Apparatus and method for training artificial intelligence-supported diagnostic assessment tool

Abstract

To provide an apparatus and method for training an artificial intelligence-supported diagnostic assessment tool.SOLUTION: An apparatus and method for training an artificial intelligence-supported diagnostic assessment tool may provide rapid and accurate prognosis determinations. The apparatus may include at least a processor configured to receive a plurality of multi-channel sensor readings of physiological data, generate training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels, train a neural network using the plurality of diagnostic labels, receive a time series input describing user physiological data from at least a sensor, input the time series input into the trained neural network, generate diagnostic data as a function of the time series input and the trained neural network, determine prognostic data as a function of the diagnostic data, and output the prognostic data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates generally to the field of medical diagnostics, and more particularly to an apparatus and method for training an artificial intelligence-assisted diagnostic assessment tool. [Background technology]

[0002] Assessment of left ventricular (LV) systolic and diastolic dysfunction (LVSD and LVDD, respectively) is essential for the evaluation and management of cardiac disease. Patients typically undergo echocardiography, which is cumbersome, expensive, not readily available, and can be inaccurate if performed incorrectly. There are no rapid, easily performed tests to assess both LVSD and LVDD to identify underlying cardiac disease. Summary of the Invention

[0003] In one aspect, an apparatus and method for training an artificial intelligence-assisted diagnostic evaluation tool can provide rapid and accurate prognosis determinations. The apparatus can include at least a processor configured to receive a plurality of multi-channel sensor readings of physiological data, generate training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels, train a neural network using the plurality of diagnostic labels, receive time series inputs describing user physiological data from at least the sensors, input the time series inputs to the trained neural network, generate diagnostic data as a function of the time series inputs and the trained neural network, determine prognostic data as a function of the diagnostic data, and output the prognostic data. In another aspect, the apparatus for artificial intelligence-assisted diagnostic evaluation can provide predictive diagnostic data based solely on electrocardiogram data with a level of accuracy similar to that of an echocardiogram.

[0004] These and other aspects and features of non-limiting embodiments of the present invention will become apparent to those skilled in the art upon review of the following description of specific non-limiting embodiments of the present invention in conjunction with the accompanying drawings. For the purpose of illustrating the invention, the drawings show aspects of one or more embodiments of the invention, it being understood, however, that the invention is not limited to the precise arrangements and instrumentalities shown in the drawings. [Brief explanation of the drawings]

[0005] [Figure 1] FIG. 1 is a system block diagram of an apparatus for artificial intelligence-assisted diagnostic evaluation. [Figure 2] FIG. 2 illustrates an exemplary machine learning module. [Figure 3] FIG. 3 illustrates an exemplary neural network. [Figure 4] FIG. 4 illustrates an exemplary node of a neural network. [Figure 5A] FIG. 5A is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5B] FIG. 5B is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5C] FIG. 5C is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5D] FIG. 5D is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5E] FIG. 5E is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5F] FIG. 5F illustrates the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5F2]FIG. 5F2 is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5F3] FIG. 5F3 is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5G] FIG. 5G illustrates the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5H] FIG. 5H is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5I] FIG. 5I is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5J] FIG. 5J is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5K] FIG. 5K is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5L] FIG. 5L is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5M] FIG. 5M illustrates the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5N] FIG. 5N is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5O]FIG. 5O is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 5P] FIG. 5P is a diagram of the methods and results of an exemplary study conducted to evaluate the performance of artificial intelligence-enabled ECG performed on left ventricular diastolic function and filling pressures. [Figure 6] FIG. 6 is a flow diagram illustrating an exemplary method for training a device for artificial intelligence assisted diagnostic evaluation. [Figure 7] FIG. 7 is a block diagram of a computing system that may be used to implement any one or more of the methods disclosed herein and any one or more portions thereof. DETAILED DESCRIPTION OF THE INVENTION

[0006] The drawings are not necessarily to scale and may be illustrated by phantom lines, schematic diagrams, and partial views. In some cases, details that are not necessary for understanding the embodiments or that make other details difficult to perceive may be omitted. At a high level, aspects of the present disclosure relate to an apparatus and method for training an artificial intelligence-assisted diagnostic assessment tool. The apparatus may include at least a processor configured to receive a plurality of multi-channel sensor readings of physiological data, generate training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels, train a neural network using the plurality of diagnostic labels, receive time-series inputs describing user physiological data from at least the sensor, input the time-series inputs to the trained neural network, generate diagnostic data as a function of the time-series inputs and the trained neural network, determine prognostic data as a function of the diagnostic data, and output the prognostic data. Aspects of the present disclosure can be used to rapidly, accurately, and cost-effectively assess and manage medical disorders. Aspects of the present disclosure can also be used to predict future prognoses, at least in part because the diagnostic tool is AI-assisted, enabling the growth and adaptation of generative predictive models. Exemplary embodiments illustrating aspects of the present disclosure are described below in the context of several specific examples.

[0007] 1 , an exemplary embodiment of an apparatus 100 for artificial intelligence-assisted diagnostic assessment is shown. In one embodiment, the apparatus 100 includes at least a processor 104 configured to receive a plurality of multi-channel sensor readings 108 of physiological data, generate training data 112 correlating each of the plurality of multi-channel sensor readings 108 with a plurality of diagnostic labels 120, train a neural network 124 using the plurality of diagnostic labels 120 and the multi-channel sensor data 116, receive time series inputs 128 describing the physiological data of a user from at least the sensors, input the time series inputs 128 to the trained neural network 124, generate diagnostic data 132 as a function of the time series inputs 128 and the trained neural network 124, determine prognostic data 136 as a function of the diagnostic data 132, and output the prognostic data 136. As used throughout this disclosure, a "diagnostic label" is closely related to the diagnostic data 132 and classifies data relevant to a patient's diagnosis. For example, without limitation, the diagnostic label may include left ventricular ejection fraction (LVEF) data, probability of left ventricular diastolic dysfunction (LVDD), LVDD within a particular time frame, etc. Alternatively, without limitation, the diagnostic label may be binary (e.g., yes / no or true / false), where the binary label may indicate a risk above a particular threshold. As used throughout this disclosure, a "time series input" is a type of data representing a series of observations or measurements recorded over a particular time interval. In a time series, each data point is associated with a particular timestamp, and the data is collected and organized in chronological order. As used in this disclosure, "diagnostic data" is data related to the diagnosis of a user's medical condition and / or health problem. Furthermore, as used in this disclosure, "prognostic data" is data associated with a prediction regarding a user's future health. Continuing with FIG. 1 , the plurality of multi-channel sensor readings 108 of physiological data may include, without limitation, sensor readings from one or more diagnostic tools, such as an echocardiogram, an electrocardiogram (ECG), and / or an electroencephalogram (EEG). As used in this disclosure, a "channel" is a distinct medium, mechanism, and / or pathway through which a signal can travel.Two separate channels may include, for example, two separate data paths for signals, such as two signals received from and / or via two sensors or leads, signals received via two separate wires, signals received by two separate antennas, two separate frequency ranges, such as corresponding to two separate passbands in radio frequency communications, two separate “tracks” in an audio or other data recording, etc. Multiple multi-channel sensor readings 108 may be correlated with each other to pair a separate multi-channel sensor reading with another separate multi-channel sensor reading. These correlations may be an embodiment of multi-channel sensor data 116. Furthermore, multi-channel sensor readings 108 may individually be an embodiment of multi-sensor data 116. For example, without limitation, ECG readings may generate ECG data and be paired with echocardiogram readings to generate echocardiogram data. This may enable the generation of training data 112 that associates particular ECG data with particular echocardiogram data. Additional multi-channel sensor readings 108 may include synthetic generation of echocardiogram data from ECG data. This may be performed without limitation, as disclosed in U.S. patent application Ser. No. 18 / 517,640, filed Nov. 22, 2023, entitled "SYSTEM AND METHOD FOR GENERATING ECHOCARDIOGRAM INFORMATION FROM ELECTROCARDIOGRAM," which is incorporated herein by reference in its entirety. Additionally, correlation may be performed between similar types of multi-channel sensor readings 108 that do not have matching channel configurations. Physiological data may include electrical signals of the body, heart rate, blood glucose, blood pressure, respiratory rate, temperature, etc.

[0008] 1 , generation of training data 112 correlating each of a plurality of multi-channel sensor readings 108 with a plurality of diagnostic labels 120 can be performed without limitation as disclosed in U.S. patent application Ser. No. 17 / 552,246, entitled "SYSTEMS AND METHODS FOR DIAGNOSING A HEALTH CONDITION BASED ON PATIENT TIME SERIES DATA," filed December 15, 2021, which is incorporated herein by reference in its entirety. Further, generation of training data 112 can be performed without limitation as disclosed in U.S. patent application Ser. No. 17 / 500,287, entitled "NONINVASIVE METHODS FOR DETECTION OF PULMONARY HYPERTENSION," filed October 13, 2021, which is incorporated herein by reference in its entirety. This can be further implemented without limitation, as disclosed in U.S. Patent Application No. 16 / 754,007, entitled "ECG-BASED CARDIAC EJECTION-FRACTION SCREENING," filed April 6, 2020, the entire contents of which are incorporated herein by reference.

[0009] Continuing with reference to FIG. 1 , in one embodiment, the neural network 124 may include a multi-output convolutional neural network (CNN). CNNs are particularly useful for computer vision tasks such as image recognition and classification because they are designed to learn spatial hierarchies of features by capturing essential features in early layers and complex patterns in deeper layers. Thus, a multi-output CNN is a predictive model that simultaneously outputs two or more sets of labels measuring different concepts. Ultimately, two or more related but separate classification problems are solved simultaneously within the same model. For example, but not limited to, the neural network 124 may be trained using exemplary inputs such as transthoracic echocardiography, echocardiography data, multi-channel sensor reading data, diagnostic labels 120, diagnostic data 132, etc., performed within 14 days and including ejection fraction, filling pressures, etc., as well as 12-lead ECG data correlated to exemplary outputs such as diagnoses of normal LV, LVSD, LVDD, prognosis data 136, diagnostic data 132, etc. The training data 112 can include any of the training data 112 described and / or incorporated herein. The training data 112 can be used iteratively, with the output used as new, additional training data 112. Training of the network 124 can occur locally and / or remotely. Similarly, retraining of the network 124 can occur locally and / or remotely. The network 124 can be further trained for various other channel variations, including, but not limited to, six-lead ECGs and / or other time-series input diagnostic tools. Further non-limiting examples of the neural network 124 can include recurrent neural networks (RNNs), multilayer perceptrons (PLPs), etc. The neural network 124 can include any of the models described throughout this disclosure. The neural network 124 can be trained on common usage lead data and / or, alternatively, reduced common usage lead data. "General purpose" refers to the typical lead counts used for any given diagnostic tool.For example, without limitation, ECG typically uses 12-lead ECG data. Similarly, general use of EEG may require up to 64 or more electrodes or leads with 10-20 lead placements. Thus, reduced lead data for general use refers to a reduction in the number of electrodes and / or leads used in acquiring lead data.

