Non-invasive methods for detecting pulmonary hypertension

Abstract

Methods and computer program products for detecting pulmonary hypertension are provided. [Solution] A method, system, and computer program product for detecting pulmonary hypertension are provided herein, comprising the steps of: receiving voltage-time data from multiple leads of an electrocardiograph of a subject; generating a feature vector from the voltage-time data; providing the feature vector to a pre-trained learning system; and receiving an indication from the pre-trained learning system whether or not the subject has pulmonary hypertension.

Description

Technical Field

[0001] Cross - reference to Related Applications This application claims the benefit of U.S. Provisional Application No. 63 / 091,715, filed October 14, 2020, the entire content of which is incorporated herein by reference.

Background Art

[0002] Embodiments of the present disclosure relate to methods for the diagnosis and treatment of pulmonary hypertension. Pulmonary hypertension (PH) is a life - threatening disease that is estimated to affect 1% of the world's population and up to 10% of patients over 65 years of age. A timely diagnosis of PH is crucial not only for effective therapeutic intervention but also for enhancing the likelihood of survival. Multiple studies have suggested that earlier diagnosis, even by a few months, can lead to dramatic improvements in quality of life and life expectancy. However, the symptoms of PH are non - specific and closely resemble those seen in other common diseases, including asthma, chronic obstructive pulmonary disease (COPD), and heart failure. This makes timely referral to a respiratory physician or cardiologist who can confirm the diagnosis highly important to reduce the suspicion of PH in an outpatient setting. Currently, diagnostic delays mainly occur due to delays in appropriate referral to a specialist, with an average period of 2.5 years (from symptom onset to diagnosis), and in some cases up to 4 years. In practice, the gold standard for a definitive diagnosis of PH is right heart catheterization (RHC), which is an invasive procedure with non - negligible risks, so physicians often hesitate to proceed until all other diseases have been sequentially excluded. Therefore, there is a great need for new and improved methods (e.g., diagnostic methods) for the early and accurate detection of PH. Accordingly, embodiments of the invention disclosed herein include algorithms applied to electrocardiograms (ECGs), which are non - invasive procedures in the precise diagnosis of PH for the detection of PH and the stratification of patients based on the risk of PH, enabling earlier diagnosis and intervention.

Summary of the Invention

[0003] Embodiments of this disclosure provide a method and a computer program product for detecting pulmonary hypertension.

[0004] In some aspects of the present invention, a method is disclosed herein that includes the steps of: receiving voltage-time data of a subject, including voltage data from a plurality of leads of an electrocardiograph; generating a feature vector from the voltage-time data; providing the feature vector to a pre-trained learning system; and receiving an indication from the pre-trained learning system whether or not the subject has pulmonary hypertension.

[0005] Aspects of the present invention disclosed herein also include a system comprising a computing node having a computer-readable storage medium

[0006] In particular aspects of the present invention, a computer program product for detecting pulmonary hypertension is disclosed herein, comprising a computer-readable storage medium having program instructions embodied therein, wherein the program instructions are executable by a processor and cause the processor to perform a method including: receiving voltage-time data of a subject including voltage data of a plurality of reads from an echocardiogram; generating a feature vector from the voltage-time data; providing the feature vector to a pre-trained learning system; and receiving an instruction from the pre-trained learning system regarding the presence or absence of pulmonary hypertension in the subject. [Brief explanation of the drawing]

[0007] [Figure 1]This is a schematic diagram of a system for detecting pulmonary hypertension according to the embodiments of this disclosure. [Figure 2] This flowchart shows a method for detecting pulmonary hypertension according to an embodiment of this disclosure. [Figure 3] A computing node according to one embodiment of this disclosure is shown. [Figure 4] The distribution of sentence sentiment for Mayo's PH cohort (positive control) is shown, where approximately 180,000 patient-related sentences containing the term "pulmonary hypertension" were classified by the model. [Figure 5] The output of the sentiment analysis used to identify patients in Mayo's PH cohort who do not have PH according to the patient's clinical notes is shown. [Figure 6] This shows the output from a diagnostic model in a potential PH cohort, where the model was run on sentences containing "pulmonary hypertension." [Figure 7] This represents the output from a logistic regression model using a combination of augmented curation results and echocardiographic measurements. [Figure 8] This shows a cohort-based diagnostic model for ECG from the diagnostic window. [Figure 9] This shows a cohort-3 diagnostic model for ECG from the preventive window. [Figure 10] This shows the performance of the three cohort diagnostic models over time, looking back five years prior to diagnosis and using a six-month window for ECG data. [Figure 11] The output of the nferX® Human Genetics application was shown, which identified 26 genes whose mutations were associated with pulmonary hypertension. Two of these, KCNK3 and KCNA5, are potassium channels. [Figure 12-1] The output of the nferX® RNA Explorer application is shown, demonstrating that expression of both KCNK3 and KCNA5 is observed in both cardiac and nervous tissue. [Figure 12-2]The output of the nferX® RNA Explorer application is shown, demonstrating that expression of both KCNK3 and KCNA5 is observed in both cardiac and nervous tissue. [Figure 13] The output of the nferX Single Cell application is shown, revealing that both KCNK3 (top row) and KCNA5 (bottom row) are strongly expressed in neuronal and cardiac cell types, respectively. KCNK3 expression, in particular, is also observed at lower levels in cardiac cell types (not shown). [Figure 14] The output of the nferX Signals application was shown, which identified numerous sources of literature evidence suggesting that mutations in both KCNK3 (left) (also known as TASK1) and KCNA5 (right) can affect cardiac electrophysiology and therefore be detectable on an electrocardiogram. [Figure 15] A network architecture diagram for an exemplary 7-layer convolutional neural network (CNN) model that does not include transformer layers is shown. [Figure 16] A network architecture diagram for an exemplary 9-layer convolutional neural network (CNN) model, including two transformer layers, is shown. [Figure 17-1] This describes data augmentation in the training set to reduce the tendency of neural networks to overfit. The data is augmented by randomly applying one of the following: (A) masking a portion of the signal over time, (B) allowing only frequencies between 0.5 and 50 Hz, (C) stretching the signal with several zoom levels, (D) shifting voltages with small voltages in different leads, or (E) lowering the frequency band at 1 and 50 Hz, or rearranging a set of small leads. [Figure 17-2]This describes data augmentation in the training set to reduce the tendency of neural networks to overfit. The data is augmented by randomly applying one of the following: (A) masking a portion of the signal over time, (B) allowing only frequencies between 0.5 and 50 Hz, (C) stretching the signal with several zoom levels, (D) shifting voltages with small voltages in different leads, or (E) lowering the frequency band at 1 and 50 Hz, or rearranging a set of small leads. [Modes for carrying out the invention]

