Systems and methods for maintaining health using a personal digital phenotype

Abstract

A system and method for identifying and treating a disease in a patient collects one or more data streams from sensors configured to detect biological signals generated within the patient's tissue over time. Patient data elements including one or more of the patient's demographic, clinical, laboratory, pathological, chemical, image, historical, genetic, and activity data are collected and processed with the data streams to generate a personalized digital phenotype (PDP). The PDP is compared to a digital taxonomy including prior data to classify the patient into one or more quantitative disease classifications to guide personalized interventions for treating the patient.

Description

[0001] Related Applications

[0002] This application claims the benefit of priority to U.S. Provisional Application No. 62 / 664,833, filed on April 30, 2018, which is incorporated herein by reference in its entirety.

[0003] Government Rights

[0004] This invention was made with government support under Grant Nos. HL83359 and HL103800 awarded by the National Institutes of Health (NIH). The government has certain rights in this invention. Technical Field

[0005] The present invention generally relates to personalized therapies for disease and health maintenance, and more specifically to a system and method for defining a digital phenotype of a disease based on data collected for a specific person (which data can reference a digital taxonomy of the disease based on population data), thereby providing identification of key abnormalities for personalized therapy. Background Art

[0006] There is a growing recognition that medical treatments are often over-generalized and can be improved through individualization. Many accepted therapies work reasonably well in populations, but in rare cases work poorly or not at all. Even among patients for whom treatments work, responses are often graded across individuals. There are often few a priori clues that a particular treatment may not work in a given patient. Indeed, a priori "predictors" of response or failure are often based on successes and failures observed after the fact, and attempts to use these predictors to improve treatments often yield incremental rather than substantial benefits.

[0007] Relative to a statistical minority, current medical practice for stated morbidity explicitly prioritizes the suitability of most individuals, because only therapies that work in most individuals are advocated. An important but neglected problem is the fate of a minority of individuals suffering from stated morbidity, and the reaction of these minority individuals to a therapy may be different from the reaction when the therapy is used on most individuals. This minority may include a large number of individuals, but therapy may usually even be abandoned because of these individuals, unless it is easy to identify (phenotype) these individuals from other individuals suffering from stated morbidity. Such "sub" phenotyping is often difficult, because otherwise, such individuals will have been separated from other individuals suffering from stated morbidity in advance and enter different subcategories.

[0008] There is a pressing need to personalize therapy: to identify a priori those patients for whom therapy is likely to work, those for whom it is unlikely to work, and ideally to optimize therapy for individuals. To meet these goals, personalized medicine is increasingly being investigated using strategies to identify the genetic cause(s) of a condition, phenotype the individual accordingly, and tailor therapy based on this phenotype.

[0009] Personalized medicine is most often applied to the symptom caused by genetic abnormality (" mechanism "), and the application of the symptom that can not prove clear genetic trait is not too frequent.Unfortunately, this includes many diseases in the most popular diseases in all organ systems.In the heart, although the symptom of genetic cases can be identified, such as coronary artery disease or cardiac rhythm disorder (including atrial fibrillation (AF)), for example, patients with familial hereditary hypercholesterolemia or isolated AF, such cases are a minority.Most cases do not have clear and provable genetic causes and are considered to be caused by multiple factors (multifactorial).Some recent studies have also failed to show genetic abnormalities even in the symptom that is regarded as hereditary (including hereditary cardiac arrest (that is, sudden arrhythmic death syndrome (" SADS ")) of young people).

[0010] Other conditions may appear partially heritable or may have a genetic cause with "incomplete penetrance." The cause of such uncertainty in disease expression or response to therapy is unknown and occurs with many atrial fibrillation therapies, for example. Such uncertainty is often attributed to "environment" and may be expressed as variation in the cellular "proteome" or "metabolome," but can be difficult to identify, is often unproven, and is rarely used to guide therapy.

[0011] Heart rhythm disorders are a common and significant cause of morbidity and mortality worldwide. Under normal circumstances, the sinoatrial node keeps the heart in sinus rhythm. Under certain circumstances, rapid activation of the normal sinoatrial node can lead to inappropriate sinus tachycardia or sinus node reentry, both of which also indicate heart rhythm disorders.

[0012] Malfunctions in the electrical system in the heart are the proximate cause of cardiac rhythm disorders. Cardiac rhythm disorders can be classified as simple or complex. To some extent, this is an arbitrary division based on whether the rhythm is currently well understood and treatable. However, simple rhythms can be viewed as having a well-defined circuit that is stable over time, with most analytical methods (mapping) being consistent. Examples include sinus rhythm (SR), atrial tachycardia (AT) or flutter (AFL), atrioventricular nodal reentrant tachycardia (AVNRT), and atrioventricular reciprocating tachycardia (AVRT). In contrast, complex rhythm disorders include atrial fibrillation (AF), ventricular fibrillation (VF), and other rhythm forms (in which the circuit changes over time), such as polymorphic ventricular tachycardia (PMVT), multiple premature atrial complexes (PAC), multiple premature ventricular complexes (PVC), and interrelated atrial macro-reentrant circuits.

[0013] Heart rhythm disorders (especially complex rhythm disorders AF, VF and VT) can be difficult to treat. Pharmacological therapy for complex rhythm disorders is not optimal. Ablation has been increasingly used for heart rhythm disorders by manipulating the sensor / probe to the heart through the blood vessels or directly during surgery and delivering energy to the position of the heart to alleviate and in some cases eliminate the heart rhythm disorder. However, in complex rhythm disorders, ablation is often difficult and ineffective because the tools for identifying and locating the cause (source) of the heart rhythm disorder are imperfect, thereby hindering attempts to deliver energy to the appropriate area of ​​the heart to eliminate the disorder. Despite the best efforts, in the medium and long term, drug therapy for complex arrhythmias has only a success rate of 30% to 60%. Ablation is increasingly used, but it is suboptimal for complex conditions. For example, in a recent multicenter trial, the success rate of a single ablation procedure for "paroxysmal" AF, which was considered the simplest form, was 65%, which dropped to 50% over time within a year. For patients with more complex persistent AF, in a recent multicenter trial by Verma et al., single-procedure success rates with “gold standard” techniques were approximately 40% to 50% in the first year after drug cessation (“Approaches to catheter ablation for persistent atrial fibrillation,” N Engl J Med. 2015;372:1812–22.) Each of these studies is consistent with the results of many other multicenter prospective investigations.

[0014] There are several unmet needs that, if addressed by new scientific breakthroughs or inventions, could help improve the success of these therapies. Many unanswered questions remain: Why does the same ablation approach work for some patients but not others despite multiple attempts? What mechanisms of rhythm disturbances are similar or different between individuals, and can the phenotypes of these rhythm disturbances be identified? Current disease classifications are not ideal for this purpose, as pulmonary vein isolation fails in 35% to 50% of cases of “simple” paroxysmal AF within 1 to 2 years but is effective in 40%-50% of cases of “advanced” persistent AF within 1 to 2 years. Why do some patients do well after ablation of the suggested AF source, while others do not? In source-driven AF, why do some individuals have only one source even for complex AF, while others have several? Why are source or driver regions associated with structural abnormalities (e.g., fibrosis on MRI or voltage analysis) in some individuals but not in others (Narayan et al., Circulation Arrhythmia / Electrophysiology 2013;6(1):58-67)? Why are cardiac rhythm pathologies such as AF associated with processes such as neural stimulation or metabolic diseases such as diabetes in some individuals but not in others?

[0015] Heart rhythm disorders are classified by their electrical patterns within the heart. Detailed mapping leads to defining conditions such as AVRT, separating them from other conditions that may otherwise have similar symptoms or ECG appearances, and then identifying the cause to develop a curative therapy. Prior art strategies for classifying the electrical causes of heart rhythm disorders use several approaches. In simple heart rhythm disorders such as atrial tachycardia, the source of the disorder can be identified by tracing the excitation to the earliest location, which can be cauterized (ablated) to alleviate and in some cases eliminate the disorder. However, even in simple heart rhythm disorders, ablating the cause of the heart rhythm disorder is challenging, and experienced practitioners often require hours to ablate simple rhythm disorders such as atrial tachycardia that exhibit a consistent beat-to-beat excitation pattern.

[0016] In complex rhythm disorders, diagnosis and treatment often involve introducing a catheter with multiple sensors / probes into the heart through the patient's blood vessels. The sensors detect the electrical activity of the heart at the sensor location in the heart. The electrical activity is typically processed as an electrogram signal representing the excitation of the heart at the sensor location. In complex disorders, the signal at each sensor location can switch on a beat-by-beat basis between one, several, and multiple deflections of various shapes. For example, when the signal for a sensor location in AF includes 5, 7, 11, or more deflections, it is difficult to identify which deflections in the signal are local to the sensor location in the heart (i.e., local excitation onset) relative to a nearby sensor location in the heart (i.e., far-field excitation onset) or are merely from noise from another part of the patient's heart, other anatomical structures, or external electronic systems. Despite recent attempts, such irregular deflections make it difficult, if not impossible, to identify the activation onset of the beat in the signal at the sensor location (Sahli et al., Ann Biomed Eng 2018;46(2):257-269; Zaman et al., Circ Arrhythm Electrophysiol 2018;11:e005258).

[0017] Several attempts have been made to identify the causes of cardiac rhythm disturbances. One approach considers the regularity of the signal at the sensor location as a surrogate for the source of complex rhythm disturbances, in which the source is considered to be more organized than adjacent sensor locations. For example, U.S. Pat. No. 7,117,030 to Berenfeld et al., U.S. Pat. No. 5,792,189 to Gray et al., Sahadevan et al. (Circulation 2004), and Sanders et al. (Circulation 2005) describe methods for identifying (multiple) sources of variable atrial fibrillation (AF) as high regularity rapid regions, often using high spectral dominant frequencies with high regularity indices. Other studies, such as Nademanee et al. (J Am Coll Cardiol 2004), suggest that complex fragmented atrial electrograms (CFAE) may be key regions of AF. Other studies, such as Cuculich et al. (Circulation 2010), use body surface imaging to show primarily disorganized transient activation patterns and rarely show source regions. In clinical use, such approaches have produced disappointing results in guiding ablation in many studies (eg, Verma et al., 2015; Wong, 2015, etc.).

[0018] Recently, the prior art has described methods and systems for identifying discrete sources of complex rhythm disturbances (including atrial fibrillation and / or ventricular fibrillation). See, for example, U.S. Patents 8,676,303 and 8,521,266 granted to Narayan et al., Narayan et al. (Journal of the American College of Cardiology 2012), Haissaguerre (Circulation 2014), and Krummen et al. (Journal of the American College of Cardiology 2014). Despite the surrounding complex collision waves and disorganization, these systems show localized organized sources, where the disorder reflects additional competing sources, noise or other factors. Online systems can now also be used to identify the source or driver of complex rhythms (Al-Husseini et al., J. Cardiovasc. Electrophysiol. Journal of Cardiovascular Electrophysiology 2017, 28: 615-622). These studies found that using the same electrogram data, some analysis systems showed stable sources, while other systems showed greater source uncertainty (Alhusseini et al., [Journal of Cardiovasc Electrophysiology] 2017; Zaman and Narayan, [Circulation Arrhythmias and Electrophysiology] 2017; 10(12): e006022). However, it is unclear which patterns work for some patients relative to others, and therefore how to optimize therapy for these different patient subtypes.

[0019] Prior art has described tools for mining biological data using machine learning techniques. For example, Guyon et al. (U.S. Pat. No. 7,542,947) teach incorporating kernel-based learning machines such as support vector machines (SVM) into a decision system that integrates data from heterogeneous existing data sources with genetics / proteomics and clinical data to facilitate identification and ranking of molecular markers for testing and treating diseases. In Guyon, heterogeneous data sets can include gene or protein expression data, mass spectrometry data, antibody levels, clinical observations and medical history, drugs, hormones, immunological tests, genetic and family history, physical and chemical measurements, online journals, online genetic databases, and other sources. The method disclosed by Guyon is not specifically directed to guiding therapy, but to the identification of molecular markers for diagnosing diseases.

[0020] U.S. Patent 8,954,339 to Schaffer discloses a clinical decision support system (CDSS) with an inference engine that applies predetermined diagnostic and treatment rules to measured physiological parameters and patient tests to make diagnostic and / or treatment recommendations based on specialized knowledge and published research. Added functionality includes identifying outliers in a data set that mark potential defects in the inference engine that may lead to erroneous recommendations. Such decision support systems provide general decisions for known conditions and do not teach or suggest solutions to problems that are not currently identified by experts. For example, regarding the question of whether a rhythm disorder will respond to ablation, methods for separating atrial fibrillation forms or ventricular fibrillation forms that can be ablated without a defibrillator are not currently known to experts and therefore will not be available in prior art CDSS. In addition, prior art CDSS may not be able to identify the clinical importance of symptoms or signs of individual patients or the presence or absence of specific features of electrical data, specific features of structural data from computed tomography or magnetic resonance imaging, blood tests, genetic tests, or specific features of any other data set.

