Method and device for measuring a process in an object

Abstract

The invention relates to a method and to a device for measuring at least one process in an object. An alternating field is irradiated into the object and received once it has passed through the object. At least one feature of the process and / or a temporal representation of the process is then generated time-dependently using a statistical model.

Description

[0001] Method and device for measuring a drain in an object

[0002] The invention relates to a method and a device for measuring at least one process in an object. An alternating field is radiated into the object and received after it has passed through it. A statistical model is then used to generate at least one feature of the process and / or a temporal representation of the process in a time-dependent manner.

[0003] When treating mechanically ventilated patients, a technical challenge lies in detecting the patient's independent (spontaneous) breathing. During mechanical ventilation, air is forced into the lungs, sometimes at high pressure. If ventilation is performed contrary to spontaneous breathing, this can lead to irreversible damage to the lungs and upper respiratory tract. The following types of damage are known to occur:

[0004] Barotrauma: Damage to the alveoli due to high ventilation pressures,

[0005] Volutrauma: Overstretching of the alveoli due to excessive ventilation volumes,

[0006] Atelectasis: Collapse of alveoli due to too low ventilation pressures or insufficient ventilation volume

[0007] However, even properly synchronized ventilation over an extended period can damage patient tissue, so it is desirable to reduce the duration of ventilation to a minimum. This requires a clear assessment of the patient's vital status, which makes the recording of additional vital parameters essential. This gives rise to further technical questions, such as how, in addition to respiratory parameters (e.g., expiratory phase, inspiratory phase), cardiologically relevant parameters (e.g., ECG data, QRS complexes, heart rates, heart rate variability, arrhythmias) can be obtained. The challenge here is to enable this using a contactless procedure in order to minimize patient discomfort and thereby increase acceptance of the measurement method.

[0008] The collection of vital data such as heart and respiratory rate, respiratory volume, depth of breath, duration and time of inspiration or expiration is useful in a wide variety of applications. For example, in the control of ventilators, in the detection of sleep apnea or respiratory diseases, in monitoring the stress level of emergency personnel, or in biometrics in vehicles. The sensor technology should be as fast, unobtrusive, and robust against interference as possible. One technology suitable for this purpose is the thorax monitor (WO 2019 / 243444 A1). The multidimensional and time-dependent data obtained here provide information about underlying dynamic changes in the test system. However, interpretation is prone to artifacts.

[0009] When recording cardiorespiratory data or vital signs in general, the majority of procedures use contact modalities or modalities that are placed close to the patient's body. This can be very stressful for patients, especially when measuring over longer periods or in everyday situations where vital sign measurement is not a priority (e.g., while driving or sleeping).

[0010] The most established method for measuring cardiac activity is certainly the electrocardiogram (ECG). This involves recording the electrical activity of the heart using electrodes attached to the body.

[0011] Many established methods combine an ECG with impedance cardiography (ICG) to additionally measure respiratory activity. These methods estimate the cardiac component and subtract it from the impedance signal to derive the respiratory component from the resulting signal. Since the recording of the respiratory parameters is indirectly derived from the estimation of the cardiac component, this necessarily depends on the quality of the cardiac model. Furthermore, the respiratory rate is primarily determined here. A direct statement about tidal volume and spontaneous breathing is not possible.

[0012] Another method for recording respiratory activity is measuring chest and abdominal expansion using respiratory inductive plethymography (RIP). Two elastic bands are placed around the patient's chest and abdomen. Coils are incorporated into the bands, which change their inductance when the bands stretch. Due to the elastic bands, this method can be uncomfortable and restrictive for the patient. A comparable method is also conceivable using strain gauges, although this method does not offer any additional advantages. There are also a number of methods that measure vital parameters without contact. Most of these methods are based on radar-based measurements. However, these radar-based methods usually only provide information about the respiratory rate or heart rate via reflections from the chest.

[0013] The object of the present invention is therefore to provide a method for measuring a process in an object that is non-invasive, non-obstructive and scalable.

[0014] The object is achieved by the method for measuring at least one process in an object according to claim 1 and the device for measuring at least one process in an object according to claim 21. The respective dependent claims specify advantageous developments of the method according to the invention and the device according to the invention.

[0015] According to the invention, a method for measuring at least one process in an object is provided. In principle, the method is applicable to all types of objects, in particular those into which an alternating electromagnetic field can penetrate. Preferably, the object can be the body of a living being, e.g., the body of a human or an animal.

[0016] When we talk about measuring the process in the object, we mean deriving the process characteristics of interest from the demodulation data. Synonyms for "measurement" can be "prediction," "predicting," or "estimating."

[0017] According to the invention, an alternating electromagnetic field is radiated into the object. The alternating field is received after it has passed through the object. The fact that the alternating field has passed through the object preferably means that it entered the object after being radiated and exited the object before being received. The alternating field has preferably entered the object by at least 1 cm, particularly preferably at least 5 cm, particularly preferably at least 10 cm, particularly preferably at least 20 cm. The alternating field is modulated by the process in the object to be measured. This can mean, in particular, that the process in the object leads to a phase shift, frequency shift and / or amplitude change of the received alternating field compared to the radiated alternating field.Modulation can advantageously be understood here as a process in which useful data is applied to a carrier signal, in this case the radiated alternating field. Useful data can be considered changes in the object that constitute the process to be measured. In this sense, these changes are modulated onto the radiated alternating field as the carrier.

[0018] The minimum penetration depth of the alternating field into the object can be determined such that the received alternating field has still undergone a modulation at the given penetration depth that exceeds the measurement accuracy of the method. Therefore, the alternating field should preferably only be considered to have reached the specified minimum penetration depth, i.e., to have penetrated the object, if the modulation at this penetration depth still leads to a distinguishable difference in the measurement of at least one process.

[0019] The received alternating field is demodulated in a step known as the demodulation step to generate multidimensional demodulation data. Advantageously, multiple receiving structures can receive the same or different alternating fields. Multidimensional demodulation data can then be demodulated from the received alternating fields of each receiving structure in a separate demodulation step and fed together into a statistical model. This allows for particularly advantageous incorporation of asymmetries of the object (e.g., right and left lungs) into the model.

[0020] Demodulation can then be understood as separating the payload data from the carrier signal. In this step, the received alternating field can be separated into the radiated alternating field on the one hand, and the changes caused by the process to be measured on the other. In this sense, the changes in the process can then be viewed as multidimensional demodulation data. The fact that the demodulation data is multidimensional can mean that it has at least two independent components, particularly advantageously at least two linearly independent components.

[0021] According to the invention, the multidimensional demodulation data and / or data derived from these multidimensional demodulation data are fed to a statistical model. The statistical model then generates, in a time-dependent manner, at least one feature of the sequence and / or a temporal representation of the sequence from the supplied multidimensional demodulation data and / or from the data derived from these data.

[0022] The fact that generation occurs in a time-dependent manner can, in particular, mean that it occurs at at least two different points in time. Advantageously, however, time-dependent generation means regular generation, in particular quasi-continuous generation or, where possible, advantageously also continuous generation. Continuous generation should be viewed as continuous in the sense in which this term is usually understood in the field of digital data processing, i.e., discretized in time, for example, with a clock frequency. In analog processing, the term continuous can also be understood in the analog sense. The data can have a continuous or quasi-continuous course.

[0023] According to the invention, a temporal representation of the process can be generated. This representation can therefore be, for example, at least one electrocardiogram, at least one impedance cardiogram, at least one ballistocardiogram, at least one seismocardiogram, the movement and / or expansion of the thorax and / or abdomen, at least one blood pressure curve over time, a respiratory volume curve, at least one respiratory pressure curve, at least one respiratory flow as a function of time, at least one oxygen saturation curve over time, temporal sequence data obtained by imaging, and / or a plethysmogram. Temporal sequences obtained from the image data are, for example, ventilation of the left and / or right lung over time, which can originate from ultrasound, electroimpedance tomography, or computed tomography.

[0024] According to the invention, one or more features of a process can also be determined, such as exceeding a threshold value and the like.

[0025] In particular, the following characteristics can be advantageously determined:

[0026] Respiratory characteristics can include, for example, one or more of the following:

[0027] Tidal volume: Tidal volume refers to the volume of air inhaled and exhaled during normal inhalation and exhalation. Measuring tidal volume can provide information about the efficiency of breathing and is relevant in the treatment of respiratory diseases.

[0028] Respiratory cycle: The respiratory cycle measures the total time required for one breath, from the beginning of inhalation to the end of exhalation. Measuring the respiratory cycle can be used to monitor respiratory rate and efficiency and is relevant in the treatment of respiratory diseases.

[0029] Inspiratory time: Inspiratory time describes the duration of a breath taken. Measuring inspiratory time can provide information about respiratory rate and efficiency and is relevant in the treatment of respiratory diseases.

[0030] Expiratory time: Expiratory time describes the length of time it takes to exhale a breath. Measuring expiratory time can provide information about respiratory rate and efficiency and is relevant in the treatment of respiratory diseases.

[0031] Inhalation-to-exhalation ratio: The inhalation-to-exhalation ratio refers to the ratio of the time taken to inhale a breath to the time taken to exhale a breath. Measuring the inhalation-to-exhalation ratio can provide information about respiratory rate and efficiency and is relevant in the treatment of respiratory diseases.

[0032] Peak inspiratory flow: Peak inspiratory flow measures the maximum volume flow during inspiration and provides information about airway performance. Measuring peak inspiratory flow can be used to monitor respiratory function and is relevant in the treatment of respiratory diseases.

[0033] Peak expiratory flow: Peak expiratory flow measures the maximum volume flow during exhalation and provides information about airway performance. Peak expiratory flow can be used to monitor respiratory function and is relevant in the treatment of respiratory diseases.

[0034] For example, one or more of the following can be advantageously determined as cardiological parameters:

[0035] Determination of position type: ECG parameters that measure the electrical activity of the heart in different directions to assess the heart rhythm and electrical activity of the heart.

[0036] ST-segment changes: Abnormalities in the ECG that may indicate possible heart disease, particularly damage to the heart muscle.

