Method and system for analyzing diastolic function using only 2D echocardiography images

Abstract

A method (100) for classifying a diastolic function of a patient, comprising: (i) receiving (120) a plurality of 2D echocardiography images of a heart of the patient from an ultrasound device (280); (ii) analyzing (150) the plurality of 2D echocardiography images of the heart of the patient by a trained diastolic function prediction algorithm to estimate a left ventricular end diastolic pressure (LVEDP); (iii) classifying (160) the diastolic function of the patient as normal or abnormal based on the estimated LVEDP; and (iv) providing (170) to a user via a user interface an indication that the patient's diastolic function is normal or abnormal.

Description

Technical Field

[0001] The present disclosure generally relates to methods and systems for classifying diastolic function of a subject. Background Art

[0002] Heart failure, which can be defined as the inability of the heart to provide adequate cardiac output while maintaining normal filling pressures, affects at least 26 million people worldwide and is estimated to increase by 46% by 2030. There are two types of heart failure: (1) heart failure with reduced ejection fraction (HFrEF) and (2) heart failure with preserved ejection fraction (HFpEF). The latter (HFpEF), which accounts for 50% of heart failure cases, is characterized by impaired relaxation of the left ventricle (LV) during diastole (diastolic dysfunction) and increased filling pressures caused by altered LV mechanical properties, most notably higher stiffness. Conditions such as cardiac amyloidosis, coronary artery disease, valvular disease, hypertrophic cardiomyopathy (HCM), pericardial disease, and hypertension can produce HFpEF.

[0003] Although guidelines exist for diagnosing diastolic dysfunction (including history, physical examination, and echocardiography, and if indicated, cardiac catheterization), these guidelines are complex and rarely followed. Therefore, definitive diagnosis of diastolic dysfunction is challenging. Conditions such as cardiac amyloidosis, coronary artery disease (CAD), valvular disease, hypertrophic cardiomyopathy (HCM), pericardial disease, and hypertension can produce diastolic dysfunction. 2D echocardiography is the primary imaging modality for diastolic dysfunction. However, it cannot be used alone to make a definitive diagnosis. Figure 1 A schematic representation of the current diagnostic processing stages for the definitive and differential diagnosis of HFpEF is shown in , recognizing the full complexity of the differential diagnosis workflow. As can be seen, ultrasound is not used in step F2, the differential diagnosis step of the process.

[0004] Currently, noninvasive assessment of LVEDP using echocardiography relies on complex guidelines and is rarely followed. A large number of patients identified as having diastolic dysfunction using noninvasive assessment are ultimately found to have normal LVEDP, and vice versa.

[0005] For borderline cases related to the diagnosis of diastolic dysfunction, the decision whether to proceed with stress testing, catheterization, or perform advanced imaging, such as positron emission tomography (PET) or cardiovascular magnetic resonance imaging (CMR), is a subjective assessment by the cardiologist. This can lead to inefficient use of resources and suboptimal patient care. Agreement exists in the cardiac imaging community that there is a strong need for standardization, guided by clinically formulated health technology assessment studies, to obtain appropriate utilization of the different available tools, with higher efficiency and efficacy of care and better patient outcomes. However, currently, there are no recognized or intelligent patient-specific methods for predicting diastolic dysfunction, particularly using only 2D ultrasound B-mode images. Therefore, there is a great need to establish an intelligent data-driven decision support tool for analyzing diastolic function and definitively diagnosing diastolic dysfunction.

[0006] WO 2021 / 209400 A1 discloses an ultrasound system suitable for applying a relevant algorithm to determine a cardiac pressure value. Summary of the invention

[0007] Therefore, there is a continuing need for methods and systems for more accurately and efficiently analyzing diastolic function. Various embodiments and implementations herein relate to methods and systems configured to classify a patient's diastolic function. A system (such as a diastolic function analysis system) receives multiple 2D echocardiographic images of a patient's heart from an ultrasound device. A trained diastolic function prediction algorithm analyzes the multiple 2D echocardiographic images of the patient's heart to estimate left ventricular end-diastolic pressure (LVEDP). The system uses the estimated LVEDP to classify the patient's diastolic function as normal or abnormal, and optionally classifies the patient's diastolic function as intermediate or uncertain. The system then provides an indication of the patient's classified diastolic function to a user via a user interface.

[0008] In general, in one aspect, a method for classifying diastolic function of a patient is provided. The method includes: (i) receiving a plurality of 2D echocardiographic images of a patient's heart; (ii) analyzing the plurality of 2D echocardiographic images of the patient's heart by a trained diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP); (iii) classifying the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and (iv) providing an indication to a user via a user interface that the patient's diastolic function is normal or abnormal.

[0009] According to another aspect, a computer program product is provided comprising computer program code instructions which, when executed by a processor, enable the processor to perform a method for classifying a patient's cardiac diastolic function, the method comprising: receiving a plurality of 2D echocardiographic images of the patient's heart; analyzing the plurality of 2D echocardiographic images of the patient's heart by a trained cardiac diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP), wherein the trained cardiac diastolic function prediction algorithm is trained to receive one or more 2D echocardiographic images as input and to generate an estimate of left ventricular end-diastolic pressure (LVEDP) as output; classifying the patient's cardiac diastolic function as normal or abnormal based on the estimated LVEDP; and providing an indication to a user via a user interface as to whether the patient's cardiac diastolic function is normal or abnormal.

[0010] According to any aspect of the present invention, any one or more of the following optional features may be advantageously implemented.

[0011] In some embodiments, the plurality of 2D echocardiographic images of the patient's heart may be received from an ultrasound device.

[0012] The plurality of 2D echocardiographic images of the patient's heart may be received from a data storage device. The images may be previously acquired images. In other words, the method may be an "offline" method, wherein the method is performed after the images have been acquired. For example, the method in this set of embodiments may not include any step of acquiring echocardiographic images.

[0013] The trained diastolic function prediction algorithm may be trained to receive one or more 2D echocardiographic images as input and to generate an estimate of left ventricular end-diastolic pressure (LVEDP) as output.

[0014] The trained cardiac diastolic function prediction algorithm may be or may include a trained machine learning model or algorithm, such as a convolutional neural network (CNN).

[0015] In some examples, the machine learning model can be a 3D CNN. This allows the model to process inputs that include multiple images instead of just a single image.

[0016] In some embodiments, one or more 2D echocardiographic images may be subjected to image preprocessing before being supplied as input to the prediction algorithm. Image preprocessing may include reducing the image size to a uniform size, such as 120x120 pixels. Image preprocessing may include normalizing the image intensity to a range of [0,1], for example, by dividing each pixel value by 255 (where the image pixels use a standard 255 value pixel range for medical images). Normalizing the image intensity ensures that the image brightness does not interfere with the learning of the model.