[0021] Some prior art methods apply machine learning and neural networks to rhythm disorders, but do not expand on what experts already know and do not address issues that are critical to treatment. For example, U.S. Patent 5,819,007 to Elghazzawi and Chinese Patent Publication CN107203692A each disclose a method for detecting AF using electrical data processed via a neural network that can provide high accuracy. As described in Chinese Patent Publication CN106066933A, this method is extended to disclose a method for adding a weighted value matrix to an artificial neural network (ANN) to identify AF from a stored database (MIT-BIH database). In other studies, trained learning machines have been used to diagnose the presence of disease (e.g., malignant skin lesions) with accuracy comparable to that of dermatologists. See, for example, Esteva et al. (Nature [Nature] 2017) (which uses a deep convolutional neural network (CNN)) or U.S. Patent 8,543,519 to Guyon et al. (which describes the use of SVM to classify images of skin lesions as malignant or benign). Hannun, Rajpurkar et al. Nature Medicine 2019 disclosed the use of CNN to detect arrhythmias from ECGs with accuracy comparable to that of cardiologists. However, this prior art has little relevance to patients who have already been diagnosed with AF because it does not improve upon what experts already know and therefore does not advance treatment from the current suboptimal level. U.S. Patent 7,171,269 B1 to Addison et al. describes the use of neural networks to adjust the timing of electric shock therapy for fibrillation, but does not address ablation or drug therapy or why neural networks are ineffective in many patients. U.S. Patent Publication 2015 / 0042646 A1 to Comaniciu et al. discloses a computational model for individual arrhythmias that incorporates electrical and structural data to plan interventions. However, this model does not take into account relatively ineffective therapies, such as pulmonary vein isolation for AF, which has a success rate of 40% to 60% or cardiac resynchronization therapy, which has a success rate of 60% to 70%.

[0022] Many experts describe the need to integrate multiple approaches to diagnose, track, or treat arrhythmias, such as combining invasive therapies with noninvasive therapies, but this approach is rarely applied objectively or titrated using quantitative tools. For example, while weight loss can improve and reduce AF, it is unclear for which patients this is most applicable or whether such individuals can be identified as different groups, as some obese individuals do not develop AF, while many non-obese individuals do. This also applies to other lifestyle changes such as exercise, which may induce AF in the context of excessive exertion, while other patients with a sedentary lifestyle will also experience AF. Inflammation may be a cause of many diseases, including AF, but it is unclear how to integrate testing and / or therapy into overall management.

[0023] It would be very beneficial for society to develop methods for personalizing therapies that work for individuals (even if these therapies do not work for all patients), thereby ideally covering a wide range of people. This can include individuals who have no or an identified genetic cause for their disease or response to therapy. This includes personal diagnostics for quantitatively guiding multiple therapy modalities and in some cases for identifying the exact cause of an individual's multifactorial disease for treatment. An exemplary embodiment is for complex cardiac rhythms. Currently, there are few, if any, systems in the prior art to achieve these goals. Summary of the invention

[0024] In an exemplary embodiment, the present methods and systems use one or more of mathematical networks utilizing algorithms, machine learning and probabilistic analysis, near real-time data sensed from sensors, and pre-stored (historical) data to generate a personal digital phenotype ("PDP") of health and disease to guide personalized strategies for treating disease and / or maintaining the health of the individual. Exemplary embodiments include cardiac applications in heart rhythm disorders, in coronary artery disease, and in heart failure.

[0025] The present technology is designed to automatically generate biologically and clinically meaningful individual phenotypes that are not currently known to experts and therefore represent a substantial improvement over prior art CDSS that approximate expert performance or may be based on expert methods or rules.

[0026] When applied to cardiac rhythm disorders, complex rhythm disorders may produce patterns of cardiac activation that are difficult for experts to understand, but can be elucidated and treated by generating a personal phenotype from electrical, structural, and clinical data inputs to find the exact cause for treatment. This hybrid approach extends beyond the traditional causes of rotational or focal drivers or "random" wavelets to include different electrical activation patterns or combinations of electrical patterns with structural or clinical data elements.

[0027] The inventive methods described herein are well suited for complex diseases, including cardiac rhythm disorders, coronary artery disease, and heart failure, each of which is a heterogeneous syndrome rather than a single disease. For example, in AF, a variety of electrical, structural, and metabolic mechanisms can contribute to its development and progression, each of which can differ between individuals. In principle, individualizing therapy ("personalized medicine") using networks built from large sets of sensed and archived input data sets (data elements), often from electronic health records, in patients with various demographic characteristics and "traditional" phenotypes, trained by response or non-response to various AF therapies, each of which has a well-defined mechanism of action, can identify therapeutic focal phenotypes and mechanisms of AF for each individual.

[0028] The lack of availability of detailed mechanism-based management for many complex organ functions due to several factors further motivates the present method. First, conceptual knowledge of the physiological contributions to cardiac rhythm disorders, heart failure or even local ischemia is limited. Second, there are inter-individual differences in the physiological contributions, for example, the effects of structural abnormalities can change atrial fibrillation or heart failure in one individual relative to another. Third, physiological interactions are often dynamic and can respond to environmental changes. Dynamic changes can occur over years, months or even weeks to reflect normal or abnormal age-related changes or disease onset or regression and can explain why initially effective management becomes ineffective. Fourth, even if defined, it is often non-trivial to regulate this pathophysiological feature. These limitations apply to multiple organ systems. Therefore, it is a major challenge to regulate the function of the heart or the rest of the body a priori from known associations or physiological patterns.

[0029] Several innovations separate the current inventive method from the prior art. First, the inventive method produces a digital bedside implementation of "personalized medicine" or "precision medicine" quantitative personal digital phenotype (PDP), which does not require cellular or genetic data that may not exist in many diseases. Secondly, the PDP represents a combination of multiple data inputs (including readily available signals). These signals are not only aggregated, but also referenced during the normal state of the individual to divide health or disease. This may not require complex data input. Thirdly, the method uses the sensory data streams and / or input data elements of the population (including those obtained from electronic health records) to generate novel digital taxonomies for a given disease or health condition. Mathematical, statistical and machine learning techniques are used to classify or divide these data sets to define the quantitative form of the disease process. The more detailed and extensive (i.e., "richer") the population data elements are, the more comprehensive the calculated digital taxonomy can be. In one embodiment, different hospitals contribute to a large digital prospective data registry that is de-identified, which will contribute to different patient groups, practice patterns and data on the results of various therapies. Fourth, the method of the present invention can adapt the PDP and classification method over time using updated data and calculated fluctuations. Fifth, the system of the present invention utilizes novel sensors and existing health care machines that can be operated separately or in combination (e.g., networked). Sixth, complex inputs (including data from specialized medical devices such as imaging systems, data streams from many sensors, individual data, or even data from other individuals) can also be merged to achieve crowdsourcing analysis. Seventh, the phenotypes generated using the method of the present invention enable customization of "personalized medicine." Finally, the systems and methods described herein provide quantitative means to provide therapy through one or a combination of lifestyle changes, medication, electrical or mechanical therapy, surgical or minimally invasive ablation, gene or stem cell therapy.

[0030] In specific application to cardiac rhythm disorders, arrhythmias can be targeted with anti-inflammatory therapies, including immunosuppression with agents such as tacrolimus, a heretofore unrecognized form of therapy for complex arrhythmias such as atrial fibrillation. Thus, the methods of the invention can improve outcomes by eliminating repeated interventions - the strategy with the highest predicted initial success rate can be designed and applied to a given individual at the outset, thereby reducing the cost and delay of treatment.

[0031] In one embodiment, the PDP is defined using multiple sensory data streams and data element streams from an individual. This is similar to a mathematical representation of similar data in a population, which constitutes a novel disease classification method, thereby enabling personalized management (including therapy for individuals). The method of the present invention utilizes direct sensory data and clinical data from various sources, thereby using deterministic and probabilistic methods (including machine learning) to digitally represent these data as a quantitative phenotype of the individual's cardiac condition, and uses an integrated device or an external device to guide therapy. The system of the present invention is able to use repeated measurements over time to update personal phenotypes to customize therapy for different conditions or predict deterioration or improvement of cardiac conditions. The wide application of the method of the present invention increases over time with the range of data and the number of updates, thereby individually forming a rich digital portrayal of the patient's health compared to a reference population. Personalized digital networks can also track dynamic balance in a given organ system to identify perturbations of the disease and normalization when therapy is successful. Data streams can be obtained invasively or non-invasively from readily measured systems, including invasive recordings of electrical activity (electrograms), blood flow and pressure (hemodynamics), wall tension (cardiac contractility and relaxation), and related indices. Sensors can record non-invasively from electrocardiograms, skin measures of neural activity, and skin reflectivity. Other types of sensed signals that can be used will be apparent to those skilled in the art.

[0032] The data types used within the methods of the present invention can be selected and categorized to represent different aspects of a biological organ or process. In the heart, the data streams include one or more of the following: electrical signals, three-dimensional anatomy and structural abnormalities, clinical data extracted from medical history and physical examination, indicators of pathophysiological comorbidities, blood and tissue biomarkers, and the genetic and cellular makeup of the individual.

[0033] The candidate individual digital phenotype is an objectified representation of the pathophysiological network, representing indicators derived from signal processing, association algorithms, data clustering (including data clustering from unsupervised machine learning), and supervised networks trained on labeled events in similar and different individuals. Using introspective analysis to pinpoint defects in disease networks and health corrections, the digital object provides a comprehensive picture of pathophysiology.

[0034] In one embodiment, the present methods and systems use machine learning to develop and learn a classification that links complex physiological and clinical inputs to patient-level outcomes. This can then be used prospectively to predict the best therapy based on the specific characteristics of the individual. Thus, the technology produces a novel patient-specific classification for disease that consists of digital phenotypes that can be quantitatively used to personalize therapy.

[0035] The application of the method of the present invention is broadened with the richness of data input. Time-invariant data streams and time-varying data streams are important inputs that can represent normal changes or disease states. Machine learning and in particular deep neural networks are considered to be most useful: when the scale of the network is large, this is a desirable feature of the present invention; and when the data repository is large, this is indispensable for generating the digital taxonomy in the present invention.

[0036] The data streams are multiple and designed to address the limitations of existing technologies: normal ranges, nomograms, and clinical studies are often performed in individuals, ethnic groups, populations, or less clear disease states that may not reflect the disease of the individual being treated. Multiple data streams are designed to help personalize management and therapy. This can include familial predisposition to disease (Mendelian or non-Mendelian), identifiable genetic loci, weight changes, or susceptibility to toxins such as tobacco or alcohol. Other phenotypes may be clinically identifiable but not tracked by biomarkers, or may have loose statistical definitions, such as racial or ethnic susceptibility.

[0037] The signal can be sensed without physical contact with the sensor. Examples include sensing the heartbeat from the electromagnetic field emitted by a magnetocardiogram (MCG) or from the infrared signature of the heart's motion.

[0038] Signals may also be sensed by physical contact with the sensor. Examples include motion sensing of chest wall movement from breathing or heartbeats, chest wall vibrations from certain types of breathing (e.g., loud obstructive breath sounds) or heart sounds (e.g., so-called "tremors" in the medical literature). Electromagnetic sensors may sense electromagnetic signals associated with electromyograms (EMGs), electroencephalograms (EEGs), electrocardiograms (ECGs), or other transmitters. Respiratory sensors may detect movement of the chest wall, abdomen, or other body parts associated with ventilation or acoustic data (sounds) associated with breathing or oxygenation associated with breathing. Chemical sensors may detect chemical signals on the skin or other membranes that reflect body chemistry (such as oxygenation and deoxygenation, metabolic acidosis, emergency, or other states that will be familiar to those skilled in the art of biochemistry). Sensors may also detect images using cameras or lenses that require contact from fingerprints or other body parts or sense movement from specific muscles or sense iris dilation or oscillation from photoelectric sensors in contact lenses. Positioning sensors can identify the position of body parts and changes over time (including gait) or provide contact sensing of the position of certain body parts at one point in time or over time (e.g., facial droop, facial tick, or other idiosyncratic movements).

[0039] Less commonly used sensors include those for detecting inflammation. In the prior art, the significance of inflammation findings is often unclear in any given person at one point or over time and from person to person. "Inflammasomes" can measure the effects of various pathological insults on inflammation at the cellular or tissue level but do not often do so, may not assess diurnal fluctuations, may not have an unclear relationship to systemic inflammatory conditions, and may have unclear differences between individuals (i.e., it is unclear how to establish a "nomogram" or normal or abnormal state). This is another area of ​​disease tracking and treatment that is well suited for personalized digital phenotyping.