[0037] T-wave changes: Changes in the shape, amplitude, or direction of the T waves on the ECG that may indicate possible heart disease or electrolyte disturbances.

[0038] QT interval prolongation: A lengthening of the time interval between the beginning of the QRS complex and the end of the T wave on the ECG, which may indicate possible cardiac arrhythmias. P-wave abnormalities: Changes in the shape or amplitude of the P waves on the ECG, which may indicate possible atrial disease.

[0039] QRS complex changes: Changes in the shape or duration of the QRS complex in the ECG that may indicate possible diseases of the heart muscle.

[0040] Bundle branch block: Disturbance in the electrical conduction in the heart, which may indicate possible heart disease.

[0041] Long-term stress ECG: ECG recordings over a longer period of time or during physical exertion to diagnose heart disease or monitor patients at risk.

[0042] Heart rate: Number of heartbeats per minute, an important indicator of heart function.

[0043] Heart rate variability: Fluctuations in heart rate, which are an important indicator of possible heart disease.

[0044] Arrhythmias: Abnormalities in heart rhythm that may indicate possible heart disease.

[0045] Ejection fraction (EF): The ejection fraction is a percentage that indicates how much blood is ejected from the left ventricle with each heartbeat. A reduced EF may indicate heart failure or other cardiac problems.

[0046] The temporal representation of the process can also be referred to as a simulation of the process. It is not mandatory that the temporal sequence be presented as a diagram. It may also be sufficient if it is determined and / or stored in such a way that it is suitable for the intended use.

[0047] In an advantageous embodiment of the invention, the alternating electromagnetic field radiated into the object can contain at least two components of different frequencies. In particular, sinusoidal oscillations with the corresponding frequencies could be considered as components. It is advantageous that different frequencies exhibit different sensitivities to disturbances. The alternating fields of different frequencies can in turn be demodulated, and the data derived therefrom can be fed into the statistical model. Advantageously, the demodulation data can be processed separately for the different frequencies. In general, the use of several different frequencies enables the suppression of disturbances by increasing the dimensionality and by varying their proportionality to the disturbance, and also advantageously improves signal quality, robustness, and reproducibility.

[0048] In a particularly advantageous embodiment, the received alternating field can be demodulated in the demodulation step using an in-phase and quadrature method to generate I data and Q data as the multidimensional demodulation data. The I and Q data can be all demodulation data, or further demodulation data can be generated in the demodulation step. The I data and the Q data, as well as any additional demodulation data, can be fed to the statistical model. Alternatively or additionally, data derived from the I data and the Q data, and optionally any additional demodulation data, can also be fed to the statistical model.

[0049] The statistical model can advantageously be viewed as a unit that has at least one input and at least one output. Data supplied to the statistical model can thus be input into the at least one input of this unit. The statistical model can then generate at least one feature of the process or the representation of the process and output it via its at least one output.

[0050] If data derived from the multidimensional demodulation data and / or the I data and / or the Q data are used, these can be one or more selected from a group containing the following data: a derivative of Q with respect to I, a derivative of I with respect to Q, a derivative of I with respect to time, a derivative of Q with respect to time, an amplitude of the received alternating field, a phase of the received alternating field, an amplitude from the derivative of Q with respect to time and the derivative of I with respect to time, a phase from the derivative of Q with respect to time and the derivative of I with respect to time, derivatives of the amplitude with respect to time, derivatives of the phases with respect to time, a phasor of the derivative of Q with respect to I, second or higher derivatives of the derived data with respect to time and / or to another derived quantity and / or Q data and / or I data, binarizations of the derived data, binarizations of Q data, binarizations of the I data,moving averages of the generated data, moving averages of the Q data, moving averages of the I data, Q data filtered by frequency filters, I data filtered by frequency filters, generated data filtered by frequency filters, derived data from past time points, Q data from past time points and / or I data from past time points.,

[0051] One or more of the following data can be used to particular advantage and has, among other things, the following advantages:

[0052] A derivative of Q with respect to I can help determine the phase shift between I and Q signals.

[0053] A time derivative of I can help identify changes in the signal's amplitude over time. It correlates directly with flow.

[0054] A time derivative of Q can help identify changes in the signal's phase over time. This value correlates directly with the flow.

[0055] An amplitude of the received alternating field can be a direct measurement of the strength of the received signal.

[0056] A phase of the received alternating field can help to determine the time shift of the received signal compared to a reference signal.

[0057] An amplitude derived from the time derivative of Q and the time derivative of I can help determine the rate of change of the signal over time. It correlates directly with the flow.

[0058] A phase from the time derivative of Q and the time derivative of I can help determine the rate of change of the signal's phase over time.

[0059] The time derivative of the amplitude can help identify changes in the signal's amplitude over time. It correlates directly with flow.

[0060] The time derivatives of the phases can help identify changes in the signal's phase over time. They directly correlate with flow.

[0061] A phasor of the derivative of Q with respect to I can provide an alternative representation of the signal as a vector in phase space.

[0062] Second or higher derivatives of the derived data with respect to time and / or another derived quantity and / or Q data and / or I data can help to capture higher orders of change in the signal.

[0063] Binarization of the derived data can help translate the signal into simpler, binary formats, leading to a reduction in complexity.

[0064] Binarization of Q data can help translate the signal into simpler, binary formats, leading to a reduction in complexity.

[0065] Binarization of the I data can help translate the signal into simpler, binary formats, leading to a reduction in complexity.

[0066] Moving averages of the generated data, moving averages of the Q data, and / or moving averages of the I data can help smooth the signal and reduce noise.

[0067] Frequency filtered Q data, frequency filtered I data, and / or frequency filtered generated data can help reduce unwanted frequency components in the signal.

[0068] Derived data from past time points, Q data from past time points, and / or I data from past time points can help track and predict changes in the signal over time.

[0069] In an advantageous embodiment of the invention, at least one input feature can be extracted from the demodulation data and / or from data derived therefrom, which is then fed to the statistical model. The at least one input feature can advantageously be a linear combination and / or a non-linear combination of the multidimensional demodulation data and / or the data derived therefrom. Such input features can also themselves be considered data derived from the demodulation data.

[0070] Advantageously, the statistical model can contain or be one or more selected from the following group: an artificial neural network, an artificial feed-forward neural network, an artificial recurrent neural network, an artificial convolutional network or a combination of the aforementioned artificial neural networks, a decision tree, a decision forest, a naive Bayes model, a nearest neighbor model, a support vector machine, a linear regression model, a non-linear regression model, an adaptive filter, a Kalman filter, a gamma filter, a Markov model, a model from a linear discriminant analysis, a model from a non-linear discriminant analysis, a model from an independent component analysis, a model from a principal component analysis and / or an ensemble model from one or more of the aforementioned models.It should be noted that in the above list, those models named with English names are commonly referred to by this English name.

[0071] For example, the following statistical models can be used with the following advantages:

[0072] Artificial neural network: Can model complex non-linear relationships in the data and thus provide high prediction accuracy for time series-based forecasts.

[0073] Artificial feed-forward neural network: A neural network in which signals flow in only one direction, without feedback. It's useful for creating simple non-linear relationships, such as predicting simple time series.

[0074] Artificial recurrent neural network: A neural network in which the output of the network also serves as input for later time points in order to access past states and is thus suitable for time series forecasting.

[0075] Convolutional neural network: A model designed to process signals by extracting and merging local features. This allows feature extraction to be performed implicitly, making it well-suited for complex time series.

[0076] Decision tree: Allows you to identify important variables and interaction patterns in the data, which is especially useful when the relationship between the variables is non-linear.

[0077] Decision Forest: Similar to a decision tree, but by using multiple decision trees, ensemble methods can be used to achieve better prediction accuracy.

[0078] Naive Bayes model: Particularly suitable for problems with a large number of variables because it makes a simple assumption about the independence of the variables.

[0079] Nearest Neighbor Model: Can be able to identify patterns and relationships in the data well, especially when the data is high-dimensional.

[0080] Support vector machine: Can also be useful in cases with high-dimensional data, especially when the data is nonlinear. Linear regression model: Particularly well-suited for problems with linear dependence between variables and can be trained quickly.

[0081] Non-linear regression model: Can also be useful in cases with non-linear dependence between variables to capture complex patterns and relationships

[0082] Adaptive filter: A model capable of adapting to changing environmental conditions and used for signal processing and prediction.

[0083] Kalman filter: A stochastic model used to estimate states of a dynamic system and is suitable for time series forecasting.

[0084] Gamma filter: Can reduce noise in the data while providing an effective trade-off between prediction accuracy and response time.

[0085] Markov model: A model based on the assumption that the future state depends only on the present state and is suitable for forecasting time series.

[0086] A model from linear discriminant analysis: Can be useful for improving category prediction in time series data.

[0087] A model from a non-linear discriminant analysis: This can provide higher predictive accuracy than linear discriminant analysis models and is particularly useful when the relationships between the variables are non-linear.

[0088] A model from independent components analysis: Can help identify hidden variables and patterns in the data by transforming the variables so that they become statistically independent. A model from principal components analysis: Can help reduce complex relationships between variables in the data and identify key trends or patterns in the data, which can be particularly beneficial when modeling time series. Furthermore, PCA can also help reduce noise in the data, thus improving the accuracy of predictions.

[0089] For neural networks (NNs), the following approaches are among those that can be considered, each with its own advantages. Some of these approaches can also be used in combination and / or consecutively. For example, an encoder-decoder network can be constructed to be Bayesian and federated.

[0090] Bayesian Neural Networks: A neural network based on Bayesian probability to quantify and reduce uncertainty in predictions by using a distribution of weights instead of a single weight. It is a good choice for modeling time series when uncertainty in predictions must be considered. It is therefore very suitable in the medical context.

[0091] Contributed neural networks: A group of neural networks trained in a decentralized manner across multiple devices to address privacy and security concerns when using sensitive data. Suitable for processing large amounts of data collected in distributed environments. This is useful for future applications where there are multiple decentralized chest monitors that all collect data and train a neural network. A good step toward generalization.

[0092] Encoder-decoder networks: A type of NN that encodes and decode input data to perform a specific task, such as prediction or translation. They are suitable for time series modeling because they are able to capture complex patterns in the data and generate predictions based on those patterns.