[0017] Preferably the same preprocessing is performed during training and during inference.

[0018] According to a set of advantageous embodiments, the prediction algorithm may take the form of a prediction model comprising both a CNN module and an image preprocessing module, wherein the image preprocessing module is adapted to apply preprocessing to the image before feeding it into the CNN. In this way, an integrated model comprising both a preprocessing part and a deep learning part is provided. This has the advantage that the prediction model can receive native images generated by the ultrasound imaging device without modifying them in advance. Such preprocessing adds a limited inference time to the model while providing flexibility and lower development costs for model integration.

[0019] According to an embodiment, the method further comprises a step of selecting a subset of the received multiple 2D echocardiographic images of the patient's heart for analysis by means of a trained image selection algorithm, wherein the image selection algorithm is trained to select 2D echocardiographic images as optimal 2D echocardiographic images for analysis, wherein the analysis step comprises analyzing the selected subset of the received multiple 2D echocardiographic images of the patient's heart.

[0020] According to an embodiment, the method further comprises the step of receiving clinical information about the subject, wherein the trained diastolic dysfunction prediction algorithm further analyzes the received clinical information to classify the patient's diastolic function as normal or abnormal.

[0021] According to an embodiment, the method further comprises the step of receiving input from a user via a user interface to initiate analysis by a trained diastolic dysfunction prediction algorithm.

[0022] According to an embodiment, when the estimated LVEDP is equal to or less than 10 mmHg, the patient's diastolic function is classified as normal. According to an embodiment, when the estimated LVEDP is equal to or greater than 15 mmHg, the patient's diastolic function is classified as abnormal. According to an embodiment, the trained diastolic function prediction algorithm is further configured to classify the patient's diastolic function as uncertain when the estimated LVEDP is between 10 mmHg and 15 mmHg.

[0023] According to another aspect, a non-transitory computer-readable storage medium comprising computer program code instructions is provided, which, when executed by a processor, enables the processor to perform the following steps: (i) receiving multiple 2D echocardiographic images of a patient's heart, for example from an ultrasound device or a data storage device; (ii) analyzing the multiple 2D echocardiographic images of the patient's heart by a trained diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP); (iii) classifying the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and (iv) providing an indication to a user via a user interface whether the patient's diastolic function is normal or abnormal.

[0024] According to another aspect, a method for classifying a patient's cardiac diastolic function is provided. The method comprises: (i) receiving a plurality of 2D echocardiographic images of the patient's heart, for example from an ultrasound device or a data store; (ii) selecting a subset of the received plurality of 2D echocardiographic images of the patient's heart for analysis by a trained image selection algorithm, wherein the image selection algorithm is trained to select a 2D echocardiographic image as the best 2D echocardiographic image for analysis; (iii) receiving clinical information about the subject, wherein the clinical information includes demographic information; (iv) analyzing the selected subset of the 2D echocardiographic images of the patient's heart and the received clinical information by a trained cardiac diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP) (v) classifying the patient's diastolic function as normal, uncertain or abnormal based on the estimated LVEDP, wherein when the estimated LVEDP is equal to or less than 10 mmHg, the patient's diastolic function is classified as normal, wherein when the estimated LVEDP is equal to or greater than 15 mmHg, the patient's diastolic function is classified as abnormal, and wherein the trained diastolic dysfunction prediction algorithm is further configured to classify the patient's diastolic function as uncertain when the estimated LVEDP is between 10 mmHg and 15 mmHg; and (vi) providing an indication to a user via a user interface as to whether the patient's diastolic function is normal, uncertain or abnormal.

[0025] According to one aspect, a system for classifying diastolic function of a patient is provided. The system includes: an ultrasound device configured to obtain a plurality of 2D echocardiographic images of the patient's heart; a trained diastolic function prediction algorithm trained to estimate left ventricular end-diastolic pressure (LVEDP) based on the plurality of 2D echocardiographic images of the patient's heart; a user interface; and a processor configured to: (i) analyze the plurality of 2D echocardiographic images of the patient's heart using the trained diastolic function prediction algorithm to estimate the LVEDP; (ii) classify the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and (iii) direct the user interface to provide an indication of whether the patient's diastolic function is normal or abnormal.

[0026] It should be understood that all combinations of the foregoing concepts and the additional concepts discussed in more detail below (assuming such concepts are not mutually inconsistent) are considered to be part of the inventive subject matter disclosed herein. In particular, all combinations of the claimed subject matter appearing at the end of this disclosure are contemplated as being part of the inventive subject matter disclosed herein. It should also be understood that terms explicitly adopted herein that may also appear in any disclosure incorporated by reference should be given the meaning most consistent with the specific concepts disclosed herein.

[0027] These and other aspects of the various embodiments will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In the accompanying drawings, the same reference numerals generally refer to the same parts throughout the different views. The accompanying drawings illustrate features and modes of implementing various embodiments and should not be construed as limiting other possible embodiments falling within the scope of the appended claims. Moreover, the drawings are not necessarily drawn to scale, but generally emphasis is placed on illustrating the principles of the various embodiments.

[0029] Figure 1 is a diagram of the diagnostic workup stages for the definitive and differential diagnosis of HFpEF according to the prior art.

[0030] Figure 2 is a flow chart of a method for classifying diastolic function of a patient according to an embodiment.

[0031] Figure 3 is a schematic diagram of a cardiac diastolic function analysis system according to an embodiment.

[0032] Figure 4 is a flow chart of a method for training a cardiac diastolic function prediction algorithm according to an embodiment.

[0033] Figure 5A is a graph of a predicted evaluation of cardiac diastolic function according to an embodiment.

[0034] Figure 5B is a table of cardiac diastolic function prediction evaluation according to an embodiment.

[0035] Figure 6 is a schematic diagram of a presentation of cardiac diastolic function analysis according to an embodiment.

[0036] Figure 7 is a flow chart of a method for classifying diastolic function of a patient according to an embodiment. DETAILED DESCRIPTION

[0037] The present disclosure describes various embodiments of systems and methods configured to generate and present a classification of diastolic function of a subject. More generally, the applicant has recognized and appreciated that it would be beneficial to provide an intelligent data-driven decision support tool for diastolic function. Therefore, a diastolic function analysis system receives multiple 2D echocardiographic images of a patient's heart from an ultrasound device. A trained diastolic function prediction algorithm analyzes multiple 2D echocardiographic images of the patient's heart to estimate left ventricular end-diastolic pressure (LVEDP). The system uses the estimated LVEDP to classify the patient's diastolic function as normal or abnormal, and optionally classifies the patient's diastolic function as intermediate or uncertain. The system then provides an indication of the patient's classified diastolic function to a user via a user interface. A medical health professional can then use the provided indication of the patient's classified diastolic function to implement medical health treatment for the subject.