[0010] With further reference to FIG. 1 , receiving a time series input 128 describing the user's physiological data from at least a sensor may include receiving data manually entered by a user. The manual input may include data of any form or embodiment of the time series input 128, as described throughout this disclosure. Alternatively, receiving the time series input 128 may include receiving data output from another machine learning model. For example, without limitation, receiving the time series input 128 may include output from a machine learning model configured to generate a decision using mismatched channel signals. This embodiment may enable the use of a diagnostic tool having a first signal set using several channels to output a transformed signal set at a target configuration higher or lower than the first signal set used in the time series input 128 using several channels. This configuration may aid in ease and cost of evaluation, since not all embodiments of a diagnostic test are in the same channel configuration. The output of the additional machine learning model used as the time series input 128 may include data of any form or embodiment, as described throughout this disclosure. As used in this disclosure, a "sensor" is any embodiment capable of detecting a phenomenon. For example, without limitation, electrodes may detect or sense electrical activity of the heart and / or brain. Sensors may be communicatively connected to device 100, and / or sensor information may be manually entered by user interaction. Additionally, sensors may reside remotely from device 100 but maintain communication with device 100, which may be physical, and / or via a remote device. A specific, non-limiting, exemplary embodiment of time series input 128 may include "Array([[24,29,5,...,20,15,24], [20,10,-10,...,20,20,20], [20,10,-10,...,44,39,39], ..., [20,39,19,...,10,24,10], [24,29,5,...,10,20,5], [20,5,-15,...,-5,15,5]], dtype=int16)".

[0011] Continuing with reference to FIG. 1 , in one embodiment, the time series input 128 can include electrocardiogram (ECG) data. An ECG is a medical method for monitoring the activity, rate, and / or rhythm of the heart. ECGs are used as tests to diagnose cardiac conditions such as cardiac arrhythmias, atrial and ventricular hypertrophy, conduction disorders, myocardial damage or necrosis, myocardial infarction, ischemic heart disease, electrolyte disorders, myocardial lesions, and some drug toxicity diagnoses. Assessment of diastolic function (DF) and filling pressure (FP) are important components in the diagnosis and management of heart failure with preserved ejection fraction (HFpEF). However, quick and easily performed tests to assess these components are lacking. Therefore, in one embodiment, the device 100 enables AI-assisted assessment of these important components through the use of ECG data. Diastolic function can be further described in terms of systolic function and / or diastolic dysfunction. Assessment can be specifically focused on the left ventricle (LV), and assessment of LV systolic dysfunction (LVSD) and / or LV diastolic dysfunction (LVDD) may be required for the evaluation and management of cardiac disease. "Systole" refers to the phase of the heartbeat when the myocardium contracts to pump blood from the heart's chambers into the body's arteries. Alternatively, "diastole" refers to the phase of the heartbeat when the myocardium relaxes, allowing the heart's chambers to fill with blood. "Dysfunction" generally refers to an abnormality or disorder in the function of a particular bodily organ or system, such as an abnormal heart rhythm. HFpEF is a clinical syndrome in which patients have clinical features of heart failure in the presence of a normal or near-normal left ventricular ejection fraction, usually defined as an ejection fraction of 50% or greater. HFpEF is not a single condition but is the result of many different pathologies, adding challenges to management. However, a common diagnostic finding is elevated diastolic filling pressure. Because management of HFpEF, especially in the later stages, is challenging, prevention and early recognition and diagnosis are important in combating the disease. The diagnosis of HFpEF generally depends on various factors, including, but not limited to, assessment of LF mass, left atrial volume, and diastolic function, which can be obtained from the interpretation of a complete echocardiogram. An "echocardiogram" is an examination of the heart's workings that uses ultrasound to produce a visual display and is used to diagnose or monitor heart disease. Furthermore, there are two forms of echocardiograms: transthoracic and / or transesophageal.A transthoracic echocardiogram, which uses ultrasound and its probe outside the chest to create images of the heart, is the most common type of echocardiogram because it is less invasive than its counterpart. However, a transesophageal echocardiogram is invasive because the test requires an endoscope to introduce an ultrasound probe into the patient's esophagus.

[0012] Continuing with reference to FIG. 1 , the device 100 can operate as a predictive model capable of determining a diagnosis based solely on LV function acquired via ECG. The time series input 128 can include 12-lead ECG data. Alternatively, the time series input 128 can include 6-lead ECG data; for example, the data can be acquired from an AliveCor 6-lead ECG. The use of "12" indicates the capture of the patient's heart's electrical activity from 12 different perspectives. Similarly, a 6-lead ECG indicates six different perspectives being measured. As used in this disclosure, a "lead" refers to one or more electrodes attached to the skin to detect the heart's electrical signals. As used in this disclosure, "ECG data" refers to electrical signals recorded from a user's heart by placing electrodes on the user's body. The user's heart signals are depicted as waves, which may then be interpreted to indicate potential and current problems with the heart's rhythm and / or function, and may further suggest a specific medical diagnosis.

[0013] Continuing with reference to FIG. 1 , in one embodiment, the processor 104 can generate diagnostic data 132 as a function of the time series input 128 and the trained neural network 124. Furthermore, the diagnostic data 132 can identify a user's condition based on the time series data by determining prognostic data 136 as a function of the diagnostic data 132. Identification can include assessment of a normal LV, LVSD, LVDD, or both LVSD and LVDD. LVSD, for purposes of this disclosure, is defined as an LV ejection fraction (EF) less than 50%. "Ejection fraction" refers to a measurement of the percentage of blood that leaves the heart with each contraction. Similarly, LVDD is defined as elevated LV filling pressure (FP). FP is considered elevated or elevated when the mean pulmonary capillary wedge pressure (PCWP) exceeds 15 mmHg or the LV end-diastolic pressure (LVEDP) exceeds 16 mmHg. The diagnostic data 132 can be used to classify users into groups. The groups may include normal LV function, LVSD only, LVDD only, and / or both LVSD and LVDD.

[0014] With further reference to FIG. 1 , the processor 104 can be configured to determine prognostic data 136 as a function of the diagnostic data 132. The prognostic data 136 can be based on a user group classification, such as, but not limited to, normal LV function, LVSD only, LVDD only, and / or both LVSD and LVDD. Such prognostic data 136 can provide a prognosis 124 for the user. A prognosis refers to a prediction of a user's future based on signs and symptoms. Several categories of prognosis are used to predict the evaluator's future vocational rehabilitation. These categories can include: good, poor, fair, good, and excellent. The device 100 simultaneously and rapidly detects LVSD and LVDD with excellent mortality prognosis information comparable to that found using echocardiography. The processor 104 can output a prognosis 124 based on the determination of the prognostic data 136 that correlates with the prognosis 124 category. Furthermore, the output prognosis 124 can further provide the user with a complete layout of data as input, use, and / or generation throughout the diagnostic process. In one embodiment, the prognosis 124 is provided via a computing device and may be displayed via a display screen and / or printed.

[0015] With further reference to FIG. 1 , apparatus 100 includes a computing device. The computing device includes a processor 104 communicatively connected to a memory. As used in this disclosure, “communicatively connected” means connected by a connection, attachment, or link between two or more data elements that allows for the reception and / or transmission of information between them. For example, without limitation, this connection may be wired or wireless, direct or indirect, and between two or more components, circuits, devices, systems, etc., allowing for the reception and / or transmission of data and / or signals between them. The data and / or signals between them may include, but are not limited to, electrical, electromagnetic, magnetic, video, audio, radio, and microwave data and / or signals, combinations thereof, among others. A communicative connection may be achieved, for example, without limitation, via wired or wireless electronic, digital, or analog communication, directly or through one or more intervening devices or components. Furthermore, a communicative connection may include electrically coupling or connecting at least an output of one device, component, or circuit to at least one input of another device, component, or circuit. For example, but not limited to, via a bus or other facility for intercommunication between elements of a computing device. A communication connection may also include an indirect connection, for example, but not limited to, via a wireless connection, wireless communication, a low-power wide area network, optical communication, magnetic coupling, capacitive coupling, or optical coupling, etc. In some cases, the term "communicatively coupled" may be used instead of communicatively connected in this disclosure.

[0016] With further reference to FIG. 1 , a computing device may include any computing device described herein, including, but not limited to, a microcontroller, microprocessor, digital signal processor (DSP), and / or system-on-chip (SoC) described herein. A computing device may be included in and / or communicate with a mobile device, including a mobile device such as a mobile phone or smartphone. A computing device may include a single computing device operating independently, or may include two or more computing devices operating in cooperation, parallel, sequentially, etc., where two or more computing devices may be included together in a single computing device or two or more computing devices. A computing device may interface or communicate with one or more additional devices via a network interface device, as described in further detail below. A network interface device may be utilized to connect a computing device to one or more of various networks and one or more devices. Examples of network interface devices include, but are not limited to, a network interface card (e.g., a mobile network interface card, a LAN card), a modem, and any combination thereof. Examples of networks include, but are not limited to, wide area networks (e.g., the Internet, enterprise networks), local area networks (e.g., networks associated with an office, building, campus, or other relatively small geographic space), telephone networks, data networks associated with telephone / voice providers (e.g., mobile communications provider data and / or voice networks), direct connections between two computing devices, and any combination thereof. Networks can use wired and / or wireless communication modes. In general, any network topology can be used.Information (e.g., data, software, etc.) can be communicated between computers and / or computing devices. Computing devices can include, for example, but are not limited to, a computing device or cluster of computing devices at a first location and a second computing device or cluster of computing devices at a second location. Computing devices can include one or more computing devices dedicated to data storage, security, traffic distribution for load balancing, etc. Computing devices can distribute one or more computing tasks, as described below, across multiple computing devices that can operate in parallel, serially, redundantly, or any other manner used to distribute tasks or memory among computing devices. Computing devices may be implemented using a "share nothing" architecture, as a non-limiting example.

[0017] With continued reference to FIG. 1 , a computing device may be designed and / or configured to perform any method, method step, or series of method steps in any embodiment described herein in any order and with any degree of repetition. For example, a computing device may be configured to repeatedly perform a single step or sequence until a desired or directed result is achieved, where the repetition of a step or series of steps may be performed iteratively and / or recursively using the output of a previous iteration as input to a subsequent iteration, aggregating the inputs and / or outputs of the iterations to generate an aggregate result, decrementing or decreasing one or more variables, such as a global variable, and / or dividing a larger processing task into a set of smaller processing tasks addressed iteratively. A computing device may perform any step or series of steps described herein in parallel, such as performing two or more steps simultaneously and / or substantially simultaneously using two or more parallel threads, processor cores, etc. The division of tasks among parallel threads and / or processes may be performed according to any protocol suitable for dividing tasks among iterations. Those skilled in the art will recognize, upon reviewing this disclosure in its entirety, various ways in which processes, sequences of processes, processing tasks, and / or data can be subdivided, shared, or otherwise processed using iterative, recursive, and / or parallel processing.

[0018] 2, an exemplary embodiment of a machine learning module 200 capable of performing one or more machine learning processes described in this disclosure is shown. The machine learning module can use the machine learning processes to perform the determination, classification, and / or analysis steps, methods, processes, etc., as described in this disclosure. As used in this disclosure, a "machine learning process" is a process that automatically uses training data 204 to generate an algorithm instantiated in hardware or software logic, data structures, and / or functions executed by a computing device / module to generate output 208 that generates given data provided as input 212, as opposed to a non-machine learning software program in which the commands to be executed are predetermined by a user and written in a programming language.

[0019] With further reference to FIG. 2 , as used herein, “training data” refers to data containing correlations that a machine learning process can use to model relationships between two or more categories of data elements. For example, without limitation, training data 204 can include multiple data entries, also known as “training examples,” each representing a set of data elements recorded, received, and / or generated together, where the data elements may be correlated by shared presence in a given data entry, proximity in a given data entry, etc. The multiple data entries in training data 204 can estimate one or more trends in correlations between categories of data elements. For example, without limitation, higher values ​​of a first data element belonging to a first category of data elements tend to correlate with higher values ​​of a second data element belonging to a second category of data elements, indicating a possible proportional or other mathematical relationship linking values ​​belonging to the two categories. Multiple categories of data elements can be associated with training data 204 according to various correlations. Correlations may indicate causal and / or predictive links between categories of data elements, which may be modeled as relationships, such as mathematical relationships, by machine learning processes, as described in further detail below. Training data 204 may be formatted and / or organized by categories of data elements, for example, by associating the data elements with one or more descriptors that correspond to the categories of the data elements. As a non-limiting example, training data 204 may include data entered in a standardized format by a person or process, such that an entry of a given data element in a given field in a form may be mapped to one or more descriptors of the category. Elements in training data 204 may be linked to descriptors of the category by tags, tokens, or other data elements.For example, but not by way of limitation, the training data 204 may be provided in a fixed-length format, a format that links the location of data to categories such as Comma-Separated Value (CSV) format, and / or a self-describing format such as Extensible Markup Language (XML) or JavaScript® Object Notation (JSON), allowing a process or device to discover the category of the data.