[0040] In the present invention, the personalized condition of inflammation can be detected by the presence of inflammatory cells in the organ system that was initially inflamed or in body fluids such as blood, urine or cerebrospinal fluid. The byproducts of inflammation can also be detected by increased concentrations of biomarkers and cytokines (such as interleukin-6, nerve growth factor, matrix metalloproteinases). On the other hand, in the case of inflammation, a variety of physiological markers are abnormal (so-called "acute phase reactants"). In addition to increased white blood cell counts, inflammation also causes abnormalities in red blood cell counts, hemoglobin concentrations, and a large number of acute phase reactants (such as C-reactive protein, erythrocyte sedimentation rate or white blood cell count). In the heart, it is well known that serum troponin (a marker of cardiac cell destruction) is an acute phase reactant, and its level decreases with inflammation ('reverse acute phase reactant').

[0041] For preferred embodiments of cardiac rhythm disorders such as atrial fibrillation, immunosuppressive therapy (used in transplants), steroidal or non-steroidal agents, or cell therapy may be effective. The rationale for this embodiment is that patients receiving immunosuppressants after heart transplantation rarely develop AF. This benefit is attributed to the isolation of the pulmonary veins by surgical scarring, whereas for other populations, isolation of the pulmonary veins provides up to 50% to 70% freedom from AF. In support of the present invention, inflammation has been implicated as a cause of AF in AF alone and AF associated with obesity, pericardial fat, or oxidative stress from reactive oxygen species. However, immunosuppression for complex rhythm disorders, including AF, is rarely, if ever, implemented. Using the present method, digital classification and PDP will identify individuals with possible inflammation-mediated arrhythmias in whom anti-inflammatory therapy, including immunosuppression, may be useful.

[0042] The system of the present invention can use novel sensors to sense data signals from body systems. The sensed data stream can be combined with other data elements sensed or detected by other systems of the individual to produce a digital network representation. The sensed signals from measurable body systems can include central and peripheral nervous systems, cardiovascular systems, respiratory systems, skeletal muscles and skin, and any other body systems. Other input data elements can come from imaging, nuclear, genetic, laboratory or other sources and can also be sensed as streams (i.e., transmitted to the system) or as value inputs at specific points in time.

[0043] In one aspect of the invention, a disease of a patient is identified and treated by: collecting at least one data stream generated by at least one sensor configured to detect biological signals generated within a patient's tissue over time; acquiring patient data elements, the patient data elements comprising one or more of the patient's demographic, clinical, laboratory, pathological, chemical, image, historical, genetic, and activity data; processing the at least one data stream and the acquired patient data elements by a processing module configured to execute a partitioning algorithm to generate a personalized digital phenotype (PDP); comparing the PDP with a digital taxonomy constructed from prior data to classify the patient into one or more quantitative disease classifications; and personalizing the patient's treatment based on the one or more quantitative disease classifications. In some embodiments, the at least one sensor is in physical contact with the patient's body, and the data stream is transmitted via one of wired or wireless communications. The sensor can be one or more of an electrode, an optical sensor, a piezoelectric sensor, an acoustic sensor, a resistive sensor, a thermal sensor, an accelerometer, a pressure sensor, a flow sensor, and an electrochemical sensor. The biosignals may include one or more of electrical heart signals, mechanical heart signals, heart rate, heart sounds, breath sounds, respiratory rate, respiratory volume, neural activity, and immunological signals. The patient data elements may include one or more of electrical signals, hemodynamic data, cardiac structures from imaging, clinical factors associated with cardiac or pulmonary conduction, neural signals, genetic profiles, metabolic state biomarkers, and patient movement.

[0044] The method of the present invention may further include applying a timestamp to each of the at least one data stream and the patient data elements before processing. The partitioning algorithm may be one or more of supervised machine learning, neural networks, correlation analysis, logistic regression analysis, decision trees, time domain analysis, frequency domain analysis, triangular transformation, logarithmic transformation, cluster analysis, and unsupervised machine learning. The digital classification method may include previous data of the patient. The one or more quantitative disease classifications may be one or more of rotational or focal excitation patterns, intermittent rotational or focal excitation patterns, incomplete excitation patterns, and the location or specific anatomical location of the specific cardiac structure of the individual. The patient tissue may be the heart, nerves supplying the area of ​​the heart, brain areas controlling the nerves, blood vessels supplying the area of ​​the heart, and tissue adjacent to the heart. In some embodiments, the disease may be a cardiac rhythm disorder, which includes one or more of atrial fibrillation, ventricular fibrillation, atrial tachycardia, atrial flutter, polymorphic or monomorphic ventricular tachycardia, ventricular flutter, or other electrical disturbances in the heart. In such embodiments, the method may further include, prior to processing, generating a map using the at least one data stream, wherein the map is an image representing actuation at locations within the heart, and wherein processing includes identifying locations of relatively higher actuation. In other embodiments, the data stream is a clinical electrogram including erroneous deflections, and wherein processing further includes generating a reconstructed electrogram using a machine learning algorithm trained on one or more reference signals associated with different heart rhythms.

[0045] In some embodiments, the personalized intervention may include modifying at least a portion of the patient's tissue by ablating one or more of: energy delivery via a contact device, energy delivery via a non-contact device, electrotherapy, thermal therapy, mechanical therapy, delivery of drug therapy, delivery of immunosuppression, delivery of stem cell therapy, and delivery of gene therapy. The method may further include generating updated personal history data of the PDP, classified one or more quantitative disease classifications, the personalized intervention, and intervention results.

[0046] In another aspect of the present invention, a system for identifying and treating a patient's disease includes: at least one sensor configured to detect biological signals generated within the patient's tissue over time and generate at least one data stream; and a computing device configured to: collect the at least one data stream; collect patient data elements, which include one or more of the patient's demographic, clinical, laboratory, pathology, chemistry, image, historical, genetic, and activity data; process the at least one data stream and the patient data elements in a processing module, which is configured to execute a partitioning algorithm to generate a personalized digital phenotype (PDP); compare the PDP with a digital classification method including stored previous data to classify the patient into one or more quantitative disease classifications; and determine a personalized intervention for treating the patient based on the classified one or more quantitative disease classifications.

[0047] In some embodiments, the at least one sensor is in physical contact with the patient's body, and the data stream is transmitted via one of wired or wireless communications. The sensor can be one or more of an electrode, an optical sensor, a piezoelectric sensor, an acoustic sensor, a resistive sensor, a thermal sensor, an accelerometer, a pressure sensor, a flow sensor, and an electrochemical sensor. The biosignals can include one or more of electrical heart signals, mechanical heart signals, heart rate, heart sounds, breath sounds, respiratory rate, respiratory volume, neural activity, and immunological signals. The patient data elements can include one or more of electrical signals, hemodynamics, cardiac structures from imaging, clinical factors associated with cardiac or pulmonary conduction, neural signals, gene maps, metabolic state biomarkers, and patient movement.

[0048] The system of the present invention further includes a computing device that applies a timestamp to each of the at least one data stream and the patient data elements before processing. The partitioning algorithm can be one or more of supervised machine learning, neural networks, correlation analysis, logistic regression analysis, decision trees, time domain analysis, frequency domain analysis, triangular transformation, logarithmic transformation, cluster analysis, and unsupervised machine learning. The digital classification method can include previous data of the patient. The one or more quantitative disease classifications can be one or more of rotational or focal excitation patterns, intermittent rotational or focal excitation patterns, incomplete excitation patterns, and the location or specific anatomical location of the specific cardiac structure of the individual. The patient tissue can be the heart, the nerves supplying the area of ​​the heart, the brain area controlling the nerves, the blood vessels supplying the area of ​​the heart, and the tissue adjacent to the heart. In some embodiments, the disease can be a cardiac rhythm disorder, which includes one or more of atrial fibrillation, ventricular fibrillation, atrial tachycardia, atrial flutter, polymorphic or monomorphic ventricular tachycardia, ventricular flutter, or other electrical disturbances in the heart. In such embodiments, the method may further include, prior to processing, generating a map using the at least one data stream, wherein the map is an image representing actuation at locations within the heart, and wherein processing includes identifying locations of relatively higher actuation. In other embodiments, the data stream is a clinical electrogram including erroneous deflections, and wherein processing further includes generating a reconstructed electrogram using a machine learning algorithm trained on one or more reference signals associated with different heart rhythms.

[0049] In some embodiments, the personalized intervention may include modifying at least a portion of the patient's tissue by ablating one or more of: energy delivery via a contact device, energy delivery via a non-contact device, electrotherapy, thermal therapy, mechanical therapy, delivery of drug therapy, delivery of immunosuppression, delivery of stem cell therapy, and delivery of gene therapy. The computing device may be further configured to generate updated personal history data of the PDP, classified one or more quantitative disease classifications, the personalized intervention, and intervention results.

[0050] A significant advantage of the method of the present invention is its ability to analyze streams from other individuals with similar or different profiles than the current patient to enable the creation and updating of digital taxonomies through crowdsourcing.

[0051] In one aspect, the system of the present invention includes a processor and a memory storing instructions that, when executed by the processor, perform operations including: detecting body signals associated with one or more body functions at one or more sensors associated with a human body; processing the body signals to produce one or more sensed signature features; processing the signature features using digital objects to determine effector responses; delivering one or more effector responses to control a body task; and monitoring the responses.

[0052] Preferred embodiments relate to cardiac rhythm disorders. Using electrical signals, the system of the present invention can identify the causes of rhythm disorders that are personalized through statistical measurements and inference data about the individual, so as to provide a preparatory therapy target that expands the prior art. Combining electrical signals with anatomical, clinical or other data, the network identifies defects in the cardiac-physiological network, thereby enabling personalized therapy to treat individual cardiac rhythm disorders. In this embodiment, the onset of cardiac rhythm disorders can be predicted, and once initiated, the source of cardiac rhythm disorders can be identified to guide invasive or non-invasive therapy.

[0053] In summary, the present invention uses multiple sensed and inputted data to generate personalized digital phenotypes that are compared to digital taxonomies from similar and different individuals (including individuals with known outcomes) and applied as digital personalized medicine strategies for maintaining health or therapies for correcting disease. Preferred embodiments, which are not intended to be limiting, include cardiac applications in heart rhythm disorders, coronary artery disease, and heart failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Some embodiments are illustrated by way of example and not limitation in the figures of the accompanying drawings in which:

[0055] Figure 1 is a block diagram that provides an overview of an embodiment of the present invention for digital taxonomies and phenotype-specific therapies.

[0056] Figure 2 is a diagram of an embodiment of the system of the present invention in which a cardiac application integrates data streaming from the heart or related organs with input data from a number of clinical sources.

[0057] Figure 3A and Figure 3B A circuit schematic diagram of an exemplary device for sensing signals for analytical processing of a personal digital phenotype (PDP) is provided.

[0058] Figure 4 is a diagram of a sample workflow using personal digital phenotypes and digital taxonomies.

[0059] Figure 5is a flow chart illustrating an exemplary process for generating a personal digital phenotype (PDP).

[0060] Figure 6 is a flow chart showing how individual phenotypes and numerical taxonomies can be used to define health, disease, or justify ongoing monitoring.

[0061] Figure 7 The process flow used to identify specific scrambled elements of an individual's digital phenotype is summarized.

[0062] Figure 8 An exemplary PDP-based therapy sequence for atrial fibrillation is presented.

[0063] Fig. 9 Sample phenotypes of complex arrhythmias are shown, including (A) fibrillatory conduction away from the source / driver; (B) disorganized activity without perturbing the source / driver; and (C) arrhythmias based on structural abnormalities.

[0064] FIG. 10A to FIG. 10C Design of interventions based on individual digital phenotypes of arrhythmias, including constraining or perturbing the spatial localization of the source of arrhythmias or terminating the source of arrhythmias, is demonstrated.

[0065] Fig.11 A process flow of one embodiment of the method of the present invention for analyzing and treating cardiac arrhythmias using electrograms (voltage-time signals) as input data is presented.

[0066] Fig. 12A A process flow of an embodiment of the method of the present invention for analyzing and treating cardiac arrhythmias using rotational activation maps from any mapping method as input data is presented. FIG. 12B to FIG. 12E Shows the use of Fig. 12A The invention provides details of arrhythmia graphs for a given patient, wherein: Fig. 12B provides an example of the challenges in identifying deflection, Fig. 12C shows a reconstruction of an embodiment of the method according to the invention, Fig.12D The application of the method of the present invention to different clinical signals is illustrated graphically, and Fig.12E The application of the scheme of the present invention to signal reconstruction is demonstrated.