[0038] According to an embodiment, a deep learning (DL) algorithm is based solely on analysis of 2D echocardiographic images to accurately identify patients with elevated LV filling pressures. All existing models for diastolic function assessment using machine learning require ultrasound parameters (i.e., ultrasound measurements) to perform the analysis. There is no deep learning-based solution that uses only images as input for classification of diastolic dysfunction. One advantage of this deep learning-based model is that if the deep learning model only requires B-mode ultrasound images as input, it can benefit institutions that do not have access to advanced ultrasound systems (for Doppler measurements, tissue Doppler, etc.). In addition, ultrasound images can provide better accuracy compared to ultrasound values ​​alone as input. Successful 2D echocardiographic classification of elevated LVEDP improves existing paradigms for non-invasive diastolic function assessment, allowing improved detection and treatment of heart failure.

[0039] According to embodiments, in some non-limiting embodiments, the systems and methods described or otherwise contemplated herein may be implemented as a component of a commercial product for ultrasound imaging or analysis, or implemented as a component of a commercial product for cardiovascular analysis, such as IntelliSpace Cardiovascular (ISCV) (available from Koninklijke Philips NV, The Netherlands), or implemented as a component of a commercial product for patient analysis or monitoring, such as the Philips Patient Flow Set (PFCS), or any suitable system.

[0040] refer to Figure 2 , in one embodiment, is a flow chart of a method 100 for classifying diastolic function of a patient using a diastolic function analysis system 200. The methods described in conjunction with the figures are provided only as examples and should be understood not to limit the scope of the present disclosure. The diastolic function analysis system may be any system described herein or otherwise contemplated. The diastolic function analysis system may be a single system or a plurality of different systems.

[0041] At step 110 of the method, a cardiac diastolic function analysis system 200 is provided. Figure 3 In the embodiment of the cardiac diastolic function analysis system 200 shown, for example, the system includes one or more of a processor 220, a memory 230, a user interface 240, a communication interface 250, and a storage device 260 interconnected via one or more system buses 212. It should be understood that Figure 3 In some respects, the abstraction is constituted, and the actual organization of the components of the system 200 may be different and more complex than shown. Additionally, the diastolic function analysis system 200 may be any system described or otherwise contemplated herein. Other elements and components of the diastolic function analysis system 200 are disclosed or otherwise contemplated herein.

[0042] At step 120 of the method, the diastolic function analysis system receives a plurality of 2D echocardiographic images of the heart of the subject. The subject may be a patient, a user, or any other individual on whom the analysis is being performed. As a result of the ultrasonic analysis of the subject's heart, a plurality of 2D echocardiographic images of the subject's heart may be received. The ultrasonic analysis of the subject's heart may be any analysis sufficient to provide ultrasonic data about the subject, which may be utilized in downstream steps of the method. The ultrasonic analysis of the subject's heart may be obtained using any ultrasonic method or device capable of providing ultrasonic data utilized in downstream steps of the method. According to an embodiment, the ultrasonic analysis of the subject's heart includes a plurality of images obtained by an ultrasonic device.

[0043] According to an embodiment, the diastolic function analysis system 200 includes an ultrasound device capable of acquiring the desired ultrasound image or analysis. According to another embodiment, the diastolic function analysis system 200 communicates wired and / or wirelessly with a local or remote ultrasound device capable of acquiring the desired ultrasound image or analysis. According to another embodiment, the diastolic function analysis system 200 communicates wired and / or wirelessly with a local or remote database storing ultrasound images or analyses. The diastolic function analysis system 200 can obtain the desired ultrasound image or analysis from one or more of these sources.

[0044] According to an embodiment, an ultrasound analysis of the subject's heart is obtained by an ultrasound imaging specialist as part of a routine analysis of the subject, or in response to a possible or known medical problem experienced by the subject. The ultrasound analysis may be performed or obtained by methods and systems described or otherwise contemplated herein for immediate or near-term analysis, or may be performed or obtained by methods and systems described or otherwise contemplated herein for future analysis. According to an embodiment, the ultrasound analysis of the subject's heart includes a 2D image or recording, a 3D image or recording from which a 2D image may be extracted, and / or both.

[0045] At optional step 122 of the method, a user of the diastolic function analysis system 200 may initiate analysis of the patient's 2D echocardiographic images by a trained image selection algorithm. For example, the analysis may be performed while the image is being acquired or after the image is acquired. After the image is acquired, the analysis may be performed immediately or shortly after the examination of the patient is performed, or the analysis may be performed using stored images at a certain time period after the examination is performed. A request or command to initiate analysis of the patient's 2D echocardiographic images is submitted to the diastolic function analysis system via a user interface. The user interface may be any device or system that allows information to be transmitted and / or received, and may include a display, mouse, and / or keyboard for receiving user commands. The user interface may be a component of the diastolic function analysis system, and / or the request or command to initiate the analysis may be transmitted to the system from another device or system via wired and / or wireless communication.

[0046] At optional step 130 of the method, a trained image selection algorithm of the diastolic function analysis system 200 analyzes the received multiple 2D echocardiographic images and identifies a subset of the images as the best images for further analysis. The subset of images may be all or less than all of the received multiple 2D images. According to an embodiment, the image selection algorithm is trained to select the image as the best for further analysis based on any of a plurality of different features. In the context of the present disclosure, the best image (or the image identified as the best image) of the multiple images preferably refers to the fact that the image is more suitable for further analysis than the other (one or more) images of the multiple images. For example, the best may be defined as including detectable features in the image that can be used by the trained diastolic function algorithm. The best may include aspects such as clarity of the image (ensuring that the image or structures or features in the image are not blurred, the structures or features are present and / or visible) and other aspects.

[0047] Once the trained image selection algorithm has selected a subset of the received plurality of 2D echocardiographic images, the subset may be used immediately or may be stored in a local or remote storage device for use in other steps of the method.

[0048] At optional step 140 of the method, the diastolic function analysis system receives clinical information about the patient. The clinical information about the subject can be any information related to or useful in any downstream step of the method, including as input to a trained diastolic function prediction algorithm to estimate the patient's left ventricular end-diastolic pressure (LVEDP). According to an embodiment, the clinical information about the subject includes one or more of the following: type of ultrasound examination, reason for ultrasound analysis, age of the subject, sex of the subject, body mass index of the subject, disease state or diagnosis, medical treatment and medical diagnosis, and many other types of clinical information. For example, age may affect the interpretation of diastolic function. Therefore, these clinical information data can be a significant factor affecting LVEDP. Therefore, the received information can be any information related to the analysis of the patient as described herein or otherwise contemplated.