[0020] Alternatively or additionally, and continuing to refer to FIG. 2, the training data 204 may include one or more elements that are uncategorized. That is, the training data 204 may be unformatted or may not include descriptors for some elements of the data. Machine learning algorithms and / or other processes may sort the training data 204 according to one or more classifications, for example, using natural language processing algorithms, tokenization, detecting correlation values ​​in the raw data, etc. Categories may be generated using correlation and / or other processing algorithms. As a non-limiting example, in a corpus of text, phrases comprising a number "n" of compounds, such as nouns modified by other nouns, may be identified according to the statistically significant prevalence of n-grams containing such words in a particular order. Such n-grams may be classified as elements of language, such as "words," that are tracked similarly to single words, and new categories may be generated as a result of statistical analysis. Similarly, in a data entry that includes some text data, a person's name may be identified by reference to a list, dictionary, or other compendium of terms, allowing for ad-hoc classification by a machine learning algorithm and / or automated association of data in the data entry with a descriptor or a given format. The ability to automatically classify data entries allows the same training data 204 to be applied to two or more different machine learning algorithms, as described in further detail below. The training data 204 used by the machine learning module 200 can correlate any input data, as described in this disclosure, to any output data, as described in this disclosure. As non-limiting illustrative examples, inputs can include multi-channel sensor readings, diagnostic labels, diagnostic data, etc., and outputs can include prognosis data, diagnostic data, etc.

[0021] With further reference to FIG. 2 , as described in further detail below, one or more supervised and / or unsupervised machine learning processes and / or models can be used to filter, sort, and / or select the training data, including, but not limited to, the training data classifier 216. The training data classifier 216 can include a “classifier,” as used in this disclosure, which is a machine learning model, as defined below, such as a data structure representing and / or using a mathematical model, neural net, or program generated by a machine learning algorithm known as a “classification algorithm,” that classifies inputs into categories or bins of data and outputs the categories or bins of data and / or their associated labels, as described in further detail below. The classifier can be configured to output at least data that labels or identifies data sets that are clustered together, that are found to be close under a distance metric, as described below. The distance metric can include any norm, such as, but not limited to, the Pythagorean norm. The machine learning module 200 can generate a classifier using a classification algorithm, which is defined as a process by which a computing device and / or any modules and / or components operating thereon derive a classifier from the training data 204. Classification can be performed using, but is not limited to, linear classifiers such as logistic regression and / or naive Bayes classifiers, nearest neighbor classifiers such as k-nearest neighbor classifiers, support vector machines, least squares support vector machines, Fisher's linear discriminant, quadratic classifiers, decision trees, boosted trees, random forest classifiers, learning vector quantization, and / or neural network-based classifiers. As a non-limiting example, the training data classifier 216 can classify elements of the training data to characterize subpopulations, such as cohorts of people and / or other analyzed items and / or phenomena, from which subsets of the training data can be selected.

[0022] 2, the processor 104 can be configured to generate a classifier using a naive Bayes classification algorithm. The naive Bayes classification algorithm generates a classifier by assigning class labels to problem instances represented as vectors of element values. The class labels are drawn from a finite set. The naive Bayes classification algorithm can include generating a family of algorithms that assume that, given a class variable, the value of a particular element is independent of the value of any other element. The naive Bayes classification algorithm can be based on Bayes' theorem, expressed as P(A / B) = P(B / A)P(A) ÷ P(B), where P(A / B) is the probability of hypothesis A given data B, also known as the posterior probability; P(B / A) is the probability of data B given that hypothesis A was true; P(A) is the probability that hypothesis A is true regardless of the data, also known as the prior probability of A; and P(B) is the probability of the data independent of the hypothesis. The naive Bayes algorithm can be generated by first converting the training data into a frequency table. The processor 104 can then calculate a likelihood table by calculating the probabilities of different data entries and classification labels. The processor 104 can utilize a Naive Bayes equation to calculate the posterior probability of each class. The class with the highest posterior probability is the predicted result. The Naive Bayes classification algorithm can include a Gaussian model that follows a normal distribution. The Naive Bayes classification algorithm can include a multinomial model used for discrete counts. The Naive Bayes classification algorithm can include a Bernoulli model that can be utilized when the vector is binary.

[0023] Continuing with reference to FIG. 2 , the processor 104 may be configured to generate a classifier using a K-nearest neighbor (KNN) algorithm. As used in this disclosure, a “K-nearest neighbor algorithm” includes a classification method that uses feature similarity to analyze how similar out-of-sample features are to training data and classify input data into one or more clusters and / or feature categories as represented in the training data. This may be performed by representing both the training data and the input data in vector form and using one or more measures of vector similarity to identify classifications in the training data and determine the classification of the input data. The K-nearest neighbor algorithm may include specifying a K value, or a number that instructs the classifier to select the k most similar entries training data for a given sample, determining the most common classifier among the entries in the database, and classifying the known sample. This may be performed recursively and / or iteratively to generate a classifier that can be used to classify the input data as further samples. For example, an initial set of samples may be run to cover initial heuristics and / or "first guesses" at outputs and / or relationships, which may be seeded using, without limitation, expert input received according to any process described herein. As a non-limiting example, the initial heuristics may include ranking associations between inputs and elements of training data. The heuristics may include selecting a few of the highest-ranked associations and / or training data elements.

[0024] Continuing with reference to FIG. 2 , an algorithm for generating a k-nearest neighbors method may generate a first vector output containing the data input clusters, generate a second vector output containing the input data, and calculate the distance between the first and second vector outputs using any suitable norm, such as cosine similarity, Euclidean distance measure, etc. Each vector output may be represented, without limitation, as an n-tuple of values, where n is at least two values. Each value in the n-tuple of values ​​may represent a measurement or other quantitative value associated with a given data category or attribute, examples of which are provided in further detail below. Vectors may be represented in n-dimensional space using, without limitation, an axis for each category of values ​​represented in the n-tuple of values, such that the vector has a geometric direction that characterizes the relative amounts of the attributes in the n-tuple compared to each other. Two vectors may be considered equivalent if their directions and / or the relative amounts of the values ​​in each vector compared to each other are the same. Thus, as a non-limiting example, a vector represented as [5, 10, 15] may be treated as equivalent to a vector represented as [1, 2, 3] for purposes of this disclosure. Vectors may be more similar if their directions are more similar, or more dissimilar if their directions are more divergent; however, vector similarity may alternatively or additionally be determined using the average of similarities between similar attributes, or any other similarity measure appropriate for any n-tuple of values, or an aggregation of numerical similarities for purposes of loss functions as described in more detail below. Any of the vectors described herein may be scaled so that each vector represents each attribute along an equivalent value scale. Each vector may be "normalized," or calculated using the Pythagorean norm: JPEG2025133736000002.jpg1447(a iThe scaling and / or normalization may be divided by a "length" attribute, such as length attribute l as derived using (where l is the attribute number i of the vector). Scaling and / or normalization can serve to make vector comparisons independent of the absolute amount of attributes, while preserving any dependence on attribute similarity. This can be advantageous, for example, when the cases represented in the training data are represented by different amounts of samples, resulting in proportionally equivalent vectors with divergent values.

[0025] With further reference to FIG. 2 , training examples for use as training data may be selected from a population of potential examples according to a cohort relevant to the analytical problem or classification task to be solved, etc. Alternatively, or additionally, training data may be selected to span a set of possible situations or inputs for the machine learning model and / or process that will be encountered upon deployment. For example, for each category of input data to a machine learning process or model that may lie within a range of values ​​within a collection of phenomena, such as, but not limited to, images, user data, process data, physical data, etc., the computing device, processor, and / or machine learning model may select training examples that represent each possible value in such range and / or a representative sample of values ​​in such range. Selecting representative samples may include, for example, selecting training examples in a proportion consistent with a statistically determined and / or predicted distribution of such values ​​according to relative frequency, such that more frequently encountered values ​​in the population of data so analyzed are represented by more training examples than less frequently encountered values. Alternatively or additionally, the set of training examples may be compared to a set of representative values ​​in a database and / or presented to a user, so that the process can detect, automatically or via user input, one or more values ​​not included in the set of training examples. A computing device, processor, and / or module can automatically generate missing training examples by receiving and / or retrieving missing input and / or output values ​​and correlating the missing input and / or output values ​​with corresponding output and / or input values ​​co-located in the data record with the retrieved values, provided by a user, other device, etc.

[0026] 2, the computer, processor, and / or module may be configured to preprocess the training data. As used in this disclosure, "preprocessing" training data refers to converting the training data from its raw form into a form that can be used to train a machine learning model. Preprocessing can include sanitizing, feature selection, feature scaling, data augmentation, etc.

[0027] With further reference to FIG. 2 , the computer, processor, and / or module may be configured to sanitize the training data. As used in this disclosure, “sanitizing” training data is a process in which training examples that prevent a machine learning model from converging and / or processing to a useful result are removed. For example, without limitation, the training examples may include input and / or output values ​​that deviate from commonly encountered values ​​so that the machine learning algorithm using the training examples is matched to less likely quantities as inputs and / or outputs. For example, values ​​that are more than a threshold number of standard deviations from the mean, average, or expected value may be eliminated. Alternatively or additionally, one or more training examples may be identified as having low-quality data, where “low quality” is defined as having a signal-to-noise ratio below a threshold. Sanitization may include processes such as removing duplicate or other redundant data, interpolating missing data, correcting data errors, standardizing data, and identifying outliers. In a non-limiting example, sanitization may include utilizing an algorithm to identify duplicate entries or a spell-checking algorithm.

[0028] As a non-limiting example, and with further reference to FIG. 2 , images used to train an image classifier or other machine learning model, and / or processes that take images as input or produce images as output, may be rejected if the image quality is below a threshold. For example, but not by way of limitation, a computing device, processor, and / or module may perform blur detection, where rejecting one or more blur detections may be performed by, by way of non-limiting example, performing a Fourier transform or approximation, such as a fast Fourier transform (FFT), of the image and analyzing the distribution of low and high frequencies in the resulting frequency domain representation of the image, where the number of high frequency values ​​below a threshold level may indicate blur. As a further non-limiting example, blur detection may be performed by convolving the image or a channel of the image, etc., with a Laplacian kernel. This may generate a numerical score that reflects some rapid changes in intensity present in the image, such that a high score indicates clarity and a low score indicates blur. Blur detection can be performed using gradient-based operators, which measure the gradient or first derivative of an image based on the hypothesis that abrupt changes indicate sharp edges in the image and therefore lower blur levels. Blur detection can be performed using wavelet-based operators that utilize the ability of discrete wavelet transform coefficients to describe the frequency and spatial content of an image. Blur detection may also be performed using statistics-based operators that utilize some image statistics as texture descriptors to calculate the focus level. Blur detection can be performed by using discrete cosine transform (DCT) coefficients to calculate the focus level of an image from its frequency content.