[0008] Convolutional neural networks provide a comprehensive method for analyzing and interpreting the vast amount of data generated in a single ECG. To screen for patients with PH (Peripheral Heart Disease), an algorithm was developed using retrospective patient-level data from the Mayo Clinic, including ECGs, procedural measurements, physician notes, and patient demographic data. Each ECG was paired with either a RHC (Raw Heart Chloride) or an echocardiogram to distinguish between PH and non-PH patients for model training. Cohorts were defined using the measurements obtained from these procedures. This yielded 65,994 unique patients (11,238 with PH and 54,756 without), of which 48% were used for model training, 12% for preliminary validation, and the remaining 40% for trial.

[0009] All models used voltage-time information from 12-read ECGs as input. The modeling techniques explored included convolutional neural networks with different structures, such as using all 12 reads as a single input, using groups of 3 reads as separate inputs, using each read converted to a spectrogram, and combinations of these methods. Furthermore, two different preliminary models were created, one in which an ECG was performed within one month of the patient's diagnosis (diagnostic model), and the other in which an ECG was performed 6 to 18 months before the date of diagnosis (preventive model). Based on their relative performance, the preliminary diagnostic model was selected for further development.

[0010] The best-performing preliminary diagnostic model was a convolutional neural network with residual connections incorporating a single input of 12 reads. The updated diagnostic model yielded an area under the curve (AUC) of 0.94 for the diagnostic validation and test sets, and an AUC of 0.90 for the validation and test sets allowed for the differentiation of PH 6–18 months prior to diagnosis. Finally, ECGs performed 3–5 years prior to diagnosis showed no significant decline in performance, with an AUC greater than 0.8466. Ultimately, these results demonstrate a strong signal within ECGs for detecting PH and suggest the potential for implementation in ECG instruments to facilitate patient diagnosis and intervention in primary and secondary care settings. Furthermore, the disease also involves underlying genetic elements, which are supported by the detection of these diagnostic signals 3–5 years prior to diagnosis. The methods disclosed herein may provide further specificity and sensitivity when combined with gene panels. Similarly, such methods may be used in conjunction with gene panels to detect novel biomarkers of the disease.

[0011] Thus, in some aspects of the present invention, a method is disclosed herein that includes receiving voltage-time data of a subject including voltage data of a plurality of leads of an electrocardiograph, generating a feature vector from the voltage-time data, providing the feature vector to a pre-trained learning system, and receiving an indication of the presence or absence of pulmonary hypertension in the subject from the pre-trained learning system. The step of generating the feature vector may include generating a spectrogram based on the voltage data of the plurality of leads. In some embodiments, the step of generating the feature vector includes grouping the voltage data of the plurality of leads into a plurality of subsets.

[0012] In some embodiments, such a method further includes receiving demographic information of the subject, wherein the step of generating the feature vector includes adding the demographic information to the feature vector. In some such embodiments, the learning system includes a convolutional neural network. Such a convolutional neural network may include at least one residual connection.

[0013] In some embodiments, the voltage-time data of the subject is received from an electrocardiograph. In further embodiments, the voltage-time data of the subject is received from an electronic medical record.

[0014] In some embodiments, the method further includes providing an instruction to an electronic health record system for storage in a health record associated with the subject. In some embodiments, the method further includes providing an instruction to a computing node for display to a user.

[0015] Next, referring to FIG. 1, a system for detecting pulmonary hypertension according to an embodiment of the present disclosure is shown. As outlined above, in various embodiments, patient information including electrocardiogram (ECG) data is provided to a learning system to determine the presence of pulmonary hypertension. Accordingly, aspects of the present invention disclosed herein include a system comprising an electrocardiograph including a plurality of leads and a computing node having a computer-readable storage medium with program instructions embodied thereon, the program instructions being executable by a processor of the computing node and including receiving voltage-time data of a subject including voltage data of a plurality of leads from an echocardiogram, generating a feature vector from the voltage-time data, providing the feature vector to a pre-trained learning system, and receiving an indication of the presence or absence of pulmonary hypertension in the subject from the pre-trained learning system. The step of generating the feature vector may include generating a spectrogram based on the voltage data of the plurality of leads. In some embodiments, the step of generating the feature vector includes grouping the voltage data of the plurality of leads into a plurality of subsets.

[0016] In some embodiments, such a system further includes receiving demographic information of the subject, wherein the step of generating the feature vector includes adding the demographic information to the feature vector. In some such embodiments, the learning system includes a convolutional neural network. Such a convolutional neural network may include at least one residual connection.

