METHOD FOR PREDICTING AND LOCALIZING STRUCTURAL CHANGES IN MYOCARDIAL TISSUE

Description

[0001] The invention relates to a method for predicting and localizing structural changes in myocardial tissue (including ischemia-related, inflammation-related, and stress-related damage to cell structure) using machine learning techniques (artificial neural network). Description and introduction of the general field of the invention

[0002] Heart attacks are one of the leading causes of death in industrialized nations and belong to the group of ischemic heart diseases. The incidence in Austria and Germany is approximately 300 heart attacks per 100,000 inhabitants annually (in Japan < 100; Mediterranean, Switzerland, and France < 200; 300 to 400 in Scandinavia; 400 to 500 in England and Hungary). In Germany, approximately 280,000 people suffer a heart attack each year. According to mortality statistics from the Federal Statistical Office, over 49,000 people died in Germany in 2015 as a result of an acute heart attack. It should be noted that up to 25% of all heart attacks cause only mild or no symptoms (so-called silent heart attacks). Structural changes in the myocardial tissue can be an indicator of ischemic heart disease, which in turn is one of the main causes of heart failure. Typically, such changes are diagnosed by, for example,Myocardial scarring can be diagnosed using high-resolution cardiac imaging such as MRI (magnetic resonance imaging). However, there are practical limitations regarding the availability of widely applicable diagnostic and risk assessment tools, as MRI-based techniques cannot be used at the bedside or on an outpatient basis due to cost and the required expertise. State of the art

[0003] To identify damaged myocardial tissue, a method based on the temporal assessment of gadolinium accumulation in the tissue (so-called late gadolinium enhancement (LGE)) was introduced in 1989. This technique uses magnetic resonance imaging (MRI) and gadolinium-based contrast agents. Since then, the procedure has become established as a routine sequence in cardiac magnetic resonance imaging, for example, to detect and define scars with high accuracy.

[0004] In addition to imaging, the presence of structural changes in myocardial tissue is traditionally assessed based on the 12-lead electrocardiogram (ECG). This assessment relies on manual analysis. Currently, no established, machine-based screening methods of sufficient accuracy exist.

[0005] Machine learning-based screening methods for detecting cardiac arrhythmias such as atrial fibrillation (AF) are already known in the state of the art. For example, the study by Hannun et al. (Hannun et al. (2019), "Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network", Nature Medicine, 25, 65-69. DOI: 10.1038 / s41591-018-0268-3) describes a method that evaluates a single-lead ECG with regard to twelve detectable cardiac arrhythmias (including atrial fibrillation and flutter, AV block type I and II and total heart block, bigeminy, trigeminy, ventricular and supraventricular tachycardias and junctional rhythm, as well as sinus rhythm and measurement error ("noise")). In this process, an artificial neural network is trained using a variety of findings to recognize and evaluate relevant patterns in the ECG signal.However, this method does not allow for direct conclusions about the condition of the heart muscle or the location of the detected abnormalities. Currently, there are no comparable methods for localizing structural changes in myocardial tissue.

[0006] Another method is described by Attia et al. (Attia et al. (2019), "Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram".

[0007] Nature Medicine 25, 70-74. DOI:10.1038 / s41591-018-0240-2). This method uses artificial intelligence to detect asymptomatic left ventricular dysfunction from ECG data. The cardiac pump function, as determined by echocardiogram, is predicted via ECG. However, this method does not provide information about the condition of individual myocardial regions. To achieve the necessary diagnostic precision and robustness for detecting structural changes in myocardial tissue, currently, costly methods are required, which also demand specialist expertise. Alternatively, other methods exist that achieve lower precision and / or are time-consuming and have therefore not been established in routine clinical practice. One example is the Selvester QRS score (Selvester et al. (1985), "The Selvester QRS Scoring System for Estimating Myocardial Infarct Size: The Development and Application of the System").Archives of Internal Medicine, 145, 1877-1881. DOI: 10.1001 / archinte.1985.00360100147024), which can quantify and localize infarct-related myocardial damage based on the ECG. This involves a high degree of manual effort (precise measurement of the ECG waveform is necessary, numerous calculations) and has proven to be impractical in everyday clinical practice.

[0008] US 2019 / 0090774 A1 describes a method for determining the arrhythmia location to identify the origin of an arrhythmia in a subject's heart by receiving the identification of the cardiac segment in which the arrhythmia originates from neural networks to which electrical data is fed. It mentions dividing the left ventricle into 17 segments according to the AHA-standardized myocardial segmentation. Task

[0009] The object of the present invention is to avoid the disadvantages of the prior art, so that ECG recordings and other clinical parameters can be used to localize structural changes in myocardial tissue using an artificial neural network (ANN). This enables a precise and reliable diagnosis, which is currently only possible with imaging techniques (e.g., MRI). Solution to the task

[0010] The solution to these problems arises from the features of the main claim. Furthermore, advantageous embodiments and further developments of the invention can be found in the dependent claims.