[0029] With continued reference to FIG. 2 , a computing device, processor, and / or module can be configured to be preconditioned on one or more training examples. For example, but not by way of limitation, if a machine learning model and / or process has one or more inputs and / or outputs that require, transmit, or receive a particular number of bits, samples, or other data units, the elements of one or more training examples used as or compared to the inputs and / or outputs can be modified to have such a number of data units. For example, a computing device, processor, and / or module can convert a smaller number of units, such as in a low pixel count image, to a desired number of units, e.g., by upsampling and interpolation. As a non-limiting example, a low pixel count image may have 100 pixels, but the desired number of pixels may be 128. The processor can interpolate the low pixel count image to convert the 100 pixels to 128 pixels. It should also be noted that, upon reading this disclosure, one of ordinary skill in the art will know various methods for interpolating a smaller number of data units, such as samples, pixels, bits, etc., to a desired number of such units. In some cases, the set of interpolation rules may be trained with highly detailed inputs and / or outputs and corresponding sets of inputs and / or outputs, downsampled to a smaller number of units, and a neural network or other machine learning model trained to predict interpolated pixel values ​​using the training data. As a non-limiting example, sample inputs and / or outputs, such as a sample picture having sample augmented data units (e.g., pixels added between original pixels), may be input to a neural network or machine learning model, which outputs a pseudo-replica sample picture having dummy values ​​assigned to pixels between the original pixels based on the set of interpolation rules.As a non-limiting example, in the context of an image classifier, a machine learning model may have a set of interpolation rules trained with highly detailed images and sets of images downsampled to a smaller number of pixels, and a neural network or other machine learning model trained using those examples to predict interpolated pixel values ​​in a facial image context. As a result, inputs with sample-expanded data units (those added between original data units, dummy values) may be run through the trained neural network and / or model, which may fill in values ​​to replace the dummy values. Alternatively or additionally, the processor, computing device, and / or module may utilize a sample expander method, a low-pass filter, or both. As used in this disclosure, a "low-pass filter" is a filter that passes signals with frequencies below a selected cutoff frequency and attenuates signals with frequencies above the cutoff frequency. The exact frequency response of the filter depends on the filter design. The computing device, processor, and / or module may fill in data units between original data units using averaging, such as luma or chroma averaging within the image.

[0030] In some embodiments, with continued reference to FIG. 2 , a computing device, processor, and / or module can downsample elements of a training example to a desired smaller number of data elements. As a non-limiting example, a high pixel count image may have 256 pixels, but the desired number of pixels may be 128. The processor can downsample the high pixel count image to convert the 256 pixels to 128 pixels. In some embodiments, the processor can be configured to perform downsampling on the data. Downsampling, also known as decimation, can involve removing every Nth entry, all but the Nth entry, etc., in a sequence of samples, a process known as “compression,” which can be performed, for example, by an N-sample compressor implemented using hardware or software. Anti-aliasing and / or anti-imaging filters, and / or low-pass filters can be used to remove compression artifacts.

[0031] 2, feature selection involves narrowing and / or filtering training data to exclude features and / or elements, or training data containing such elements, and / or sets of features and / or elements, or training data containing such elements, that are not relevant to the purpose for which the trained machine learning model and / or algorithm is being trained, based on their relevance or usefulness to the intended task or purpose of the trained machine learning model and / or algorithm. Feature selection can be performed using any process described in this disclosure, including, but not limited to, the use of training data classifiers, outlier exclusion, etc.

[0032] 2, feature scaling can include, but is not limited to, normalization of data entries, which can be accomplished by dividing a numeric field by its norm, for example, as is performed for vector normalization. Feature scaling can include absolute maximum scaling, where each quantitative data is divided by the maximum absolute value of all quantitative data in a set or subset of quantitative data. Feature scaling can include min-max scaling, where each value X has the minimum value X in the set or subset of values ​​subtracted from it. min and the result is divided by the range of values ​​to give the maximum value in the set or subset: JPEG2025133736000003.jpg1982. Feature scaling can include mean normalization, which involves using the mean value of a set and / or subset of values, X mean has a maximum and minimum value: JPEG2025133736000004.jpg1859. Feature scaling is X and X mean This may involve standardization, where the difference between is divided by the standard deviation σ of the set or subset of values: JPEG2025133736000005.jpg1752. Scaling is done by using a set or subset X median and / or the median with the interquartile range (IQR), which represents the difference between the 25th and 50th percentile values ​​(or the nearest value depending on the rounding protocol), as follows: JPEG2025133736000006.jpg1959. Those skilled in the art will recognize, upon review of this disclosure in its entirety, various alternative or additional techniques that may be used for feature scaling.

[0033] With further reference to FIG. 2 , the computing device, processor, and / or module may be configured to perform one or more processes of data augmentation. As used in this disclosure, “data augmentation” refers to adding data to a training set using elements and / or entries already present in the dataset. Data augmentation can be achieved using, but is not limited to, interpolation, generating modified copies of existing entries and / or examples, and / or one or more generative AI processes, for example, using deep neural networks and / or generative adversarial networks. The generative process may alternatively be referred to in this context as “data synthesis” and creating “synthetic data.” Augmentation may include performing one or more transformations on the data, such as a geometric transformation, a color space transformation, an affine transformation, a brightness transformation, a cropping transformation, and / or a contrast transformation of the image.

[0034] With further reference to FIG. 2 , the machine learning module 200 may be configured to execute a lazy learning process 220 and / or protocol, which may alternatively be referred to as a “lazy loading” or “call on demand” process and / or protocol, in which machine learning is performed upon receiving inputs that are converted into outputs by combining the inputs and training set and deriving an algorithm used to generate outputs on demand. For example, an initial set of simulations may be run to cover initial heuristics and / or “first guesses” at outputs and / or relationships. As a non-limiting example, the initial heuristic may include ranking associations between inputs and elements of the training data 204. The heuristic may include selecting several highest-ranking associations and / or training data 204 elements. Lazy learning may implement any suitable lazy learning algorithm, including, but not limited to, a K-nearest neighbor algorithm, a lazy naive Bayes algorithm, etc. Upon reviewing this disclosure in its entirety, those skilled in the art will recognize a variety of lazy learning algorithms that may be applied to generate outputs as described in this disclosure, including, but not limited to, lazy learning applications of machine learning algorithms as described in more detail below.

[0035] Alternatively or additionally, and continuing to refer to FIG. 2 , the machine learning model 224 can be generated using a machine learning process such as those described herein. As used herein, a “machine learning model” is a data structure that represents and / or instantiates a mathematical and / or algorithmic representation of a relationship between inputs and outputs, generated and stored in memory using any machine learning process, including, but not limited to, any of the processes described above. Once generated, inputs are submitted to the machine learning model 224, which generates an output based on the derived relationship. For example, without limitation, a linear regression model generated using a linear regression algorithm may calculate a linear combination of the input data using coefficients derived during the machine learning process to calculate the output data. As a further non-limiting example, the machine learning model 224 may be generated by creating an artificial neural network, such as a convolutional neural network, that includes an input layer of nodes, one or more hidden layers, and an output layer of nodes. Connections between nodes can be created through a process of "training" the network, where elements from a training data 204 set are applied to the input nodes, and then an appropriate training algorithm (e.g., Levenberg-Marquardt, conjugate gradient, simulated annealing, or other algorithm) is used to adjust the connections and weights between nodes in adjacent layers of the neural network to produce desired values ​​at the output nodes. This process is sometimes referred to as deep learning.

[0036] With further reference to FIG. 2 , the machine learning algorithm can include at least a supervised machine learning process 228. As defined herein, at least a supervised machine learning process 228 includes an algorithm that receives a training set relating some inputs to some outputs and attempts to generate one or more data structures that represent and / or instantiate one or more mathematical relationships relating the inputs to the outputs, each of which is optimal according to some criteria specified to the algorithm using some scoring function. For example, a supervised learning algorithm can include inputs, such as those described above, any outputs, such as those described above, as inputs, and a scoring function that represents a desired type of relationship to be found between the inputs and outputs. The scoring function can, for example, maximize the probability that a given input and / or combination of component inputs is associated with a given output and minimize the probability that a given input is not associated with a given output. The scoring function can be expressed as a risk function that represents the “expected loss” of the algorithm with respect to the input to the output, where the loss is calculated as an error function that represents the degree to which the predictions generated by the relationships are inaccurate when compared to a given input-output pair provided in the training data 204. Those skilled in the art will recognize, upon reviewing this disclosure in its entirety, various possible variations of at least the supervised machine learning process 228 that may be used to determine the relationship between inputs and outputs. The supervised machine learning process may include a classification algorithm as defined above.

[0037] With further reference to FIG. 2 , training a supervised machine learning process can include iteratively updating coefficients, biases, and weights based on, but not limited to, an error function, an expected loss, and / or a risk function. For example, outputs generated by a supervised machine learning model using example inputs in the training examples may be compared with example outputs from the training examples. An error function can be generated based on the comparison, which can include any error function suitable for use in any machine learning algorithm described in this disclosure, including, for example, the squared difference between one or more sets of comparison values. Such an error function can be used to update one or more weights, biases, coefficients, or other parameters of the machine learning model via any suitable process, including, but not limited to, a gradient descent process, a least-squares process, and / or other processes described in this disclosure. This may be done iteratively and / or recursively to gradually adjust such weights, biases, coefficients, or other parameters. The updating can be performed using one or more backpropagation algorithms in a neural network. The iterative and / or recursive updating of weights, biases, coefficients, or other parameters as described above can be performed until the currently available training data is exhausted and / or a convergence check is passed, where a "convergence check" is a test of a condition selected as indicating that the model and / or its weights, biases, coefficients, or other parameters have reached a certain degree of accuracy. The convergence check can, for example, compare the difference between two or more consecutive error or error function values, with a difference below a threshold amount being taken to indicate convergence. Alternatively or additionally, one or more error and / or error function values ​​evaluated in a training iteration can be compared to a threshold.

[0038] With further reference to FIG. 2 , a computing device, processor, and / or module may be configured to perform the methods, method steps, series of method steps, and / or algorithms described with reference to this figure in any order and with any degree of iteration. For example, a computing device, processor, and / or module may be configured to repeatedly execute a single step, sequence, and / or algorithm until a desired or directed result is achieved. The repetition of a step or series of steps may be performed iteratively and / or recursively using the output of a previous iteration as input to a subsequent iteration, aggregating the inputs and / or outputs of the iterations to generate an aggregate result, decreasing or decrementing one or more variables, such as global variables, and / or dividing a larger processing task into a set of smaller processing tasks addressed iteratively. The computing device, processor, and / or module may execute any step, series of steps, or algorithm in parallel, such as performing two or more steps simultaneously and / or substantially simultaneously using two or more parallel threads, processor cores, etc. The division of tasks among parallel threads and / or processes may be performed according to any protocol suitable for dividing tasks among iterations. Those skilled in the art will recognize, upon reviewing this disclosure in its entirety, various ways in which processes, sequences of processes, processing tasks, and / or data can be subdivided, shared, or otherwise processed using iterative, recursive, and / or parallel processing.

[0039] 2 , the machine learning process can include at least an unsupervised machine learning process 232. As used herein, an unsupervised machine learning process is a process that draws inferences within a dataset without regard to labels. As a result, the unsupervised machine learning process is free to discover any structure, relationships, and / or correlations provided in the data. The unsupervised process 232 may not require a response variable. The unsupervised process 232 can be used to find patterns of interest and / or inferences between variables, such as to determine the degree of correlation between two or more variables.