[0017] In some embodiments, the voltage-time data of the subject is received from an electrocardiograph. In further embodiments, the voltage-time data of the subject is received from an electronic medical record.

[0018] In some embodiments, the system further includes the step of providing instructions to an electronic health record system for storing information in a health record associated with the subject. In some embodiments, the system further includes the step of providing instructions to a computing node for display to a user.

[0019] Patient data may be received from an electronic health record (EHR) 101. An electronic health record (EHR), or electronic medical record (EMR), may refer to a systematic collection of electronically stored health information of patients and populations in digital format. These records can be shared across different healthcare environments. Records may be shared through networked enterprise-wide information systems or other information networks and exchanges. An EHR may include a range of data, including demographic data, medical history, medication and allergy records, immunization records, clinical test results, radiographic images, vital signs, age and weight, and billing information. The EHR system may be designed to store data and capture the patient's condition over time. In this way, the need to search for the patient's past paper medical records is eliminated.

[0020] Electrocardiogram (ECG) data may be received directly from the electrocardiogram device 102. In an exemplary 12-lead ECG, 10 electrodes are attached to the surface of the patient's limbs and chest. The overall magnitude of the cardiac potential is then measured from 12 different angles (leads) and recorded over a period of time (usually 10 seconds). In this way, the overall magnitude and direction of the cardiac electrical depolarization are captured at each moment throughout the entire cardiac cycle.

[0021] Further datastores 103 may include further patient information as described herein. Preferred datastores include databases, single-layer files, and other structures known in the art.

[0022] Naturally, ECG data may be stored in the EHR for later retrieval. Also, naturally, ECG data may be cached rather than sent directly to the learning system for further processing.

[0023] The learning system 104 receives patient information from one or more of the EHR 101, ECG 102, and further data stores 103. As described above, in some embodiments, the learning system includes a convolutional neural network. In various embodiments, the input to the convolutional neural network includes voltage-time information relating to the ECG, which in some embodiments is paired with further patient information such as demographic data or genetic information.

[0024] The learning system 104 may be pre-trained with suitable population data, as described in the embodiments, to generate an indication of the presence or absence of pulmonary hypertension. In some embodiments, this indication is binary. In some embodiments, this indication is a probability value indicating the likelihood of pulmonary hypertension, taking into account the input patient data.

[0025] In some embodiments, the learning system 104 provides pulmonary hypertension indications for memorization as part of the EHR. In this way, a computer-aided diagnosis is provided that can be referenced by a clinician. In some embodiments, the learning system 104 provides pulmonary hypertension indications to a remote client 105. For example, the remote client may be a health application, cloud service, or another consumer of diagnostic data. In some embodiments, the learning system 104 is integrated into an ECG device for immediate feedback to the user during testing.

[0026] In some embodiments, a feature vector is provided to the learning system. Based on the input features, the learning system generates one or more outputs. In some embodiments, the output of the learning system is a feature vector.

[0027] In some embodiments, the learning system includes an SVM. In other embodiments, the learning system includes an artificial neural network. In some embodiments, the learning system is pre-trained with training data. In some embodiments, the training data is retrospective data. In some embodiments, the retrospective data is stored in a data store. In some embodiments, the learning system may be further trained by manual curation of previously generated outputs.

[0028] In some embodiments, the learning system is a trained classifier. In some embodiments, the trained classifier is a random decision forest. However, naturally, a variety of other classifiers, including linear classifiers, support vector machines (SVMs), or neural networks such as recurrent neural networks (RNNs), are suitable for use in accordance with this disclosure.

[0029] Suitable artificial neural networks include, but are not limited to, feedforward neural networks, radial basis function networks, self-organizing maps, learning vector quantization, recurrent neural networks, Hopfield networks, Boltzmann machines, echo-state networks, long- and short-term memory, bidirectional recurrent neural networks, hierarchical recurrent neural networks, probabilistic neural networks, modular neural networks, associative neural networks, deep neural networks, deep belief networks, convolutional neural networks, convolutional deep belief networks, mass memory and retrieval neural networks, deep Boltzmann machines, deep stacking networks, tensor deep stacking networks, spike-and-slab restricted Boltzmann machines, complex hierarchical deep models, deep coding networks, multilayer kernel machines, or deep Q networks.

[0030] In machine learning, a convolutional neural network (CNN) is a class of feedforward artificial neural networks applicable to analyzing visual images and other natural signals. A CNN consists of input and output layers as well as multiple hidden layers. The hidden layers of a CNN typically consist of convolutional layers, pooling layers, fully connected layers, and normalization layers. A convolutional layer applies a convolutional operation to the input and passes the result to the next layer. Convolution mimics the response of individual neurons to a stimulus. Each convolutional neuron processes data only for its receptive field.

[0031] Convolutional operations allow for a reduction in free parameters compared to fully connected feedforward networks. In particular, tiling a given kernel makes it possible to learn a fixed number of parameters regardless of image size. This also reduces the memory footprint for a given network.

[0032] The parameters of the convolutional layer consist of a set of learnable filters (or kernels) that have small receptive fields but extend through the entire depth of the input volume. During the forward pass, each filter is convolved across the width and height of the input volume, and the dot product between the filter's entry and the input is calculated to create a two-dimensional activation map of that filter. As a result, the network learns filters that are activated when certain types of features are detected at certain spatial locations in its input.