[0067] Fig.13A An exemplary design of a machine learning method using CNN is shown; Fig. 13B An exemplary process flow of a supervised machine learning approach for personal digital phenotyping is shown.

[0068] FIG. 14A to FIG. 14D Exemplary output information generated for cardiac rhythm disorders in an embodiment of the system of the present invention is provided, showing the following: Fig.14APlots showing the accuracy of the neural network on the AF map segmented into rotational and non-rotational regions and the test set arrhythmia map of AF; Fig. 14B A sample embodiment of machine learning for AF graph interpretation is shown; Fig. 14C Demonstrating the accuracy of interpretability analysis for localizing the cause of cardiac rhythm disturbances; and Fig.14D Possible displayed content of a user interface / mobile device is shown.

[0069] Fig.15 is a block diagram of an exemplary embodiment of a general purpose computer system for implementing the present invention. DETAILED DESCRIPTION

[0070] For the purposes of this disclosure, the following definitions apply:

[0071] "Associative learning" means the process of linking input data to measurable physiological or clinical outcomes. Associative learning can be iterative, enabling associations to be modified ("learned") based on changing patterns between inputs and measured outputs (physiological or clinical endpoints).

[0072] "Biological signal" means a signal generated by the body and may reflect one or more body systems. For example, heart rate reflects cardiac function, autonomic tone, and other factors. See also non-biological signals.

[0073] "Biometric signal" means a signal that provides a measure of a human characteristic. A biometric identifier may be physiological or behavioral. Physiological biometrics include, but are not limited to, DNA, fingerprints or palm prints, oral swabs, tissue or urine samples, retinal images, facial recognition, geometry of the hands or feet, identification of an individual's iris or odor / breath. This may also apply to signals such as vital signs, ECG, EEG, EMG, etc. Behavioral biometrics include patterns such as gait during walking or a cadence. Embodiments of the present invention use dynamic patterns of combined physiological and behavioral biometrics over time that are adaptable to individual changes and are therefore robust to forgery from previous "versions" of a person's signature features.

[0074] "Body" refers to the physical structure of single-celled organisms, multi-celled organisms, viruses, and prions. Organisms include animals (such as but not limited to humans and other mammals), plants, bacteria, etc.

[0075] "Consumer device" means a device that is intended for direct use by a consumer without a medical prescription. Historically, such devices have generally not been regulated by medical regulatory authorities or agencies (such as the US Food and Drug Administration or similar regulatory agencies in other countries), however, recently, some devices have been approved by the FDA. Consumer devices may include hardware, software, or a combination thereof. Consumer devices are generally not medical devices, which are defined as instruments, apparatus, appliances, machines, tools, implants, in vitro reagents, or other similar or related articles (including component parts or accessories) intended for use in the diagnosis of disease or other conditions in humans or other animals or in the cure, mitigation, treatment, or prevention of disease. The definition of medical device excludes medical decision support software.

[0076] An "effector" is a means of performing a bodily task and may include a physical device, a prosthesis, a mechanical or electronic device. A physical device may enhance bodily function, such as a device for moving a limb or moving the diaphragm to enhance breathing during sleep, or a splint for keeping the airway open during sleep, or one or more signals for stimulating a bodily function (such as electrical stimulation of the phrenic nerve to enhance breathing during sleep), or an artificial prosthesis (such as a cybernetic limb or implanted circuitry for the peripheral or central nervous system).

[0077] "Data stream" or "data stream(s)" means biological data sensed by one or more sensors that can provide real-time or near real-time information about the sensed biological process. Sensors in the heart can provide streams including electrocardiogram (ECG), pulse rate, pulse waveform, and therefore cardiac hemodynamics. Other data streams may include cardiac acoustics, including analysis of heart sounds, murmurs, and complex analysis of cardiac-related hemodynamics. Pulmonary function can be sensed as chest movement associated with breathing, auscultatory sounds, and neural discharges. Gastrointestinal diseases can be sensed as sounds (belly noise), movements on the abdominal wall, and electrical signals related to intestinal smooth muscle activity. Central and peripheral nervous system activity can be sensed as neural activity on the scalp (electroencephalogram, EEG), away from the scalp but still reflecting the EEG, and away from peripheral nerve discharges.

[0078] As used herein, "demographic characteristics" means clinically relevant personal information, which may include, but is not limited to, age, gender, family medical history, and the presence of comorbidities.

[0079] "Numeric taxonomies" means the classification of different diseases or health conditions based on quantitative indicators. A traditional disease classification may state that "atrial fibrillation is more common in elderly individuals, individuals with cardiac comorbidities (such as valvular disease or heart failure), and individuals with metabolic syndrome." In contrast, numeric taxonomies are designed to be quantitative, describing individual health for a specific disease in terms of quantifiable primary and secondary data elements (i.e., data vectors). Disease entity D n The probability of existing in a particular individual is given by the probability p(D n ) to roughly estimate:

[0080]

[0081] where m is the number of available data input types, n is the disease under consideration, and p(V n,i ) is the data vector V n,i The probability that input i causes disease n, and k n is a weighting constant for disease n. These elements are estimated by a digital classification method that calculates the specific probability that a certain data input causes the disease. The probability can be obtained from group data, where a specific person is matched with the closest individual in the group to estimate this probability. It is also possible to obtain the probability only from this specific individual at a possible or determined healthy time and ideally at a possible or determined disease time. These calculations can be performed by traditional estimation equations, but can also be formed by machine learning. In this way, the digital classification method probabilistically represents the disease entity from the aggregation of anomalies in multiple related data inputs. Therefore, this process is dynamic because the classification equation for any given disease will change when more data inputs are available and when the data stream changes (especially if the determined disease condition is updated accordingly). Therefore, this digital classification method is very suitable for analyzing large amounts of data from individual wearable devices or massive data from several individuals as a crowdsourcing paradigm.

[0082] "Historical data" means stored data that may include reports from medical imaging (e.g., magnetic resonance imaging (MRI), computed tomography (CT), radiology, or other scans of organs), data from genetic testing analysis (e.g., the presence of one or more genomic variants), previously obtained ECG reports, pathology, cytology, and other laboratory reports, and clinical demographic characteristics such as age, sex, family history, and the presence of comorbidities. Historical data may further include other personal history details that may be relevant to the generation of a PDP, such as mental illness, employment in a high-pressure occupation, number of pregnancies (for women), engagement in high-risk behaviors such as smoking, drug abuse, or alcohol abuse.

[0083] "Input data" or "data input(s)" means data that is not directly sensed by the physical components of the system, but is utilized by the processing unit in conjunction with the sensed data to generate the PDP and digital taxonomy. Input data from data sources may include data streams detected using other systems (e.g., external ECG or EEG systems), clinical, laboratory, pathology, chemistry or other data, or data from medical imaging equipment, which data is transmitted to the processing unit.

[0084] "Machine learning" means a range of analytical methods and algorithms that can learn from and make predictions about data by building models rather than following static programming instructions. Machine learning is often classified as a branch of artificial intelligence and focuses on developing computer programs that can change when exposed to new data. In the present invention, machine learning is a tool that can be used to generate a digital network that links sensed signature features to each individual's physical tasks (i.e., for personalized solutions to maintain health and diagnose disease). Machine learning techniques include supervised learning, transfer learning, unsupervised learning, or reinforcement learning. There may be several other classifications, but these classifications mainly reflect the following concepts:

[0085] “Unsupervised machine learning” includes methods such as cluster analysis, which can be used to identify internal connections between data, potentially such as connections between clinical data (diagnosis of atrial fibrillation), family history, data from physical examination (irregular pulse), data from sensors, electrical data (irregular atrial signals on ECG), structural imaging data (enlarged left atrium), biomarkers, available genetic and tissue data.

[0086] “Supervised machine learning” includes methods that can classify a series of related or seemingly unrelated inputs into one or more output categories without explicitly modeling the inputs, i.e., without assuming potentially incorrect (“partial”) mechanistic assumptions.

[0087] "Reinforcement learning" is a form of machine learning related to psychology that focuses on how software agents take actions in a particular environment to maximize cumulative rewards. Reinforcement learning is often used in game theory, operations research, swarm intelligence, and genetic algorithms and has other names such as approximate dynamic programming. One implementation of machine learning is through formulation as a Markov decision process (MDP). Reinforcement learning differs from supervised machine learning in that reinforcement learning does not require matching inputs and labeled outputs, and actions that result in suboptimal rewards are not explicitly corrected (unlike supervised learning, which can correct for suboptimal rewards by, for example, the backpropagation algorithm in a perceptron).

[0088] “Medical device” means an instrument, apparatus, implement, machine, tool, implant, in vitro reagent, or other similar or related article (including component parts or accessories) intended for use in the diagnosis of disease or other conditions, or in the cure, mitigation, treatment, or prevention of disease in man or other animals. The definition of medical device excludes medical decision support software.

[0089] "Neural network" means a self-learning network of interconnected nodes loosely modeled after the human brain that can be used to recognize patterns. Artificial neural networks can combine heuristics, deterministic rules, and detailed databases.

[0090] "Personal Digital Phenotype" ("PDP") means a digital combination of multiple data inputs and streams representing an individual's disease entity or health aspect. Data elements can directly or indirectly represent the health status of an individual, weighted by the probability of disease or health of the individual of a certain age, gender and morbidity caused by the data elements. Digital combination and weighting are performed by algorithmic analysis, which parameterizes different elements into functional groups through calculated and documented probabilities of influence on health and disease. The combination can include deterministic algorithms and machine learning forms. For example, the cardiac rhythm phenotype will mainly consider heart rate and electrogram signals (surface ECG and intracardiac). These data elements will be given a higher mathematical weight. Data streams from other (indirect) organ systems can include changes in respiratory rate with heart rate (i.e., lung sensors), changes in neural discharges with heart rate (i.e., neural function). Other important data elements include abnormal cardiac ejection fraction, position of the heart, and presence of structural abnormalities. Historical data including age, gender, and family history can also affect the overall digital personal phenotype.

[0091] "Index individual" means a patient or target for study or estimation for which a personal digital phenotype can be generated.

[0092] As used herein, "population data" is a key determinant of the accuracy of the methods of the present invention. If the reference population is very different from the index individual, the digital taxonomy may not include individuals of this phenotype and may or may not represent the individual. In this case, the population data and the digital taxonomy are primarily derived from the individual's previous judgment data, that is, data points when the individual was healthy and when the individual was unwell. If the reference population is broad and inclusive, but is not adequately phenotyped or adequately labeled for the data elements that cause the disease, the taxonomy will be useless. Therefore, the ideal data set is not a 'large-scale' or 'big data' collection using ubiquitous sensors that are neither adequately labeled nor in a well-defined population. Instead, the ideal data set includes parallel data streams that can be self-organized by clustering to divide different categories within the digital taxonomy.

[0093] "Sensor" includes a device that can detect biological signals from an individual's body. The sensor can be in direct contact with the body or can be remote. When applied to a group of individuals, the sensor can represent all or part of the defined group. The sensor can detect electrical information, such as voltage or current representing ECG, EEG, EMG or nerve discharge. The term "sensor" (especially when describing certain cardiac applications of the present invention in which electrical information is detected) can be used interchangeably with "electrode", "electrode catheter" or "catheter". Electrical sensors can also detect bioimpedance, such as reduced trans-skin conductivity when a person sweats, which can occur during the time when the sympathetic nervous system is dominant. Sensors can also detect other chemical changes through electric current. Sensors also include devices that detect temperature, such as thermistors or other thermal detectors. Sensors can detect light, such as changes in the color of light reflected from pulsating heart activity (photoplethysmography), changes in peripheral oxygenation (e.g., cyanosis, anemia, skin vasodilation). Sensors can detect sound through microphones. Sensors can detect other vibrations or movements through piezoelectric elements. Sensors can directly detect chemicals using specialized sensors for hormones, drugs, bacteria, and other elements, which are typically transduced into electrical signals on the device. In an exemplary embodiment of the system of the present invention, multiple sensors can be used, which communicate with a central computing device or can form a network via WiFi TM or other protocols to form the Internet of Things (IoT) of biosensors.

[0094] "Signal" includes electronic, electromagnetic, digital or other information that can be sensed or acquired. Sensed signals are detected, which are not changed from their natural form (i.e., recorded) without transformation. The sensed signals are typically biological signals. The sensed signals can be detected by humans (e.g., sound, vision, temperature) or by machines such as microphones, audio recorders, cameras, thermometers, etc. The acquired signals are detected under transformed conditions (such as ECG recordings). Such signals can be biological (because the heart bioelectricity generates ECG) or non-biological signals (e.g., vibrations sensed after applying acoustic or ultrasonic energy) or tactile signals transduced from sensed electrical signals, acoustic signals, or other signals.