[0093] Attention-Based Networks: A neural network that optimizes the weighting of input variables through an attention function to identify and focus on important information while suppressing irrelevant information. They are suitable for processing long time series with many variables because they can selectively direct attention to specific variables or patterns.

[0094] Generative models (Adversarial Networks (GANs) and Variational Autoencoders (VAEs)): NNs capable of generating new data that follow the statistics of the input data. They are suitable for generating simulated time series that match the properties of real time series. They are good for data augmentation and thus a step toward generalization. The generated data can then be fed back into another NN.

[0095] Reinforcement Learning Networks: A neural network that interacts with an environment and learns how to perform a task through trial and error, receiving positive and negative rewards. Suitable for processing time-dependent data because it is capable of making decisions based on history and current state.

[0096] LSTMs (Special Form of RNNs): An RNN with a special architecture capable of capturing and storing long-standing dependencies in the data. Suitable for modeling time series because of their ability to effectively store and process past states. Optionally, an LSTM can be combined multiple times in the encoder-decoder architecture. An LSTM can also optionally follow a CNN consecutively (e.g., for feature extraction).

[0097] Deep Belief Networks: A type of NN that has multiple layers of connected neurons used for feature extraction and output data prediction. Suitable for time series modeling because of its ability to capture and exploit complex dependencies between data.

[0098] Echo State Networks (RNN type): A special type of RNNs that have a random, fixed connection between neurons in the inner layer. They are suitable for modeling time series because they are able to effectively store and process past states, which is crucial for predicting future states.

[0099] Gated Recurrent Unit (RNN type): A special type of RNN that allows for the control of forgetting information from previous time points by using a gating mechanism in the architecture. They are suitable for modeling time series because they are able to retain important information from the past while discarding irrelevant information.

[0100] Transformer Networks: An architecture based on the use of multi-head attention mechanisms to model relationships between different parts of the input variable. They are suitable for processing long time series because they are able to focus selective attention on important parts of the input variable and suppress irrelevant information.

[0101] One or more of the following models can also be used optionally:

[0102] Autoregressive Integrated Moving Average (ARIMA): A statistical model based on the combination of autoregressive and moving average processes, capable of modeling nonlinear relationships in a time series. They are suitable, for example, for predicting short-term trends in time series, such as arrhythmias.

[0103] Seasonal Autoregressive Integrated Moving Average (SARIMA): An extension of ARIMA that considers the seasonal component in a time series and is capable of capturing seasonal patterns in the data series. They are suitable for predicting seasonal trends in time series, such as chronic diseases.

[0104] Exponential Smoothing (ETS): A statistical model based on the estimation of trends and seasonal effects and capable of modeling both short-term and long-term trends in a time series. This is suitable for predicting short-term and long-term trends in time series. Application, for example, as described above. Vector Autoregression (VAR): A statistical model that can model multiple variables in a time series by taking into account the dependencies between the variables. Suitable for modeling time series with multiple variables and for predicting relationships between the variables. In this case, the dependency is Q / l ■

[0105] A particularly advantageous embodiment of the invention provides that at least one feature of the sequence and / or the temporal representation of the sequence is generated in real time. "Real time" is understood in particular to mean that the time required for processing in the generation step, and advantageously also in the demodulation step, is shorter than a predetermined threshold. The threshold is selected to suit the given application. In particular, it can be selected so that it is smaller than the smallest time scale of the application in question. Advantageously, rapid generation or processing in the generation step can be achieved by the following measures, which can be advantageous for real-time evaluation, but also for other purposes:

[0106] In an advantageous embodiment, a statistical model, e.g., a deep neural network, with an inference latency of less than or equal to 200 ms, preferably less than or equal to 150 ms, preferably less than or equal to 100 ms, preferably less than or equal to 50 ms, preferably less than or equal to 30 ms, preferably less than or equal to 20 ms, preferably less than or equal to 10 ms, preferably less than or equal to 5 ms, preferably less than or equal to 1 ms can be used. 0 ms can, of course, be considered the lower limit, since a minimum inference latency is normally not required.

[0107] Advantageously, a deep neural network with a depth of less than or equal to 150, preferably less than or equal to 50, preferably less than or equal to 20, preferably less than or equal to 10, and preferably less than or equal to 3 layers can be used. At the same time, however, the network should be sufficiently complex to be able to perform the relevant generation with sufficient accuracy. The model preferably has two or more layers. In general, for fast or real-time processing, the inference latency is the more important parameter for characterizing the network.

[0108] For fast generation, especially for real-time generation in the generation step, one or more of the following optimization techniques can be advantageously used:

[0109] Quantization: It can reduce the accuracy of the weights and activations in the network, which can significantly reduce the memory requirements and inference time while maintaining accuracy.

[0110] In an advantageous embodiment, pruning can be used. This allows unimportant connections or neurons to be removed from the network, thereby reducing the size of the network and significantly reducing memory requirements and inference time while maintaining accuracy.

[0111] In an advantageous embodiment, knowledge distillation (transfer learning) can be used. This allows a smaller network to be trained to mimic the behavior of a larger, more complex network. This can significantly reduce memory requirements and inference time while maintaining accuracy.

[0112] In an advantageous embodiment, compression algorithms can be used. For example, Huffman coding can be used advantageously. This is a lossless compression technique that encodes frequently used symbols, in this case, neuron weights, in a dataset with shorter bit sequences to reduce the overall size of the network, thereby significantly reducing memory requirements and inference time while maintaining accuracy.

[0113] In a further advantageous embodiment, Lempel-Ziv compression can be used. This is a lossless compression technique that detects redundant data in a dataset—in this case, neuron weights—and replaces it with references to previously occurring data to reduce the overall size of the network, thereby significantly reducing memory requirements and inference time while maintaining accuracy.

[0114] Learning-based compression (see, for example, https: / / ar-xiv.org / pdf / 1802.03494.pdf) can also be used advantageously. This is a data compression method that uses reinforcement learning to generate a compression policy that determines which parts of the model can be reduced. This can significantly reduce memory requirements and inference time while maintaining accuracy.

[0115] In an advantageous embodiment, NAS (Neural Architecture Search) can be used for generation. This approach uses automated techniques to search for optimal network architectures that strike a balance between size, i.e., inference latency, and accuracy. In contrast to the other compression techniques mentioned, which attempt to optimize a pre-trained statistical model, NAS directly searches for an optimal architecture that meets the user's needs—in this case, inference latency and accuracy.

[0116] The generation step can also be accelerated by parallelizing the model. This includes running the calculations on multiple processors or GPUs to increase prediction speed.

[0117] It is also possible to use multiple models in parallel within the ensemble. Inference can run in parallel using several smaller, lower-latency models, followed by majority voting or averaging of the inferences.

[0118] Furthermore, the sampling rate and / or the dimension of the values ​​of the device or method according to the invention could be adaptively changed to adaptively change the complexity of the input data and thereby adapt the inference latency of the statistical model to the situation. For example, one device according to the invention (also called a "thorax monitor") can generate 2 signals, two devices can generate 4 signals, or 3 devices can generate 6 signals.

[0119] In an advantageous embodiment of the invention, the alternating electromagnetic field can be radiated into the object via at least one transmitting structure by a signal generator generating an alternating voltage, which is applied to the at least one transmitting structure. Furthermore, the alternating field can be received by at least one receiving structure after it has passed through the object.

[0120] Advantageously, the alternating voltage and / or the radiated alternating field is generated at a frequency of greater than or equal to 10 MHz, preferably greater than or equal to 30 MHz, preferably greater than or equal to 100 MHz and / or less than or equal to 1,000 MHz, preferably less than or equal to 500 MHz, preferably less than or equal to 300 MHz.

[0121] It is particularly advantageous if the at least one transmitting structure and the at least one receiving structure are arranged on the same side of the object. For example, the transmitting structure and the receiving structure can be arranged in a backrest, at least one siderest, and / or at least one armrest of a seat or in a reclining surface. In this way, a measurement is possible, for example, when a person is sitting on a seat or lying on a recliner.

[0122] Advantageously, the at least one transmitting structure and the at least one receiving structure can be arranged in textiles worn on the body and / or close to the body. These can be, for example, jackets or blankets. Also possible are arrangements in sports equipment, in medical devices such as a wheelchair, in household appliances such as a stove for monitoring the water level, or in industrial plants, for example, for monitoring the condition of machines and systems.

[0123] In an advantageous embodiment of the invention, the object can be a human body. In this case, the at least one transmitting structure can be arranged, for example, in the immediate vicinity of and / or on a thorax and / or an abdomen of the body. The at least one receiving structure can likewise be arranged in the immediate vicinity of and / or on the thorax and / or the abdomen. In this case, an arrangement on the corresponding body part should preferably be regarded as a constellation in which the transmitting structure or the receiving structure is arranged in sufficient proximity to the corresponding body part so that measurement with the desired accuracy is possible. It is of course also possible to provide devices which hold the transmitting structure or the receiving structure on the corresponding body part. Alternatively, the transmitting structure orThe receiving structure can also be arranged in a fixed location, and the person can move their body into a corresponding position near the transmitting structure or the receiving structure. This configuration should also be considered optionally as the arrangement of the transmitting structure or the receiving structure on the corresponding body part.

[0124] In an advantageous embodiment, the object to be measured is located in the near field of at least one transmitting and / or receiving structure. The near field can be understood as the area around the transmitting and / or receiving structure in which the alternating electromagnetic field has not yet completely detached itself from the transmitting and / or receiving structure and is not yet radiated into free space as a plane wave. In the near field, the alternating electromagnetic field is still coupled to the transmitting and / or receiving structure. The IEEE definition of the near field can also optionally be used as a basis. Here, the outer limit is defined as the distance X / (2n) from the antenna surface, where X is the wavelength in free space (cf. IEEE Standard for Definitions of Terms for Antennas, IEEE Std 145-2013). Consequently, the distance between the transmitting and / or receiving structure and the object to be measured is advantageously less than or equal to X / (2n).