[0049] The diastolic function analysis system may receive patient clinical information from a variety of different sources. According to an embodiment, the diastolic function analysis system communicates with an electronic medical record database from which the patient clinical information may be obtained or received. According to an embodiment, the diastolic function analysis system includes an electronic medical record database or system 270, which may optionally communicate directly and / or indirectly with the system 200. According to another embodiment, the diastolic function analysis system may obtain or receive information from an instrument or medical healthcare professional that obtains the information directly from the patient.

[0050] The patient clinical information received by the diastolic function analysis system can be processed by the system according to methods for data handling and processing / preparation, including but not limited to methods described herein or otherwise contemplated. The patient clinical information received by the diastolic function analysis system can be utilized immediately before or after processing, or the patient clinical information received by the diastolic function analysis system can be stored in a local or remote storage device for use in other steps of the method.

[0051] At step 150 of the method, the trained diastolic function prediction algorithm analyzes the received plurality of 2D echocardiographic images of the patient's heart or a selected subset of the plurality of 2D echocardiographic images of the patient's heart to estimate the left ventricular end-diastolic pressure (LVEDP) of the patient's heart. According to one embodiment, the trained diastolic function prediction algorithm may also analyze the received clinical information about the patient, but this is not required. Other inputs to the trained diastolic function prediction algorithm are possible.

[0052] According to an embodiment, the trained diastolic function prediction algorithm estimates the LVEDP of the patient's heart using a variety of different classifiers and / or machine learning algorithms as described herein or otherwise contemplated. According to an embodiment, the trained diastolic function prediction algorithm of the diastolic function analysis system may be trained according to a variety of methods and measures. As an example, the algorithm may include a neural network measure.

[0053] refer to Figure 4 , in one embodiment, is a flow chart of a method 300 for training a diastolic function prediction algorithm of a diastolic function analysis system. At step 310 of the method, the system receives a training data set including training data (such as historical patient data) about a plurality of patients. For example, the training data may include input for previous or historical patients for whom data about diastolic function was obtained. For example, the training data may include information about patients who underwent 2D echocardiography and left heart catheterization with LVEDP measurement (with various different results). Thus, the diastolic function prediction algorithm can be trained to predict LVEDP based on the 2D echocardiogram. The training data may also include demographic information about previous or historical patients, and the training data may also be used to train the diastolic function prediction algorithm. The training data may be stored in and / or received from one or more databases. The database may be a local and / or remote database. For example, the diastolic function analysis system may include a database of training data.

[0054] According to an embodiment, the cardiac diastolic function analysis system may include a data preprocessor or similar component or algorithm configured to process the received training data. For example, the data preprocessor analyzes the training data to remove noise, bias, errors and other potential problems. The data preprocessor can also analyze the input data to remove low-quality data. Many other forms of data preprocessing or data point identification and / or extraction are possible.

[0055] At step 320 of the method, the system trains a machine learning algorithm, which will be an algorithm for analyzing input information, as described or otherwise contemplated. The machine learning algorithm is trained using a training data set according to known methods for training machine learning algorithms. According to an embodiment, the algorithm is trained using the processed training data set to analyze a plurality of 2D echocardiographic images of a patient's heart to estimate a left ventricular end-diastolic pressure (LVEDP) of the patient's heart, and thereby classify the patient's cardiac diastolic function as normal, uncertain, or abnormal based on the estimated LVEDP.

[0056] At step 330 of the method, the trained diastolic function prediction algorithm of the diastolic function analysis system is stored for future use. Depending on the embodiment, the model may be stored in a local or remote storage device.

[0057] According to an embodiment, the actual truth for LVEDP can be collected from the clear results of diagnostic tests for multiple patients, combining 2D echocardiography and left heart catheterization with LVEDP measurements. For example, the actual truth for the diastolic function status can be collected from invasive catheterization for LVEDP measurements, which can optionally be analyzed by a team of cardiologists for some or all patients. Based on the above inputs and actual truth about diastolic function, an algorithmic model based on deep learning can be implemented. According to an embodiment, this supervised learning method can be an institution-independent tool for the task of diastolic function classification. The accuracy of the AI-based learning network can become stronger over time by adding more data to it (such as a self-learning algorithm).

[0058] Further details regarding the architecture and training of an example prediction algorithm compatible with any of the embodiments described in this application will now be discussed.

[0059] According to one or more embodiments, the prediction algorithm comprises a machine learning algorithm operable to receive as input one or more 2D echocardiographic images of the patient's heart and to generate as output an estimate of left ventricular end-diastolic pressure (LVEDP).

[0060] The prediction algorithm is preferably operable to receive as input a plurality of 2D echocardiographic images of the patient's heart and to generate as output an estimate of left ventricular end-diastolic pressure (LVEDP).

[0061] According to one or more embodiments, the machine learning algorithm is a convolutional neural network (CNN).

[0062] In a preferred set of embodiments, the CNN is a 3D CNN. The 3D CNN is capable of processing a set of images as a cohesive unit. For example, the multiple images may be images acquired at different points in time, such as a series of images over time, or may be a set of image planes at different heights or depths in a region.

[0063] In a particularly advantageous set of embodiments, the use of a 3D CNN is proposed, which comprises a standard convolution operation (i.e., to process the image and filter the features), followed by batch normalization (for improving the performance and stability of the neural network) and a ReLU (rectified linear unit) activation layer, which introduces non-linearity. The convolutional layers are further followed by global average pooling layers (for compressing the information) and fully connected layers (for combining all the information). In this sense, the proposed CNN takes the form of a VGG-network architecture style, complemented by batch normalization. The CNN can be a 7-layer 3D CNN.

[0064] This type of neural network is highly effective in automatically and adaptively learning the spatial hierarchy of features. It includes input and output layers and multiple hidden layers, wherein the hidden layers include the aforementioned convolutional layers, pooling layers, fully connected layers, and normalization layers.

[0065] Preprocessing of the images can be performed before inputting them into the neural network. Preprocessing can include resizing each input image to a uniform size (if necessary), such as 120x120 pixels. Optionally, further preprocessing steps such as color normalization or data augmentation can be performed.

[0066] Considering the architecture in more detail, in some embodiments, in summary, a CNN includes several convolutional blocks (which help create a multi-dimensional image), a global pooling layer that transforms the image into a single vector, and a fully connected layer that reduces the dimensionality of the single vector.

[0067] In more detail, the proposed machine learning model may start with an initial convolutional layer with, for example, 64 feature maps and with, for example, a stride of 2 (i.e., the filter moves two pixel values ​​at a time). This may be followed by three resolution levels, each containing two blocks. Each block may include a 2D convolution operation, a batch normalization operation, and a ReLU operation. At the end of each block, a convolution with a stride of 2 may be included to halve the spatial dimension and double the number of feature maps.