[0040] With further reference to FIG. 2 , the machine learning module 200 can be designed and configured to create the machine learning model 224 using techniques for developing linear regression models. The linear regression model can include ordinary least squares regression, which aims to minimize the square of the difference between predicted and actual results according to an appropriate norm for measuring such difference (e.g., a vector space distance norm). To improve the minimization, the coefficients of the resulting linear equation can be modified. The linear regression model can include a ridge regression method, in which the function to be minimized includes a least-squares function and a term that multiplies the square of each coefficient by a scalar to penalize large coefficients. The linear regression model can include a least absolute shrinkage and selection operator (LASSO) model, in which ridge regression is combined with multiplying the least-squares term by a coefficient equal to 1 divided by twice the number of samples. The linear regression model can include a multitask lasso model, in which the norm applied to the least-squares term of the lasso model is the Frobenius norm, which is equivalent to the square root of the sum of the squares of all terms. The linear regression model may include an elastic net model, a multitask elastic net model, a minimum angle regression model, a LARS lasso model, an orthogonal matching pursuit model, a Bayesian regression model, a logistic regression model, a stochastic gradient descent model, a perceptron model, a passive aggression algorithm, a robustness regression model, a Huber regression model, or any other suitable model that may occur to one of ordinary skill in the art upon reviewing this disclosure in its entirety. In one embodiment, the linear regression model may be generalized to a polynomial regression model, whereby a polynomial (e.g., quadratic, cubic, or higher order equation) that provides the best predicted output / actual output fit is found. As will be apparent to one of ordinary skill in the art upon reviewing this disclosure in its entirety, methods similar to those described above can be applied to minimize the error function.

[0041] With continued reference to FIG. 2 , the machine learning algorithm may include, but is not limited to, linear discriminant analysis. The machine learning algorithm may include quadratic discriminant analysis. The machine learning algorithm may include kernel ridge regression. The machine learning algorithm may include support vector machines (including, but not limited to, support vector classification-based regression processes). The machine learning algorithm may include stochastic gradient descent algorithms, including classification and regression algorithms based on stochastic gradient descent. The machine learning algorithm may include nearest neighbor algorithms. The machine learning algorithm may include various forms of latent space regularization, such as variational regularization. The machine learning algorithm may include Gaussian processes, such as Gaussian process regression. The machine learning algorithm may include cross-decomposition algorithms, including partial least squares and / or canonical correlation analysis. The machine learning algorithm may include naive Bayes methods. The machine learning algorithm may include decision tree-based algorithms, such as decision tree classification or regression algorithms. The machine learning algorithm may include ensemble methods, such as bagging meta-estimators, forests of randomized trees, AdaBoost, gradient tree boosting, and / or voting classifier methods. The machine learning algorithms may include neural net algorithms, including convolutional neural net processes.

[0042] With further reference to FIG. 2 , machine learning models and / or processes can be deployed or instantiated by incorporating them into a program, device, system, and / or module. For example, without limitation, machine learning models, neural networks, and / or some or all of their parameters can be stored and / or deployed in any memory or circuit. Parameters such as coefficients, weights, and / or biases can be stored as circuit-based constants, such as arrays of wires and / or binary inputs and / or outputs set to logic “1” and “0” voltage levels in a logic circuit to represent numbers according to any suitable encoding system, including two's complement, or can be stored in any volatile and / or non-volatile memory. Similarly, mathematical operations and data inputs and / or outputs to or from models, neural network layers, etc. can be instantiated in hardware circuits and / or in the form of instructions in firmware, machine code such as binary opcode instructions, assembly language, or any higher-level programming language. Any technique for hardware and / or software instantiation of memory, instructions, data structures, and / or algorithms may be used to instantiate the machine learning process and / or model, including, but not limited to, the manufacture and / or configuration of non-reconfigurable hardware elements, circuits, and / or modules such as ASICs; the manufacture and / or configuration of reconfigurable hardware elements, circuits, and / or modules such as, but not limited to, FPGAs; non-reconfigurable and / or configurable non-rewriteable memory elements, circuits, and / or modules such as, but not limited to, rewriteable ROMs; reconfigurable and / or rewriteable memory elements, circuits, and / or modules such as, but not limited to, other memory technologies described in this disclosure; and / or any combination of the manufacture and / or configuration of any computing device and / or components thereof described in this disclosure.Such deployed and / or instantiated machine learning models and / or algorithms may receive inputs from, and generate outputs to, any other processes, modules, and / or components described in this disclosure.

[0043] Continuing with reference to FIG. 2 , any process of training, retraining, deployment, and / or instantiation of any machine learning model and / or algorithm can be performed and / or repeated after initial deployment and / or instantiation to modify, refine, and / or improve the machine learning model and / or algorithm. Such retraining, deployment, and / or instantiation can be performed as a periodic or periodic process, such as retraining, deployment, and / or instantiation at regular elapsed periods, after some measure of quantity such as the number of bytes or other measure of data processed, the number of uses or runs of the processes described in this disclosure, and / or according to a software, firmware, or other update schedule. Alternatively or additionally, retraining, deployment, and / or instantiation can be event-based, triggered by, without limitation, user input indicating suboptimal or other problematic performance, and / or by an automated on-site inspection and / or audit process, and the output of the machine learning model and / or algorithm, and / or its error and / or error function, can be compared to any threshold value, convergence check, etc., and / or the output of the processes described herein can be compared to similar threshold value, convergence check, etc. Event-based retraining, deployment, and / or instantiation may alternatively or additionally be triggered by the receipt and / or generation of one or more new training examples, which may be compared to a preconfigured threshold, and exceeding the preconfigured threshold may trigger retraining, deployment, and / or instantiation.

[0044] With further reference to FIG. 2 , retraining and / or additional training can be performed using any of the processes for training described above, using any current or previously deployed version of the machine learning model and / or algorithm as a starting point. Training data for retraining may be collected, preprocessed, screened, classified, sanitized, or otherwise processed according to any of the processes described in this disclosure. Training data can include, but is not limited to, training examples including inputs and correlation outputs used, received, and / or generated from any version of any system, module, machine learning model or algorithm, device, and / or method described in this disclosure. Such examples can be modified and / or labeled according to user feedback or other processes to indicate desired results and / or can have actual or measured results from the process being modeled and / or predicted by the system, module, machine learning model or algorithm, device, and / or method as “desired” results to be compared with the output of the training process as described above.

[0045] The rearrangement may be performed using any reconfiguration and / or rewriting of reconfigurable and / or rewritable circuit and / or memory elements, or may be performed by the generation of new hardware and / or software components, circuits, instructions, etc., which may be added to and / or replace existing hardware and / or software components, circuits, instructions, etc.

[0046] With further reference to FIG. 2 , one or more processes or algorithms described above can be performed by at least a dedicated hardware unit 232. For purposes of this figure, a “dedicated hardware unit” is a hardware component, circuitry, etc., other than a main control circuit and / or processor, that performs the method steps described in the present disclosure, specifically designated or selected to perform one or more particular tasks and / or processes described with reference to this figure, such as, but not limited to, preconditioning and / or sanitizing training data and / or training a machine learning algorithm and / or model. The dedicated hardware unit 232 can include, but is not limited to, a hardware unit that can efficiently use pipelining, parallel processing, etc. to perform iterative or mass calculations, such as matrix-based calculations for updating or adjusting parameters, weights, coefficients, and / or biases of a machine learning model and / or neural network. Such a hardware unit can be optimized for such processes by including dedicated circuitry for matrix and / or signal processing operations, including, for example, multiple arithmetic and / or logic circuit units, such as multipliers and / or adders, that can operate simultaneously and / or in parallel. Such special-purpose hardware units 232 may include, but are not limited to, graphical processing units (GPUs), special-purpose signal processing modules, FPGAs, or other reconfigurable hardware configured to instantiate parallel processing units for one or more specific tasks, and the computing device, processor, apparatus, or module may be configured to instruct one or more special-purpose hardware units 232 to perform one or more operations described herein, such as evaluation of model and / or algorithm outputs, one-time or iterative updates to parameters, coefficients, weights, and / or biases, and / or any other operations, such as vector and / or matrix operations, described in this disclosure.

[0047] Referring now to FIG. 3, an exemplary embodiment of a neural network 300 is shown. A neural network 300, also known as an artificial neural network, is a network of "nodes," or data structures having one or more inputs, one or more outputs, and a function that determines the output based on the inputs. Such nodes can be organized into networks such as, but not limited to, convolutional neural networks, including an input layer of nodes 304, one or more hidden layers 308, and an output layer of nodes 312. Connections between nodes can be created through a process of "training" the network, in which elements from a training data set are applied to the input nodes, and then an appropriate training algorithm (e.g., Levenberg-Marquardt, conjugate gradient, simulated annealing, or other algorithm) is used to adjust the connections and weights between nodes in adjacent layers of the neural network to produce desired values ​​for the output nodes. This process is sometimes referred to as deep learning. Connections may run solely from input nodes to output nodes in a "feedforward" network, or the output of one layer may be fed back to the input of the same or a different layer in a "recurrent network." As a further non-limiting example, a neural network can include a convolutional neural network that includes an input layer of nodes, one or more hidden layers, and an output layer of nodes. As used in this disclosure, a "convolutional neural network" is a neural network in which at least one hidden layer is a convolutional layer that convolves the input to that layer with a subset of the input known as a "kernel," along with one or more additional layers, such as a pooling layer, a fully connected layer, etc.

[0048] Referring now to Figure 4, an exemplary embodiment of a node 4400 of a neural network is shown. The node may include multiple inputs xi, which may receive numerical values ​​from other nodes and / or inputs to the neural network that includes the node. The node may perform one or more activation functions to generate its output given one or more inputs, including, but not limited to, computing a binary step function that compares the input to a threshold and outputs either a logic 1 or logic 0 output or the like; a linear activation function, where the output is directly proportional to the input; and / or a non-linear activation function, where the output is not proportional to the input. A non-linear activation function may be, but is not limited to, an activation function of the form x given an input x: Sigmoid function of JPEG2025133736000007.jpg1945, format JPEG2025133736000008.jpg1523 tanh (hyperbolic tangent) function, f(x)=tanh 2 tanh derivatives such as f(x), normalized linear unit functions such as f(x)=max(0,x), "leaky" and / or "parametric" normalized linear unit functions such as f(x)=max(ax,x), for some value of α (this function may be replaced and / or weighted by its own derivative in some embodiments). JPEG2025133736000009.jpg22101 exponential linear unit function such as, when the input to the instantaneous layer is xi f(x) = x*sigmoid(x) for some values ​​of a, b, and r, which can include softmax functions such as JPEG2025133736000010.jpg1941, and swish functions such as f(x) = x*sigmoid(x). a Gaussian error linear unit function, such as JPEG2025133736000011.jpg16102, and / or JPEG2025133736000012.jpg25104. In principle, there are no restrictions on the nature of the functions of the inputs x i that can be used as activation functions. As a non-limiting illustrative example, a node may have a function for each input x i Weight w multiplied by i A weighted sum of the inputs can be performed using a weighted sum of the inputs. Additionally or alternatively, a bias b may be added to the weighted sum of the inputs such that an offset is added to each unit in the neural network layer that is independent of the input to the layer. The weighted sum can then be input to a function φ, which can generate one or more outputs y. The weights w applied to the inputs xi i can indicate whether an input is "excitatory," e.g., by having a large corresponding weight, indicating that the input has a strong influence on one or more outputs y, and / or "inhibitory," e.g., by having a small corresponding weight, indicating that the input has a weak influence on another input y. i The value of may be determined by training the neural network using training data, which may be performed using any suitable process, such as those described above.

[0049] Specific Implementations for LVDD Research Implementation 5A-5E, diagrams of the methods and results of an exemplary study conducted to evaluate the performance of an implemented artificial intelligence-enabled ECG prediction model for left ventricular diastolic function and filling pressure are shown. In this study of a specific implementation of AI-assisted diagnostic assessment, the diagnostic assessment was based on a 12-lead ECG to predict elevated diastolic function and filling pressures determined by echocardiography, and then its association with all-cause mortality was evaluated.