[0033] In the exemplary convolution, the kernel contains multiple weights w1…w9. Naturally, the sizes provided here are illustrative, and any kernel dimensions may be used as described herein. The kernel is applied to each tile of the input (e.g., an image). The result for each tile is an element of the feature map. Naturally, multiple kernels may be applied to the same input to generate multiple feature maps.

[0034] Feature maps are stacked for all kernels to form the entire output volume of the convolutional layer. Therefore, every entry in the output volume can also be interpreted as the output of a neuron that shares the same feature map and parameters as the neuron, looking at a small region in the input.

[0035] Convolutional neural networks can be implemented on a variety of hardware, including hardware CNN accelerators and GPUs.

[0036] Referring next to Figure 2, a flowchart is provided illustrating a method for detecting pulmonary hypertension according to an embodiment of the present disclosure. In step 201, voltage-time data of the subject is received. The voltage-time data includes voltage data from multiple leads of an electrocardiograph. In step 202, a feature vector is generated from the voltage-time data. In step 203, the feature vector is provided to a pre-trained learning system. In step 204, an indication of the presence or absence of pulmonary hypertension in the subject is received from the pre-trained learning system.

[0037] Next, referring to Figure 3, a schematic diagram of an example of a compute node is shown. Compute node 10 is merely one example of a suitable compute node and does not imply any limitation on the scope of use or functionality of the embodiments described herein. Nevertheless, compute node 10 can implement and / or perform any of the functionalities described above.

[0038] Computing node 10 contains computer systems / servers 12 that operate with numerous other general-purpose or dedicated computing system environments or configurations. Examples of well-known computing systems, environments, and / or configurations that may be suitable for use with computer systems / servers 12 include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments that include any of the above systems or devices.

[0039] The computer system / server 12 may be described in the general context of computer system executable instructions, such as program modules, that are executed by the computer system. Generally, a program module may include routines, programs, objects, components, logic, and data structures that perform a specific task or implement a specific abstract data type. The computer system / server 12 may run in a distributed cloud computing environment where tasks are performed by remote processing units linked over a communication network. In a distributed cloud computing environment, program modules may reside in both local and remote computer system storage media, such as memory storage devices.

[0040] As shown in Figure 3, the computer system / server 12 of the computing node 10 is shown in the form of a general-purpose computing device. The components of the computer system / server 12 include, but are not limited to, one or more processors or processing units 16, system memory 28, and a bus 18 that connects various system components such as the system memory 28 to the processor 16.

[0041] Bus 18 represents one or more of several bus structures, such as a memory bus or memory controller, peripheral bus, accelerated graphics port, and a processor or local bus using one of various bus architectures. Examples of such architectures include, but are not limited to, the Industrial Standard Architecture (ISA) bus, Microchannel Architecture (MCA) bus, Extended ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, Peripheral Interconnect (PCI) bus, Peripheral Interconnect Express (PCIe), and Advanced Microcontroller Bus Architecture (AMBA).

[0042] The computer system / server 12 typically includes various computer system-readable media. Such media may be any available media accessible by the computer system / server 12, and include both volatile and non-volatile media, and removable and non-removable media.

[0043] The system memory 28 may include computer system-readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. The computer system / server 12 may further include other removable / non-removable volatile / non-volatile computer system storage media. As just one example, the storage system 34 may be provided for reading and writing to a non-removable non-volatile magnetic medium (not shown, typically referred to as a “hard drive”). Not shown, a magnetic disk drive may be provided for reading and writing to a removable non-volatile magnetic disk (e.g., a “floppy disk”), and an optical disk drive may be provided for reading and writing to a removable non-volatile optical disk, such as a CD-ROM, DVD-ROM, or other optical medium. In such cases, each may be connected to the bus 18 by one or more data medium interfaces. As further shown and described below, the memory 28 may include at least one program product having a set of program modules (e.g., at least one) configured to perform the functions of the embodiments of the present disclosure.

[0044] A program / utility 40 having a set (at least one) of program modules 42 may, as an example but not limited to, be stored in memory 28 as well as in the operating system, one or more application programs, other program modules and program data. Each or several combinations of the operating system, one or more application programs, other program modules and program data may include an implementation of a networking environment. The program modules 42 generally perform the functions and / or methodologies of the embodiments described herein.

[0045] The computer system / server 12 may also communicate with one or more external devices 14, such as a keyboard, pointing device, and display 24; one or more devices that enable a user to interact with the computer system / server 12; and / or any devices that enable the computer system / server 12 to communicate with one or more other computing devices (e.g., a network card, modem, etc.). Such communication can be performed via the input / output (I / O) interface 22. Furthermore, the computer system / server 12 may communicate with one or more networks, such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet), via the network adapter 20. As illustrated, the network adapter 20 communicates with other components of the computer system / server 12 via the bus 18. It should be understood that other hardware and / or software components, not shown in the illustration, can be used with the computer system / server 12. Examples, but not limited to, include microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archiving systems.

[0046] This disclosure may be embodied as a system, method and / or computer program product. For example, in some aspects of the present invention, a computer program product for detecting pulmonary hypertension is provided herein, comprising a computer-readable storage medium having program instructions embodied therewith, wherein the program instructions are executable by a processor, and the computer program product causes the processor to perform a method comprising: receiving voltage-time data of a subject including voltage data of a plurality of reads from an echocardiogram; generating a feature vector from the voltage-time data; providing the feature vector to a pre-trained learning system; and receiving an indication from the pre-trained learning system whether or not the subject has pulmonary hypertension. The step of generating the feature vector may include generating a spectrogram based on the voltage data of the plurality of reads. In some embodiments, the step of generating the feature vector includes grouping the voltage data of the plurality of reads into a plurality of subsets.