[0068] With respect to the pipeline, according to at least one set of embodiments, it is proposed to first apply cardiac view recognition to each of the plurality of images. The plurality of images may include only AP4 and / or AP2 views. It has been found that these specific views or angles of the heart provide the most informative visual data for the application. Preliminary testing has shown improved performance by including only these vertex views. View recognition may be followed by cropping, preprocessing and / or enhancement.

[0069] As described above, the image can be preprocessed before being input to the CNN. Image preprocessing can include reducing the image size to a uniform size, such as 120x120 pixels. Image preprocessing can include normalizing the image intensity to a range of [0,1], such as by dividing each pixel value by 255 (where the image pixels use the standard 255-value pixel range for medical images). Normalizing the image intensity ensures that the image brightness does not interfere with the learning of the model.

[0070] Preferably the same preprocessing is performed during training and during inference.

[0071] According to a set of advantageous embodiments, the prediction algorithm can take the form of a prediction model comprising both a CNN module (e.g., any of the features described above) and an image preprocessing module that applies preprocessing to the image before feeding it into the CNN. In this way, an integrated model comprising both a preprocessing part and a deep learning part is provided. This has the advantage that the prediction model can receive native images generated by the ultrasound imaging device without modifying them in advance. Such preprocessing adds limited inference time to the model while providing flexibility and lower development costs for model integration.

[0072] It should also be noted that alternatively, hybrid approaches are also possible. For example, the preprocessing can be included in an interoperable or cross-platform model, such as an ONNX model, which is then run on the CPU of the computing device, while the deep learning component can be included in a second interoperable model, such as an ONNX model, which can be run on the GPU of the computing device.

[0073] Regarding model training, as already discussed, typically this involves the process of feeding training data to the CNN, and adjusting internal parameters to learn the characteristics of the data, such as by minimizing the error between the CNN predictions and the ground truth.

[0074] For example, the model can be trained using stochastic gradient descent optimization. As an example, the Adam optimizer can be employed, which provides first-order gradient-based optimization. Cosine-based learning rate decay without learning rate restarts is preferably used. Learning rate decay gradually reduces the learning rate over time, which results in more precise adjustments. Training is typically performed on 20-35 iterations or "epochs."

[0075] In some embodiments, during training, online enhancements are further applied to increase model robustness. These enhancements or modifications to the image may include, for example, any one or more of: modifying the brightness of the image, modifying the contrast of the image (commonly referred to as "gamma"), adjusting the position of the image (shift enhancement), flipping the image, and rotating the image. This improves the versatility of the model in dealing with a wide variety of images and helps avoid overfitting to the training data.

[0076] Example 1 - Training Data

[0077] According to a non-limiting example, the training data includes a database containing 10,000 digital imaging and communications in medicine (DICOM) images from 632 studies of patients who underwent 2D echocardiography and left heart catheterization with LVEDP measurement on the same day at the University of Chicago Medical Center in a retrospective study. The images were randomly divided into training (80%) and test (20%) data sets. A trained DL algorithm (convolutional neural network (CNN)) was developed using the training data set, which automatically classified AI-derived LVEDP into two categories using 2-D apical 4-chamber images (n=5,212 from 626 studies): elevated or abnormal (LVEDP≥15mmHg) and normal (LVEDP≤10mmHg), as verified by true invasive LVEDP. It is worth noting that other ranges can be used, including using 10-15mmHg as uncertainty. In the training set, automatic DL showed a high level of diagnostic accuracy reflected by a high AUC value of 1.00 and an overall accuracy of 1.00. In the test data set, the computational performance of the DL-based model remained strong (AUC 0.81, accuracy 0.76), such as Figure 5A and 5B According to an embodiment, the training data may also be combined with Doppler images and / or measurement results.

[0078] Return to Figure 2In method 100 of step 160, a diastolic function analysis system having a diastolic function prediction algorithm or another algorithm utilizes an estimated left ventricular end-diastolic pressure (LVEDP) of the patient's heart (determined by the trained diastolic function prediction algorithm in step 150) to classify the patient's diastolic function. The LVEDP-based classification may be determined by comparing the estimated LVEDP to a predetermined or learned range or threshold or by other methods.

[0079] For example, according to one embodiment, when the estimated LVEDP is equal to or less than 10 mmHg, the patient's diastolic function is classified as normal. According to an embodiment, when the estimated LVEDP is equal to or greater than 15 mmHg, the patient's diastolic function is classified as abnormal. According to an embodiment, when the estimated LVEDP is between 10 mmHg and 15 mmHg, the patient's diastolic function is classified as uncertain. Although these thresholds are provided as examples, other thresholds are possible. As described herein or otherwise contemplated, the thresholds can be predetermined, or they can be learned during initial or subsequent training of the diastolic function prediction algorithm.

[0080] At step 170 of the method, a classification of the patient's diastolic function is provided to the user via a user interface of the diastolic function analysis system 200, which may be normal, uncertain, abnormal, or other possible classifications. An indication of the classification of the patient's diastolic function may be transmitted via text or via visualization. For example, the indication may be transmitted via a simple text indication such as "normal," "abnormal," "uncertain," etc. The indication may include additional information, such as information about the patient, including optional demographic information provided to and / or used by the system, the patient's name, and / or treatment or intervention recommendations, and many other types of information. The indication may also include an estimated LVEDP determined by a diastolic function prediction algorithm, such as "8 mmHg," "13 mmHg," and "18 mmHg," as a few non-limiting examples.

[0081] According to an embodiment, the information may be transmitted to a user interface and / or another device via wired and / or wireless communications. For example, the system may transmit information to a mobile phone, a computer, a laptop, a wearable device, and / or any other device configured to allow display and / or other transmission of reports. The user interface may be any device or system that allows transmission and / or reception of information, and may include a display, a mouse, and / or a keyboard for receiving user commands.

[0082] According to an embodiment, the provided classification may be conveyed by a visualization such as a panel or other possible visualization. Figure 6, in one embodiment, is a schematic diagram of a possible visualization of the classification of a patient's diastolic function. The panel provides information including the patient's name, examination date, and details about the current ultrasound examination. The panel also provides information about the diastolic function analysis performed by the diastolic function prediction algorithm, including the estimated LVEDP (16 mmHg in this example) and the diastolic function classification (abnormal in this example). The panel also provides a recommendation for treatment or intervention or follow-up ("recommend intervention" in this example, although more specific treatment or intervention information is possible).