[0050] Left ventricular diastolic function can be abnormal in myocardial disease. Deteriorating diastolic function can lead to elevated filling pressures even while the ejection fraction is preserved. Recording elevated left ventricular filling pressures can be used to diagnose heart failure with preserved ejection fraction. Furthermore, elevated diastolic filling pressures can be associated with heart failure symptoms and higher mortality in patients with myocardial infarction, valvular disease, and / or cardiomyopathies. Diastolic function can be assessed by echocardiography. However, such assessment of diastolic function may require skilled sonographers and / or cardiologists with advanced training in echocardiographic interpretation. The results of diastolic function assessments may be equivocal in a significant proportion of patients. 12-lead electrocardiography (ECG) is less expensive and more widely available.

[0051] Research method Data Research Group For an exemplary study of a specific implementation of an AI-assisted diagnostic assessment tool, the following methodology was followed. Within studies, adults (age 18 years or older) who had at least one ECG and transthoracic echocardiogram with echocardiographic assessment of diastolic function performed within 14 days of the ECG between September 2001 and April 2023 were identified from the Mayo Clinic Unified Data Platform. No exclusion criteria were applied. All ECGs were measured at 250 or 500 Hz sampling rates using a GE-Marquette machine for standard 10-second, 12-lead ECGs and stored on a GE-MUSE system (Marquette, Wisconsin, USA). ECGs with an original sampling rate of 250 Hz were upsampled to 500 Hz before analysis. The final cohort (n = 219,462) was divided into training (n = 98,736, 45%), validation (n = 21,963, 10%), and testing (n = 98,763, 45%) sets. The final model was tested in 55,248 patients with indeterminate diastolic function by echocardiography.

[0052] Filling pressure and diastolic function grading assessment For an exemplary study of a specific implementation of the AI-assisted diagnostic assessment tool, the following method was followed. As recommended by the 2016 ASE / EACVI diastolic function guidelines, the following four parameters were used to assess diastolic function: e', E / e', tricuspid regurgitation velocity, and left atrial volume index, with minor modifications (see Figure 5P). If three or all of the above four parameters were abnormal, filling pressure was determined to be elevated. This group was then separated into Grade 2 or 3 based on an E / A ratio of 2.0. If three or all parameters were normal, filling pressure was determined to be normal. These patients were further divided into normal diastolic function or Grade 1 according to an E / A ratio dividing a value of 0.8. If the four parameters were divided into two normal and two abnormal, diastolic function was assessed as indeterminate. Diastolic function and filling pressure were labeled for all subjects based on the above algorithm.

[0053] Overview of the AI ​​model Regarding the overview of the AI ​​model in this illustrative study of a specific implementation of an AI-supported diagnostic assessment tool, the following methodology was followed. The primary goal of developing an AI-enabled ECG was to predict left ventricular filling pressure and diastolic function grade using a 12-lead ECG. A ResNet-1831 convolutional neural network was implemented as the model architecture. Each ECG had a 12x5000 matrix consisting of a 12-lead ECG sampled at 500 Hz for 10 seconds. For network input, the ECG was divided into 2-second segments, and the output values ​​from the five segments were averaged. The network was trained with an Adam optimizer using a learning rate of 0.001 for 20 epochs. Validation performance converged before the 20th epoch. The final model was selected according to the AUC value from the validation set for elevated filling pressure. The model was trained as a multiclass model with four outputs representing four grades of diastolic function, with the sum of the four outputs being 1. Normal and grade 1 were considered normal filling pressures, and grades 2 and 3 were considered increased filling pressures. The model outputs four values, where the sum of the normal and grade 1 outputs represents the normal filling pressure output, and the sum of the grade 2 and grade 3 outputs represents the increased filling pressure output. Similarly, the sum of the normal and increased filling pressure outputs was 1. The sum of the two classes was converted into a binary model, and existing metrics were applied to the final output value. Similarly, aggregate outputs and labels were created for grade 1 or more, grade 2 or higher, and grade 3, respectively, to evaluate performance on an ordinal scale of diastolic function grade. Furthermore, deep neural networks were trained with the same architecture on single-lead ECGs and single-lead median beats. Lead 1 is used for both models because it is most commonly measured on wearable devices. The single-lead median beat is a representative beat over a 10-second period, with a duration of 1.2 seconds. Holdout test set results from selected models are reported.

[0054] statistical analysis Regarding statistical analysis in exemplary studies of certain embodiments, the following methods were followed: Model performance was assessed by calculating the AUC, sensitivity, specificity, PPV, NPV, and accuracy of the ROC curve. Two-sided 95% confidence intervals were calculated. Kaplan-Meier estimates were used to assess whether the model discriminated the risk of all-cause mortality and compared with the log-rank test. Multivariate Cox proportional hazards models were developed. Hazard ratios were adjusted for age, sex, and comorbidities (diabetes, hypertension, obesity, myocardial infarction, congestive heart failure, cerebrovascular disease, chronic pulmonary disease, and renal disease). For continuous variables, Student's t-tests were used to compare groups. For categorical variables, chi-square tests were used. A two-sided P value of <0.001 was considered significant, but its interpretation was cautious and comprehensive due to the large sample size.

[0055] Research results With further reference to specific embodiments, 274,710 patients who underwent ECG and echocardiographic diastolic function assessment within 14 days were identified without any exclusion criteria. Echocardiographic determination of diastolic function was possible in 219,462 patients (80%) and indeterminate in 55,248 patients (20%). Baseline patient characteristics were similar among the training, validation, and testing groups (Table 3). There were 20,264 patients with a left ventricular ejection fraction <0.001 (Table 1). Similarly, patients identified by AI-ECG as having elevated filling pressures had more comorbidities (p<0.001, Table 4). [Table 1] [Table 2] [Table 3] [Table 4]

[0056] AI-enabled ECG classification performance Continuing with reference to specific embodiments, the test set, AI-enabled ECG to predict echocardiography, determined elevated filling pressures, had a receiver operating characteristic (ROC) area under the curve (AUC) of 0.911 (95% CI: 0.909-0.914), sensitivity of 83.2%, specificity of 82.9%, positive predictive value (PPV) of 58%, and negative predictive value (NPV) of 94.5%, with a threshold of 0.26 and prevalence of 22.2% (Figure 1a and Table 2). The AUCs of the AI-enabled ECG for grade ≥1, grade ≥2, and grade 3 were 0.847 (95% CI: 0.844-0.85), 0.911 (95% CI: 0.909-0.914), and 0.943 (95% CI: 0.938-0.948) at thresholds of 0.443, 0.264, and 0.058, respectively (see ROC plot 508 in Figure 5 and Table 2). The median incremental filling pressure output from the model was significantly higher in echocardiographic diastolic function grades 2 and 3 compared with normal and grade 1 (Figure 5B). The model showed higher specificity in younger patients and tended to decrease in patients with more comorbidities (see Figures 5F-5F3 and Table 3). Echocardiographic diastolic parameters were significantly different between patients identified by the model as having elevated filling pressures on examination and those with normal filling pressures (see Figure 5H). These diastolic parameters were nearly identical in patients with normal filling pressures determined by AI-ECG and echocardiography. These values ​​in patients with elevated filling pressures by both AI-ECG and echocardiography were consistent with grade 2–3 diastolic dysfunction and significantly different from those in patients with normal filling pressures. In the indeterminate group, all echocardiographic diastolic parameters except e' velocity were significantly different between patients with normal and elevated filling pressures determined by AI-ECG (see Figure 5I). The AI-enabled ECG trained exclusively by ECG lead I had AUCs of 0.804 (95% CI: 0.801-0.807), 0.875 (95% CI: 0.872-0.878), and 0.915 (95% CI: 0.909-0.921) for grade ≥1, grade ≥2, and grade 3, respectively.The AI-enabled ECG trained by the median beat of ECG lead I had AUCs of 0.763 (95% CI: 0.76-0.766), 0.834 (95% CI: 0.83-0.837), and 0.877 (95% CI: 0.87-0.884), respectively. The AUCs of the AI-enabled ECG before and after the median year of echocardiography, i.e., 2014, were 0.856 and 0.839 for grade ≥1, 0.91 and 0.913 for grade ≥2, and 0.944 and 0.942 for grade 3, respectively (see Figure 5J). [Table 5]

[0057] Survival analysis With further reference to specific embodiments, death from any cause was observed in 20,223 of 98,763 patients (20.5%) in the tested group and 18,224 of 55,248 patients (33.0%) in the indeterminate group over a median follow-up of 5.9 years (IQR 2.7, 10.2) and 5.7 years (IQR 2.6, 9.9), respectively. Mortality was significantly higher in patients with elevated filling pressures compared with patients with normal filling pressures predicted by AI-enabled ECG after adjusting for age, sex, and comorbidities (hazard ratio (HR) 1.7, 95% CI 1.645-1.757; Figure 3a). This was similar to mortality predicted by echocardiographically determined filling pressures (HR 1.65, 95% CI: 1.597-1.705; see Figure 5C, echocardiography (test) 516). All-cause mortality was also predicted by ECG-determined diastolic function classification, with significantly higher mortality in patients with grade 2 and 3 diastolic dysfunction compared with patients with normal or grade 1 diastolic dysfunction (HR 1.299, 95% CI 1.279-1.319, see Figure 5D, AI-matched (test) 512). Diastolic function classification based directly on echocardiographic parameters had similar prognostic value (HR 1.298, 95% CI 1.277-1.32), even after adjusting for age, sex, and comorbidities. Because some patients had discordance between AI-ECG and echocardiography determinations of diastolic function, we divided the tested patients into four groups for each category of diastolic dysfunction: true positive (TP; AI-ECG (+) and echocardiography (+); true negative (TN); AI-ECG (-) and echocardiography (-); false positive (FP); AI-ECG (+) and echocardiography (-); and false negative (FN; AI-ECG (-) and echocardiography (+)). TP had the worst mortality rate in all three diastolic dysfunction groups, and TN was the best. FP and FN groups had similar mortality rates for grades ≥ 1 or ≥ 2. However, FP was found to have the same mortality rate as TP, which was significantly worse than that of FN for grade 3 (HR 1.402, 95% CI 1.281-1.535) after adjusting for age, sex, and comorbidities.The risk of death was also higher among patients in the indeterminate group who had higher filling pressures than predicted by the AI-enabled ECG (HR 1.34, 95% CI 1.298-1.383). Among patients with normal filling pressures by the AI-enabled ECG, grade 1 dysfunction had a worse survival rate than normal grade in both the test and indeterminate groups. Among patients with grade 1 diastolic dysfunction by echocardiography, 54.7% were classified as normal by the AI-enabled ECG. Patients labeled as normal had a lower risk of death than patients labeled as grade 1 dysfunction by the AI-enabled ECG. The AI-enabled ECG successfully discriminated the risk of death among specific age groups (<50 years, 50<70 years, and ≥70 years), even after adjusting for age, sex, and comorbidities.

[0058] Research drawings 5A, AI-enabled ECG ROC curves for diastolic function grade and filling pressure are shown. Specifically shown are ROC plots for detecting increased filling pressure 504 and for detecting diastolic function grade using an ordinal scale 508.

[0059] Referring now to Figure 5B, the AI-enabled ECG output distribution for elevated filling pressures by estimated diastolic function grade is shown. This distribution is depicted as a box plot with a kernel density plot. The box plot can show the median and the first and third quartiles with outliers as 1.5 times the IQR.

[0060] Referring now to Figures 5C-E, all-cause mortality rates are shown using Kaplan-Meier curves with 95% point-by-point confidence intervals. Figure 5C shows the Kaplan-Meier curves for the patient test group with filling pressures predicted by AI-enabled ECG 512 and the Kaplan-Meier curves for the patient test group with filling pressures by echocardiography 516. Figure 5D shows the Kaplan-Meier curves for the test group with diastolic function grade predicted by AI-enabled ECG 520 and the Kaplan-Meier curves for the diastolic function grade by echocardiography 524. Figure 5E shows the Kaplan-Meier curves for the indeterminate group with diastolic function grade predicted by the deep learning model 528 and the Kaplan-Meier curves for the test group with normal diastolic function grade and echocardiographic grade 1 predicted by AI-enabled ECG 532. Log-rank tests were used for p-values.