[0011] First, a definition is given of how some terms relevant to the method according to the invention are to be understood within the scope of this invention: ECG:

[0012] ECG stands for electrocardiography and refers to an examination method in which the electrical activity of the heart is measured on the body's surface. Nerve and muscle cells communicate via electrical and chemical signals. Regular electrical impulses also control the heartbeat. These impulses originate in the sinoatrial node in the right atrium of the heart and spread across the heart muscle like small electrical shocks. This causes the atria and then the ventricles to contract. The propagation of these electrical impulses in the heart muscle can also be measured on the skin. An ECG measures these electrical potential differences at various points on the body and displays them as curves. These ECG curves are called electrocardiograms.

[0013] When the heart beats regularly, the typical ECG pattern emerges: The first deflection (P wave) shows how the electrical impulse (excitation) spreads across the atria. The atria contract, pumping blood into the ventricles, and then immediately relax again. The excitation then reaches the ventricles. On the ECG, this is visible as Q, R, and S waves, the so-called QRS complex, which is associated with the contraction of the ventricles. The T wave then indicates that the excitation is dissipating and the ventricles are relaxing again.

[0014] The temporally ordered data of the excitation process of the heart are referred to below, in simplified terms, as ECG (measurement) data. MRI:

[0015] MRI stands for magnetic resonance imaging, abbreviated MRI or MR. This is an imaging technique primarily used in medical diagnostics to create spatial representations of the structure and function of tissues and organs in the body. It is physically based on the principles of nuclear magnetic resonance (NMR), particularly gradient-field NMR, and is therefore also called nuclear spin tomography (occasionally shortened to nuclear spin). MRI can generate cross-sectional images of the human (or animal) body, allowing for the assessment of organs and many pathological changes. The data generated, as well as the interpretive data derived from them, are referred to below as MRI data. Myocardium:

[0016] Myocardium: The term myocardium refers to the heart muscle. The myocardium forms the largest part of the heart. The heart muscle is formed and surrounded externally by the epicardium and internally by the endocardium. The heart muscle is a hollow muscle with a macroscopic (loop-like, interconnected) structure specific to its contraction, which reduces the volume of the heart chamber. The muscle fibers of the heart chambers run superficially (subepicardially) towards the apex of the heart, beat inwards at the vortex cordis, and then extend back towards the heart skeleton as a deep (subendocardial) muscle layer. Myocardial pathology:

[0017] The term myocardial pathology encompasses various pathological changes of the myocardium, including scarring, perfusion deficits, inflammation, and hypertrophy. These can affect one or more segments of the myocardium. Localization:

[0018] For diagnostic purposes, the heart can be divided into several regions. The 17-segment model proposed by the American Heart Association (AHA) is widely used. This model focuses exclusively on the left ventricle, which constitutes the most significant part of the myocardium. In this model, the ventricle is divided into equal segments perpendicular to the heart's longitudinal axis. These basal, mid-cavitary, and apical segments are then further subdivided into smaller segments. Each segment is named based on two criteria: its position on the longitudinal axis and its circumferential position. These segments can be viewed as a bull's-eye diagram (see [reference]). Figure 1a and Table 1). Based on this segmentation, the likely supply areas of the respective coronary artery can also be defined, which leads to the also under Figure 1b The specified approximation leads to... Table 1: Basal segments (third near the valve) Midventricular segments (middle third) Apical segments (near the apex third) 1. basal anterior 7. mid-anterior 13. apical anterior 2. basal anteroseptal 8. middle roseptal 14. apical septal 3. basal inferoseptal 9. midinferoseptal 15. apical inferior 4. basal inferior 10. middle inferior 16. apical lateral 5. basal inferolateral 11. midinferolateral 17. apex 6. basal anterolateral 12. mid-anterolateral Clinical parameters:

[0019] The interpretation of ECG measurements can be improved by including additional clinical parameters. These relate in particular to classic cardiovascular risk factors (see Yusuf et al. (2019), "Modifiable risk factors, cardiovascular disease, and mortality in 155722 individuals from 21 high-income, middle-income, and low-income countries (PURE): a prospective cohort study". The Lancet, Sep 03, 2019. DOI: 10.1016 / S0140-6736(19)32008-2).

[0020] A major risk factor is the presence of atherosclerosis. Atherosclerosis is the underlying cause of the majority of ischemic cardiovascular diseases, such as myocardial infarction. This progressive disease is characterized by an accumulation of lipids and fibrous elements in the large arteries. This is due, in part, to an inflammatory response that leads to alteration and dysfunction of the arterial endothelium. This results in the formation of plaques within the arterial walls. These plaques reduce the arterial lumen and pose a risk of rupture, which can lead to severe thrombotic occlusion of the affected artery. This ultimately leads to an infarction of the dependent tissue. Atherosclerosis is a complex disease in which several processes interact and influence one another.