[0095] "Smart data" means application-specific information obtained from a source that can be used to identify and / or act on normal or abnormal functions in an application. Therefore, smart data is different from the term "big data". "Smart data" is customized for an individual and is customized to solve a specific task or application, such as maintaining health and alertness or detecting and treating diseases such as sleep-disordered breathing. The customization is based on knowledge about which systems can affect the task in question. Such knowledge can be based on physiological, engineering or other principles. In contrast, "big data" often focuses on extremely large data sets with the goal of identifying statistical patterns or trends without individually customized links. In machine learning terms, smart data may be generated by supervised learning of a data set to a known output, while big data only refers to the amount of data without necessarily implying any knowledge of the importance of a particular data set.

[0096] Other biological terms use their standard definitions, such as heart failure, tidal volume, sleep apnea, obesity, etc.

[0097] The following description and accompanying drawings provide examples of applications of the systems and methods of the present invention for generating personal digital phenotypes (PDPs) of health and disease, compared to the digital taxonomies disclosed herein for enabling personalized strategies to treat an individual's disease and / or maintain the individual's health. The examples described herein are intended to be illustrative only. As will be apparent to those skilled in the art, additional variations and combinations may be formed using the principles of the invention disclosed herein.

[0098] Figure 1 Flow within an exemplary system for defining personal digital phenotypes, comparing these personal digital phenotypes to digital taxonomies, and designing personalized therapies is shown. Personal I / O data flow 100 obtains time-varying (dynamic) data from an index individual 10 (patient) through signals sensed at one or more sensors 102, which can be external devices (solid lines) or internal (implanted) devices (dashed lines). Implanted devices can be inserted explicitly for the purpose of developing / maintaining PDPs or implanted for different purposes. Such devices can communicate using the Internet of Things (IoT), where time-stamped data is sent to an input unit 110. Such data can be sent in a connected or wireless manner and can be signals acquired in a continuous, near-continuous, real-time, near real-time, or some other time format.

[0099] Deviations from normal people (beyond pre-specified "tolerance limits" compared to their own data) can be quantified in the index individual and compared to different groups. In a preferred embodiment, this is achieved by repeatedly sensing the data stream 100 or acquiring data. The data stream can be acquired using novel sensors, consumer sensors, or medical-grade sensors. Data can be input from several sources during periods when the organ system of the person is judged to be "healthy" and similarly during periods when the health / "disease" is reduced. The accumulated data helps future supervised learning to verify digital phenotypes. Different states can be detected for changing health or disease conditions or levels (e.g., exercise vs. rest) in individuals and between individuals. This approach is different from current medical practice, in which "group" ranges of "normal" and "disease" are applied across multiple patients, with little room to customize these ranges according to individuals. This aspect of the invention enables "personalized medicine" or "precision medicine."

[0100] Input data 105 may include demographic characteristics, laboratory, chemical and image data. For example, important data inputs of people may include "static" stored data such as date of birth (age), gender and race. Input data 105 may also include near real-time data, such as patient movement from a separate device (e.g., a treadmill, a motion sensor in a building), patient ECG or heart information from a separate device (e.g., hospital telemetry, ICU bed monitor), respiratory sensors, delayed or time series data from a separate device (e.g., regular blood glucose counts from a blood glucose meter) or other data inputs. Input data 105 is also sent to input unit 110 via a wired or wireless connection along with a timestamp.

[0101] Conventional cloud-based computing / storage 115 may optionally be used to store data (as a supplement or backup to the data stored on the device in 100 or 105) and / or perform processing of the data. The raw data and analysis results saved or generated in the cloud-based computing / storage may be transmitted to an external server (not shown) connected via the Internet, respectively. For example, independent recipients may include research facilities, clinical trial managers, or other recipients authorized by the patient.

[0102] The group database 120 provides a mathematical reference for the personal data and can include group stored data and optional data streams. The data in this database can take the form of time-varying stream data, but can also include accumulated and stored data from the database.

[0103] Processing unit 125 is programmed to execute algorithms including deterministic formulas as well as neural networks (or other learning machines) and other distributed representations to generate a personal digital phenotype (PDP) in step 130 and compare the PDP to a digital taxonomy in step 135 that includes data from previously determined time points for the index individual 10 and data from the population database 120. In one embodiment, machine learning is used to process the input data, develop and learn taxonomies that link complex physiological and clinical inputs to patient-level outcomes (i.e., develop PDPs) and compare these taxonomies to quantitative traits in related populations (digital taxonomies of health or disease) and prospectively design "optimal" or "personalized" therapies based on the individual's specific characteristics relative to previous observations of the individual, previous observations in a comparator population, or mathematically inferred predictions.

[0104] The comparison performed in step 135 is used in step 145 to identify and / or track the patient's state, i.e., health or disease. In step 150, the state information is transmitted to a unit for display, such as a smartphone app, a dedicated device, or an existing medical device. The health / disease information can be used to guide therapy through a therapy unit 155. In an embodiment, the therapy unit 155 for use with cardiac rhythm disorders includes delivering electrical intervention (pacing) or destructive energy (ablation) to tissue that is directed to the electrical and / or structural targets of the arrhythmia determined by the personalized digital phenotype. Other therapy units can use an infusion pump (thereby delivering anti-inflammatory therapy, because inflammation may be the proximate cause of cardiac rhythm abnormalities such as fibrillation) or gene or stem cell therapy delivery to deliver anti-arrhythmic drugs. In another embodiment, a therapy for delivering mechanical constraints can be delivered to relieve stretching that may trigger arrhythmias.

[0105] Figure 2 Exemplary systems are presented for defining personal digital phenotypes in the heart using sensors 200 or input streams 205 including clinical, laboratory, genetic or other data, particularly for cardiac rhythm, comparing these personal digital phenotypes to digital classifications, and delivering personalized arrhythmia therapy.

[0106] The heart 210 can be measured in many ways, including electrical signal sensors or electromagnetic signal sensors, such as electrode catheters 225 in the esophagus, electrodes 230 in the right atrium 230, electrodes 220 in the septum or left atrium, or through the great cardiac vein to various locations in the coronary sinus (electrodes 235, 240, 245), to the precardiac veins entering the left or right ventricles (electrode 215), or directly to any of these chambers. ECG electrodes can be applied to the body surface 250. Electrodes can also reside in other blood vessels or other locations near the heart 210. These electrodes are configured to detect cardiac excitation information in the heart and transmit the detected cardiac excitation information to the process controller 285 through the electrode controller. Data can be transmitted by wire or wirelessly.

[0107] In some embodiments, one or more of the sensors may be located outside the patient's heart. For example, sensor 250 detects cardiac agitation (e.g., an electrocardiogram (ECG)) via the surface of the patient. Other sensors (not shown) may detect cardiac agitation remotely without contact with the patient (e.g., a magnetocardiogram). As another example, some sensors may also obtain cardiac agitation information from cardiac motion of non-electrical sensing devices (e.g., echocardiograms, Doppler blood flow signals, scans with red blood cell labels). Such sensors may generally be classified as "external sensors" to distinguish these sensors from catheters and electrodes (i.e., "internal sensors") inserted into the patient's body, in or near the heart (or other organs). In various embodiments, various external sensors may be used individually or in different combinations, and further, these individual or different combinations of these external sensors may also be used in combination with one or more internal sensors.

[0108] Additional sensors may be provided to detect abnormal inflammatory or immune conditions of the body, thereby enabling such conditions to be modulated to maintain, enhance or correct biological rhythms. In an embodiment of cardiac applications, sensors are provided for detecting abnormal immune conditions and modulating them to prevent or treat complex cardiac rhythm disorders of atrial fibrillation or ventricular fibrillation.

[0109] The steps of the present invention may include, but are not limited to, sensing direct and indirect measures of "immune status," iteratively sensing these measures, comparing individual measures to normal and abnormal values ​​for that individual rather than population-based measures, and directing interventions and therapies to maintain normal immune balance or correct an impaired balance. In this case, the "sensing" of the measures goes beyond the traditional interpretation of raw signals and data collected from detection equipment and may include data generated from other testing procedures (e.g., clinical, laboratory, chemical, etc.) and inputting the test data into the process controller 285. As Figure 2As shown, information may be generated by a device 200 that senses electrical signals (e.g., an ECG or bioimpedance sensor) in combination with a transmitter for transmitting the detected signals via wired or wireless transmission to a process controller 285. Other sources of stream or input data 205 obtained from clinical systems, hospital databases, hospital equipment, or laboratory equipment may also provide input to the process controller 285 via wired or wireless transmission.

[0110] Direct measures of inflammation / immune balance include, but are not limited to, counts of inflammatory cells or concentrations of cytokines in body fluids or in affected organs. Indirect measures of inflammation / immune balance represent the variable effects of inflammation on various organ system abnormalities in resting and diurnal measures of body temperature, body fluid composition, heart rhythm, neural firing rates, and electroencephalograms.

[0111] The sensors 215 to 245 positioned at the sensor locations in the heart of the index individual can further detect cardiac activation information at the corresponding locations of these sensors and can further deliver energy to ablate the heart at the sensor locations. It should be noted that the sensors 215 to 245 can also detect cardiac activation information from overlapping areas of the heart (e.g., the right atrium and the left atrium 220).

[0112] The process controller 285 may include a digital signal processor for interfacing with output elements configured to stimulate different parts / nerves of the body or cause mechanical actions in external machines. Such elements may include integrated circuits with isolation or networking (e.g., cloud computing), biocomputing, integrated biological / artificial devices (cybernetics), or traditional computing machines that utilize unused biological capacity to perform specific directed tasks. Therapeutic devices may include direct electrical outputs, piezoelectric devices, visual / infrared or other stimulation systems, neural stimulation electrodes, or even virtualized data (such as avatars in virtual world interfaces or elements in large databases that can be queried), as well as other effector elements that are obvious to those skilled in the art.

[0113] The process controller 285 is configured to process (e.g., clarify and amplify) the cardiac activation information generated by the various input sources (sensor 200, (multiple) input data sources 205, sensors 215 to 245) into electrical, mechanical or other functional cardiac signals and provide the processed signals to the process controller 285 for analysis according to the methods disclosed herein. Figure 1, the functions corresponding to the input unit 110 and the processing unit 125 will be in the computational operations performed within the process controller 285. When processing the cardiac activation information from the sensors 200, 215 to 245 and the input data source 205, the processing controller 285 can subtract the cardiac information from the overlapping areas of the heart to provide a processed signal for analysis. In some embodiments, the process controller 285 is configured to analyze unipolar signals; in other embodiments, the process controller 285 analyzes bipolar signals, and in other embodiments, analyzes motion, optical, other electromagnetic signals.

[0114] The process controller 285 is programmed to implement the functional modules to generate a personalized digital phenotype of the individual's heart (rhythm) (PDP module 260, which corresponds to Figure 1 130 in step 130), using data from the population database 290 to match or compare the PDP to a digital taxonomy (taxonomy module 265, corresponding to Figure 1 ), designing a personalized intervention (design module 270) and guiding the delivery of the intervention through intervention module 275.

[0115] In some embodiments, the process controller 285 analyzes the signal according to readily available methods or methods disclosed herein so that (multiple) graphs ((multiple) representations) of cardiac activation indicators that can be used to locate (multiple) sources of cardiac rhythm disorders and eliminate the (multiple) sources can be generated. The (multiple) graphs can then be displayed on an output device and saved in a database with the PDP.

[0116] Population data 290 is configured to support or assist process controller 285 in analyzing the signals. In some embodiments, database 290 may store maps of potential AF source locations for other individuals whose digital phenotypes are known to be de-identified, which maps may be used to guide therapy for this individual (if the phenotype is similar) or to guide therapy away from this area (as suggested by the phenotype and classification). Population database 290 may also be used to store intermediate data.

[0117] Once an appropriate therapy has been designed in the design module 270, the therapy can be delivered. For example, using personalized intervention, instructions will be generated by the intervention module 275, causing the process controller 285 to deliver pacing from the pacing module 255. In another example, existing internal electrodes (e.g., sensors 215, 230 to 245) can be used to deliver ablation energy from an energy generator 280 as controlled by the process controller 285 and the intervention module 275. Other forms of energy, such as heating, cooling, ultrasound, laser, can also be delivered using appropriate therapy devices controlled by the process controller and intervention module 275. The therapy is delivered to tissues that are electrical and / or structural targets of arrhythmias determined by the personalized digital phenotype. Other therapy units can deliver anti-arrhythmic drugs using an infusion pump (thereby delivering anti-inflammatory therapy, as inflammation may be a proximate cause of abnormal cardiac rhythms such as fibrillation) or gene or stem cell therapy delivery. In another embodiment, a therapy for delivering mechanical constraints can be delivered to relieve stretching that may trigger arrhythmias. Therapy may also include external radiation, allowing for a completely non-invasive therapy where energy is focused at key arrhythmia targets inside the heart using high-energy electromagnetic radiation.