[0083] According to an embodiment of the diastolic function analysis system 200, the system may include a user interface to facilitate the methods described herein or otherwise contemplated. Thus, the user interface may include a "Decision Support Tools for Diastolic Function" button or activator that appears on an ultrasound scanner touch panel or workspace (such as the Philips IntelliSpace Cardiovascular (ISCV) platform) for the user to launch the application.

[0084] exist Figure 2 At optional step 180 of method 100 depicted in , a medical care professional may utilize the provided indication of the patient's estimated LVEDP and / or classification of diastolic function to implement a medical care treatment for the subject. For example, a clinician or other decision maker utilizes the patient's estimated LVEDP and / or classification of diastolic function to make patient care decisions. For example, a medical care recommendation may include a recommendation to initiate, continue, or stop a specific treatment configured to address abnormal diastolic function, based on the determined classification of the estimated LVEDP and / or diastolic function. Implementations may include a prescription, an order, additional testing, and / or another implementation. Many other implementations are possible.

[0085] According to an embodiment, the diastolic function analysis system 200 provides a user-friendly and intelligent patient-specific decision support tool for diastolic function analysis. Currently, there is no intelligent non-invasive decision tool for diastolic dysfunction determination for cardiologists. Subjective decisions about the need for advanced follow-up invasive catheterization and stress examinations lead to misdiagnosis, increased costs, inefficient use of resources, and suboptimal patient care. The diastolic function analysis system provides an intelligent patient-specific deep learning-based tool to notify cardiologists of the presence of diastolic dysfunction by utilizing 2D echocardiographic images of the left ventricle. Therefore, the system enables clinicians to have a decision support tool for diastolic dysfunction diagnosis, so that doctors can better plan patient care. The system utilizes information retrieved and / or calculated from a 2D ultrasound image of the left ventricle (LV) of the subject in the examination, wherein the ultrasound image can be established in a conventional 2D image acquisition from the LV of the heart. According to one possible embodiment, the preferred 2D image is taken with an apical 4-chamber view, without LV projection reduction and with good image quality.

[0086] refer to Figure 7 , in one embodiment, is a schematic diagram showing the inputs and outputs of an AI-based network for estimating diastolic dysfunction. The model input is an echocardiogram image (or a sequence of stacked images per study / per patient), and the output is a binary classification of diastolic function status (normal, uncertain, abnormal). Ground truth values ​​for diastolic function status are collected from invasive catheterization for LVEDP measurements, which are analyzed for each patient by a panel of cardiologists. Based on the above inputs and ground truth values ​​regarding diastolic function, a deep learning based algorithmic model can be implemented.

[0087] According to an embodiment, the diastolic function analysis system can utilize additional inputs to the system and / or trained diastolic function prediction algorithm. In addition to ultrasound images, additional inputs can also be considered for improving the deep learning network. These additional inputs include, but are not limited to, patient demographics (age, gender, BMI, history of diastolic dysfunction), other ultrasound biomarkers: Comprehensive ultrasound examination requires the collection of several quantitative parameters (such as global longitudinal strain (GLS), ejection fraction, left atrial volume index (LAVI), mitral propagation velocity (Vp), mitral inflow at the beginning of diastolic phase (E), tissue velocity at the beginning of diastolic phase (e'), tricuspid regurgitation (TR) velocity, relative wall thickness (RWT), LV thickness, septum, valve and RV thickness.

[0088] According to an embodiment, the diastolic function analysis system may also provide a saliency map along with the classification of the patient's diastolic function and / or the estimated LVEDP determined by the diastolic function prediction algorithm. For example, the saliency map may show the region(s) in the input image that the model pays particular attention to.

[0089] refer to Figure 3 , is a schematic diagram of a cardiac diastolic function analysis system 200. The system 200 may be any system described herein or otherwise contemplated, and may include any components described herein or otherwise contemplated. It should be understood that Figure 3 Some aspects are abstracted, and the actual organization of components of system 200 may be different and more complex than shown.

[0090] According to an embodiment, the system 200 includes a processor 220 that can execute instructions stored in the memory 230 or the storage device 260 or otherwise process data to, for example, perform one or more steps of the method. The processor 220 can be formed by one or more modules. The processor 220 can take any suitable form, including but not limited to a microprocessor, a microcontroller, multiple microcontrollers, a circuit, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a single processor, or multiple processors.

[0091] Memory 230 may take any suitable form, including non-volatile memory and / or RAM. Memory 230 may include various memories, such as L1, L2 or L3 cache or system memory. Thus, memory 230 may include static random access memory (SRAM), dynamic RAM (DRAM), flash memory, read-only memory (ROM) or other similar memory devices. The memory may store an operating system, etc. RAM is used by the processor for temporary storage of data. According to an embodiment, the operating system may include code that controls the operation of one or more components of system 200 when run by the processor. It will be apparent that in an embodiment where the processor implements one or more functions described herein in hardware, software described as corresponding to such functions in other embodiments may be omitted.

[0092] The user interface 240 may include one or more devices for implementing communication with the user. The user interface may be any device or system that allows transmission and / or reception of information, and may include a display, mouse, and / or keyboard for receiving user commands. In certain embodiments, the user interface 240 may include a command line interface or a graphical user interface that may be presented to a remote terminal via the communication interface 250. The user interface may be located together with one or more other components of the system, or may be located away from the system and communicated via a wired and / or wireless communication network.

[0093] The communication interface 250 may include one or more devices for implementing communication with other hardware devices. For example, the communication interface 250 may include a network interface card (NIC) configured to communicate according to the Ethernet protocol. Additionally, the communication interface 250 may implement a TCP / IP stack for communicating according to the TCP / IP protocol. Various alternative or additional hardware or configurations of the communication interface 250 will be apparent.

[0094] The storage device 260 may include one or more machine-readable storage media, such as a read-only memory (ROM), a random access memory (RAM), a magnetic disk storage medium, an optical storage medium, a flash memory device, or a similar storage medium. In various embodiments, the storage device 260 may store instructions for execution by the processor 220 or data that the processor 220 may operate on. For example, the storage device 260 may store an operating system 261 for controlling various operations of the system 200.

[0095] It will be apparent that various information described as being stored in storage device 260 may additionally or alternatively be stored in memory 230. In this regard, memory 230 may also be considered to constitute a storage device, and storage device 260 may be considered to be a memory. Various other arrangements will be apparent. In addition, both memory 230 and storage device 260 may be considered to be non-transitory machine-readable media. As used herein, the term non-transitory will be understood to exclude transient signals, but include all forms of storage devices, including both volatile memory and non-volatile memory.