[0021] Other important clinical parameters are described below.

[0022] Dyslipidemia, also known as a lipid metabolism disorder, is characterized by an abnormal composition of lipids (often referred to simply as fats in this context) in the blood. Cholesterol is an important lipid. Lipids perform vital functions, such as energy storage, cell membrane formation, and act as signaling molecules. A high lipid level can promote the formation of atherosclerotic plaques. The lipid level in the blood is regulated by absorption in peripheral tissues. The liver is responsible for regulating lipid levels through secretion and absorption. If this mechanism is impaired or overworked, it leads to elevated lipid levels. This can cause more atherosclerotic plaques to form and further increase the risk of rupture.

[0023] Overweight individuals have excessive adipose tissue, which leads to an increase in pro-inflammatory cytokines. These cytokines promote the progression of atherosclerosis and insulin resistance. Furthermore, obesity has negative effects on arterial hypertension and diabetes mellitus, making it a particularly important clinical parameter.

[0024] In general, arterial hypertension, diabetes mellitus and dyslipidemia are often associated with overweight in the so-called metabolic syndrome, with these factors reinforcing each other.

[0025] Another important clinical parameter is the presence of arterial hypertension (high blood pressure). It increases the stress on the endothelium. Arterial hypertension places additional strain on the myocardium. Thus, patients often suffer from both atherosclerosis and elevated blood pressure. As a result, the myocardium has to cope with less oxygen on the one hand and increased physical stress on the other. This leads to a discrepancy between oxygen supply and demand, which can result in ischemia and necrosis.

[0026] Smoking is another important factor. Smoking damages the endothelium and increases platelet aggregation and fibrinogen concentration in the blood (a condition known as hypercoagulability). Furthermore, smoking reduces the amount of high-density lipoproteins, which are responsible for transporting excess cholesterol back to the liver. This promotes the development of plaques and increases the risk of rupture and blockage of affected arteries.

[0027] Another important parameter is the presence of diabetes mellitus. This is a complex metabolic disorder associated with an increased risk of, for example, suffering a myocardial infarction. A distinction is made between type 1 and type 2 diabetes. Type 1 is based on a chronic autoimmune disease that prevents insulin from functioning adequately. Type 2 is also referred to as acquired diabetes, as it is dependent on and influenced by lifestyle factors (diet, activity). Damage occurs through both direct and indirect influences. Due to reduced sensitivity to insulin or low insulin secretion, blood glucose levels remain chronically elevated. This exposes cells to increased osmotic stress, which can damage their structure and potentially lead to necrosis of the affected tissue. This can have similar effects to a heart attack caused by atherosclerosis.Diabetes mellitus also has a negative impact on the endothelium and promotes atherogenesis, especially when it occurs together with elevated lipid levels.

[0028] Genetic predisposition has a proven influence on cardiovascular diseases and is therefore an important clinical parameter. Several studies have shown that individuals with first-degree relatives who suffered a myocardial infarction before the age of 55 have an increased risk of developing structural changes in myocardial tissue. Genome-wide association analysis is used to identify the occurrence of genetic defects or mutations that could be responsible for an increased risk of developing certain diseases. Several genes have been identified that are associated with increased susceptibility to diabetes mellitus and accelerated atherogenesis.It has been shown that a predisposition to left main coronary artery stenosis, calcification, and lesions (aneurysm / ectasia) is hereditary, directly influencing the risk of coronary heart disease and myocardial infarction. Furthermore, it has been established that these genetic factors are independent of other risk factors.

[0029] Another important parameter is the patient's sex, as sex-specific differences also exist in the prevalence of cardiovascular diseases. Based on various studies, it has been observed that the first myocardial infarction occurs on average nine years later in women than in men. Premenopausal women show a reduced prevalence of cardiovascular events compared to men of the same age. This difference decreases with menopause. This can be explained by the fact that estrogens are vasoactive hormones that have a direct effect on the arterial wall. They protect the endothelium by increasing the production of nitric oxide (NO), which reduces oxidative stress. They also indirectly regulate blood lipid levels.

[0030] Age affects the properties of the endothelium and blood in various ways. Endothelial function is impaired with increasing age. Aging reduces the production of nitric oxide (NO), which leads to elevated blood pressure and impairs endothelial function. Due to age-related metabolic changes, the blood's clotting ability, as well as the risk of thrombosis, is increased.

[0031] This list only provides examples of important clinical parameters for identifying cardiac pathologies and is not exhaustive. Artificial Neural Networks:

[0032] Artificial Neural Networks, also known as Artificial Networks, or in short: KNNArtificial neural networks (ANNs) are networks made up of artificial neurons. They are based on the interconnection of many individual artificial neurons, usually arranged in layers. The strength of the connections between the neurons is controlled by so-called weights. The topology and the controlling parameters of a network are chosen depending on the problem at hand.