[0118] One type of sensing signal comes from the heart. This is a preferred embodiment of the present invention and may include several types of cardiac signals. Electrical activity may be sensed directly from the heart using sensors that may be placed on the heart and in contact or not, near other body regions (e.g., esophagus, bronchi and airways, mediastinum), on the body surface or not in contact with the body (e.g., magnetocardiography that can sense the magnetic field generated by the electrical activity of the heart). Sensors may measure cardiac function, motion, electrical excitation patterns, or the presence of ischemic areas by detecting cardiac motion or blood movement through the heart. Cardiac motion may be sensed from the movement of cardiac motion over body regions (ballistic cardiography), from changes in electrical impedance due to changes in the volume of cardiac chambers using non-electronic devices (e.g., echocardiography or ultrasound). Blood flow may be detected from imaging methods such as Doppler echocardiography, 4D flow MRI, and labeling of carriers such as red blood cells. In various embodiments, these sensors may be used alone or in different combinations, and further, these alone or in different combinations may also be used in combination with sensors inserted into the patient's heart.

[0119] Neural activity is an example of another sensed signal that can be used in embodiments of the invention, with metrics such as rate and periodicity of discharge, periodicity during the day and between days, type and pattern of neural discharge, and spatial distribution of these metrics. In one embodiment, non-invasive recordings are made from a skin patch, but other embodiments may use electroneurography (ENG) in which electrodes are inserted into the skin to record from nearby neural tissue. Invasive methods may be suitable for inpatient care, but are not suitable for continuous recording or consumer applications. Sensors can record from different regional nerves if placed in different areas, for example, electrodes on the chest can measure neural activity related to the heart or its nerves, electrodes on the neck or head can measure neural signals (including neural signals that control the heart), or other locations familiar to those skilled in the art.

[0120] Pulmonary (lung) functional activity is another example of a type of sensed signal and may vary independently of the heart or due to changes in the heart. This may be measured by sensors for breath sounds, chest wall movement, oxygenation, electrical activity of the phrenic nerve, or other sensors.

[0121] Figure 3A and Figure 3B Examples of devices that may be used in conjunction with analytical elements and personal digital phenotypes in embodiments of the present invention are provided. Figure 3A A simple circuit diagram of an exemplary sourcing 300 and sinking 330 circuit is provided. For the sourcing circuit 300, the power source is preferably low power for wearable or portable devices. In the example shown, an integrated photosensor 305 is included to detect optically detectable signals, however, as will be apparent to those skilled in the art, other sensor types 320 (e.g., piezoelectric sensors, acoustic sensors, resistive sensors, thermal sensors, accelerometer sensors, pressure sensors, flow sensors, electrochemical sensors, or other sensor types) may be used to measure chemical, light, skin activity / moisture levels, pressure, movement, and other parameters relevant to PDP development. In each hardware embodiment, the sensors may be interchangeable or fixed. Selecting appropriate component values ​​(resistors, capacitors, etc.) and circuit performance characteristics and adding supporting components / circuitry (filters, amplifiers, etc.) would be within the skill level of the art and are not described herein. The voltage output 307 provides a signal to an input channel for streaming data, for example, to a Figure 1 100 in the I / O data stream. For the sink circuit 330, a photosensor 340 is shown, wherein the output voltage 335 is transmitted to the I / O data stream.

[0122] Figure 3BFurther details of a preferred embodiment of the device are provided. This embodiment connects the sensors to logic circuits that process the combination of the (multiple) signals for optimal use in the personal digital phenotype and transmit the raw or processed signals to another computing device. The sensor output 350 includes the outputs from the sensors 1 ... n (e.g., Figure 3A The outputs 307 and 335 in FIG. 1 ). These outputs form inputs (v_input 1...n) 353 to a logic unit 355 (e.g., a microprocessor), RAM 358, a battery / power supply 362, and other components (including buses and I / O ports, such as a USB port 360 for wired connection via a USB cable 367). The USB link can be used for power and / or data transfer. The device preferably includes a wireless communication module 365 for wireless transmission, for example, Data and control signals are sent to another central processing unit (CPU) 375, which may include fast, parallel processors including a graphics processing unit (GPU) 378 that may be used to perform more demanding calculations.

[0123] The calculations performed by the device may include reading, compressing or storing the voltage output from the sensor, filtering the signal, or performing more advanced calculations to produce a personal digital phenotype or processing the signal based on the PDP. The device can then transmit the raw signal or the processed signal or control signal (e.g., a logical decision based on the sensed signal) to another device using USB or wireless methods (Bluetooth or WIFI). The Internet of Things (IoT) 366 can be used to link (multiple) sensor and / or computing device groups. Alternatively, the data can be passed directly to a secure cloud 370 for cloud storage and cloud computing services (via wired or wireless communication). The cloud engine 370 will be able to run mathematical operations and calculations at optimal speeds and allow communication from any Internet connection.

[0124] Figure 4 A workflow for using a personal digital phenotype to guide personalized therapy and monitor the effectiveness of therapy is summarized. Two forms of personal input are taken: a data stream 400 that can be updated over time and personal stored data 420. The data stream 400 can include sensory data from novel or existing sensors 402 (such as from wearable devices or consumer products), implanted devices 404, invasive sensing signals 406 from existing or specialized devices (including minimally invasive products such as skin, nose, cornea, cheek, anal or auditory probes), from non-invasive sensors 408 (which can provide data including motion and temperature from infrared probes), and from transmitted data 410 (such as telemetry from existing medical equipment).

[0125] The stored data 420 may include: static data, such as imaging data 422 (ideally including detailed coordinates of areas of scarring, fibrosis, ischemia, reduced contractile function, and potential border zone tissue), laboratory values ​​424 (including biochemical levels but also genetic, proteomic, and metabolomic data) (when available); demographic data 426 and elements from the patient's medical history (such as the presence of diabetes or hypertension, the size of the left atrium from echocardiography); outcome data 428, which may include subjective symptoms of whether the patient feels well, such as "healthy" or "not so healthy". The outcome data 428 ideally also includes objective evidence, such as acute endpoints of therapy, such as resolution of fever with antibiotics, or in a preferred embodiment, termination of atrial fibrillation with ablation. Objective evidence may also include chronic endpoints, such as the absence of infection or the absence of atrial fibrillation recurrence at long-term follow-up. As described below, the method of the present invention uses a combination of population data including population stored data 449, population data streams 452 and domain knowledge 455 to define a disease classification method 446 to distinguish between health and disease, divide data categories based on health conditions, and compare population categories to individuals.

[0126] Step 440 shows the continuous updating of the individual's digital phenotype by the machine. This can be performed periodically at predetermined time points or continuously. Step 443 performs a comparison of the individual's digital phenotype with an externally determined digital disease taxonomy 446 generated from one or more of the population stored data 449, the population data stream 452, and the domain knowledge 455. These steps are described in detail in detail below. Figure 5 Detailed description is given in further detail in .

[0127] In step 460, therapy is customized to the individual digital phenotype, and in step 463, the effectiveness of therapy is iteratively monitored using the data stream in the context of the data already stored in step 440. Finally, in step 466, an interactive interface for reporting data is provided. Display 470 provides an example of the various types of data that can be displayed by an application on a computer or mobile device. (In the example shown, a smartphone is shown.) An embodiment of an application ("app") is produced. In the illustrated display screen, a sample diagram of a complex arrhythmia with ablation targets, a 3D image of the heart, numerical coordinates, textual descriptions, and quantitative scores are shown. The inventive method generates personal phenotypes from personal and population data that can provide management and treatment decisions without the guidance of expert rules.

[0128] According to the present invention, the phenotype is based on the basic facts of whether the patient is sick, how sick the patient is, and how best to treat these patients to maintain health or treat the disease. At steps 428, 446 to 455, these clinical and biological related operations are included. The present invention acquires novel data in steps 402 to 410 to generate personalized phenotypes using data types that may not always correspond to the data types in the stored data of the individual (420 to 428) or the comparator group 446 to 455. Such data types are actively acquired by the system so that the personal phenotype can better guide therapy, and such data types include data types that contain electrical information or cardiac structure.

[0129] Figure 5 is a detailed description of the process for generating a PDP. Step 500 obtains personal data at a time vector [T] including one or more time points. This personal data includes a data stream 502 (corresponding to Figure 4 400) and similar to the step 400 from diagram 504 Figure 4 The personal stored data 420 in the personal stored data includes biologically or clinically meaningful metrics such as acute or chronic outcomes.

[0130] Group data stream 506 ( Figure 4 Such data streams may include data from people with the same or different sensed data to the target individual. To provide a few examples, this data may include telemetry data from a patient in an ICU ward, an individual using a similar wearable device, or a soldier in the field who is being monitored for various vital signs.

[0131] Use the weakening statistical association to merge the data from the population storage database 508 ( Figure 4 For example, when the population is evaluated statistically, AF is significantly associated with obesity, but there are also lean individuals who develop AF and many obese individuals who do not develop AF. Similarly, there is an association between atrial scar density or atrial scar burden and AF, but many individuals with AF have minimal scars, while some individuals without AF have considerable scars. However, the actuarial associations are obtained from the literature, and mathematical weighting is used to adjust for individual demographic characteristics using multivariate estimation.

[0132] In step 510 (similar to Figure 4In step 446 of the process, the population data from the database 508 is integrated, and the population data includes biologically or clinically significant indicators such as acute or chronic results used as references for diagnostic or therapeutic efficacy. The first step is to characterize the data in step 514. This step is usually designed as a feature (variable) simplification step and is designed to solve the dimensionality problem in machine learning. Feature simplification techniques and feature extraction techniques are well known in the art. Possible operations that can be used for this feature simplification step 526 include time domain analysis, such as averaging, integration, area analysis, correlation, and principal component analysis. Frequency domain analysis includes Fourier analysis, wavelet transform, and time-frequency analysis. Frequency analysis can be performed using selected parameters associated with the multiple coordinate pairs (or cycles) to generate a spectrum. The selected parameters can be amplitude (e.g., voltage), angle, vector, area, and derivative. In some embodiments, the peak may include the fundamental frequency; in other embodiments, the peak may include the fundamental frequency and one or more harmonics of the fundamental frequency. In other embodiments, the peak may include only one or more harmonics of the harmonics of the fundamental frequency. Polynomial fitting can also summarize the data set as polynomial coefficients. Other general characterization steps may be used using widely available libraries such as TSFresh (Time Series Feature Extraction). In parallel, in step 534, the population data is parameterized using similar or different operations.

[0133] Step 518 partitions the data into categories that represent numerical 'disease phenotypes' in individuals or numerical 'disease taxonomies' in populations. The goal is to better separate the data (clinical data) as well as granular invasive data points and laboratory tests, into partitions of individuals that may appear similar but have different outcomes (successes vs. failures) with a given therapy. Mathematically, this is accomplished by constructing 'hyperplanes' in k-parameter space that separate patients who have one outcome from those who do not. For example, for a preferred embodiment of arrhythmias, it is not clear why two patients with similar profiles of 'paroxysmal' AF may respond completely differently to medication or pulmonary vein ablation. Personal phenotypes encode observations from multiple patients into crowdsourced partitions ('numerical taxonomies') to explain why some patients develop AF and not others, reflecting regions of origin / drivers, structural abnormalities, neural components, and metabolic comorbidities (including obesity). These factors cannot be predicted by traditional classifications of 'paroxysmal' or 'persistent' AF. Using individual digital phenotypes, inferential methods including statistics and machine learning are used to compare the data with reference populations to infer optimal management.

[0134] Step 530 divides the data into classification methods. The division can be performed by many techniques known in the art, including but not limited to cluster analysis and other types of unsupervised learning or supervised learning methods (including support vector machines (SVM), logistic regression, naive Bayes, decision trees or other methods). This division is completed for the individual data (step 518) and in the parallel group data (step 538). It should be noted that the division technique used can be different for each step.

[0135] In step 530, cluster analysis (a known unsupervised learning technique) can be used to group unlabeled data (e.g., data streams from multiple sensors, input data, etc.) into sets of items that are "similar" to one cluster but different from other clusters. This can occur even when there are no obvious natural groupings, which is often the case in these applications because typical phenotypes rarely include clinical data, imaging, and continuous data streams. Clustering is a powerful tool in the present invention, but because the final cluster pattern depends on the initial clusters, any ambiguity in identifying the initial cluster pattern may lead to bias. The results of the clustering are later verified in step 558.