[0125] Particularly preferably, the at least one transmitting structure and / or the at least one receiving structure is arranged laterally, dorsally or ventral to the abdomen and / or laterally, dorsally or ventral to the thorax of the body.

[0126] In an advantageous development of the invention, the alternating electromagnetic field can be transmitted and received via at least two transmit / receive pairs. Each of the transmit / receive pairs can contain a transmitting structure and a receiving structure. The alternating field transmitted by a transmitter of one of the transmit / receive pairs and received by the receiver of this pair can be viewed as a single alternating field, and the alternating field transmitted by the transmitters and receivers of other transmit / receive pairs can each be viewed as separate alternating fields. Each transmit / receive pair can thus perform the transmission and reception processes with its own alternating field.On the other hand, since the totality of all alternating fields generated in this way can also be viewed as a common alternating field with a specific spatial distribution, this can also be formulated such that the electromagnetic alternating field is radiated into the object via the at least two transmit / receive pairs and is received after it has passed through the object. This configuration allows for the suppression of interference variables and / or an improvement in the estimation or determination of the characteristic of the process below the temporal representation of the process by increasing the dimensionality of the demodulation data or the model. The signal quality, robustness, and reproducibility can also be improved in this way. The use of an unequal number of receive and transmit structures is also possible. The use of more receive structures than transmit structures is particularly advantageous in this case.For example, the at least one transmitting structure can radiate the alternating field into the object, and the at least two receiving structures can receive the alternating field after it has passed through the object. Advantageously, the at least two receiving structures can be arranged at different positions around the object or on the object.

[0127] In an advantageous embodiment of the invention, the method can be conducted in such a way that it continuously monitors the process and continuously generates at least one feature of the process and / or a temporal representation of the process. Continuous is defined here as any regular repetition of the method, i.e., execution of the method at discrete points in time, at quasi-continuous points in time, or continuously in the sense of digital data processing.

[0128] The characteristic determined in this way or the representation of the process determined in this way can be compared with previous values ​​of the characteristic or previous representations of the process. In particular, an event can be triggered if the most recently determined characteristic deviates from a previous value of the characteristic by more than a predetermined threshold and / or if a characteristic of the representation of the process deviates significantly from an earlier embodiment of this characteristic of the process. An alarm, for example, can be issued as an event. Possible applications for such a process control include, for example, monitoring vital functions, where a change can indicate a critical condition.

[0129] For this purpose, the method is preferably executed regularly and repeatedly, and a value of the at least one process is regularly and repeatedly compared with at least one previous value of the feature. A warning event can then be triggered if the value of the feature deviates from the previous value of the feature by more than a predetermined threshold.

[0130] In a particularly advantageous embodiment, parameters of the statistical model and / or the statistical model can be adjusted based on a difference between the value of the generated feature and the corresponding previous value of the feature if the value of the generated feature deviates from the previous value by less than the threshold. In this way, the model can adapt to the current state of the person, for example, so that, for example, an alert event is only triggered if the deviation between the last value of the feature and the corresponding previous values ​​is greater than the threshold.

[0131] In an advantageous embodiment of the invention, the at least one feature of the process can be a trigger event and / or an exceeding or falling below a predetermined threshold value. If the process is, for example, a respiratory volume curve or respiratory flow curve, and the method is intended to support the breathing process, for example, the trigger event can be considered to be the exceeding of a certain respiratory volume or respiratory flow during inhalation. This trigger event can then, for example, trigger an inhalation support process. Advantageously, the at least one feature of the process and / or the temporal representation of the process can describe the process continuously, quasi-continuously, or at a plurality of, preferably equidistant, points in time. In particular, the at least one feature and / or the temporal representation can describe the process at regularly recurring points in time.

[0132] In many embodiments of the invention, it can be advantageous if the statistical model is trained before the actual measurement of the process begins. In this case, training would therefore preferably take place before the alternating electromagnetic field is radiated. The statistical model can be trained in a training process in which an alternating electromagnetic training field is radiated into at least one training object. Advantageously, the alternating training field can share one or more properties with the alternating electromagnetic field later radiated for measurement, or particularly preferably can be identical. The training object can advantageously be an object that shares one or more properties of interest with the object in which the process is to be measured. For example, the training object can be a living being of the same species as the object to be measured.For individualized procedure management, the training object can also be the object itself in which the process is to be measured. If the object is a living being, such as a human or an animal, the training object can be an object that is representative, for example, of a cohort to which the living being to be measured later falls.

[0133] As described above, a federated approach is advantageous here. There are, among others, the following options for creating a federated trained model from several individual models:

[0134] Model fusion: A method in which multiple models are combined into a single model to achieve better predictions. This is useful because different models have different strengths and weaknesses, and combining these models can lead to a more powerful model. Transfer learning: A method in which a pre-trained model is used as a starting point for training a new model to reduce the amount of training data required. This is useful because it can reduce the effort required to train models and improve model accuracy.

[0135] Ensemble methods: A method in which multiple models are combined into a group and decisions are made by majority vote to achieve better predictions. This is useful because the combination of multiple models can lead to better forecast accuracy than any single model.

[0136] Knowledge Distillation: A method that uses a pre-trained model as a teacher to train a smaller model as a student. This can be useful because it can help a smaller model achieve similar performance to the larger model while requiring fewer computational resources.

[0137] For the training process, the alternating electromagnetic training field is radiated into at least one training object, and the alternating training field is then received after it has passed through the at least one training object. The received alternating training field can then be demodulated to generate multi-dimensional demodulation data. The multi-dimensional demodulation data and / or training data derived from it can then be fed to the statistical model. The statistical model can then generate at least one training feature of the process in a time-dependent manner from the supplied multi-dimensional demodulation data and / or the training data derived from it. The at least one training feature can then be compared in a time-dependent manner with a measured feature of the process. The measured feature can, for example, be generated by a direct measurement with a suitable measuring device or sensor.The statistical model and / or parameters of the statistical model can then be adjusted based on the result of this comparison. The adjustment can preferably be directed in such a way that the difference between the training feature and the measured feature is reduced.

[0138] In an optional embodiment, it may also be possible to use the following instead of data generated by the thorax monitor itself:

[0139] Generative models (adversarial networks (GANs) and variational autoencoders (VAEs)) are neural networks capable of generating new data that follow the statistics of the input data. These can be suitable for generating simulated time series that match the properties of real time series. This is useful for data augmentation and thus a step toward generalization. The generated data can then be fed back into another neural network.

[0140] Data augmentation is a technique to increase the quantity and diversity of training data by creating artificial variations:

[0141] As already mentioned above, the following networks can also be used for data augmentation:

[0142] Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU): LSTM and GRU are special types of recurrent neural networks (RNNs) that are particularly well-suited for modeling sequential data such as time series. They can detect patterns in the data and generate new data accordingly.

[0143] Other variants from stochastics that can be used optionally are:

[0144] Markov Chain Monte Carlo (MCMC) method: MCMC is a stochastic method for generating data from an unknown distribution using a Markov chain that converges to the target distribution. MCMC methods such as Metropolis-Hastings and Gibbs sampling can be used to generate time series that follow a specific distribution. Manual or automatic data augmentation:

[0145] Phase Shift (Rotation): o The time series is cyclically shifted along the time axis to generate new variations in the data. This can help train the model for different phase shifts within the time series data.

[0146] Smooth or dynamic shift: Data can be multiplied by a factor to change the amplitude of the signals. This can help train the model for different amplitude ratios.

[0147] Add noise: o Random noise (e.g. Gaussian noise) can be added to the time series data to increase the robustness of the model to noise in the input data.

[0148] Time Warping: o The speed at which the time series runs can be changed by stretching or compressing data points along the time axis.

[0149] Mirroring: o Mirroring the time series along the time or amplitude axis to create variations in the order or direction of the time series data.

[0150] Frequency-domain manipulation: Transforms in the frequency domain, e.g., using the Fourier transform. The frequency spectrum can also be manipulated by amplifying, attenuating, or shifting specific frequency bands. This can help train the model for different frequency characteristics.

[0151] Random Permutations: o Random permutations of time series segments can be performed. This can help train the model for different orderings of events within the time series.

[0152] Regarding the alternating training field, the training object, the received alternating training field, the demodulation data, and the training data derived from it, the statements made regarding the alternating electromagnetic field, the object, the demodulation data, the derived data, and the features used for the measurement apply analogously. Indeed, it is advantageous if the training process is carried out identically to the measurement process, as this achieves the best possible agreement between the measurement result and the values ​​of the feature or process that would be generated by direct measurement with a suitable measuring device.

[0153] Advantageously, the training process can be performed on a large number of training objects that are representative of the object. For example, the training process can be performed on a large number of living beings in a cohort, including the living being in which the process is to be measured. A large number of training objects makes the model more robust, and a generalized statistical model can be created. However, it is also advantageous to use the object on which the measurement is to be performed as the training object. The training process is then performed on this object.

[0154] The method according to the invention is particularly suitable for controlling a ventilation process. For this purpose, a breathing process can be measured as the process to be measured. The onset of a spontaneous breathing process, for example, can then be determined as at least one feature of the process. The ventilation process can then be controlled in such a way that no ventilation occurs against spontaneous breathing.

[0155] According to the invention, a device for measuring at least one process in an object is also specified. The statements regarding the method apply analogously to the object. The device has at least one transmitting structure with which an alternating electromagnetic field can be radiated into the object. A signal generator generates an alternating voltage, which is applied to the at least one transmitting structure. The alternating voltage can be sinusoidal or contain a superposition of several frequencies. The device also has at least one receiving structure with which the alternating field can be received after it has passed through the object. The statements regarding the penetration depth apply analogously.

[0156] The device according to the invention also comprises a demodulation unit with which the received alternating field can be demodulated to generate multidimensional demodulation data. Furthermore, the device comprises a modeling unit to which the demodulation data and / or data derived therefrom can be supplied, and with which, by means of a statistical model, at least one feature of a sequence and / or a temporal representation or simulation of the sequence can be generated from the supplied multidimensional demodulation data and / or the data derived therefrom, in a time-dependent manner.

[0157] Preferably, a method as described above can be carried out with the device.