[0096] Although system 200 is shown as including one of each described component, various components may be replicated in various embodiments. For example, processor 220 may include multiple microprocessors configured to independently run the methods described herein, or configured to perform the steps or subroutines of the methods described herein, so that multiple processors collaborate to implement the functions described herein. In addition, in the case where one or more components of system 200 are implemented in a cloud computing system, various hardware components may belong to separate physical systems. For example, processor 220 may include a first processor in a first server and a second processor in a second server. Many other variations and configurations are possible.

[0097] According to an embodiment, the electronic medical record database 270 is an electronic medical record database from which information about the patient, including clinical information, can be obtained or received. The electronic medical record database 270 can also be a database from which training data can be obtained or received. The electronic medical record database can be a local or remote database and communicate directly and / or indirectly with the cardiac diastolic function analysis system 200. Therefore, according to an embodiment, the cardiac diastolic function analysis system includes an electronic medical record database or system 270.

[0098] According to an embodiment, the system includes one or more ultrasound devices 280 capable of acquiring the desired ultrasound images or analysis. According to another embodiment, the diastolic function analysis system 200 communicates wired and / or wirelessly with a local or remote ultrasound device 280 capable of acquiring the desired ultrasound images or analysis. According to another embodiment, the diastolic function analysis system 200 communicates wired and / or wirelessly with a local or remote database 280 storing ultrasound images or analysis. The diastolic function analysis system 200 can obtain the desired ultrasound images or analysis from one or more of these sources.

[0099] According to an embodiment, the storage device 260 of the system 200 may store one or more algorithms, modules and / or instructions to perform one or more functions or steps of the methods described herein or otherwise contemplated. For example, the system may include a trained diastolic function prediction algorithm 262 and / or reporting instructions 263 and other instructions or data.

[0100] According to an embodiment, the trained diastolic function prediction algorithm 262 is configured to analyze a plurality of received 2D echocardiographic images of the patient's heart or a selected subset of the plurality of 2D echocardiographic images of the patient's heart to estimate the left ventricular end-diastolic pressure (LVEDP) of the patient's heart. According to one embodiment, the trained diastolic function prediction algorithm may also analyze received clinical information about the patient, but this is not required. Other inputs to the trained diastolic function prediction algorithm are possible. According to an embodiment, the trained diastolic function prediction algorithm estimates the LVEDP of the patient's heart using a variety of different classifiers and / or machine learning algorithms as described herein or otherwise contemplated. According to an embodiment, the trained diastolic function prediction algorithm of the diastolic function analysis system may be trained according to a variety of methods and measures. As an example, the algorithm may include a neural network measure. The trained diastolic function prediction algorithm 262 is trained using a training data set as described herein or otherwise contemplated.

[0101] According to an embodiment, the report instructions 263 direct the system to generate and provide information to the user via a user interface, the information including a classification of the patient's diastolic function (which may be normal, uncertain, abnormal, or other possible classifications) and optionally providing an estimated LVEDP determined by a diastolic function prediction algorithm. The information provided may include additional information, such as information about the patient, including optional demographic information provided to and / or used by the system, the patient's name and / or treatment or intervention recommendations, and many other types of information. The information may be transmitted via a user interface of the system or another device by wired and / or wireless communication. For example, the system may transmit information to a mobile phone, a computer, a laptop computer, a wearable device, and / or any other device configured to allow display and / or other transmission of a report. The user interface may be any device or system that allows transmission and / or reception of information, and may include a display, a mouse, and / or a keyboard for receiving user commands.

[0102] Thus, in the context disclosed herein, aspects of the embodiments may take the form of a computer program product embodied in one or more non-transitory computer-readable media having computer-readable program code embodied thereon. Thus, according to an embodiment, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium comprising computer program code instructions, the computer program code instructions, when executed by a processor, enabling the processor to perform a method comprising: (i) receiving a plurality of 2D echocardiographic images of a patient's heart; (ii) analyzing the plurality of 2D echocardiographic images of the patient's heart by a trained diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP); (iii) classifying the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and, (iv) providing an indication to a user via a user interface that the patient's diastolic function is normal or abnormal. The program code may be executed entirely on a user's computer, partially on a user's computer, as a stand-alone software package, partially on a user's computer and partially on a remote computer, or entirely on a remote computer or server.

[0103] According to an embodiment, the diastolic function analysis system is configured to process thousands or millions of data points in the input data used to train the system, as well as to process and analyze the received multiple 2D echocardiographic images. For example, generating a functional and skilled trained system using automated processes such as feature recognition and extraction and subsequent training requires processing millions of data points from the input data and the generated features. This would require millions or billions of calculations to generate a novel trained system from these millions of data points and millions or billions of calculations. Therefore, based on the input data and the parameters of the machine learning algorithm, the trained system is novel and different, and thus improves the functionality of the diastolic function analysis system. Therefore, generating a functional and skilled trained system includes a process with a large number of calculations and analyses that the human brain cannot complete in a lifetime or multiple lifetimes. By providing improved patient analysis, this novel diastolic function analysis system has a huge positive impact on patient diagnosis and care compared to prior art systems.

[0104] All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and / or ordinary meanings of the defined terms.

[0105] The indefinite articles "a" and "an" as used in this specification and the claims should be understood to mean "at least one" unless expressly stated otherwise.

[0106] The phrase "and / or" as used herein in the specification and claims should be understood to mean "either or both" of the elements so combined, i.e., elements that are present in combination in some cases and separately in other cases. Multiple elements listed with "and / or" should be interpreted in the same manner, i.e., "one or more" of the elements so combined. In addition to the elements specifically identified by the "and / or" clause, other elements may optionally be present, whether related or unrelated to those elements specifically identified.

[0107] As used herein in the specification and claims, "or" should be understood to have the same meaning as "and / or" as defined above. For example, when separating items in a list, "or" or "and / or" should be interpreted as inclusive, that is, including at least one of multiple elements or lists, but also including more than one, and optionally, additional unlisted items. Only terms that clearly indicate the opposite, such as "only one" or "exactly one", or when used in the claims, "consisting of..." will refer to including exactly one element in a list of multiple elements or elements. Generally, the term "or" as used herein will only be interpreted as indicating an exclusive alternative (i.e., "one or the other but not both") when preceded by an exclusive term, such as "either", "one", "only one" or "exactly one".

[0108] As used herein in the specification and claims, the phrase "at least one" with respect to a list of one or more elements should be understood to mean at least one element selected from any one or more elements in the list of elements, but not necessarily including at least one of each element specifically listed within the list of elements, and not excluding any combination of elements in the list of elements. This definition also allows that elements may optionally be present in addition to the elements specifically identified within the list of elements to which the phrase "at least one" refers, whether related or unrelated to those elements specifically identified.

[0109] It should also be understood that in any method claimed herein that includes more than one step or action, the order of the steps or actions of the method is not necessarily limited to the order in which the steps or actions of the method are listed unless explicitly stated to the contrary.