[0061] Referring now to Figures 5F-5F3, AUC, sensitivity, specificity, PPV, NPV, and odds ratios (OR) are shown with 95% confidence intervals across age, sex, and comorbidity subsets. Figure 5F specifically illustrates grade 1 or more. Figure 5F2 specifically illustrates grade 2 or higher (elevated filling pressures). Finally, Figure 5F3 specifically illustrates grade 3.

[0062] Referring now to FIG. 5G, the distribution of left atrial volume index, e', E / e', tricuspid regurgitation velocity, and E / A according to filling pressures from our AI-ECG and echocardiography in the test set is shown. Each point represents the mean value with a 95% confidence interval. Normal filling pressures include normal and grade 1, and elevated filling pressures include grades 2 and 3.

[0063] Referring now to Figure 5H, the distribution of left atrial volume index, e', E / e', tricuspid regurgitation velocity, and E / A according to filling pressures by our AI-ECG in the indeterminate set is shown. Each point represents the mean with a 95% confidence interval. Normal filling pressures include normal and grade 1, and elevated filling pressures include grades 2 and 3. Student's t-test was used for p-values.

[0064] Referring now to Figure 5I, the ROC curves for AI-ECG before and after the median year of echocardiography in the test set are shown. ROC plot for detecting diastolic function grade using an ordinal scale.

[0065] Referring now to Figure 5J, all-cause mortality using Kaplan-Meier curves with 95% confidence intervals and p-values ​​from the log-rank test for true positive (TP), false positive (FP), false negative (FN), and true negative (TN) for grade 1 or more, grade 2 or higher, and grade 3 is shown.

[0066] Referring now to Figure 5K, all-cause mortality using Kaplan-Meier curves with 95% confidence intervals and p-values ​​from the log-rank test by age is shown. Figure 5M(a) shows the Kaplan-Meier curve for the patient group (age 50 or younger) according to filling pressure measured by our AI_ECG (HR 1.413, 95% CI 1.329-1.503). Figure 5M(b) shows the Kaplan-Meier curve for the patient group (age 50 < < 70) according to filling pressure measured by our AI_ECG (HR 1.35, 95% CI 1.313-1.388). Figure 5M(c) shows the Kaplan-Meier curve for the patient group (age 70 or older) according to filling pressure measured by our AI_ECG (HR 1.235, 95% CI 1.211-1.26). HRs were calculated for the four grades after adjusting for age, sex, and comorbidities.

[0067] Referring now to FIG. 5L, the distribution of left atrial volume index, e', E / e', tricuspid regurgitation velocity, and E / A according to diastolic filling pressure by AI-ECG for our echocardiographic grade 1 patients is shown. Each point represents the mean value with a 95% confidence interval. Normal filling pressures include normal and grade 1, and elevated filling pressures include grades 2 and 3. Student's t-test was used for p-values.

[0068] Referring now to FIG. 5M, the ROC curves and performance of AI-ECG in normal and low ejection fractions, defined as left ventricular ejection fraction, are shown.

[0069] Referring now to FIG. 5N, an exemplary case of a heart failure patient present in an exemplary study of certain embodiments is shown.

[0070] Referring now to Figure 5O, a flowchart of dataset construction and patient distribution is shown: Figure 5O(a) is a flowchart illustrating patient selection; Figure 5O(b) is a pie chart of dataset partitioning.

[0071] Referring now to FIG. 5P, an algorithm for assessment of diastolic filling pressures and function is shown.

[0072] Consideration Continuing with reference to Figures 5A-5P and a specific embodiment of an apparatus for an AI-assisted diagnostic assessment tool, we developed a deep neural network of AI-enabled ECGs to detect patients with elevated filling pressures and grade diastolic dysfunction. This specific embodiment demonstrates three main findings. First, the AI-enabled ECG model can identify patients with elevated left ventricular filling pressures with high accuracy. Second, the model's predictions of diastolic function and filling pressures are well aligned with echocardiographically determined diastolic function and estimated diastolic filling pressures. Furthermore, ECG-based diastolic function and filling pressure assessments can be associated with mortality. This specific embodiment of the model was designed to predict filling pressures and diastolic function grades using ECGs alone, which is much more cost-effective and scalable to a broader population. It is noteworthy that the echocardiographic parameters of patients with different diastolic function grades predicted by AI-enabled ECGs were comparable to those of the corresponding grades determined by echocardiography. Progression of diastolic dysfunction accompanied by elevated filling pressures as determined by echocardiography may predict increased mortality in various cardiovascular disorders, including heart failure with preserved ejection fraction (HFpEF). Similarly, mortality risk identified by AI-enabled ECG and echocardiography-based stratification may also be predictive. This may be explained by characteristic ECG features shaped by delayed myocardial relaxation and elevated filling pressures detected by a neural network. The model predicted mortality based on elevated filling pressures. This data suggests a promising role for AI-enabled ECG as a screening test to identify or rule out cardiac causes in patients presenting with dyspnea or other signs of heart failure. HFpEF is a common and growing public health problem, particularly affecting the elderly. Demonstration of elevated left ventricular filling pressures may support the diagnosis of HFpEF, but echocardiography or even invasive evaluation may be required in most cases. This AI-enabled ECG, in conjunction with AI-ECG determination of left ventricular ejection fraction, can identify patients with a high likelihood of HFpEF. It may also be possible to identify relatively asymptomatic patients with early-stage HFpEF.AI-ECG, used for diastolic function with a high negative predictive value, may rule out cardiac etiologies in patients with unexplained dyspnea if the AI-enabled ECG indicates normal diastolic function and / or filling pressures. Preliminary retrospective data from 2,000 patients with dyspnea evaluated in the emergency department of our medical center showed that AI-ECG determinations of normal and increased filling pressures were significantly associated with non-cardiac and cardiac dyspnea, respectively. In patients with known heart failure, this model can monitor their response to management to optimize filling pressures. AI-enabled ECGs trained on single-lead or median-beat ECGs showed promising performance and could be incorporated into patients' wristwatches to adjust heart failure medications. Echocardiography is an established diagnostic tool with well-documented reliability in determining diastolic function and filling pressures and was used as the reference technique in this study. These data suggest that AI-enabled ECG may also provide incremental value in the echocardiographic assessment of diastolic function. This model can predict diastolic filling pressures when echocardiography cannot assess diastolic function, a relatively frequent situation in clinical practice. Furthermore, the risk of death was robustly predicted by specific implementation of the model in patients with indeterminate echocardiographic diastolic function assessment. Echocardiographic parameters reflecting diastolic function and filling pressures (the ratio of early mitral inflow velocity (E) to early mitral annular diastolic velocity (e') (E / e'), left atrial volume index, tricuspid regurgitation velocity, and the ratio of early mitral inflow velocity to late mitral inflow velocity (E / A)) differed between normal and increased filling pressures predicted by ECG in the indeterminate group, similar to those in patients with a definite echocardiographic assessment of diastolic function. The main reason for indeterminate echocardiographic diastolic classification may be discordance between the four variables used to assess diastolic function. The recommended normal values ​​of echocardiographic diastolic parameters may be specific for diastolic dysfunction / elevated filling pressures but may not be highly sensitive. The differences in echocardiographic diastolic parameters between the elevated and normal filling pressure groups were smaller in the indeterminate group than in the test group.It is noteworthy that 54.7% of patients with grade 1 echocardiographic diastolic dysfunction were identified as normal by our AI-enabled ECG. The main difference between normal and grade 1 dysfunction is delayed myocardial relaxation, but both have normal filling pressures. Grade 1 echocardiographic diastolic dysfunction may comprise a heterogeneous population. This may be a common finding of normal aging, but it can also be observed in patients with compensated heart failure. Patients predicted as normal by our model were younger (66.1 ± 10.9 vs. 71.6 ± 9.1 years) and had fewer comorbidities compared with patients with grade 1 diastolic dysfunction by AI-ECG and echocardiography (Table 6). Left atrial volume index, e', E / e', and E / A, were significantly different between the two groups (p < 0.001). Furthermore, the group with grade 1 diastolic dysfunction by echocardiography and normal diastolic function by our model had a lower mortality rate than the group with grade 1 diastolic dysfunction by both echocardiography and the model. One-third of patients with grade 1 dysfunction at rest developed elevated filling pressures with exercise. The AI-ECG model had no exclusion criteria and included patients with reduced ejection fraction or heart disease characterized by both diastolic dysfunction and abnormal ECG patterns, such as hypertrophic cardiomyopathy, cardiac amyloidosis, and / or aortic valve stenosis. The majority of patients with grade 3 diastolic dysfunction by AI-ECG had cardiac amyloidosis or reduced left ventricular ejection fraction, followed by moderate to severe mitral regurgitation and hypertrophic cardiomyopathy (Table 7). There were 9,072 patients with low ejection fraction in the test group, of whom 3,007 (33%) were predicted to have normal diastolic function or grade 1 by AI-ECG, but they had a lower NPV compared with patients with a normal ejection fraction. Within the study, there was a specific case of a 30-year-old woman who developed heart failure due to postpartum cardiomyopathy with a LVEF of 30%. AI-ECG at that time showed a reduced LVEF and elevated filling pressures. Heart failure treatment made her asymptomatic, but her LVEF remained reduced. Follow-up AI-ECG showed a reduced LVEF and normal LV filling pressures. [Table 6] [Table 7]

[0073] conclusion In conclusion, this study demonstrates that application of an AI-assisted diagnostic assessment tool, a specific embodiment of a device for AI-enabled ECG, allows for grading of diastolic dysfunction and estimation of filling pressures, with robust prognostic value similar to that of comprehensive echocardiography. These data suggest that AI-enabled ECG may be useful for improving the diagnostic assessment of regional diastolic dysfunction and disorders associated with elevated filling pressures, including heart failure with preserved ejection fraction.

[0074] Referring now to FIG. 6, a flow diagram illustrating a method for training an apparatus for artificial intelligence-assisted diagnostic assessment 600 is shown. Method 600 may be implemented and / or modified in any manner as described within this disclosure. At step 605, a computing device receives a plurality of multi-channel sensor readings of physiological data, which may be implemented without limitation as described above in connection with FIGS. 1-5A-5C. At step 610, the computing device generates training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels, which may be implemented without limitation as described above in connection with FIGS. 1-5A-5C. At step 615, the computing device trains a neural network using the plurality of diagnostic labels, which may be implemented without limitation as described above in connection with FIGS. 1-5A-5C. At step 620, the computing device receives time series input describing the user's physiological data from at least the sensor, which may be implemented without limitation as described above in connection with FIGS. 1-5A-5C. In step 625, the computing device inputs the time series input to the trained neural network, which may be performed, without limitation, as described above in connection with Figures 1-5A-5C. In step 630, the computing device generates diagnostic data as a function of the time series input and the trained neural network, which may be performed, without limitation, as described above in connection with Figures 1-5A-5C. In step 635, the computing device determines prognostic data as a function of the diagnostic data, which may be performed, without limitation, as described above in connection with Figures 1-5A-5C. In step 640, the computing device outputs the prognostic data, which may be performed, without limitation, as described above in connection with Figures 1-5A-5C.

[0075] It should be noted that any one or more of the aspects and embodiments described herein may be conveniently implemented using one or more machines (e.g., one or more computing devices utilized as user computing devices for electronic documents, one or more server devices such as document servers, etc.) programmed in accordance with the teachings herein, as would be apparent to one skilled in the computer arts. As would be apparent to one skilled in the art, appropriate software coding could be readily produced by skilled programmers based on the teachings of the present disclosure. The above-described aspects and implementations using software and / or software modules may also include appropriate hardware to assist in implementing the machine-executable instructions of the software and / or software modules.