[0033] In a training phase, KNNs are trained using input data (input signal) for which expected results (target signal) are available, by comparing these with the calculated results (outcome signal).

[0034] The difference between the output signal and the target signal is called the loss. The loss indicates how well the expected and the actual result match. To improve the quality of processing, this loss must be taken into account for future iterations. This adjustment of the processing is called backpropagation or weight adjustment. Here, the weights between the neurons are adjusted so that they more closely approximate the desired result in the future.

[0035] Based on this, the learning ability of KNNs can be defined as the ability to compare a given result with an expected result, followed by an adjustment according to the difference between these values.

[0036] KNNs are therefore able to learn complicated nonlinear functions using an iterative training process.

[0037] A widely used principle for training KNNs is supervised learning. Here, the neural network is trained and iteratively adapted using a set of tuples of related input (x) and output signals (y) according to the principle described above. The set of tuples (x, y) used for training is called training data (D training). The network's performance on unfamiliar data is determined using validation data (D validation). This set of tuples (x, y) contains data with which the network was not trained. D training ∩ D validation = Ø. Inventive method:

[0038] The ECG training measurement data TD1 and clinical parameters D2 obtained in a previous step (which is not part of the method according to the invention) are used as input signals of the method according to the invention.

[0039] The clinical parameters D2 are optional and are included in the processing after possible imputation. Imputation means that missing data for parameters D2 (incomplete datasets) in statistical surveys—the so-called dropouts—are added to the data matrix. This reduces the silence bias caused by dropouts. Various methods for adding missing values ​​are well known. A broad distinction is made between single and multiple imputation.

[0040] For training the neural network, at least one clinical parameter D2, an ECG segment TD1, and the diagnostic information D3 derived from an MRI scan are preferably used. The clinical parameters and the ECG segment represent the input signals (TD1, D2), while the diagnosis D3 provides the target values ​​for the network's output signals. The diagnostic dataset D3 includes indicators for various tissue changes, including their location. This set of tuples (x, y) constitutes the training data (D training). Here, x corresponds to a tuple of the data TD1 and preferably D2 (ECG segment, clinical parameters), and y corresponds to a vector containing the diagnostic information D3. The ECG segment is a matrix consisting of #LeadsECG and #SequenceLengthECG, while the clinical parameters D2 are provided as a vector with the parameters as entries.

[0041] From the input signals x, a result signal y' is determined via the KNN, which is compared with the true y as described above and used to calculate the lot. To train the neural network:

[0042] By comparing the actual values ​​calculated from TD1 and preferably D2 with the target values ​​D3 of the training data, the weights contained in the artificial neural network are adjusted. This iterative process results in a self-contained neural network N.

[0043] This enables the network N to recognize patterns. Therefore, when using N for prediction, no access is made to specific data sets from the training data, but only to the learned weights. The prediction is thus based solely on these weights learned through training. Training phase

[0044] Before the actual implementation of the method according to the invention, the KNN must go through a training phase as a step 0.

[0045] This step 0 comprises the following sub-steps: 0a. Receiving the ECG training measurement data TD1. 0b. Preprocessing of the ECG training measurement data TD1 to preprocessed ECG training measurement data TD1* and transfer of this preprocessed ECG training measurement data TD1* to the artificial neural network N. Preprocessing during the training phase:

[0046] To use high-quality data for network training, the data used, such as ECG data TD1 or clinical parameters D2, must be preprocessed into ECG data TD1* or preprocessed clinical parameters D2*. This is also referred to as preprocessing. This is shown schematically in Figure 2 shown.

[0047] This requires setting up a so-called data pipeline. This defines the individual steps of the preprocessing. This data pipeline serves to transform the heterogeneously stored data into cleaned datasets that can be used without problems in the further process.

[0048] The preprocessing takes into account suitable measures for data augmentation.

[0049] The preprocessing in the training phase includes the following steps: a. Extraction

[0050] The ECG data of individual patients are extracted from the raw data. The ECG data is typically in XML file format.

[0051] The optional clinical parameters of each patient, as well as the diagnoses made based on the MRI, are extracted from the raw patient data files. These are typically in CSV file format. Subsequently, the ECG data, clinical parameters, and diagnostic information for each patient are combined into a tuple (x, y). b. Selection

[0052] In this step, unusable datasets are excluded from further processing (e.g., incomplete clinical parameters that cannot be imputed). Only complete datasets should be used for training to optimize the learning of correlations.