[0047] In some embodiments, such a computer program product further includes the step of receiving demographic information of interest, where the step of generating feature vectors includes adding the demographic information to the feature vectors. In some such embodiments, the learning system includes a convolutional neural network. Such a convolutional neural network may include at least one residual connection.

[0048] In some embodiments, the voltage-time data in question is received from an electrocardiograph. In further embodiments, the voltage-time data in question is received from an electronic medical record.

[0049] In some embodiments, the computer program product further includes the step of providing instructions to an electronic health record system for storing information in a health record associated with the subject. In some embodiments, the system further includes the step of providing instructions to a computing node for display to the user.

[0050] The computer program products provided herein may include a computer-readable storage medium having computer-readable program instructions for causing a processor to execute aspects of this disclosure.

[0051] A computer-readable storage medium may be a tangible device capable of holding and storing instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any preferred combination thereof. A non-exclusive list of more specific examples of computer-readable storage media includes portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital multipurpose disks (DVDs), memory sticks, floppy disks, mechanically encoded devices such as punch cards or raised structures in grooves on which instructions are recorded, and any preferred combination thereof. The computer-readable storage media used herein should not be construed as themselves being transient signals such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses passing through fiber optic cables) or electrical signals transmitted through wires.

[0052] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computer / processor, or to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include copper transmission cables, optical transmission fibers, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface within each computer / processor receives computer-readable program instructions from the network and transfers them for storage on a computer-readable storage medium within each computer / processor.

[0053] The computer-readable program instructions for performing the operations of this disclosure may be assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, such as smalltalk or object-oriented programming languages ​​like C++ and conventional procedural programming languages ​​like the C programming language or similar languages. The computer-readable program instructions may run entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer via any type of network, such as a local area network (LAN) or wide area network (WAN), or the connection may be to an external computer (for example, via the Internet using an Internet service provider). In some embodiments, electronic circuits such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) may execute computer-readable program instructions by individualizing the electronic circuit using state information of computer-readable program instructions in order to carry out aspects of the present disclosure.

[0054] Aspects of the present disclosure are described herein with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present disclosure. Naturally, each block in a flowchart and / or block diagram, as well as any combination of blocks in a flowchart and / or block diagram, can be executed by computer-readable program instructions.

[0055] A machine may be constructed by providing these computer-readable program instructions to a general-purpose computer, a dedicated computer, or a processor of a programmable data processing device, so as to create means for implementing functions / operations specified in the blocks of a flowchart and / or block diagram, which are executed via the processor of a computer or other programmable data processing device. Alternatively, these computer-readable program instructions may be stored in a computer-readable storage medium that can direct a computer, a programmable data processing device, and / or other device to function in a particular way, so as to include a product containing instructions that implements the modes of functions / operations specified in the blocks of a flowchart and / or block diagram.

[0056] Alternatively, computer-readable program instructions may be loaded into a computer, another programmable data processing device, or another device that causes a series of actions to be executed on a computer, another programmable device, or other device, so that the instructions executed on the computer, another programmable device, or other device perform the functions / actions specified in the blocks of the flowchart and / or block diagram, thereby generating a process to be implemented on a computer.

[0057] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions containing one or more executable instructions for implementing a specified logical function. In some other implementations, the functions described in a block may be performed in an order other than that shown in the figure. For example, two blocks shown consecutively may actually be executed substantially simultaneously, or they may sometimes be executed in reverse order depending on the relevant functionality. It should also be noted that each block in a block diagram and / or flowchart diagram, and combinations of blocks in a block diagram and / or flowchart diagram, can be executed by a dedicated hardware-based system that performs the specified function or operation or combinations of dedicated hardware and computer instructions.

[0058] While descriptions of various embodiments of this disclosure have been provided for illustrative purposes, they are not intended to be exhaustive or to limit the embodiments disclosed. Many modifications and variations that do not deviate from the scope and spirit of the embodiments described will be apparent to those skilled in the art. The terminology used herein has been selected to best describe the principles of the embodiments, their practical applications, or any technological improvements that surpass the technology available on the market, and to enable other persons skilled in the art to understand the embodiments disclosed herein. [Examples]

[0059] Example 1: Cohort identification and expanded curation Using a combination of mean pulmonary artery pressure (mPAP) measured during RHC and tricuspid regurgitation velocity (TRV) measured during echocardiography, we generated diverse sets of cohorts for the development of a preliminary model (Table 1). RHC is the gold standard for PH diagnosis, with mPAP ≥ 21 mmHg indicating PH (recently lower than the historical threshold of 25 mmHg). TRV measurements are not as definitive; a TRV of ≤ 2.8 m / s indicates the absence of PH, while a TRV of ≥ 3.4 m / s indicates its presence. However, there is an intermediate range where diagnosis is not definitive using TRV alone, and other measurements must also be considered. [Table 1]

[0060] Furthermore, a cohort was generated using diagnoses extracted from clinical notes and combined with echocardiographic measurements to test the potential for extended curation (Cohort 7). This cohort was generated using a subset of patients with echocardiographic measurements, consisting of a smaller number of PH patients. As a first step toward this objective, physicians in Mayo provided a positive control cohort of 1,630 patients, which will hereafter be referred to as the Mayo PH cohort. To expand this cohort, an additional 19,504 patients with the term "pulmonary hypertension" included in their notes were identified, which will hereafter be referred to as the potential PH cohort. A BERT model was trained to classify sentiments regarding the PH diagnosis. Using this model, true positive and negative patients were identified in the Mayo PH cohort and the potential PH cohort, and such a model could be extended to further diseases and characteristics.

