2 to 7 demonstrate the high accuracy of the inventive method using CNN in predicting the blood parameters total hemoglobin concentration (ctHb) ( Fig. 2 ), hematocrit ( Fig. 3 ), oxygenated hemoglobin (O2Hb) ( Fig. 4 ), deoxygenated hemoglobin (HHb) ( Fig. 5 ), carboxyhemoglobin (COHb) ( Fig. 6 ) and methemoglobin (Hi) ( Fig. 7 Each cross represents a sample from the test data set, the solid line the optimum with complete agreement with the reference, and the dashed line the adjustment function after application of the device and method according to the invention.

[0093] To evaluate the prediction accuracy of the method according to the invention under conditions that occur during measurements with mobile measuring devices, in particular in-situ measurements and resulting instrument drift, the following procedure was followed.

[0094] The predictions according to the invention, using a CNN, were compared with those of k-OPLS, which can be considered prior art. For this purpose, the concentration of total hemoglobin (ctHb), the hematocrit, and the fractional amounts of oxygenated hemoglobin (O₂Hb), deoxygenated hemoglobin (HHb), carboxyhemoglobin (COHb), and methemoglobin (Hi) were predicted on the test dataset, while the detected light intensity varied by ±10%. The drift of the detected intensity from -10% to +10% was simulated by multiplying the transmission by values ​​between 0.9 and 1.1 across all wavelengths for all observations in the test dataset. The prediction error was determined using the coefficient of determination (R²< ) metric. Fig. 8 ), square root of mean square deviation (RMSE) ( Fig. 9 ) and mean absolute error (MAE) ( Fig. 10) quantified. The crosses show the accuracy curve under instrument drift in the prior art (k-OPLS), the dots show the accuracy curve under instrument drift in the inventive method (CNN). For R 2<, higher values ​​are better, for RMSE and MAE, lower values ​​are better.

[0095] The clear superiority of the inventive method (CNN) over the current state of the art (k-OPLS) is evident. With a deviation of -10% in the detected light intensity, the error with k-OPLS is up to a factor of 5 higher than with CNN (ctHb, MAE). The use of the inventive method consistently reduces the prediction error for all parameters.

[0096] To evaluate the prediction accuracy under spectral shift of the measured transmission spectra, this was simulated by a 1D interpolation of the measured transmission of all observations onto a wavelength vector shifted between -2 nm and +2 nm. The error was quantified using the coefficient of determination (R²< ) metric. Fig. 11 ), square root of mean square deviation (RMSE) ( Fig. 12 ) and mean absolute error (MAE) ( Fig. 13 For R²<, higher values ​​are better; for RMSE and MAE, lower values ​​are better.

[0097] The assumed linearity of the shift across the entire spectral range is based on experimental measurements on grating spectrometers subjected to temperature drift. The simulated spectral shift of +2 nm corresponds to a temperature change of approximately 40 K. This also demonstrates a clear superiority of the inventive method (CNN) over the current state of the art (k-OPLS). The error with k-OPLS is up to a factor of 8 higher than with CNN (COHb, MAE). The predictive accuracy with k-OPLS is even negative when predicting COHb. The inventive method also consistently reduces the prediction error for all parameters in this case.

[0098] To further compare the predictions according to the invention using a CNN with those of k-OPLS, which can be considered prior art, the predictions of ctHb, O₂Hb, and COHb under the influence of instrument drift are shown as examples. For each influence, drift of the detected intensity by ±10% ( Fig. 14 ) and a spectral shift of ±2 nm ( Fig. 15 The most extreme boundary conditions are shown. Each cross represents a sample from the test dataset.

[0099] List of characters: Figure 1 : Scheme of the device according to the invention Figure 2 Accuracy of the inventive method in predicting ctHb Figure 3 Accuracy of the inventive method in predicting hematocrit Figure 4 Accuracy of the inventive method in predicting O₂Hb Figure 5Accuracy of the inventive method in predicting HHb Figure 6 : Accuracy of the inventive method in predicting COHb Figure 7 Accuracy of the inventive method in predicting Hi Figure 8 : Course of accuracy under instrument drift in the form of fluctuations in the detected light intensity when predicting the concentration of ctHb, hematocrit, as well as the fractional proportions of O 2 Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the measure of determination R 2< . Figure 9: Course of accuracy under instrument drift in the form of fluctuations in the detected light intensity when predicting the concentration of ctHb, hematocrit, as well as the fractional proportions of O 2 Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the mean squared deviation (RMSE). Figure 10 : Course of accuracy under instrument drift in the form of fluctuations in the detected light intensity when predicting the concentration of ctHb, hematocrit, as well as the fractional proportions of O 2 Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the mean absolute deviation (MAE). Figure 11: Course of accuracy under instrument drift in the form of spectral shift in prediction of the concentration of ctHb, hematocrit, as well as the fractional fractions of O 2 Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the measure of determination R 2< . Figure 12 : Accuracy trend under instrument drift in the form of spectral shift when predicting the concentration of ctHb, hematocrit, and the fractional fractions of O2Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the mean squared deviation (RMSE). Figure 13: Accuracy trend under instrument drift in the form of spectral shift when predicting the concentration of ctHb, hematocrit, and the fractional fractions of O2Hb, HHb, COHb and Hi in the prior art (k-OPLS) (crosses) and inventive methods (CNN) (dots), characterized by the mean absolute deviation (MAE). Figure 14 : Maximum error in the prediction of ctHb, O2Hb and COHb under instrument drift in the form of fluctuations in the detected light intensity in the prior art (k-OPLS) (left column) and inventive methods (CNN) (right column). Figure 15 : Maximum error in the prediction of ctHb, O2Hb and COHb under instrument drift in the form of spectral shift in the state of the art (k-OPLS) (left column) and inventive methods (CNN) (right column). Figure 16 : Characteristic handling of incoming signals in the CNN Reference symbol list:

[0100] 1 Light source 2 Lens 3 Aperture 4 Measuring cuvette 5 Sample 6 Integrator 7 Optical fiber 8 Spectrometer 9 Microcontroller 10 Output 11 Memory 12 Predictive model with CNN

Claims

1. A spectroscopic device for determining blood parameters in non-hemolyzed blood samples (5), having at least one light source (1) for emitting radiation in the wavelength range between 300 nm and 1000 nm, a measurement cuvette (4) that receives the blood sample during a measurement, which is optically transparent at least in a measurement region through which the radiation passes at least intermittently during the measurement, an integrator (6) configured to receive light scattered forward by the irradiated blood sample (5), a spectrometer (8), to which the radiation received by the integrator (6) is fed, and a data processing unit comprising a microcontroller (9), wherein the spectrometer (8) is configured to generate a measurement signal based on at least one property of the radiation and to transmit it at least intermittently to the data processing unit via an interface, and wherein the data processing unit and / or the microcontroller (9) is configured to generate measurement data based on the measurement signal and to evaluate the measurement data using at least one predictive model (12) based on a convolutional neural network (CNN) trained with training data that at least partially contains augmented data, and to transmit a result of the evaluation to an output (10) via an interface.

2. The device according to claim 1, characterized in that the integrator (6) comprises an Ulbricht sphere.

3. The device according to claim 1 or 2, characterized in that an optical fiber (7) for at least partially transmitting the radiation is arranged between the integrator (6) and the spectrometer (8).

4. The device according to any one of the preceding claims, characterized in that the integrator (6) is adjacent to the spectrometer (8).

5. The device according to any one of the preceding claims, characterized in that the data processing unit comprises a data storage device (11) or is connected, via a wired or wireless data transmission link, to a data storage device (11), on which the predictive model (12) trained with a convolutional neural network (CNN) is stored at least temporarily.

6. The device according to any one of the preceding claims, characterized in that the predictive model (12) has been trained by a convolutional neural network (CNN) comprising at least one convolutional layer and / or comprises at least one file containing filter kernel weights trained by the convolutional neural network (CNN).

7. The device according to claim 6, characterized in that the data processing unit is configured to load the trained filter kernel weights into the predictive model (12) using a software framework.

8. A method for evaluating measurement data generated by an optical setup based on a transmission measurement on a non-hemolyzed blood sample (5), wherein the measurement data is evaluated by means of a predictive model (12) implemented as a Convolutional Neural Network (CNN) that has been trained using training data at least partially containing augmented data.

9. The method according to claim 8 for determining blood parameters in non-hemolyzed blood samples (5), comprising the steps of: - emitting radiation with a light source (1) in the wavelength range between 300 nm and 1000 nm towards a measurement area, - arranging a measurement cuvette (4), that is optically transparent at least partially in the wavelength range from 300 nm to 1000 nm, containing a blood sample in a measurement region through which the radiation is passed at least intermittently during the measurement, - receiving the light scattered forward by the irradiated blood sample (5) with an integrator (6), - feeding the radiation received by the integrator (6) to a spectrometer, which generates at least one measurement signal based on at least one property of the radiation, and transmits the measurement signal at least intermittently to a data processing unit with a microcontroller (9) via an interface, wherein the data processing unit and / or the microcontroller (9) generates, based on the measurement signal, at least one measurement datum or measurement data, which is evaluated using a predictive model (12), which comprises at least one convolutional neural network (CNN) that has been trained using training data at least partially containing augmented data, and a result of the evaluation of the measurement datum or the measurement data is transmitted to an output (10) via an interface.

10. The method according to claim 9, characterized in that - a reference spectrum IR is determined as part of a transmission measurement using a detector, without a medium absorbing and / or scattering in a wavelength range from 300 nm to 1000 nm in a beam path in the measurement range, without the blood sample (5) being positioned in the beam path in the measurement range, - the blood sample (5) is introduced, at least partially, into the beam path in the measurement range, - a sample spectrum IP is determined as part of a transmission measurement performed in the measurement range using the detector, - the transmission through the sample is measured at various detector integration times to generate a measurement signal based on the incident radiation, while the light source (1) used is switched off, in order to obtain dark spectra IRD and IPD, - I0 is determined by subtracting IRD from IR and I by subtracting IPD from IP, - I0 and I are normalized, so that both have an identical time scale, - I is divided by I0 to determine an input vector X for at least one predictive model, - a mean and a variance of X are adjusted using stored coefficients, - X is applied to the at least one stored predictive model (12), which is based on at least one convolutional neural network (CNN) trained using training data that at least partially includes augmented data, and that - information regarding at least one blood parameter of the blood sample (5) is output based on at least one result value obtained by applying the trained predictive model (12).

11. The method according to claim 9 or 10, characterized in that at least one pooling layer is used in the convolutional neural network (CNN).

12. The method according to any one of claims 9 to 11, characterized in that, during the transmission measurement for determining the reference spectrum, distilled water, air and / or a flushing solution optically transparent in the wavelength range of 300 nm to 1000 nm, particularly in a wavelength range of 450 nm to 700 nm, is arranged in the beam path.