[0053] All ECG leads can be used; there is no need to select specific leads. c. Cleanup

[0054] The ECG recordings are trimmed to eliminate technical artifacts, e.g., at the beginning and / or end. d. Augmentation

[0055] To increase the amount of data available for training and thereby improve network performance, distinct sections are extracted from each ECG recording. A fixed-size window is moved across the recording step by step, generating a new recording each time. This allows the data volume to be increased, for example, by a factor of 100. e. formatting

[0056] To reduce high-frequency noise and the size of the ECG recordings, downsampling is performed for each data set. This uses an average pooling-based method to reduce recording details while preserving their significant patterns. This allows the data size to be reduced by, for example, a factor of 4. f. Save

[0057] The data is divided into equally sized groups. Depending on the desired ratio between training and validation data (ideally 80% to 20%), a specific number of groups (e.g., 6) are used for training, and the remainder for validation. This combination of training and validation splits is each stored as a separate data set variant. This process is repeated until either enough variants have been created or all possible combinations of groups have been encountered in training and validation. This allows for the training of identically structured models with different data. In combination, these models (e.g., 6) deliver higher performance on unknown data than a single model from this set. This method is called ensembling. 0c. Determination of the probability of the predictive value of at least one cardiac pathology for at least one segment of the myocardium by the Artificial Neural Network N. 0d . Input of MRI diagnosis information D3.

[0058] The MRI diagnostic information D3 serves as reference data for training the KNN. Alternatively, data from other imaging procedures can also be used as reference data. 0e Iterative target-actual comparison of the determined probabilities with the MRI diagnostic information D3, followed by adjustment of the weights G in the artificial neural network N. 0f . Completing the trained artificial neural network N with the adjusted weights G.

[0059] In a further development of the training procedure, an optional additional step 0a* is performed before step 0b, in which at least one clinical parameter D2 is received. In this case, an additional step 0b* then follows, which takes place between step 0a* and step 0c. Step 0b* comprises preprocessing the at least one clinical parameter D2 into a preprocessed clinical parameter D2* and transferring this preprocessed clinical parameter D2* to the artificial neural network N. Including the clinical parameters D2 increases the reliability of the training phase.

[0060] After the training phase, the method according to the invention is carried out.

[0061] The method according to the invention comprises the following steps: I. Providing the ECG measurement data D1

[0062] The first step involves providing ECG measurement data. ECG measurement data refers to the data available after the ECG measurement has been completed on the patient. This data provision only begins after the examination involving data collection on the human or animal body has finished, i.e., only after the measurement on the patient is complete.

[0063] No measurements are taken on the patient themselves during the procedure. The procedure can even be performed if the patient has already died. II. Preprocessing of the ECG measurement data D1 and transfer of this data to an artificial neural network N Pre-processing in the inventive method:

[0064] For the inventive method, the data used, such as ECG data D1 and, if available, clinical parameters D2, must also be processed into pre-processed ECG data D1* and pre-processed clinical parameters D2* so that they can be passed to an artificial neural network. This is also referred to as pre-processing.

[0065] This requires setting up a so-called data pipeline. This defines the individual steps of the preprocessing. This data pipeline serves to transform the heterogeneously stored data into cleaned datasets that can be used without problems in the further process.

[0066] The preprocessing takes into account suitable measures for data augmentation.

[0067] The preprocessing includes the following steps: a. Extraction of the ECG data of the individual patient

[0068] The individual patient's ECG data is extracted from the raw data. The ECG data is typically in XML file format.

[0069] The optional clinical parameters of each patient are extracted from the raw patient data files, which are typically in CSV format. Subsequently, the ECG data, clinical parameters, and diagnostic information for each patient are combined into a tuple (x, y). b. Selection

[0070] In this step, unusable datasets are excluded from further processing (e.g., incomplete clinical parameters that cannot be imputed). Only complete datasets should be used for the procedure to optimize its use.

[0071] All ECG leads can be used; there is no need to select specific leads. c. Cleanup

[0072] The ECG recordings are trimmed to eliminate technical artifacts, e.g., at the beginning and / or end. d. Augmentation

[0073] To increase the amount of usable data and thereby improve network performance, distinct sections are extracted from each ECG recording. A fixed-size window is moved across the recording step by step, generating a new recording each time. This allows the data volume to be increased, for example, by a factor of 100. e. Formatting

[0074] To reduce high-frequency noise and the size of the ECG recordings, downsampling is performed for each recording. This uses an average-pooling-based method to reduce the details of the recording while preserving its significant patterns. This allows the data size to be reduced by, for example, a factor of 4. f. Saving the pre-processed ECG data D1*

[0075] The pre-processed ECG data D1* are finally saved so that they are available for the application of an Artificial Neural Network. III. Application of the trained artificial neural network N to the preprocessed ECG data D1* to determine the result data E, which include at least one prediction value for at least one cardiac pathology for at least one segment of the myocardium.