[0061] Example 2: Training a BERT model for diagnosis As the first step toward creating a BERT model for diagnosis, the top 250 phenotypes most closely associated with "pulmonary hypertension" were determined using the nference® Signals application, and sentences from the Mayo corpus of clinical notes were extracted for these phenotypes. The sentences were manually categorized, e.g., positive (YES), negative (NO), suspected (MAYBE), and other context (OTHER), examples of which are shown in Table 2. [Table 2]

[0062] Optionally, further categories may be added to this training set to support improved model granularity, for example, separating disease risks arising from family history and / or medication (both of which are included by OTHER as illustrated above).

[0063] To improve efficiency in sentence tagging, we developed a Deep Model Builder application with a user interface that tracks changes made across multiple users. A first model was generated using 11,433 sentences, which had an overall accuracy of 0.85, where the percentage of accurately predicted models for all sentences was defined as the percentage of labels. Deep Model Builder not only allows users to review tagged sentences where the model misclassified, but it can also be used to run the model with sets of untagled sentences, further improving efficiency downstream of the "extended curation" process. As shown in Table 3, multiple cycles of extended curation improved the model's accuracy from 0.85 to 0.936. [Table 3]

[0064] The above model was trained using 250 different PH-related phenotypes, so the sentences used to train this model primarily discussed diseases related to cardiovascular, respiratory, and metabolic disorders. Considering the breadth of phenotypes already incorporated by the model, it suggests that the model is robust enough to be extended to further therapeutic areas, from COVID-19 to tumors, with minimal retraining (approximately 1000-3000 sentences). Some areas will require further curation to incorporate specific language or context within their particular domains.

[0065] To extract new features or more specific associations, it is necessary to train a new model. In this case, augmented curation becomes important because it allows for the flexibility to create specific improved models for any use case. Therefore, other methods, which require pre-defining the necessary fields by structured text data for manual curation, cannot keep pace with the speed at which a new model can be trained for any desired feature by the disclosed method.

[0066] Example 3: Use of a diagnostic model for cohort selection Before running the BERT model on a potential PH cohort to identify further PH patients, we ran it on the Mayo PH cohort to evaluate the distribution of sentence sentiment analysis for positive controls. Here, approximately 180,000 sentences for these patients containing the term "pulmonary hypertension" were classified by the model. As shown in Figure 4, on average, 68% of sentences were classified as YES sentiment, only 2% as NO, 7% as MAYBE, and 23% as OTHER, demonstrating excellent validation of our model against a positive cohort.

[0067] The sentiment analysis shown above was also used to identify patients without PH in the Mayo PH cohort according to their clinical notes. Of the 1,630 patients for whom clinical notes were provided, sentiment analysis and subsequent manual review identified 35 patients without PH in this cohort. An example of this semi-automated workflow is shown in Figure 5. Here, the distribution includes PH-negative patients, resulting in a longer tail in the NO classification. For 25 patients within this specific tail, an inference application built within the Mayo environment was used to examine each entry of "pulmonary hypertension" in their notes, resulting in 7 patients having PH, 2 suspected of having PH, and 16 not having PH. The remaining 19 patients without PH in this cohort were identified in previous iterations. By excluding these 35 patients, who represent more than 2% of the Mayo PH cohort, we were able to demonstrate a significant difference in model performance.

[0068] After validating the diagnostic model against Mayo's PH cohort, we ran the model on 19,504 patients in a potential PH cohort using sentences containing "pulmonary hypertension." As shown in Figure 6, the mean YES sentiment was lower at 58% than in Mayo's PH cohort, but this result can be mainly explained by the 30% of patients who did not have a YES sentence. Similarly, nearly 80% of patients did not have a sentence containing a NO sentiment, which means that the inference could potentially increase the PH-positive control set by an order of magnitude.

[0069] To automate the identification of positive and negative PH patients in these cohorts, various logistic regression models were tested using augmented curation results and / or combinations of echocardiographic measurements, i.e., TRV and estimated right atrial pressure (RAP). Features used to describe patients by augmented curation included the percentage of sentences with sentiments YES, NO, MAYBE, and OTHER, as well as the number of PH occurrences per note. Features used for TRV and RAP included the mean, median, minimum, maximum, and standard deviation for each measurement. A positive control cohort of 1556 patients with a positive diagnosis and echocardiographic measurements was generated from Mayo's PH cohort. A negative control cohort was generated by manual curation of patient records with TRV and RAP measurements. Models were evaluated using 10-fold cross-validation and a 90:10 training / test split.

[0070] As shown in Figure 7, augmented curation performs better when combined with echocardiographic measurements than when either is used alone. Furthermore, augmented curation performs significantly better than echocardiographic measurements alone. This was expected, as the goal of augmented curation is to incorporate the physician's interpretation of the overall outcome of the patient's records.

[0071] 200 patients were randomly sampled as a holdout set, and their records were manually curated to determine whether or not they were diagnosed with PH. One patient withdrew their consent and was subsequently excluded. Of the remaining 199 patients, 191, or 95.9%, were accurately classified using the logistic regression model.

[0072] Example 4: Algorithm Development and Results For each cohort, the preliminary model was evaluated in two different time windows: either one month before or after the date of diagnosis (diagnosis window) and 6 to 18 months before diagnosis (prevention window). In the preliminary model, all ECGs from negative patients were examined. In the updated model, ECGs from negative patients were limited to those taken prior to the last procedure used to classify the patients. All ECGs measured when the patient was younger than 18 years of age were excluded. For each cohort, patients were divided into training (48%), trial (40%), and validation (12%) sets.