[0158] The device can advantageously comprise an I / Q demodulator, into which the alternating voltage applied to the transmitting structure is fed, and into which a received signal generated by the receiving structure from the received alternating field can be introduced. The I / Q demodulator can then be configured to perform I / Q demodulation of the received signal.

[0159] In an advantageous embodiment, the at least one transmitting structure and the at least one receiving structure can be arranged next to one another and / or one below the other and / or diagonally to one another in a contact surface, preferably a backrest or a reclining surface. For this purpose, the transmitting structure and / or the receiving structure can be configured, for example, as flat coils.

[0160] Advantageously, the device according to the invention or a device for carrying out the method according to the invention can have one or more transmitting and / or receiving antennas as transmitting and / or receiving structures. The number of I / Q signal pairs can then result from the number of transmit signals emitted by one or more antennas (e.g., split across multiple frequency bands or by different signal modulations) multiplied by the number of receiving antennas. The different transmit signals can be transmitted identically on all transmitting antennas, or, for example, each transmitting antenna can transmit its own separate transmit signal that is different from the other transmitting antennas. Any combination of these is also possible.

[0161] Advantageously, the device can have N transmitting antennas and IVI receiving antennas, where N is a natural number greater than or equal to 1 and IVI is a natural number greater than or equal to 2, wherein the N transmitting antennas are supplied with alternating voltages L of different frequencies, where L is a natural number greater than or equal to 1. Preferably, L can also be less than or equal to N. If L is equal to N, each transmitting antenna can be supplied with an alternating voltage of different frequency. If L is less than N, at least one of the transmitting antennas can be supplied with alternating voltages of several different frequencies.

[0162] In an advantageous variant, no transmitting antenna is supplied with several alternating voltages of different frequencies, so each transmitting antenna is supplied with an alternating voltage of exactly one frequency.

[0163] In a further advantageous variant, each of the receiving antennas can be configured to receive exactly one of the L frequencies or to receive several of the L frequencies. If each receiving antenna is configured to receive exactly one frequency, the number IVI of receiving antennas is preferably equal to the number L of different frequencies of the radiated alternating voltages. If one or more of the receiving antennas are configured to receive multiple frequencies, the number of receiving antennas can be less than L.

[0164] Optionally, the number L of transmit signals with different transmit frequencies multiplied by the number IVI of receive antennas can yield the maximum number of I / Q signal pairs for the evaluation / algorithm. However, fewer I / Q signals can also be used, for example, if not all receive antennas are sensitive to all transmit frequencies. Optionally, a very broadband receive antenna can also be used that does not have a specific frequency selectivity.

[0165] In particular, the following combinations of antennas in the transmitting and receiving structure are advantageous:

[0166] One transmitting antenna and three receiving antennas. This results in three I / Q signal pairs.

[0167] Two transmitting antennas, both preferably supplied with the same transmit signal, i.e., the same AC voltage, and five receiving antennas. This results in five I / Q data pairs.

[0168] Two transmit antennas (preferably with different transmit signals, e.g., at different frequencies) and three receive antennas. This results in six I / Q signal pairs.

[0169] The advantages of these multiple transmit / receive antennas lie in data redundancy, if the signal quality is better or worse due to different antenna positioning for each antenna pair (the algorithm can select the best transmit / receive signal combination). Alternatively, the algorithm can use all or any combination of the I / Q signal pairs to more specifically calculate out interference effects (e.g., motion artifacts) or to separate the actual useful information (respiration) from the interference effects and thus filter it out. In an advantageous embodiment, the selection of the I / Q signal pairs used can also be preceded by a signal quality estimation preceding the evaluation algorithm, on the basis of which the most suitable signal pairs are selected for further processing.

[0170] In particular, it may be advantageous if the at least one transmitting structure and / or the at least one receiving structure has or is a broadband antenna. In this case, the bandwidth of the antenna can preferably be greater than or equal to one-eighth of the carrier frequency, preferably greater than or equal to one-quarter of the carrier frequency, preferably greater than or equal to half the carrier frequency, and / or the bandwidth can be less than or equal to the carrier frequency.

[0171] In an advantageous embodiment, the device according to the invention can comprise an adaptation device with which a transmission frequency band or a transmission frequency of the transmission structure and / or a reception frequency band or a reception frequency of the reception structure can be adapted. Such an adaptation device can advantageously comprise or be at least one capacitance diode. In this way, the at least one transmission structure and / or at least one reception structure can be adapted in such a way that it is optimally suited for the intended measurement.

[0172] The transmitting structure and / or the receiving structure can each be connected to the signal generator, the demodulation unit, or another component of the device via a cable. Advantageously, these cables can be positively balanced, for example, by means of a terminating resistor together with an impedance converter and / or by means of at least one current-compensated choke and / or at least one ferrite.

[0173] Advantageously, the at least one transmitting structure and / or the at least one receiving structure can be supplied with the alternating voltage at a first connection and with an alternating voltage shifted by 180° relative to the alternating voltage at a second connection. For this purpose, the device can preferably have at least one balun that generates two alternating voltages from the first alternating voltage that have a phase shift of 180° from one another. The input alternating voltage can be the one generated by the signal generator. As already described with regard to the method, it can be advantageous if the measuring device has a plurality of transmitting structures, i.e. two, three, four or more than four transmitting structures, and / or a plurality of receiving structures, i.e. two, three, four or more than four receiving structures. This allows a more precise measurement result to be achieved.

[0174] Advantageously, the signal generator can be configured to generate as alternating voltage a superposition of a plurality of sinusoidal alternating voltages, which can advantageously also have time-varying frequencies.

[0175] The method and device presented here enable a timely derivation of dynamic states of the target system. With the present algorithm, the data of a non-invasive, non-obstructive and scalable (simultaneous or serial use of multiple sensor sub-elements at a spatial distance) and surface and volume-sensitive measurement technology (alternating field penetrates at least partially into the target structure) based on the time-dependent complex convolution of at least two electromagnetic fields (sent and received) are processed at runtime, interpreted and converted, for example, into cardiothoracic vital data or other physiodynamic data of a target structure. For this purpose, the system can be trained, at least initially, to compare the changes in the measurement system data resulting from the physiodynamic changes of the target structure or a generic model of the target structure with reference system data (e.g.cardiothoracic vital data). The following may be possible, for example:

[0176] Detection of spontaneous breathing in mechanically ventilated patients

[0177] Recording of the expiratory and inspiratory phases Monitoring of patients

[0178] Alarm in case of failure of specific vital parameters Contactless derivation of ECG curves

[0179] Determination of heart rate variability and / or respiratory volumes volume flows and / or changes

[0180] Prediction of complications during surgery by measuring vital signs and monitoring data streams, analysis of cardiovascular data to predict cardiovascular diseases,

[0181] Tracking patient progress in rehabilitation.

[0182] The invention can also be used for one or more of the following applications:

[0183] Fitness tracking: Monitoring physical activities and analyzing fitness or training progress.

[0184] In sports medicine, for example, for performance analysis and training planning.

[0185] In health monitoring, for example to monitor chronic diseases, e.g. long-term ECG.

[0186] For respiratory gating during radiotherapy of patients.

[0187] In stress management, for example to monitor stress levels in humans, but also in mice, for example.

[0188] In telemedicine, for example, for remote monitoring of elderly or chronically ill patients.

[0189] The invention will be explained below by way of example with reference to a few figures. Like reference numerals denote like or corresponding features. The features described in the examples can also be combined between the examples and implemented independently of the respective example.

[0190] It shows:

[0191] Fig. 1 is a schematic representation of an apparatus for carrying out a method according to the invention,

[0192] Fig. 2 is a schematic representation of an exemplary sequence of a method according to the invention,

[0193] Fig. 3 exemplary curves of I-data and Q-data as well as of derivatives dQ / dl during a breathing process,

[0194] Fig. 4 exemplary demodulation data and data generated from them,

[0195] Fig. 5 shows an example of preprocessing of I and Q data, Fig. 6 shows an example of a training process for a statistical model,

[0196] Fig. 7 shows an exemplary configuration for obtaining data for training in Fig. 6,

[0197] Fig. 8 shows an exemplary model used for classification,

[0198] Fig. 9 an exemplary model used for regression,

[0199] Fig. 10 shows an example of using a neural network as a statistical model for estimating respiratory flow,

[0200] Fig. 11 shows the use of an FIR filter as a statistical model for estimating respiratory flow,

[0201] Fig. 12 Combinations of an electromagnetic transmission measurement to a soft sensor system,

[0202] Fig. 13 an exemplary circuit of a transmitting antenna with a tunable matching network,

[0203] Fig. 14 shows an example of using buffer circuits with termination to bridge a cable without mismatch, and

[0204] Fig. 15 shows a detailed example of a process according to the invention.

[0205] Fig. 1 schematically shows a device according to the invention for measuring a process in an object 5, with which the inventive method for measuring a process in an object can also be carried out. The device has a transmitting structure 4, with which an alternating electromagnetic field can be radiated into the object 5 by a signal generator 2, here a frequency generator 2, generating an alternating voltage which is applied to the transmitting structure 4. The device also has a receiving structure 6 with which the alternating field can be received after it has passed through the object 5. The device according to the invention also has a demodulation unit 8 with which the received alternating field can be demodulated to generate multi-dimensional demodulation data. In addition, the device has a control unit 1, which functions in particular as a modeling unit.The demodulation data and / or the data generated from them are fed to this modeling unit 1. Using a statistical model, the modeling unit 1 can then generate at least one feature of a process and / or a temporal representation of the process in a time-dependent manner from the supplied multidimensional demodulation data and / or the data generated from them.

[0206] In the example shown in Fig. 1, the device also comprises an optional transmit amplifier 3, to which the signal generated by the signal generator is fed, which amplifies it, and then outputs it to the transmit structure 4. Furthermore, in the example shown, the device comprises an optional receive amplifier 7, into which the signal received by the receive structure is fed, which amplifies it, and feeds the amplified signal to the demodulation device 8.

[0207] In the device shown, the alternating voltage generated by the frequency generator 2 can also be fed to the demodulation unit 8, which can thereby carry out an I / Q demodulation of the received signal.