[0136] Step 546 generates a disease taxonomy from the population data using a mathematical model to integrate the data stream from the population and the stored data (database 508), data reduction schemes (step 534), and data partitioning (step 538). Data from the domain knowledge database 574 is incorporated to filter the mathematical relationships. For example, the mathematical weighting may be minimized for rare male breast cancer or for rare juvenile AF, while the mathematical weighting may be increased for common male aging with coronary artery disease. Such traditional domain knowledge may be obtained from epidemiological data and population statistics, and may also be available from stored population data and easily converted into mathematical weightings.

[0137] Step 550 generates a population numeric phenotype from the disease classification. In other words, the population numeric phenotype corresponds to partitions of the data that form self-consistent disease categories from the quantitative data. These can be clinically obvious or clinically ambiguous, for example, in some studies there is a link between low magnesium levels and atrial fibrillation. These partitions are each expressed statistically with confidence intervals and will be used for comparison with the individual numeric phenotype.

[0138] Step 522 uses the personal disease partitions from step 518 as a basis to generate a prototype personal digital phenotype. Step 542 compares the personal digital phenotypes (PDPs) to find the best matching population digital phenotype. In step 554, the candidate personal phenotypes are given by a vector matrix [P]. This vector matrix includes multiple data elements, data types, some ordinal elements, some vector elements, and some time correlation elements.

[0139] In a preferred embodiment, supervised learning is used to refine the digital phenotype to predict a defined outcome. This involves feature selection, choice of network architecture, and appropriate data for training and testing.

[0140] By intentionally generating sparse input "vectors", features are identified and "fine-tuned" for machine learning to avoid overfitting (i.e., poor generalization to future unseen inputs). The present invention eliminates redundant features and maximizes the diversity of input features to comprehensively cover the underlying input data.

[0141] In one embodiment, features are grouped into three types: (a) traditional clinical variables (demographics, comorbidities, biomarkers); (b) electrical signals (12-lead ECG and intracardiac signals, where signal processed parameters can separate AF phenotypes; and (c) imaging data, including but not limited to 2-D echocardiography images (atrial geometry), 3-D CT data (geometry), 3-D MRI data (fibrosis areas, geometry), and 3-D electroanatomical shells of voltage and complex electrogram distributions generated under EP studies. Clustering (unsupervised learning) will help with data simplification and serve as additional input features. To understand the importance of features, filtering and regularization are used. The present invention eliminates variables that are not associated with the response categories in training. One method uses the least absolute shrinkage and selection operator (LASSO) combined with filtering methods and wrapping methods to minimize prediction errors and include variables that contribute to regression analysis in the final model.

[0142] Missing data in each feature group will be handled by inserting (inputting) the following: a) median value; b) predicted value using multiple imputation (statistical technique); c) expected value for this data type from the literature; and d) constant representing missing data. Each method will be compared during the training of various network models.

[0143] Formatting of input images and signals. The present invention formats each 3-D MRI, CT or pseudo-color atrial electroanatomical image (representing anatomical structure, distribution of a specific type of electrogram, such as voltage) into a 3-D matrix. Time series data (12-lead ECG, bipolar coronary sinus electrogram, unipolar intracardiac electrogram) will be processed using feature extraction, cluster analysis and pre-processing networks before entering the final network.

[0144] The results used to train the phenotype will vary with each application. For a preferred embodiment of cardiac rhythm disorders, several results can be used to train the phenotype. One result can be a high voltage versus a low voltage (e.g. <0.1 mV) electrogram signal; a phenotype associated with a high voltage signal can have a higher therapeutic effect. Another potential result is the presence of clear AF spatial maps that show consistent rotors or focal sources / drivers; these sites can be effective therapeutic targets. Another desirable outcome is the termination of AF by drug therapy or ablation or long-term success with drug therapy or ablation. Both can be retrospectively determined in a reference population to form a numerical classification system and then used to identify matching individual phenotypes.

[0145] In one embodiment, supervised learning (often implemented as an artificial neural network ("ANN")) is used to represent various input data and data streams for individuals and groups. ANNs typically include three elements. First, the connection pattern between different layers of nodes (artificial neurons), forming a network with a variable number of layers, each network containing multiple nodes in each layer. The implementation can be as simple as a perceptron, an adaptive linear network, or many other designs (including deep learning networks). The actual network design can adapt to the complexity of the specific task and the data partitioning. Secondly, the connection weights between nodes can be changed and updated according to a variety of learning rules. Third, the activation function: determines how each weighted input is converted to its output. The activation function f(x) is typically a composite function of other functions g(x), which in turn can be expressed as a composite function of other functions. A nonlinear weighted sum can be used, i.e., f(x)=K(∑ i w i g i (x), where K (activation function) can be a sigmoid function, a hyperbolic function, or other functions.

[0146] Various connection patterns, weights, and mathematical activation functions may be selected, and various update functions are possible for any embodiment. Specific forms are optimal for different disease conditions and tasks. For example, a machine for detecting abnormal heart rate during known atrial fibrillation will be less complex than a machine for identifying ablation sites in atrial fibrillation, for predicting the onset of atrial fibrillation, for predicting the exacerbation of heart failure, or for predicting the onset of coronary ischemia.

[0147] Alternative forms of learning include supervised and unsupervised methods including linear logistic regression, support vector machines, decision trees in "if-then-else" statements, random forests, and k-nearest neighbor analysis. Such formulations can be applied independently or as part of machine learning to increase or generate boundaries between desired decisions such as the presence or absence of a source of atrial fibrillation or other associations linking input data to clinical or physiological outcomes for an individual. Several other forms of machine learning can be applied and will be apparent to those skilled in the art.

[0148] Various connection patterns, weightings, node activation functions, and update schemes can be selected, and depending on the data input, a specific form is optimal for different applications. For example, imaging input and continuous data sequences (e.g., electrogram signals) can be represented by different networks optimized using a large amount of training data in a preferred embodiment as each data set in a given reference population. Therefore, depending on the application, the present invention can be customized to best represent EEG data, cardiac and respiratory signature features, body weight, skin impedance, respiratory rate, and cardiac output. Recurrent neural networks are data structures that enable analysis of how networks achieve their trained conclusions. Fully connected layers can be used to merge manually engineered scalar features (e.g., clinical data elements). Characterized time series (i.e., 12-lead ECGs or so-called 'electroglymograms' from inside the heart) are processed by convolutional neural networks. Standard techniques for loss, batch normalization, and hyperparameter fine-tuning are used to avoid overfitting.

[0149] An important feature of machine learning approaches is that they do not require prior knowledge of the details of human pathophysiology, but rather learn patterns in sensory signals and input data in health and deviations in disease. Therefore, these machine learning approaches are well suited to augment current disease classifications based on current mechanistic hypotheses that may be suboptimal.

[0150] Step 558 determines which candidate phenotype(s) can be validated, i.e., which candidate phenotype predicts any robust outcome measure. For AF, this candidate phenotype may be the site of ablation terminating AF. For coronary artery disease, this candidate phenotype may be a clinical constellation predicting severe stenosis of the epicardial coronary arteries, i.e., a late coronary risk score. By extension, such candidate phenotypes may be defined for non-cardiac diseases. If a match is not achieved, the process is repeated 522 or the acceptable tolerance X is expanded (box 562). If a match is achieved within an acceptable tolerance (vector X), then in step 566 the candidate phenotype becomes a personal digital phenotype P within tolerance X at time T. The phenotype is then used to update the individual's personal history data in database 570 for the results of the marking used to validate the phenotype. This step is used to validate the clusters defined in the previous step and for supervised training.

[0151] Figure 6 Health and disease are defined using personal digital phenotypes. Step 600 obtains each personal digital phenotype at each time T in step 600 and checks the signal type in the phenotype for each personal digital phenotype 603. Mathematical and network analysis 606 is used to identify abnormalities, which are first compared with personal phenotypes stored in database 609 (i.e., data determined as results from previous times, including when individuals feel well or unwell (symptoms) or have objective evidence of disease presence (e.g., AF attack or not) or disease worsening (e.g., progression from paroxysmal AF to persistent AF)). Step 612 generates a portrayal of personal health or disease based on this analysis.

[0152] Now, step 615 determines if this is healthy for the individual within accepted tolerances 618. If this is "out of range" for the individual's health, then the individual may have entered a potential disease condition 621.

[0153] Mathematical and network analysis 624 is performed and compared to the population disease classification 627 to determine if the individual's abnormality also falls within the "outside healthy range" of the population. Based on the comparison to the population fixed and variable definition of statistically defined abnormalities (step 630), the present invention now asks if a disease is present in step 633. If "yes", the disease is declared 639; if not, the patient continues to be under careful monitoring in an observation step 636. In either case, the process repeats for continued monitoring.

[0154] Figure 7A novel interpretability analysis of supervised networks is outlined, i.e., identifying which signal type and input type contributes most to an abnormal personal digital phenotype at which point in time. Interpretability is the main novel feature of the method of the present invention, addressing the main traditional criticism that machine learning is a "black box". Expert knowledge can be used to characterize massive data (i.e., enhanced by adding important data without overwhelming the network with "dimensions"). Domain knowledge is also used to assist manual interpretation. Step 700 analyzes all signals that are part of the personal phenotype in turn. Data elements 703, 706...709 consider each data element (input data or data stream) in turn. In step 715, several methods are used to interpret the classification performed by the neural network 712, i.e., which (which) components cause the classification of 'disease' and 'health'.

[0155] Several interpretation techniques can be used, and these interpretation techniques will be familiar to those skilled in the art. The first method includes using an attention layer in a recurrent neural network. Alternatively, the present invention can use local interpretable model agnostic explanations (LIME) to explain predictions by approximating an interpretable model. LIME can be used for one-dimensional data, such as ECG or electrical signals (electrograms) within the heart, digital features, or images. A third method is gradient weighted class activation mapping (Grad-CAM), which identifies the most critical nodes as the maximum weight multiplied by the back-propagation pooling gradient downstream of the final convolutional layer. Another embodiment specifies features that should or should not be part of the model, including spatial domains in the image (e.g., the size of the atrial driver region or the spatial amplitude of ventricular conduction velocity or human heart fibrosis), thereby enabling customized interpretations for domain electrophysiology "concepts" to ensure that the model does not converge on unrelated concepts. An example of this is the method of testing with concept activation vectors (TCAV). This method can check specific features that should or should not be part of the model (e.g., the size of the AF driver region), thereby enabling the present invention technology to customize the interpretability of the accepted "concepts". As another example, predictions of AF outcomes (e.g., ablation success or failure) can be tested with interpretable models (e.g., presence of fibrosis near the right atrium). This approach attempts to ensure that the numerical model is relevant to the prediction and that the model does not converge on unrelated concepts. Thus, interpretable features of the predicted outcome will be quantitatively identified. Clinical rationales can then be added 574 with domain knowledge, for example, determining that obesity predicts a negative outcome of ablation or drug therapy, while hair color predicts that a positive outcome may not occur. Data on populations that can use Class IC antiarrhythmic drugs (AADs) can also be included.

[0156] A key feature of the digital taxonomy is that it encodes inconsistent situations (i.e., where the neural network cannot predict the actual outcome). For example, in a patient who has failed ablation therapy (the patient's profile includes atrial scar on MRI), the present invention would be trained to associate the location of the scar with the location of the ablation lesion and the outcome. One potential output from a trained network could be that ablation that missed the scar area is likely to produce a poor outcome. Domain knowledge (physiological interpretation) is used to provide rationality for any trained network to mathematically ensure that unreliable (or impossible) connections are not constructed and the network is corrected accordingly. This combined mechanistic / machine learning approach is a novel advantage of the present invention that is often overlooked in machine learning systems that do not check data representations against known domain knowledge. Well-known errors in existing networks include adversarial examples; applications in image recognition where changing one pixel can change the classification from 'cat' to 'dog'. The present invention prevents such trained networks from evolving in this medical space where errors must be minimized.

[0157] Therefore, the present invention focuses on developing, testing and revising increasingly interpretable data structures. The model combines statistical analysis with expert interpretation of case failure / success. Simple statistical tests and linear models can help identify the correlation between different variables in the system, but may not capture the potential complexity and nonlinearity of these complex clinical problems studied. Decision trees (CART) can allow for greater interpretability of the importance of each extracted feature from the layers of the network. The input to the decision tree will be features extracted from images and time series signals.

[0158] Another approach in the present invention is called "network clamping." In step 715, starting with a trained baseline "healthy" version of the network, the inputs are shuffled individually or in batches, and the network 712 is rerun to identify which abnormal input combination causes the network to most closely recapitulate the "disease condition."

[0159] These steps are evaluated in step 718 to identify a constellation of inputs that characterize health, disease, and outcome conditions in step 721. This result is then used to update the personal history data in the database 724.