[0073] Each model was evaluated using two performance metrics: area under the patient curve (AUC) and age-sex AUC. For patient-related AUC, one ECG was randomly sampled from each patient, and the mean of 50 random runs was reported. Patient-related AUC ensures that patients with more ECGs, i.e., those potentially more disease-prone, are not over-included. For age-sex AUC, four negative ECGs were randomly sampled for each age- and sex-matched positive ECG at the time the ECGs were performed. If four negative ECGs were unavailable, positive ECGs were undersampled to maintain a 1:4 positive-to-negative ECG ratio. Again, the mean of 50 random runs is reported. The advantage here is that the age and sex distribution is maintained between positive and negative cohorts.

[0074] We developed an algorithm to test single-branch, quad-branch, and 12-branch 1D convolutional neural networks (CNNs) using a 12-lead voltage-time signal as a single input, with four groups of three leads each, and with each individual lead. Various other parameters included age and sex, additional 2D spectrograms, residual connectivity, and window size (i.e., a 10-second window or overlapping 2-second windows) as inputs (see Table 4 for results from preliminary models). The optimal model architecture was found to be a single-branch 1D CNN using residual connectivity and overlapping 2-second windows (Figure 8: shown for the updated model). Age and sex were not required as inputs, and including 2D spectrograms did not dramatically improve performance (data not shown). We also trained the model using ECGs containing only sinus rhythm, or by excluding patients with pacemakers, but neither modification dramatically improved performance (see Tables 5 and 6 for results from preliminary models). Finally, the inventors tested the diagnostic model on ECGs from the prevention window (Figure 9: shown for the updated model) and found that the diagnostic model performed better than the prevention model for these ECGs (data not shown). The preliminary diagnostic model trained on Cohort 3 was one of the best performing models and was used for further research. Using the updated diagnostic model, the inventors tested ECGs from 0 to 5 years prior to diagnosis within a 6-month window (Figure 10: shown for the updated model). [Table 4] [Table 5] [Table 6]

[0075] Example 5: Identification of putative genetic biomarkers that can explain early ECG signals associated with pulmonary hypertension. The finding that ECG-based models for diagnosing pulmonary hypertension maintain their performance in identifying patients with pulmonary hypertension up to five years before the date of diagnosis suggests that ECG signals provide information about a patient's long-term susceptibility to pulmonary hypertension. While we do not want to be bound by theory, one possible explanation for this is that certain germline gene mutations that predispose patients to pulmonary hypertension also regulate cardiac electrophysiology and are therefore detectable in ECG well before a diagnosis of pulmonary hypertension. Using the nferX® platform, we identified gene mutations associated with pulmonary hypertension and then triangulated evidence supporting the idea that these genes regulate ECG signals. We identified six candidate genes, including two potassium channels, KCNK3 and KCNA5, as well as four additional genes, including CAV1, SMAD4, GJB2, and TBX4 (see Table 7). [Table 7]

[0076] The nferX Human Genetics application has revealed that 26 genes are associated with pulmonary hypertension (Figure 11). Of these genes, two, KCNK3 and KCNA5, are potassium channels expressed in cardiac tissue (Figure 12). Since potassium channels are known contributors to cellular action potentials, and cardiac action potentials are the basis of signals observed by ECG, we prioritized KCNK3 and KCNA5 for further investigation. The nferX Single Cell application has demonstrated that KCNK3 and KCNA5 are strongly expressed at the single-cell level in neuronal and cardiac cell types, respectively (Figure 13). At lower expression levels, KCNK3 expression has also been observed in cardiac cell types (not shown). Finally, the nferX Signals application has confirmed literature evidence that both KCNK3 and KCNA5 influence cardiac electrophysiology (Figure 14).

[0077] The inventors further investigated non-ion channel genes identified by nferX Human Genetics as containing mutations associated with pulmonary hypertension. Eight of these genes, CAV1, SMAD4, GJB2, ACVRL1, SMAD4, BMPR1A, BMPR2, and TBX4, have a significant association with related terms associated with cardiac electrophysiology, such as “electrocardiogram,” “ECG,” “cardiac action potential,” and “cardiac conduction.” Using nferX RNA Explorer and nferX Single Cell applications, the inventors accumulated evidence linking these genes to the heart. These genes, in addition to potassium channel genes, may serve as part of a gene panel of at least 10 genes that further support the diagnosis of pulmonary hypertension, considering positive ECG-based predictive tests.

[0078] Example 6: Time Series Convolution Neural Network Architecture Without a Transformer Layer We designed a 7-layer "Convolution Transformer" neural network architecture that receives 12-channel ECG and outputs an indication of the presence of pulmonary hypertension as described above (see Figure 15).

[0079] Example 7: "Convolution Transformer" Neural Network Architecture Including a Transformer Layer To increase the interaction between different parts of the ECG signal compared to the Time Series Convolution model in Example 6, another neural network architecture called "Convolution Transformer" is provided. It uses a convolutional neural network to generate fixed-size encodings for smaller portions of the ECG waveform. A series of such generated encodings are fed into the transformer network to generate predictions as shown in the network architecture diagram in Figure 16. These longer interactions were made possible by using a 5-second clipping instead of a 2-second clipping.