[0208] The control unit 1 can handle the configuration of the entire system. For example, it can set the transmission frequency and transmission power. The signal or frequency generator 2 can generate an alternating voltage with the configured frequency and optionally power. For this purpose, it has two synchronous outputs. The first output is connected to the optional transmission amplifier 3 or directly to the transmission structure. The second output serves as a reference path for the I / Q demodulator 8, which acts as a receiver. The alternating voltage applied to the transmission structure 4 is radiated by it as an electromagnetic wave or electromagnetic field. The wave or field is transmitted through the object 5, for example a patient, and after transmission is converted back into an alternating voltage by the reception structure 6. During transmission, the electromagnetic wave orthe electromagnetic field is modulated by processes inside the patient, such as cardiorespiratory activity.

[0209] The signal received by the receiving structure 6 is forwarded to the optional receiving amplifier 7 or directly to the I / Q demodulator 8. This can convert the received signal to baseband using the reference signal from the frequency generator 2. The I / Q demodulator can generate a complex output signal with an in-phase channel (abbreviated to I) and a quadrature channel (abbreviated to Q). These signals correspond to the modulation of the real and imaginary parts of the wave transmitted through the object 5. I and Q are then digitized by a data sink 1, which here is part of the control unit 1.

[0210] In other embodiments of the invention, it may also be possible to increase the number of transmitting structures 4 and receiving structures 6. This can be done symmetrically or asymmetrically. Thus, the same number of transmitting structures 4 and receiving structures 6 can be provided, or a different number. It may be advantageous to use multiple receiving structures 6 per transmitting structure 4. Mounting the transmitting and receiving structures 4, 6 in a coplanar arrangement, e.g., in a backrest or a mattress, can be particularly advantageous.

[0211] Since both the transmitting structures 4 and the receiving structures 6 can be mounted very close to the body when used on people, the transmitting structures 4 and the receiving structures 6 can experience a detuning of their resonant frequencies. To counteract this detuning and still achieve good transmission and reception performance, the following configurations of the transmitting and receiving structures can be advantageous:

[0212] It is advantageous to use a very broadband antenna or generally an antenna with a high robustness against detuning.

[0213] A tunable matching network can advantageously be used upstream of the actual transmit and receive structure. One possible design here could use variable-capacitance diodes. Fig. 13 shows an example of such a tunable matching network for the transmit case. As soon as the object enters the near field of the transmit and receive structure, the detuning of the transmit or receive structure can be compensated using control voltages. The magnitude of the receive signal, for example, can be used as a metric for the quality of the compensation.

[0214] The circuit shown in Fig. 13 is a matching network. This allows the antenna impedance to be transformed to a different impedance at the TX port. The ideal case is a power match where the impedance of the circuit components coming from the TX is equal to the impedance of the matching network with the antenna. This compensates for antenna mismatches that can lead to power losses and reflections in the signal generator or IQ demodulator.

[0215] The matching network shown in Fig. 13 corresponds to an n-circuit. With this circuit topology, theoretically any antenna impedance can be mapped to any impedance at TX. The basic impedance matching is achieved by the inductance L_antenna and the two capacitances Cpl and Cp2. However, as soon as the antenna impedance changes, e.g. due to an object in the near field of the antenna, the matching network consisting of L_antenna, Cpl and Cp2 is no longer correct. For this reason, additional varactor diodes Dl, D2 and D3 are introduced. By reducing the control voltages Vtunel and Vtune2, the capacitance of the varactor diodes Dl and D2 can be increased and thus the effective parallel capacitances of the matching network can be increased, or increased by increasing the control voltages. The capacitance of D3 can be varied using the control voltage Vtune3, thus allowing the inductance of L_antenna to be varied for a specific operating frequency.

[0216] The high-frequency part of the circuit is separated from the control voltages by the inductors L_dc, which should preferably be significantly larger than L_antenna to avoid a significant impact on the matching network. To prevent the control voltages from affecting other circuit components, they are separated from them by the capacitances C_dc. The capacitances C_dc must be chosen to be much larger than the capacitances Cpl and Cp2 to avoid a significant impact on the matching network. This type of adaptable matching network is advantageous for both transmit and receive structures.

[0217] Detuning of the transmit / receive structure can lead to a mismatch between the transmit / receive amplifier or the frequency generator / IQ demodulator. With long cable lengths between the transmit / receive structure and the actual signal source or sink, the mismatch can cause standing waves. This is especially the case when using cables. These standing waves can generate artifacts, but also unwanted sensitive areas along the cable. To counteract these artifacts, the matching network described above and / or the following configurations can be advantageous.

[0218] Forced balancing can be performed between the forward and return conductors of the cable to the at least one transmitting structure and / or to the at least one receiving structure, e.g., via current-compensated chokes, by means of a terminating resistor together with an impedance converter, by means of at least one current-compensated choke and / or by means of ferrites.

[0219] It is also possible to use differential transmit and receive structures and / or baluns.

[0220] It is also possible to terminate the cables with the characteristic impedance of the cable (typically 'Q) and use buffer circuits to tap the voltage at the termination with high impedance and pass it on with low impedance to the next segment in the signal chain. This is shown as an example in Fig. 14.

[0221] In Fig. 14, matching resistors are shown together with impedance converters as a buffer circuit. At the non-inverting inputs of the operational amplifiers U1 and U2, there is a terminating resistor Rt2 and Rt4, respectively, connected to ground. Since the inputs of the operational amplifiers generally have a much higher impedance than the terminating resistors, only the terminating resistor is relevant for considering the total impedance of the terminating resistor and the operational amplifier input. The terminating resistor can thus be selected so that it exactly matches the characteristic impedance of the preceding circuit section to prevent mismatches due to impedance jumps. Thus, the terminating resistor Rt2 in Fig. 14 has the same impedance as the preceding coaxial cable with a characteristic impedance of 50 ohms.Likewise, the series resistors Rtl and Rt3 can be used to adjust the output resistances of the operational amplifiers U1 and U2 to match the characteristic impedance of the downstream circuit components. Since the impedance of the operational amplifier outputs is significantly lower than the resistors Rtl and Rt3, the total resistance of the operational amplifier output and the series resistance is formed primarily by the series resistance. Thus, the series resistor Rt3 in Fig. 14 has the same impedance as the subsequent coaxial cable with a characteristic impedance of 50 ohms. The values ​​of the terminating resistors and the series resistors can differ from each other, for example, to connect two circuit components with different impedances without mismatch.

[0222] For the described differential transmit and receive structure, the at least one transmit structure 4 and / or the at least one receive structure 6 can be supplied with a first AC voltage at a first terminal and with a second AC voltage shifted by 180° relative to the first AC voltage at a second terminal. When using a balun, this can generate two AC voltages from an input AC voltage, which have a phase shift of 180° from each other.

[0223] Fig. 2 shows an exemplary sequence of a method according to the invention for measuring at least one process in an object. The sequence can be carried out, for example, in the device shown in Fig. 1. An alternating field is received by the receiving structure 6, which in the example shown can be, for example, an RF signal. This is fed to the I / Q demodulation 21, which in the example shown generates I data and Q data per antenna. These are fed to an optional preprocessing step 22, which can include, for example, filtering and / or transformation. In the preprocessing step 22, data can be generated from the multidimensional demodulation data, here the I and Q data per antenna, which can also be referred to as derived quantities. The data generated by the preprocessing can then be fed to a statistical model 24.Additionally or alternatively, it is possible to first perform feature extraction on the generated variables to determine features of the multidimensional demodulation data and / or the data generated from it. These features can then be fed into Model 24. Such feature generation is particularly useful when the multidimensional demodulation data is unsuitable or poorly suited for direct use in Model 24.

[0224] The model generates, from the supplied multidimensional modulation data, here I and Q, and / or from the data generated from these, at least one feature of the process and / or a temporal representation of the process, e.g., as predicted values, such as a respiratory flow curve, in a time-dependent manner. This result can optionally be fed to a curve interpretation 25. In this way, for example, a real-time carrier, volumes, and / or a comparison to a reference can be generated.

[0225] Fig. 3 shows, as an example, the curves of I and Q in the left part of the image, as well as the derivatives of dQ / dl in the right part. The curve shown is characteristic of one breath. In the left part of the image, the lower left section of the curve corresponds to inhalation, and the upper right section to exhalation. During one respiratory cycle, Q and I thus run counterclockwise along the curve shown.

[0226] In the right part of the image, the dQ / dl gradient is shown as arrows on the curve shown in the left part. The change in the direction of the gradient when changing from inspiration to expiration is clearly visible. Based on the gradient, it can be determined in real time, for example, whether an inhalation or exhalation is taking place.

[0227] Fig. 4 shows examples of preprocessing of data that can be obtained in the demodulation step. This preprocessing can, for example, take place in step 22 in Fig. 2. In the sub-images of Fig. 4, a 3 x 3 pair plot of the data in its raw state is shown on the left. In the pair plots, the non-diagonal plots show the dependence of the raw data on other raw data, for example to illustrate correlation. The diagonal plots show histograms of the raw data. The right-hand sub-image shows the pair plots of the corresponding data at corresponding positions after preprocessing. The preprocessing can, for example, include filtering and / or a transformation.

[0228] Preprocessing may also include, for example, an inspection of data, including the investigation of statistical relationships between the input variables, e.g., via a generalized scatterplot matrix (GSPLOM), as shown in Fig. 4. Subsequent preprocessing may include, for example, smoothing, mean removal, and / or scaling.

[0229] Such preprocessing can be used both in measuring the process and in training the statistical model.

[0230] Fig. 5 shows an example of a preprocessing step 22. Here, I and Q are fed to the preprocessing step 22 as multidimensional demodulation data. This preprocessing step can then generate one or more of the variables shown on the right in Fig. 5 from this data. The variables shown on the right are merely examples, and some of these variables, all of these variables, and even additional variables can be generated.

[0231] Fig. 6 shows an example of a process with which the statistical model 24 can be trained in a training process. During the training process, the statistical model 24 is trained based on the data. During operation, the model 24 can then be applied to new data. In particular, a distinction can be made between two supervised learning tasks: classification and regression.