[0110] In the claims, as well as in the above description, all transitional phrases such as "comprising," "including," "carrying," "having," "containing," "involving," "maintaining," "consisting of," etc. should be understood as open-ended, i.e., meaning including but not limited to. Only the transitional phrases "consisting of" and "consisting essentially of" should be closed or semi-closed transitional phrases, respectively.

[0111] Although several inventive embodiments have been described and illustrated herein, a person of ordinary skill in the art will readily envision various other devices and / or structures for performing functions and / or obtaining results and / or one or more advantages described herein, and each of these changes and / or modifications is considered to be within the scope of the inventive embodiments described herein. More generally, it will be readily understood by those skilled in the art that all parameters, dimensions, materials, and configurations described herein are intended to be exemplary, and that actual parameters, dimensions, materials, and / or configurations will depend on one or more specific applications for which the inventive teachings are used. Those skilled in the art will recognize, or will be able to determine many equivalents of the specific inventive embodiments described herein using only routine experiments. Therefore, it should be understood that the foregoing embodiments are presented only as examples, and within the scope of the appended claims and their equivalents, inventive embodiments may be practiced in a manner different from that specifically described and claimed. The inventive embodiments of the present disclosure relate to each individual feature, system, article, material, kit, and / or method described herein. Additionally, any combination of two or more such features, systems, articles, materials, kits, and / or methods (if such features, systems, articles, materials, kits, and / or methods are not mutually contradictory) is included within the scope of the invention disclosed herein.

Claims

1. A computer program product comprising computer program code instructions which, when executed by a processor, enable the processor to perform a method (100) for classifying a patient's cardiac diastolic function, the method comprising: receiving (120) a plurality of 2D echocardiographic images of the patient's heart; analyzing (150) the plurality of 2D echocardiographic images of the patient's heart by a trained diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP), wherein the trained diastolic function prediction algorithm is trained to: receive one or more 2D echocardiographic images as input and generate an estimate of left ventricular end-diastolic pressure (LVEDP) as output; classifying (160) the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and An indication is provided (170) to a user via a user interface as to whether the patient's diastolic function is normal or abnormal.

2. The computer program product according to claim 1, wherein: The method further comprises the steps of selecting (130) a subset of the received plurality of 2D echocardiographic images of the patient's heart for analysis by a trained image selection algorithm, wherein the image selection algorithm is trained to select the 2D echocardiographic images as optimal for analysis, wherein the step of analyzing comprises analyzing the selected subset of the received plurality of 2D echocardiographic images of the patient's heart.

3. The computer program product according to claim 1 or 2, wherein: The method further includes the step of receiving (140) clinical information about the subject, wherein the trained diastolic dysfunction prediction algorithm further analyzes the received clinical information to classify the patient's diastolic function as normal or abnormal.

4. The computer program product according to any one of claims 1 to 3, wherein: The method further includes the step of receiving (122) input from a user via a user interface to initiate analysis by the trained diastolic dysfunction prediction algorithm.

5. The computer program product according to any one of claims 1 to 4, wherein: When the estimated LVEDP is equal to or less than 10 mmHg, the patient's diastolic function is classified as normal.

6. The computer program product according to any one of claims 1 to 5, wherein: When the estimated LVEDP is equal to or greater than 15 mmHg, the patient's diastolic function is classified as abnormal.

7. The computer program product according to any one of claims 1 to 6, wherein: The trained diastolic dysfunction prediction algorithm is also configured to classify the patient's diastolic function as uncertain when the estimated LVEDP is between 10 mmHg and 15 mmHg.

8. The computer program product according to any one of claims 1 to 7, wherein: The computer program product comprises a non-transitory computer-readable storage medium comprising computer program code instructions which, when executed by a processor, enable the processor to perform the method according to any one of claims 1-7.

9. A system (200) for classifying cardiac diastolic function of a patient, comprising: an ultrasound device (280) configured to obtain a plurality of 2D echocardiographic images of the patient's heart; a trained cardiac diastolic function prediction algorithm (262) trained to estimate left ventricular end-diastolic pressure (LVEDP) based on the plurality of 2D echocardiographic images of the patient's heart, wherein the trained cardiac diastolic function prediction algorithm is trained to: receive one or more 2D echocardiographic images as input and generate an estimate of left ventricular end-diastolic pressure (LVEDP) as output; User interface (240); A processor (220) configured to: (i) analyze the plurality of 2D echocardiographic images of the patient's heart to estimate LVEDP using the trained diastolic function prediction algorithm; (ii) classify the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and (iii) direct the user interface to provide an indication of whether the patient's diastolic function is normal or abnormal.

10. The system according to claim 9, wherein: The processor is also configured to select a subset of the multiple 2D echocardiographic images received of the patient's heart for analysis using a trained image selection algorithm, wherein the image selection algorithm is trained to select the 2D echocardiographic images as optimal for analysis, wherein the step of analyzing includes analyzing the selected subset of the multiple 2D echocardiographic images received of the patient's heart.

11. The system according to claim 9 or 10, wherein: The processor is further configured to receive clinical information about the subject, wherein the trained diastolic function prediction algorithm further analyzes the received clinical information to classify the patient's diastolic function as normal or abnormal.

12. The system according to any one of claims 9 to 11, wherein: The processor is further configured to receive input from a user via the user interface to initiate analysis by the trained diastolic dysfunction prediction algorithm.

13. The system according to any one of claims 9 to 12, wherein: When the estimated LVEDP is equal to or less than 10 mmHg, the patient's diastolic function is classified as normal.

14. The system according to any one of claims 9 to 13, wherein: When the estimated LVEDP is equal to or greater than 15 mmHg, the patient's diastolic function is classified as abnormal.

15. The system according to any one of claims 9 to 14, wherein: The trained diastolic function prediction algorithm is further configured to classify the patient's diastolic function as uncertain when the estimated LVEDP is between 10 mmHg and 15 mmHg.

16. A method (100) for classifying cardiac diastolic function of a patient, the method comprising: receiving (120) a plurality of 2D echocardiographic images of the patient's heart; analyzing (150) the plurality of 2D echocardiographic images of the patient's heart by a trained diastolic function prediction algorithm to estimate left ventricular end-diastolic pressure (LVEDP), wherein the trained diastolic function prediction algorithm is trained to: receive one or more 2D echocardiographic images as input and generate an estimate of left ventricular end-diastolic pressure (LVEDP) as output; classifying (160) the patient's diastolic function as normal or abnormal based on the estimated LVEDP; and An indication is provided (170) to a user via a user interface as to whether the patient's diastolic function is normal or abnormal.

Images