[0080] Example 8: Data Expansion Data augmentation was used during the training phase to reduce the neural network's susceptibility to overfitting. The training dataset was augmented by randomly applying one of the parameters A-F below, with a 40% probability during training alone. See also Figure 17(A-E). A. Masking a portion of the time in the entire signal. B. Only frequencies between 0.5 and 50 Hz are permitted. C. Stretching the signal using several zoom levels. D. Shifting voltage with a small voltage across different leads. Lowering the frequency band at E.1 and 50Hz. F. Rearranging the set of small reeds Table 8 shows the results of extended CNN tests with and without a transformer layer. [Table 8]

[0081] All publications (including patents, patent application publications and sequence acceptance numbers referenced herein) are incorporated herein by reference in whole, as if each individual publication were specifically and individually indicated for incorporation herein by reference. In case of any conflict, the application containing any definition shall prevail.

[0082] Those skilled in the art will recognize, or can verify by routine experimental methods, many equivalents to the specific embodiments of the present invention described herein. Such equivalents are intended to be covered by the following claims.

Claims

1. A process of receiving target voltage-time data, including voltage data from multiple leads of an electrocardiograph, A step of generating a feature vector from the voltage-time data, The process of providing the aforementioned feature vector to a pre-trained learning system, A step of receiving instructions from the pre-trained learning system regarding the presence or absence of pulmonary hypertension in the subject. A method that includes this.

2. The method according to claim 1, wherein the step of generating the feature vector includes generating a spectrogram based on the voltage data of the plurality of leads.

3. The method according to claim 1, wherein the step of generating the feature vector includes grouping the voltage data of the plurality of leads into a plurality of subsets.

4. The method according to claim 1, further comprising the step of receiving the demographic information of the subject, wherein the step of generating the feature vector includes adding the demographic information to the feature vector.

5. The method according to claim 1, further comprising the step of receiving the target genome information, wherein the step of generating the feature vector includes adding the genome information to the feature vector.

6. The method according to claim 1, wherein the learning system includes a convolutional neural network.

7. The method according to claim 6, wherein the convolutional neural network includes at least one residual connection.

8. The method according to claim 1, wherein the voltage-time data of the target is received from an electrocardiograph.

9. The method according to claim 1, wherein the voltage-time data of the target is received from an electronic medical record.

10. The method according to claim 1, further comprising the step of providing instructions to an electronic health record system for storing information in a health record associated with the subject.

11. The method according to claim 1, further comprising the step of giving instructions to a computing node for display to a user.

12. An electrocardiograph with multiple leads, A computing node having a computer-readable storage medium having program instructions that are materialized therein, wherein the program instructions are executable by the processor of the computing node, A step of receiving target voltage-time data, including voltage data from the multiple leads from the echocardiogram, A step of generating a feature vector from the voltage-time data, The process of providing the aforementioned feature vector to a pre-trained learning system, A step of receiving instructions from the pre-trained learning system regarding the presence or absence of pulmonary hypertension in the subject. The processor is made to perform a method that includes the following: Computing nodes and A system equipped with these features.

13. The system according to claim 12, wherein the step of generating the feature vector includes generating a spectrogram based on the voltage data of the plurality of leads.

14. The system according to claim 12, wherein the step of generating the feature vector includes grouping the voltage data of the plurality of leads into a plurality of subsets.

15. The system according to claim 12, further comprising the step of receiving the demographic information of the subject, wherein the step of generating the feature vector includes adding the demographic information to the feature vector.

16. The system according to claim 12, further comprising the step of receiving the target genome information, wherein the step of generating the feature vector includes adding the genome information to the feature vector.

17. The learning system according to claim 12, wherein the learning system includes a convolutional neural network.

18. The system according to claim 16, wherein the convolutional neural network includes at least one residual connection.

19. The system according to claim 12, wherein the voltage-time data of the target is received from an electrocardiograph.

20. The system according to claim 12, wherein the voltage-time data of the target is received from an electronic medical record.

21. The system according to claim 12, further comprising the step of providing instructions to an electronic health record system for storing information in a health record associated with the subject.

22. The system according to claim 12, further comprising the step of giving instructions to a computing node for display to a user.

23. A computer program product for detecting pulmonary hypertension, It also includes a computer-readable storage medium having program instructions that are materialized therein, The aforementioned program instructions are executable by the processor, A step of receiving target voltage-time data, including voltage data from the multiple leads from the echocardiogram, A step of generating a feature vector from the voltage-time data, The process of providing the aforementioned feature vector to a pre-trained learning system, A step of receiving instructions from the pre-trained learning system regarding the presence or absence of pulmonary hypertension in the subject. The processor is made to perform a method that includes the following: Computer program products.

24. The computer program product according to claim 23, wherein the step of generating the feature vector includes generating a spectrogram based on the voltage data of the plurality of leads.

25. The computer program product according to claim 23, wherein the step of generating the feature vectors includes grouping the voltage data of the plurality of leads into a plurality of subsets.

26. The computer program product according to claim 23, further comprising the step of receiving the demographic information of the subject, wherein the step of generating the feature vector includes adding the demographic information to the feature vector.

27. The computer program product according to claim 23, further comprising the step of receiving the target genome information, wherein the step of generating the feature vector includes adding the genome information to the feature vector.

28. The computer program product according to claim 23, wherein the learning system includes a convolutional neural network.

29. The computer program product according to claim 27, wherein the convolutional neural network includes at least one residual connection.

30. The computer program product according to claim 23, wherein the voltage-time data of the subject is received from an electrocardiograph.

31. The computer program product according to claim 23, wherein the voltage-time data of the subject is received from an electronic medical record.

32. The computer program product according to claim 23, further comprising the step of providing instructions to an electronic health record system for storing information in a health record associated with the subject.

33. The computer program product according to claim 23, further comprising the step of giving instructions to a computing node for display to a user.

Images