[0232] During the training process, model 24 receives, in addition to the demodulation data generated from a received alternating field, additional measured values ​​from at least one reference sensor (ground truth). The at least one reference sensor can measure the quantity to be measured or predicted using the statistical model. By comparing the measured values ​​of the reference sensors with the values ​​of the process feature predicted by the model and / or the temporal representation of the process, an error value can be calculated, for example using an optimization algorithm, in order to train model 24 accordingly. The training can, for example, be generalized using an entire representative population or it can be individual, for example, patient-specific. During patient-specific training, the reference sensors can initially be connected to the patient in order to individually train the mathematical model.After training, the reference sensors can be removed.

[0233] As shown in Fig. 6, the demodulation data or data generated from them that are to be used to measure the process in the object during operation are fed to the model 24 for training. In Fig. 6, these are, for example, I, Q, an amplitude, a phase, and gradients of the received alternating field. From these, the model 24 generates an estimated value, which is fed to the quality metric 62. The quality metric 62 compares the estimated value with the value measured by the reference sensors and generates an error value. This error value is fed to an optimization algorithm 61, which then trains the model 24 such that the agreement between the measured value and the value estimated by the model 24 is improved.

[0234] Fig. 7 shows an example of an arrangement for teaching or training the model 24 in Fig. 6. In the example shown in Fig. 7, patient-specific training is carried out on a patient 71. The patient is exposed to an alternating field via the transmitting structure 4, for example by means of the device 72 shown in Fig. 1, which is received by the receiving structure 6 after it has passed through the patient 71. From the received alternating field, as described for Fig. 2, a feature of a positive sequence and / or a temporal representation of the sequence can be generated. In the example shown, the sequence to be measured is the patient's cardiorespiratory activity. An electrocardiogram 73 and a flow meter 74 for directly measuring respiratory flow are provided as reference sensors. The sensors 73 and 74 generate the measured values, which are fed to the quality metric 62.This compares the values ​​of the interesting features of the process with those predicted by the model and thus generates the said error value.

[0235] During operation, Model 24 no longer receives any reference data, but instead estimates the characteristics of the process and / or the temporal data of the process based on the various input data. The system can act as a soft sensor here, since the target variable to be determined is not measured directly, but is estimated using correlated variables and a trained model.

[0236] Two types of estimation can be distinguished, both of which are advantageous. In one case, a classification of the demodulation data and / or the data generated from it can be performed. In this case, quantities with a discrete value range, usually binary, but sometimes also ternary, can be distinguished. For example, the data could be classified as inspiration, expiration, or optionally also as a pause. Based on the current demodulation data, it is possible to distinguish whether inspiration, expiration, or a pause is currently taking place. It is also possible to determine, for example, whether an R-wave occurred in the previous time slice.

[0237] Fig. 8 shows such a classification again schematically. The aforementioned demodulation data and / or data generated from them serve as input data for model 24. The model classifies the specific constellation of this data into classification results such as inspiration, expiration, presence of an R-wave, etc.

[0238] The second type of estimation is regression. In regression, the model generates quasi-continuous time response variables with a continuous or quasi-continuous range of values. Such a response variable can be, for example, the current respiratory flow or an ECG excitation curve.

[0239] Fig. 9 schematically shows the model as a regression. The demodulation data or data generated from them are input, and the model generates, for example, respiratory flow, an ECG, or a similar curve.

[0240] Fig. 10 shows an example of an embodiment of model 24 as a trained neural network, in this case a multi-layer perceptron, which estimates the respiratory flow based on the input variables I and Q and the features derived from I and Q. The left-hand sub-image shows an example of a course of I and Q over time, which is fed to the neural network 24 as a model, and the right-hand sub-image shows a comparison of a predicted flow with an actual flow, as determined using a reference sensor. It can be seen that the trained neural network 24 predicts the respiratory flow very well. The neural network has an input layer, four hidden layers with 200, 50, 60, and another 60 neurons, and an output layer.

[0241] Fig. 11 shows an example of respiratory flow estimation, where the statistical model 24 is a trained adaptive FIR filter. The input variables are the I-data and the Q-data, as shown in the left panel. The result is the respiratory flow, as shown in the right panel compared to the measured respiratory flow.

[0242] Fig. 12 shows another example of a soft sensor system. The method shown in Fig. 12 begins with an electromagnetic transmission measurement, for example, using the arrangement shown in Fig. 1. The data is again fed to preprocessing 22, which is also optional here. The data thus generated is then fed to model 24, which was previously trained with a reference system 124. Optionally, the data can be fed to a post-processing step 122 and ultimately output as an estimated value in output step 123. This soft sensor system, in turn, does not directly measure the output variable, but rather estimates the output variable via the mathematical model 24 using several input variables correlated with the output variable.

[0243] Fig. 15 shows a more detailed example of a method implementation according to the invention. Flow data is imported in a step 151, which is subjected to linear interpolation in a step 152. The data generated by the linear interpolation is subjected to an exemplary Butterworth bandpass filter in a step 153, whose lower limit in the example shown is 0.11 Hz and whose upper limit is 1.1 Hz. The data thus filtered are then fed to a Yeo-Johnson power transform 154, which generates the reference flow as ground truth 165. This reference flow is used as reference sensor data for training the statistical model 24, which here is a multi-layer perceptron regressor.

[0244] The reference data determined in steps 151, 152, 153, 154, and 165 are used for training purposes only. These steps are omitted when used as a process measurement procedure.

[0245] During operation and for estimation during training, demodulation data, such as Q and I, are generated in a step 155. These are subjected to linear interpolation in a step 156. An amplitude, for example, is calculated from the interpolated data in a step 157. The result of step 157 is fed to a Butterworth bandpass filter 158, and the filtered signal is in turn fed to a Yeo-Johnson power transform 159. Numerous features of the curve to be measured can now be determined from the data transformed in the Yeo-Johnson power transform 159. The data generated in 159 may, for example, be fed to a shift operator 164, whereby shifted data may be determined, for example Q, I, amplitude, dQ / dt, dl / dt, dAmplitude / dt, binary dQ / dt, binary dl / dt and / or binary dAmplitude / dt, as shown in 167.The data generated by the transformation 159 can also be used directly as features Q, I, amplitude as in 171.

[0246] Furthermore, gradients can be generated in step 160, which in turn can be subjected to a Yeo-Johnson power transform 161, thereby generating the time-dependent change in Q, I, and amplitude. This is shown in 170. The data generated by the transform 161 can also be subjected to a gradient binarization 162, thus generating the binary dQ / dt, dI / dt, and dAmplitude / dt shown in 169.

[0247] The data generated in step 162, together with the data generated in transformation 161, can also be fed to a moving average method. This can generate the data shown in 168. The features 166 can then be fed to the statistical model 24. This model generates the regression function.

[0248] Due to its contactless nature, the invention is also suitable for patients who have easily irritated, burned or inflamed skin.

[0249] It is also suitable for very small patients who do not have much space to attach sensors, e.g. premature or newborn babies.

[0250] The invention's contactless design minimizes the likelihood of poor measurement results due to user error. For example, with ECGs, electrodes can be incorrectly or poorly placed. When measuring respiration with flow or pressure sensors, leaks can occur that can distort both flow and pressure measurements. For example, leaks often occur with face masks in people with beards.

[0251] Only minimal patient compliance is required for the use of the invention. The electromagnetic transmission measurement is imperceptible to patients and can be easily integrated into a chair back or mattress, for example, allowing them to assume a comfortable position. The feeling of tightness that may occur with methods such as RIP bands or ECG chest straps is completely eliminated. Since the invention can also measure through thick textiles, patients do not need to undress.

[0252] Due to the contactlessness, the risk of contamination and wear of the transmitting and receiving structures is also significantly reduced.

[0253] By implementing a soft sensor, the invention can essentially emulate conventional sensors such as ECGs or flow meters. This allows the invention to be seamlessly integrated into existing sensor chains that normally use a conventional sensor.

[0254] With this algorithm, the data of a non-invasive, non-obstructive, and scalable (simultaneous or serial use of multiple sensor sub-elements at spatial distances) and surface- and volume-sensitive measurement technique (fields penetrate at least partially into the target structure) based on the time-dependent complex convolution of at least two electromagnetic fields (sent and received) can be processed, interpreted, and converted into cardiothoracic vital data or other physiodynamic data of a target structure at runtime. For this purpose, the system can be trained, at least initially, to map the changes in the measurement system data resulting from the physiodynamic changes in the target structure or a generic model of the target structure with reference system data (e.g., cardiothoracic vital data).

[0255] The invention is particularly suitable for the following applications:

[0256] Sensor for artificial respiration, monitoring for sudden infant death syndrome, monitoring the progression of cardiac and respiratory diseases (comparable to a long-term ECG), tracking of a person's current condition (awake, asleep, pathological), contactless patient monitoring in intensive care units, contactless recording of an ECG, contactless tracking of pacemakers, contactless determination of respiratory rate, integration into wearables, mattresses, car seats -> stress display or similar, sleep assessment, measuring training parameters e.g. VOZmax, monitoring of sleep apnea, use for biofeedback, input data for artifact or failure compensation of vital data sensors, signal generator for dead man's switch, determination of performance data, use of the "procedure" for evaluating the data of the thorax monitor, measurement methods beyond the derivation of pure cardiothoracic vital data (e.g.technical time-varying conditions such as fill levels, approach, compression of bulk material or porous structures or biological such as the tension of muscle tissue, peristalsis, movement in general of humans as well as animals).

Claims

Patent claims 1. Method for measuring at least one process in an object, wherein an alternating electromagnetic field is radiated into the object, the alternating field is received after it has passed through the object and has been modulated by the process in the object, in a demodulation step the received alternating field is demodulated to generate multidimensional demodulation data, the multidimensional demodulation data and / or data generated from it are fed to at least one statistical model, and in a generation step at least one statistical model generates a time-dependent feature of the process and / or a temporal representation of the process from the fed multidimensional demodulation data and / or the data generated from it.

2. A method according to the preceding claim, wherein the alternating electromagnetic field contains at least two components of different frequencies.