[0076] In this process, the preprocessed ECG data D1* is used across the entire course of an ECG measurement, and not just individual typical peaks, as is typically the case with manual evaluation by physicians. Thus, the entirety of the ECG data contributes to the determination of the results. The sampling rate for ECG devices is at least 50 Hz. Preferably, however, they have a sampling rate of at least 500 Hz. Since all measurement points of the preprocessed ECG data D1* are used, the sampling rate of the preprocessed ECG data D1* is also at least 50 Hz. Preferably, the sampling rate of the preprocessed ECG data D1* is at least 500 Hz.

[0077] This enables a particularly precise and reliable determination of result data. IV. Output of the results data E

[0078] In step III, the trained artificial neural network N is applied to the pre-processed ECG data D1* with weights G adjusted during a training phase. The pre-processed data D1* and, if applicable, at least one pre-processed clinical parameter D2* are passed as input to the network N, which then calculates a prediction p (0<=p<=1) – i.e., a one-dimensional probability value – for the presence and, optionally, the location of a cardiac pathology by applying the trained network structure, and outputs this prediction as result data E.

[0079] The outcome data E includes at least one predictive value for at least one cardiac pathology in at least one segment of the myocardium, such that these outcome data E allow for the prediction and localization of structural changes in the myocardial tissue. This predictive value for the location of a pathology is a diagnostically relevant intermediate result, but it still needs to be interpreted by medical professionals within a medical context to arrive at a diagnosis. A predictive value in itself, therefore, does not constitute a diagnosis.

[0080] The MRI diagnostic information D3 is used only during the training phase but not during the method according to the invention.

[0081] Prediction values ​​are preferably determined for all 17 segments of the myocardium.

[0082] When evaluating the test results, it is crucial to know the probability that a positive test result actually indicates a cardiac pathology. This predictive value of a positive test is also referred to as the (positive) predictive value. The test results are preferably presented in the form of these predictive values. The interpretation of the predictive values ​​to establish a specific diagnosis is reserved for medical professionals and is not part of the procedure.

[0083] The threshold values ​​can be selected separately and independently for each pathology and each cardiac segment. The selection of the threshold values ​​takes place after the training phase and before step III of the inventive method and depends on the type of pathology and its effects on the corresponding cardiac segment. These threshold values ​​are determined by medical professionals based on their training and experience and are not part of the inventive method. A typical threshold value for the presence of a pathology in a specific cardiac segment might, for example, be 0.5, meaning that values ​​of 0.5 or greater are identified as pathologies.

[0084] In a further development of the method according to the invention, an additional step 1a is optionally performed before step II, in which at least one clinical parameter D2 is received. In this case, in a subsequent additional step IIa, which takes place between step 1a and step III, the at least one clinical parameter D2 is preprocessed into a preprocessed clinical parameter D2*, and this preprocessed clinical parameter D2* is then transferred to the trained artificial neural network N. This is optimally performed, however, if the at least one clinical parameter D2 has already been used in a training phase.

[0085] The reliability of the procedure can be increased by including clinical parameters D2.

[0086] Within the framework of the inventive method, the result of a diagnostic method that is only available to a limited extent (e.g. MRI) is predicted on the basis of a widely available diagnostic method (ECG), thereby expanding the diagnostic possibilities of the ECG. Device:

[0087] The device 1 according to the invention for predicting specific cardiac pathologies based on ECG recordings and other clinical parameters comprises at least the following elements (see Figure 3 ): One input unit 10:

[0088] The input unit 10 is designed to allow the input of ECG measurement data D1 and the data of the clinical parameters D2, which are then transmitted to the evaluation unit 20 for further processing. The input unit 10 can be a keyboard or a graphical user interface (GUI). Data transmission can occur via a fixed data connection (e.g., an internal data line, local area network (LAN), etc.) or wirelessly (e.g., via wireless local area network (WLAN), Bluetooth, etc.). A data interface 50:

[0089] The data interface 50 is designed to receive ECG measurement data D1 and clinical parameter data D2 from an external data processing system and to forward the results data to an external data processing system. Data transmission can occur via a fixed data line (e.g., via a LAN) or wirelessly (e.g., via WLAN, Bluetooth).

[0090] Furthermore, the data interface 50 is designed to transmit ECG measurement data D1 and the clinical parameter data D2 to the evaluation unit 20. Data transmission can occur via a fixed data line (e.g., via a LAN, etc.) or wirelessly (e.g., via WLAN, Bluetooth, etc.). The transmission of the result data E to external data processing units can also occur via the data interface 50.

[0091] The 50 data interface is typically designed for exchange with various external data processing systems; common interfaces for data exchange include USB, card reader, LAN, WLAN, and Bluetooth.