[0076] Such software may be a computer program product using a machine-readable storage medium. A machine-readable storage medium may be any medium that can store and / or encode a set of instructions for execution by a machine (e.g., a computing device) and cause the machine to perform any one of the methodologies and / or embodiments described herein. Examples of machine-readable storage media include, but are not limited to, magnetic disks, optical disks (e.g., CDs, CD-Rs, DVDs, DVD-Rs, etc.), magneto-optical disks, read-only memory "ROM" devices, random-access memory "RAM" devices, magnetic cards, optical cards, solid-state memory devices, EPROMs, EEPROMs, and any combination thereof. As used herein, machine-readable medium is intended to include a single medium as well as a collection of physically separate media, such as a compact disc in combination with computer memory or a collection of one or more hard disk drives. As used herein, machine-readable storage medium does not include a transitory form of signal transmission.

[0077] Such software may also include information (e.g., data) carried in a data signal on a data carrier, such as a carrier wave. For example, machine-executable information may be contained in a data carrier signal embodied in a data carrier, the signal encoding a sequence of instructions, or portions thereof, for execution by a machine (e.g., a computing device), and any associated information (e.g., data structures and data) that cause the machine to perform any one of the methodologies and / or embodiments described herein.

[0078] Examples of computing devices include, but are not limited to, e-book reading devices, computer workstations, terminal computers, server computers, handheld devices (e.g., tablet computers, smartphones, etc.), web appliances, network routers, network switches, network bridges, any machine capable of executing a set of instructions that specify actions that the machine should take, and any combination thereof. In one example, a computing device may include and / or be included in a kiosk.

[0079] 7 illustrates a schematic diagram of one embodiment of a computing device in the exemplary form of a computer system 700 upon which a set of instructions for causing a control system to perform any one or more of the aspects and / or methodologies of the present disclosure may be executed. It is also contemplated that multiple computing devices may be utilized to implement a specially configured set of instructions for causing one or more of the devices to perform any one or more of the aspects and / or methodologies of the present disclosure. Computer system 700 includes a processor 704 and memory 708, which communicate with each other and with other components via a bus 712. Bus 712 may include any of several types of bus structures, including, but not limited to, a memory bus, a memory controller, a peripheral bus, a local bus, and any combination thereof, using any of a variety of bus architectures.

[0080] Processor 704 may include any suitable processor, such as, but not limited to, a processor incorporating logic circuitry for performing arithmetic and logical operations, such as an arithmetic logic unit (ALU), which may be coordinated with a state machine and directed by operational input from memory and / or sensors. Processor 704 may be organized according to the Von Neumann and / or Harvard architectures, as non-limiting examples. Processor 704 may include, incorporate, and / or be integrated into, but is not limited to, a microcontroller, a microprocessor, a digital signal processor (DSP), a field programmable gate array (FPGA), a complex programmable logic device (CPLD), a graphical processing unit (GPU), a general-purpose GPU, a tensor processing unit (TPU), an analog or mixed-signal processor, a trusted platform module (TPM), a floating-point unit (FPU), a system-on-module (SOM), and / or a system-on-chip (SoC).

[0081] Memory 708 may include a variety of components (e.g., machine-readable media), including, but not limited to, random-access memory components, read-only components, and any combination thereof. In one example, a basic input / output system 716 (BIOS), containing the basic routines that help to transfer information between elements within computer system 700, such as during start-up, may be stored in memory 708. Memory 708 may also include instructions (e.g., software) 720 (e.g., stored on one or more machine-readable media) that embody any one or more of the aspects and / or methodologies of the present disclosure. In another example, memory 708 may further include any number of program modules, including, but not limited to, an operating system, one or more application programs, other program modules, program data, and any combination thereof.

[0082] Computer system 700 may also include storage 724. Examples of storage (e.g., storage 724) include, but are not limited to, hard disk drives, magnetic disk drives, optical disk drives combined with optical media, solid-state memory devices, and any combination thereof. Storage 724 may be connected to bus 712 by an appropriate interface (not shown). Exemplary interfaces include, but are not limited to, SCSI, Advanced Technology Attachment (ATA), Serial ATA, Universal Serial Bus (USB), IEEE 1394 (FIREWIRE®), and any combination thereof. In one example, storage 724 (or one or more components thereof) may be removably interfaced with computer system 700 (e.g., via an external port connector (not shown)). In particular, storage 724 and associated machine-readable media 728 may provide nonvolatile and / or volatile storage of machine-readable instructions, data structures, program modules, and / or other data for computer system 700. In one example, software 720 may reside, completely or partially, within machine-readable medium 728. In another example, software 720 may reside, completely or partially, within processor 704.

[0083] Computer system 700 may also include input device(s) 732. In one example, a user of computer system 700 may input commands and / or other information into computer system 700 via input device(s) 732. Examples of input device(s) 732 include, but are not limited to, an alphanumeric input device (e.g., a keyboard), a pointing device, a joystick, a gamepad, an audio input device (e.g., a microphone, a voice response system, etc.), a cursor control device (e.g., a mouse), a touchpad, an optical scanner, a video capture device (e.g., a still camera, a video camera), a touch screen, and any combination thereof. Input device(s) 732 may interface with bus 712 via any of a variety of interfaces (not shown), including, but not limited to, a serial interface, a parallel interface, a game port, a USB interface, a FIREWIRE® interface, a direct interface to bus 712, and any combination thereof. Input device(s) 732 may include a touch screen interface, which may be part of or separate from display 736, as described further below. The input device 732 may be utilized as a user selection device for selecting one or more graphical representations within the graphical interface, as described above.

[0084] A user may also input commands and / or other information into computer system 700 via storage device 724 (e.g., a removable disk drive, a flash drive, etc.) and / or network interface device 740. A network interface device, such as network interface device 740, may be utilized to connect computer system 700 to one or more of various networks, such as network 744, and one or more remote devices 748 connected thereto. Examples of network interface devices include, but are not limited to, a network interface card (e.g., a mobile network interface card, a LAN card), a modem, and any combination thereof. Examples of networks include, but are not limited to, a wide area network (e.g., the Internet, an enterprise network), a local area network (e.g., a network associated with an office, building, campus, or other relatively small geographic space), a telephone network, a data network associated with a telephone / voice provider (e.g., a mobile communications provider data and / or voice network), a direct connection between two computing devices, and any combination thereof. A network, such as network 744, may use wired and / or wireless communication modes. In general, any network topology may be used. Information (eg, data, software 720 , etc.) can be communicated to and from computer system 700 via network interface device 740 .

[0085] Computer system 700 may further include a video display adapter 752 for communicating displayable images to a display device, such as display device 736. Examples of display devices include, but are not limited to, a liquid crystal display (LCD), a cathode ray tube (CRT), a plasma display, a light emitting diode (LED) display, and any combination thereof. Display adapter 752 and display device 736 may be utilized in combination with processor 704 to provide graphical representations of aspects of the present disclosure. In addition to a display device, computer system 700 may include one or more other peripheral output devices, including, but not limited to, audio speakers, a printer, and any combination thereof. Such peripheral output devices may be connected to bus 712 via peripheral interface 756. Examples of peripheral interfaces include, but are not limited to, a serial port, a USB connection, a FIREWIRE® connection, a parallel connection, and any combination thereof.

[0086] The foregoing has been a detailed description of exemplary embodiments of the present invention. Various modifications and additions may be made without departing from the spirit and scope of the present invention. Features of each of the various embodiments described above may be combined with features of other described embodiments as needed to provide multiple feature combinations in related new embodiments. Moreover, while the above describes several separate embodiments, what has been described herein is merely illustrative of the application of the principles of the present invention. Furthermore, while certain methods herein may be illustrated and / or described as being performed in a particular order, the order may be highly variable within the skill of those skilled in the art to achieve the methods, systems, and software according to the present disclosure. Accordingly, this description is intended to be taken by way of example only, and not to limit the scope of the present invention.

[0087] Exemplary embodiments are disclosed above and shown in the accompanying drawings. Those skilled in the art will understand that various modifications, omissions, and additions can be made to what is specifically disclosed herein without departing from the spirit and scope of the present invention.

Claims

1. 1. An apparatus for artificial intelligence assisted diagnostic evaluation, comprising: The system includes at least a processor, the processor comprising: receiving a plurality of multi-channel sensor readings of physiological data; generating training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels; training a neural network using said plurality of diagnostic labels; receiving time series inputs describing physiological data of a user from at least a sensor; inputting the time series inputs into the trained neural network; generating diagnostic data as a function of the time series inputs and the trained neural network; determining prognostic data as a function of said diagnostic data; and outputting the prognostic data.

2. The apparatus of claim 1 , wherein the time series input comprises electrocardiogram (ECG) data.

3. The plurality of multi-channel sensor readings further includes a plurality of ECG readings, and the plurality of training data further includes a plurality of ECG readings, and generating the plurality of training data receiving echocardiogram data correlated to each ECG reading of the plurality of ECG readings; The apparatus of claim 2 , further comprising: generating the plurality of diagnostic labels using the echocardiogram data.

4. The apparatus of claim 1 , wherein the neural network is a multi-output convolutional neural network.

5. The apparatus of claim 1 , wherein the neural network is trained using general usage lead data.

6. The apparatus of claim 1 , wherein the neural network is trained using reduced general use lead data.

7. 2. The device of claim 1, wherein the neural network is trained using data collected from 12-lead ECGs performed within 14 days of a transthoracic echocardiography, and the plurality of multi-channel sensor readings of physiological data include 12-lead ECG data and echocardiography data correlated to each 12-lead ECG data.

8. The apparatus of claim 1 , wherein the diagnostic data includes an assessment of LVDD.

9. The diagnostic data includes classification under a plurality of groups, the groups being: normal LV function, LVSD only, LVDD only, and The device of claim 1 including both an LVSD and an LVDD.

10. The apparatus of claim 1 , wherein the output of the prognostic data is displayed on a display device.

11. 1. A method for training an artificial intelligence assisted diagnostic assessment tool, comprising: receiving a plurality of multi-channel sensor readings of physiological data; generating training data correlating each of the plurality of multi-channel sensor readings with a plurality of diagnostic labels; training a neural network using the plurality of diagnostic labels; receiving a time series input describing physiological data of a user from at least a sensor; inputting the time series inputs into the trained neural network; generating diagnostic data as a function of the time series inputs and the trained neural network; determining prognostic data as a function of said diagnostic data; and outputting the prognostic data.

12. The method of claim 11 , wherein the time series input comprises electrocardiogram (ECG) data.

13. The plurality of multi-channel sensor readings further includes a plurality of ECG readings, and the plurality of training data further includes a plurality of ECG readings, and generating the plurality of training data receiving echocardiogram data correlated to each ECG reading of the plurality of ECG readings; The method of claim 12 , further comprising generating the plurality of diagnostic labels using the echocardiogram data.

14. The method of claim 11 , wherein the neural network is a multi-output convolutional neural network.

15. The method of claim 11 , wherein the neural network is trained using general usage lead data.

16. 12. The method of claim 11, wherein the neural network is trained using reduced general use lead data.

17. 12. The method of claim 11, wherein the neural network is trained using data collected from 12-lead ECGs performed within 14 days of a transthoracic echocardiography, and the plurality of multi-channel sensor readings of physiological data includes 12-lead ECG data and echocardiography data correlated to each 12-lead ECG data.

18. The method of claim 11 , wherein the diagnostic data includes an assessment of both LVSD and LVDD.

19. The diagnostic data includes classification under a plurality of groups, the groups being: normal LV function, LVSD only, LVDD only, and The method of claim 11 , including both an LVSD and an LVDD.

20. The method of claim 11 , wherein the output of the prognostic data is displayed on a display device.

Images