3. Method according to one of the preceding claims, wherein the received alternating field is demodulated in the demodulation step in an ln-phase & quadrature method to generate I-data, I, and Q-data, Q, as the multidimensional demodulation data, and the I-data and the Q-data and / or data generated from them are fed to the statistical model.

4. A method according to any of the preceding claims, wherein the data generated from the multidimensional demodulation data and / or the I-data, I, and the Q-data, Q, include at least one from the group comprising: a derivative of Q with respect to I, a derivative of I with respect to time, a Derivative of Q with respect to time, an amplitude of the received alternating field, a phase of the received alternating field, an amplitude from the derivative of Q with respect to time and the derivative of I with respect to time, a phase from the derivative of Q with respect to time and the derivative of I with respect to time, derivatives of the amplitude with respect to time, derivatives of the phase with respect to time, a phasor of the derivative of Q with respect to I, the first derivative of the generated data with respect to time, of another generated quantity, of the Q data and / or the I data, the second or higher derivative of the generated data with respect to time, of another generated quantity, of the Q data and / or the I data, the first or a higher integral of the generated data, of the Q data and / or the I data with respect to time, of another generated quantity, of the Q data and / or the I data, binarizations of the generated data, binarizations of Q data, binarizations of I data, moving Averages of the generated data,Moving averages of Q-data, moving averages of I-data, Q-data filtered by frequency filters, I-data filtered by frequency filters, generated data filtered by frequency filters, generated data from past times, Q-data from past times and / or I-data from past times.

5. Method according to one of the preceding claims, wherein at least one input feature is extracted from the demodulation data and / or from data generated therefrom and fed to the statistical model, wherein the at least one input feature is preferably a linear combination and / or a non-linear combination of the multidimensional demodulation data and / or the data generated therefrom.

6. Method according to any of the preceding claims, wherein the object is a body, preferably a human or animal body.

7. A method according to any of the preceding claims, wherein the process is a physiological process in a human or animal body.

8. A method according to any of the preceding claims, wherein the statistical model comprises or is an artificial neural network, an artificial neural feedforward network, an artificial neural recurrent network, an artificial convolutional network or a combination of the aforementioned artificial neural networks, a decision tree, a decision forest, a Naive Bayes model, a nearest neighbor model, a support vector machine, a linear regression model, a non-linear regression model, an adaptive filter, a Kalman filter, a gamma filter, a Markov model, a model from a linear discriminant analysis, a model from a non-linear discriminant analysis, a model from an independent component analysis, a model from a principal component analysis and / or an ensemble model from one or more of the aforementioned models.

9. Method according to one of the preceding claims, wherein the alternating electromagnetic field is radiated into the object via at least one transmitting structure by generating an alternating voltage by a signal generator, with which the at least one transmitting structure is acted upon, and the alternating field is received by at least one receiving structure after it has passed through the object.

10. Method according to the preceding claim, wherein the at least one transmitting structure and the at least one receiving structure are arranged on the same side of the object.

11. Method according to one of the two preceding claims, wherein the object is a human body and the at least one transmitting structure is arranged in the immediate vicinity, preferably at a distance of 20 cm or less, preferably at a distance of 10 cm or less, and / or on a thorax and / or an abdomen of the body, and the at least one receiving structure is arranged in the immediate vicinity, preferably at a distance of 20 cm or less, preferably at a distance of 10 cm or less, and / or on the thorax and / or the abdomen. wherein preferably the at least one transmitting structure and / or the at least one receiving structure is arranged laterally, dorsally or ventrally to the abdomen and / or laterally, dorsally or ventrally to the thorax.

12. A method according to any of the preceding claims, wherein the alternating electromagnetic field is radiated into the object via at least two transmit-receive pairs, each comprising a transmit structure and a receive structure, and is received after passing through the object, wherein advantageously at least one of the at least one transmit structures may also be part of several of the transmit-receive pairs, and / or at least one of the at least one receive structures may also be part of several transmit-receive pairs, wherein particularly advantageously separate demodulation data are generated for each of the receive structures in a demodulation step, wherein the totality of the demodulation data of all receive structures is then understood as the multidimensional demodulation data according to the preceding claims.

13. Method according to one of the preceding claims, wherein the method is regularly repeated, wherein a value of at least one feature of the process is regularly repeated and compared with at least one previous value of the feature, and an alert event is triggered when the value of the feature deviates from the previous value of the feature by more than a predetermined threshold, wherein preferably parameters of the statistical model and / or the statistical model is adjusted based on a difference between the value of the generated feature and the corresponding previous value of the feature when the value of the generated feature deviates from the previous value by less than the threshold.

14. Method according to any of the preceding claims, wherein the procedure involves at least one electrocardiogram, at least one impedance cardiogram, at least one ballistocardiogram, at least one seismocardiogram, movement and / or expansion of a chest cavity. and / or an abdomen, at least one blood pressure over time, at least one tidal volume curve, at least one tidal pressure curve, at least one respiratory flow, at least one oxygen saturation over time curve, temporal sequence data obtained by imaging, and / or a plethysmogram. A method according to any of the preceding claims, wherein the at least one feature of the sequence is a trigger event or an exceeding or falling below a predetermined threshold. A method according to any of the preceding claims, wherein the at least one feature of the sequence and / or the temporal representation of the sequence describes the sequence continuously, quasi-continuously, or at a plurality of, preferably equidistant, time points. A method according to any of the preceding claims, wherein the statistical model is trained in a training process prior to irradiation with the alternating electromagnetic field.by radiating an electromagnetic training alternating field into at least one training object, receiving the training alternating field after it has passed through the training object and been modulated by a process in the training object, demodulating the received training alternating field to generate multidimensional demodulation data, feeding the multidimensional demodulation data and / or training data generated from it into the statistical model, generating at least one training feature of the process from the fed multidimensional demodulation data and / or the data generated from it in a time-dependent manner using the statistical model, comparing at least one training feature with a measured feature of the process in a time-dependent manner, and adjusting parameters of the statistical model and / or the statistical model itself based on the result of this comparison.

18. Method according to the preceding claim, wherein the training process is carried out for a plurality of training objects that are representative of the object, or wherein the training process is carried out for the object as the training object.

19. Method for controlling a ventilation process, wherein in a method according to one of the preceding claims a breathing process is measured as the sequence and at least one feature of the sequence is the onset of a spontaneous breathing process, wherein the ventilation process is controlled in such a way that no ventilation against spontaneous breathing takes place.

20. Method according to any of the preceding claims, wherein the statistical model is a neural network with an inference latency of less than or equal to 200ms, preferably less than or equal to 150ms, preferably less than or equal to 100ms, preferably less than or equal to 50ms, preferably less than or equal to 30ms, preferably less than or equal to 20ms, preferably less than or equal to 10ms, preferably less than or equal to 5ms, preferably less than or equal to 1ms.

21. Device for measuring at least one process in an object, comprising at least one transmitting structure with which an alternating electromagnetic field can be radiated into the object by generating an alternating voltage through a signal generator, with which the at least one transmitting structure is acted upon, at least one receiving structure with which the alternating field can be received after it has passed through the object, further comprising a demodulation unit with which the received alternating field can be demodulated to generate multidimensional demodulation data, further comprising a modeling unit to which the demodulation data and / or data generated from these can be supplied, and with which a statistical model is generated from the supplied multidimensional Demodulation data and / or the data generated from these, at least one feature of the process and / or a temporal representation of the process can be generated in a time-dependent manner. Device according to the preceding claim, wherein the device comprises an I / Q demodulator into which the alternating voltage on the one hand and a received signal generated by the receiving structure from the received alternating field on the other can be introduced, and which is configured to perform I / Q demodulation of the received signal. Device according to one of the two preceding claims, wherein the at least one transmitting structure and the at least one receiving structure are arranged in a contact surface, preferably in a backrest, at least one side armrest, at least one armrest, and / or a lying surface, next to each other and / or one above the other and / or diagonally to each other.A device according to any one of claims 21 to 23, wherein the at least one transmitting structure and / or the at least one receiving structure comprises a broadband antenna, wherein the bandwidth of the antenna is preferably greater than or equal to one-eighth of the carrier frequency, preferably greater than or equal to one-quarter of the carrier frequency, preferably greater than or equal to half of the carrier frequency, and / or the bandwidth is less than or equal to the carrier frequency. A device according to any one of claims 21 to 24, comprising a matching device with which a transmitting frequency band of the transmitting structure and / or a receiving frequency band of the receiving structure can be matched, wherein the matching device advantageously comprises or is at least one varactor diode.Device according to one of claims 21 to 25, wherein at least one cable to which the transmitting structure and / or the receiving structure is connected is forcibly balanced, preferably by means of a termination resistor together with an impedance converter. and / or at least a current-compensated choke and / or ferrites.

27. Device according to one of claims 21 to 26, wherein the at least one transmitting structure and / or the at least one receiving structure are supplied at a first terminal with the alternating voltage as a first alternating voltage and at a second terminal with a second alternating voltage shifted by 180° relative to the first alternating voltage, wherein preferably the device has at least one balun which generates two alternating voltages from an input alternating voltage which have a phase shift of 180° relative to each other.

28. Device according to any one of claims 21 to 27, wherein the device has two, three, four or more than four of the transmitting structures and / or has two, three, four or more than four of the receiving structures.

29. Device according to any one of claims 21 to 28, wherein the device comprises N transmitting antennas and IVI receiving antennas, wherein N is a natural number greater than or equal to 1 and IVI is a natural number greater than or equal to 2, wherein the N transmitting antennas are supplied with alternating voltages L of different frequencies, wherein L is a natural number greater than or equal to 1 and preferably less than or equal to N, wherein preferably no transmitting antenna is supplied with several of the alternating voltages of different frequencies and / or wherein each of the receiving antennas is configured to receive exactly one of the L frequencies or to receive several of the L frequencies.

30. Device according to one of claims 21 to 29, wherein the signal generator is configured to generate the alternating voltage with different and / or time-varying frequencies.

31. Device according to one of claims 21 to 30, wherein the device is capable of performing a method according to one of claims 1 to 20.

Images