[0092] The data interface 50 is designed in such a way that a user with any web-enabled device (e.g. tablet, PC, smartphone) in the local network can access the device 1 via the data interface 50. One evaluation unit 20:

[0093] The evaluation unit 20 is designed to receive the ECG measurement data D1 and the clinical parameter data D2 from the input unit 10 or a data interface 50. Data transmission can be via a fixed data line (e.g., an internal data line, LAN, etc.) or wirelessly (e.g., via WLAN, Bluetooth, etc.).

[0094] Evaluation unit 20 is further configured to calculate the probability and spatial localization of myocardial pathology using the artificial neural network, based on the ECG measurement data D1, the clinical parameters D2, and the weights learned via training data. The result of this calculation is the outcome data E. Evaluation unit 20 is an electronic data processing device, such as a programmable microprocessor integrated into device 1.

[0095] The evaluation unit 20 is further configured to transmit the result data E to an output unit 30. Data transmission can occur via a fixed data line (e.g., an internal data line, LAN, etc.) or wirelessly (e.g., via WLAN, Bluetooth, etc.).

[0096] The evaluation unit 20 uses an artificial neural network N for pattern recognition. Device 1 calculates prediction values ​​(a list of probability values ​​for the presence of cardiac pathologies in the various segments of the myocardium) as output data E. This data indicates the probability of the presence or absence of a specific cardiac pathology in a particular area of ​​the heart. This output data E is displayed via the output unit 30 in the form of visual and / or audible outputs. Alternatively, it is forwarded to the data interface 50. One output unit 30:

[0097] Output unit 30 is designed to receive and then output the result data E from evaluation unit 20. The output can be visual, audible, haptic, and / or in the form of a printout.

[0098] The input unit 10 or the output unit 30 of the device can include, for example, buttons, LEDs, (touch) display and speech input / output.

[0099] In a first alternative embodiment, the device 1 is designed such that completely encapsulated data processing can take place, excluding any network access (LAN, WLAN). For this purpose, all network components are blocked for any data exchange that does not occur between input unit 10, evaluation unit 20, and output unit 30.

[0100] Medical devices (e.g., measuring instruments) contaminated with pathogens can be a source of infection in humans. Therefore, the use of such medical devices requires prior reprocessing, which must meet defined requirements.

[0101] In a second alternative embodiment, the device 1 therefore comprises a housing for the evaluation unit 20, which is designed to be washable. In this embodiment, the input unit 10 and the output unit 30 are also designed to be washable. For this purpose, the housing 10, the input unit 10, and the output unit 30 preferably have a coating of Parylene or waterproof silicone. This enables particularly hygienic use.

[0102] Washable means that the surface is completely sealed and can be cleaned according to the guidelines of the Association for Applied Hygiene (VAH) or the German Society for Hygiene and Microbiology (DGHM) Treat with all common disinfectants.

[0103] Preferably, the housing, input unit 10, and output unit 30 are sterilizable according to DIN EN 285. According to DIN EN 285, all surfaces of the sterilized items must be exposed to pure, saturated steam at a temperature of 134 °C for at least three minutes. A suitable surface material for this purpose is, for example, PEEK (polyetheretherketone).

[0104] One aspect of using a dedicated device is data protection, as it allows for completely isolated data processing, excluding any network access (LAN, WLAN) when necessary. This ensures the protection of sensitive patient data. The device should also support hygiene and ease of use in everyday clinical practice. A PC with a keyboard or a touch-based device (tablet, smartphone) cannot be used everywhere. This is due to both hygienic and practical reasons. A device dedicated to the detection of a specific disease offers the advantages of compactness, user-friendliness, hygiene, and data protection, as it only contains the input and output units necessary for the application.A safe, easy-to-clean and easy-to-use device therefore offers important and necessary aspects of a medical device, which is why such a device is also part of the invention. Application example (here for a 12-lead ECG) Inputs:

[0105] D1: 12-lead ECG (10 seconds) as XML file: patient123.xml D2: Clinical parameters of the patient: Gender: female Age: 59 years Weight: 82 kg Size: 164 cm Diabetes mellitus: No High blood pressure: Yes Elevated cholesterol: Yes Genetic predisposition to heart disease: No Smoking habits: ex-smoker

[0106] The examination focused on the following cardiological pathologies: scarring, perfusion deficit, inflammation, and hypertrophy. The examination was performed in segments 1 to 17, as shown in the figure. Fig. 1a shown.

[0107] The following table (Table 2) shows the predictive values ​​E for the various cardiological pathologies in segments 1 to 17, as determined by applying the method according to the invention using a neural network. This predictive value for the location of a pathology is an intermediate result relevant for diagnosis and must still be interpreted by medical professionals for a diagnosis within a medical context. In this application example, the threshold for the presence of a cardiological pathology in a specific segment is a predictive value of 0.5. Values ​​of 0.5 or greater are identified as pathologies. This threshold was determined by medical professionals based on their experience and is not part of the method according to the invention. Table 2: scar Perfusion deficit inflammation hypertrophy Segment 1 0,03 0,66 0,01 0,04 Segment 2 0,02 0,81 0,10 0,03 Segment 3 0,00 0,00 0,00 0,00 Segment 4 0,00 0,00 0,00 0,00 Segment 5 0,00 0,00 0,00 0,00 Segment 6 0,00 0,00 0,00 0,00 Segment 7 0,05 0,77 0,02 0,06 Segment 8 0,01 0,82 0,02 0,08 Segment 9 0,00 0,00 0,00 0,00 Segment 10 0,00 0,00 0,00 0,00 Segment 11 0,00 0,00 0,00 0,00 Segment 12 0,00 0,00 0,00 0,00 Segment 13 0,05 0,75 0,03 0,07 Segment 14 0,01 0,81 0,01 0,09 Segment 15 0,00 0,00 0,00 0,00 Segment 16 0,00 0,00 0,00 0,00 Segment 17 0,01 0,81 0,01 0,10 Aggregated (here with max) 0,05 0,82 0,10 0,10 Output (here dichotomous) 0 1 0 0 Result:

[0108] In the example shown, there is a perfusion deficit in segments 1, 2, 7, 8, 13, 14, and 17. These segments are highlighted in bold in Table 2. Figure captions and list of reference symbols

[0109] The Figure 1 shows a schematic division of the areas of the human heart.

[0110] The Figure 2 The diagram schematically illustrates the preprocessing sequence. The MRI diagnostic information D3 is used only during the training phase, but not during the inventive method.

[0111] The Figure 3 Figure 1 schematically shows the device according to the invention. Reference sign

[0112] 1 device 10 Input unit 20Evaluation unit 30 Output unit 50 Data interface TD1 ECG training measurement data TD1* pre-processed ECG training measurement data D1 ECG data D1* pre-processed ECG data D2 clinical parameters D2* pre-processed clinical parameters D3 MRI diagnostic information E Results data N trained neural network

Claims

1. Method for predicting and localising structural changes in myocardial tissue, selected from the group comprising: scarring, perfusion deficits, inflammation, hypertrophy, and ischaemia-induced damage to the cell structure, comprising the following steps: I. Providing the ECG measurement data (D1), wherein the provision takes place either via an input unit (10) and / or via a data interface (50), wherein the provision only begins after the examination has been completed, once the measurement on the patient has been concluded II. Pre-processing the ECG measurement data (D1) into pre-processed ECG measurement data (D1*) so that it can be transferred to an artificial neural network, and transferring the pre-processed ECG data (D1*) to a trained artificial neural network (N). III. Application of the trained artificial neural network N to the pre-processed ECG data (D1*) to determine the result data (E), whereby all measurement points of the pre-processed ECG data (D1*) are used and the sampling frequency of the pre-processed ECG data (D1*) is at least 50 Hz, wherein the result data (E) comprise at least one prediction value for at least each cardiac pathology from the group: scarring, perfusion deficits, inflammation, hypertrophy for all 17 segments of the myocardium in accordance with the 17-segment model of the American Heart Association (AHA), such that a distinction is made between scarring, perfusion deficit, inflammation and hypertrophy is made by specifying a prediction value for all these specific structural changes in all 17 segments, wherein the result data (E) enable the prediction and localisation of a structural change in the myocardial tissue, wherein the complete pre-processed ECG data (D1*) is used over the entire course of an ECG measurement, whereby the pre-processing of the ECG measurement data (D1) and the application of the trained artificial neural network N are carried out via an evaluation unit (20). IV. Output of the result data (E), wherein the output is performed via an output unit (30).

2. Method for predicting and localising structural changes in myocardial tissue according to claim 1, characterised in that step II comprises at least the following substeps for converting the ECG data (D1) into pre-processed ECG data (D1*): a. Extraction of ECG data for a single patient from the ECG measurement data (D1) b. Selection of suitable data sets from the ECG data from step a, whereby unusable data sets are excluded from further processing c. Cleaning of the data sets from step b, whereby these are trimmed to eliminate technical artefacts d. Augmentation of the data sets from step c, whereby distinct sub-segments are extracted from each ECG recording by moving a window of fixed size step by step across the recording, which generates a new recording in each instance. e. Formatting the data sets from step d, whereby down-sampling is performed for each data set to reduce high-frequency noise and the size of the ECG recordings f. Saving the pre-processed ECG data (D1*) from step e3. Method for predicting and localising structural changes in myocardial tissue according to claim 1 or 2, further comprising a step la, which takes place prior to step II and in which at least one clinical parameter (D2) is received, and further comprising a step IIa, which takes place between step la and step III and in which pre-processing of the at least one clinical parameter (D2) into a pre-processed clinical parameter (D2*) and a transfer of this at least one pre-processed clinical parameter (D2*) to an artificial neural network (N) takes place.