[0160] Figure 8 An example of using an abnormal personal digital phenotype to treat AF is provided. As previously described, step 800 includes personal data and step 803 includes population data. Step 806 finds the best match between the personal AF phenotype and the digital taxonomy. Step 809 uses the data from Figure 7 The key element analysis (step 718) is used to identify markers of the AF process in the individual. Steps 812 to 824 identify different AF features that can be identified by mapping. These steps refer to Figures 9 to 11Step 830 updates the personal history data 724 in the database.

[0161] Step 812 identifies patient phenotypes that exhibit electrical patterns of intact rotation (rotors) or focal origin without structural abnormalities. In this case, treatment 827 may include ablation of the driving source.

[0162] Step 815 identifies a patient phenotype that exhibits electrically intact rotations (rotors) or focal sources (associated near structural abnormalities marked by low voltage or potentially by abnormalities on delayed enhancement MRI. In this case, treatment 827 may include ablation of the driver source(s), with the ablation extended to areas of scar or fibrosis that are non-conducting.

[0163] Step 818 identifies a patient phenotype with structural abnormalities marked by potential abnormalities on low voltage or delayed enhancement magnetic resonance imaging, but may not show classic electrical rotations (rotors) or focal sources. In this case, the representative network will find other electrical patterns in the patient that are targets for therapy, such as partial rotations or repetitive excitations. Such atypical patterns may be useful for patients with specific comorbidities, such as those consistent with advanced disease. In this case, treatment 827 may include ablation at these (multiple) atypical power sources, with the lesion extended to areas of scar or fibrosis that are non-conductive.

[0164] Step 821 provides options for disease prediction. In this step, the present technology identifies patient phenotypes that do not exhibit AF but may be at risk due to specific patterns of structural abnormalities marked by underlying abnormalities on low voltage or delayed enhancement magnetic resonance imaging. In this case, the present invention provides AF prediction. In this case, ideal input data may include granular imaging data showing MRI abnormalities or granular data on low voltage areas (for enabling the network to non-invasively detect structural risk maps to provide prognosis or potentially provide targets for therapy). Treatment 827 may include ablation for connecting these areas of scar or fibrosis.

[0165] Step 824 is another option for prediction. The present invention identifies patient phenotypes that do not exhibit AF, but may be at risk due to specific electrical patterns marked by electrical propagation that changes at varying rates (thus indicating rate-dependent block) or oscillations in conduction to specific areas indicating refractory or alternating conduction. In this case, ideal input data may include body surface potential signals for enabling the network to non-invasively detect electrical risk maps. In this case, treatment 827 may include ablation to eliminate these areas of electrical vulnerability.

[0166] Fig. 9Examples of complex arrhythmias are shown, describing how sample personalized phenotypes can be used to customize therapy for each of the cardiac rhythm disorders. Unorganized excitation can have many effects on the source. On the left side of the figure, diagram 900 shows unorganized excitation emanating from the source. This behavior is often referred to as "fibrillatory conduction" from a rotor, focal source, or other driving mechanism. In this case, the present invention can use one of several mapping systems (including freely available mapping systems) to determine whether the pattern of predominantly rotational or focal excitation drives AF (i.e., within the core indicated by the sequence 1, 2, 3, 4, 5). Alternatively, in diagram 910 on the right, the unorganized excitation can surround the source ("external unorganization"), but be directed inward (such as from a second source). If this external disorganization does not perturb the driver, the present invention is able to show that rotational or focal excitation (again, sequence 1, 2, 3, 4, 5) can continue to drive AF even if periodically interrupted by the peripheral disturbance (as shown by human optical mapping by Lin, Fedorov et al. [Circulation] 2016 and clinically by Kowalewski et al., [Circulation Arrhythmias / Electrophysiology] 2018). A second source can be a rotor or focal source of a complex rhythm where there is disorganization at an interface known as a "collision" or "fusion." Third, disorganized excitation can cluster near the site of the anatomical abnormality 930 without obvious rotational or focal activity. Here, (multiple) structural obstructions 930 can maintain the rhythm disorder. The present method can determine that a specific element, structural element, or excitation of its interaction showing partial rotational or partial focal excitation can be a driver. Further, the present scheme can determine that the area of ​​interest controlling a larger area of ​​the atrium is a productive area for therapy. This determination is made for each individual based on detailed cardiac mapping using machine learning and other processing elements involved in personalized digital phenotyping.

[0167] FIG. 10A to FIG. 10C are examples of how therapy can be tailored for personalized digital phenotypes for dysrhythmias, each showing a personalized phenotype that includes an electrical driver. For simplicity, the dysrhythmias are shown in each case as rotating, but can be focal or another pattern (e.g., repetitive, partially focal, or partially rotating). Fig. 10A In, interventions such as pacing (1005, 1010) can modify the key driver by limiting the spatial location of the key driver, blocking at electrical and structural sites to limit the movement of the fibrillating core while the key driver continues to operate. The source can precess in a limited spatial region 1020, thereby stabilizing the source. Therapy can now be customized to limit source precession so that the source is fixed, thereby converting a complex rhythm disorder to a simple rhythm disorder, such as a focal atrial tachycardia that is easier to treat. In Fig. 10B In, therapy is tailored to intentionally change the spatial location of a source, thereby pushing the key driver to a boundary or region where it can no longer be sustained (e.g., an inert region of the heart or even outside the heart), therefore destabilizing the locus. Pacing 1025 passes through areas of functional block (electrical) or anatomical block (structural) that protect the key driver from surrounding collision waves. Incoming pacing waves cause the source to migrate 1035 to a site that may be more amenable to treatment or encounter an obstruction and terminate. In Fig. 10C In the pacing 1040 passes through areas of functional block (electrical) or anatomical block (structural) that protect the key drivers from the influence of the surrounding waves and thus the intrusive source, thus causing their termination. This termination can be permanent or temporary.

[0168] Fig.11 Example operations for cardiac arrhythmias are shown. This preferred embodiment analyzes a raw voltage-time data stream from an intracardiac electrogram or ECG detected by a wearable or implantable device or during an electrophysiology study to guide therapy. Electrograms for simple arrhythmias 1100 and complex arrhythmias 1105 are shown. As examples, the cardiac rhythm disorder can be a complex rhythm disorder AF, VF, and polymorphic VT or another cardiac rhythm disorder.

[0169] In this example, the signals in 1100 and 1105 show deflections 1103 that represent the onset of excitation (e.g., for a heartbeat) and deflections that represent noise 1108. In AF, multiple colliding waves mean that any one electrode can detect local activity plus activity or noise from other sites. The main challenge is to identify AF deflections that represent "local" cardiac excitation to that site. This information is then used to generate an AF map that can point to areas of organized rotation or focal excitation that can drive surrounding disordered activity. However, if true excitation is detected by a different method, the map may obscure these features. Some signals do not show an identifiable onset of excitation (e.g., for a heartbeat), such as noise 1108. Such signals include multiple deflections of short duration caused by cardiac rhythm disturbances that can make it difficult to discern the onset of excitation (depolarization) and have little clear ground truth. Traditional methods use features such as sharp inflection points and high depolarization slopes to indicate the true location of cardiac activation, while the duration of the gentle, low-deviation repolarization slope is typically 100 to 250 milliseconds. This is addressed below by the inventive method, which has been trained using ground truth values ​​of activation time and repolarization time from monophasic action potential recordings.

[0170] At step 1120, the real-time data stream forms the input to the system with its time stamp. In step 1123, other non-streaming data for this individual is input. Step 1126 generates initial data partitions or data vectors given by the matrix P, including different data types, values ​​(in this case, values ​​of electrogram data). The partitions from step 1126 are linked to the results from the individual data or group data. The results can be subjective (symptoms) or objective (acute termination by ablation or chronic termination for long-term freedom from AF). This is different from the prior art.

[0171] Steps 1129 to 1147 correspond to individual phenotypes to population phenotypes (ie, numerical taxonomies). Step 1129 scans all individual phenotypes and all population taxonomies (phenotypes) input from step 1132.

[0172] Step 1138 determines that the individual digital phenotype matches the population digital phenotype within the acceptance limits defined for the application (electrogram analysis) and the phenotype-tuned vector X. Figure 5 The matching is performed by the corresponding step of step 542 in . Several methods can be used, but the preferred embodiment uses machine learning. If a match occurs, step 1144 defines a therapeutic target from the data of the reference digital taxonomy and database 1147.

[0173] Phenotypes that may be candidates for AF ablation targets include learned mathematical combinations of electrogram features plus clinical comorbidities (e.g., high body mass index, diabetes, hypertension), demographic characteristics (e.g., age, sex, prior ablation or not), and (if available) genetic, metabolic, and biomarker information.

[0174] Applications of the present invention use machine learning to improve identification of local electrical activation to produce better AF maps and use a personal digital phenotype to identify features other than rotational or focal activity that may be critical to that person's AF. The specific digital taxonomy for this application includes electrograms at sites where ablation terminated and did not terminate AF. These data are used as input-output pairs to compare maps produced by various methods. In competitive machine learning, methods that produce maps that explain the site of AF termination (e.g., showing focal or rotational sources or low voltages) are weighted more highly. These electrogram features and optimal mapping can be different for patients of different atrial size, wall thickness (and therefore voltage), and degree of scarring. This can help distinguish local activation from far-field activation in elements 1103 and 1108.

[0175] Potential electrogram targets include low voltage, rotating electrogram patterns (exemplified by pattern 1205 in FIG. 12 ), focal activation patterns (i.e., sites of consistent centrifugal firing), partial rotating patterns, partial focal patterns, high dominant frequency sites, high signal complexity sites, and each of these sites that can control large areas of the atrium in AF (supporting a causal role). If consistent, each site can indicate an area that is critical to maintaining the rhythm disorder. The present invention combines these electrogram features with clinical data into a personal electrogram profile.

[0176] By analyzing beyond 'traditional' targets, individual digital phenotypes can be used to identify novel electrogram targets. For example, studies have shown that targets such as repetitive patterns or transient rotations or focal patterns or intermittent rotations or focal patterns may be critical for maintaining arrhythmias in some individuals. This preferred embodiment of the invention defines these electrogram features by determining which aspects of an individual patient may be associated with a favorable outcome. This then becomes a numerical classification within a digital taxonomy as data from more individuals is labeled and accumulated.

[0177] Still refer to Fig.11 , steps 1150 to 1159 provide some therapy options based on the personalized electrogram-based phenotype. Therapy option 1150 identifies tissue areas that are directed to a clear electrical target. Depending on the patient, this electrical target can be a rotational or focal source / driver or other electrical feature, regardless of its structure. Option 1153 targets tissue areas that are directed to structural abnormalities. Intermediate phenotypes can exist in the phenotype of a specific individual (electrical and structural, which can change dynamically with changes in, for example, health status). In addition, multiple forms of electrical patterns can be co-located with this structural element, and the present invention will store electrical signals associated with these sites to update personal and population databases (1147). In both cases, option 1155 provides therapies including destroying tissue by surgery or minimally invasive ablation to regulate tissue by electrical pacing or mechanical pacing or gene therapy or stem cell therapy or drug therapy. Medication can include Class I agents for reducing atrial conduction velocity or Class III agents for delaying refractory. AF ablation can not only eliminate tissue, but also target areas adjacent to fibrosis or areas of electrical vulnerability. Therapy can also target the associated tissues of these areas, their nerve supply or other regulatory biological systems.

[0178] Option 1159 provides other AF therapy options, including lifestyle changes (such as weight loss, control of ventricular afterload, or other factors). This personalized approach can improve outcomes, eliminate repeat interventions by selecting the strategy with the highest expected initial success rate for a given individual, and thus reduce costs and delays in treatment.

[0179] Fig. 12AAnother example of a preferred embodiment of arrhythmia is shown, which is presented for AF maps that are increasingly available from freely available sources and proprietary systems. The AF map 1200 is colored (shaded) and annotated to indicate features in regions 1206 and 1207. Several features can be key drivers, including timing, shape, and complexity. This includes rotational circuits (shown) that can be identified using a rotational activation time series 1205 of sites 1 to 5 in the putative circuit, but can also be identified using phase angle information represented as θ. Other novel features include machine-learned partial rotation, partial focality, and repetitive spatial features. The maps can be generated and used as input using software available from commercial vendors, from freely available online mapping tools (Alhusseini, 2017), or other new systems. These maps can be generated by electrophysiological recorders, dedicated mapping equipment, online resources, smartphone apps (e.g., see Fig.14E ) is generated and streamed to the system as input 1220. Fig.11 In step 1223, personalized input data such as age, gender, clinical, imaging, genetics can be entered.

[0180] Step 1226 generates data partitions based on the AF map. An advantage of the present invention is the ability to use data that extends beyond the cardiac rhythm to identify one or more sources of complex rhythm disorders in an individual. Complex cardiac rhythm disorders may appear chaotic, but may be caused by excitations from a small number (1 to 5) of sources. Sources may include rotating circuits (rotors), from which waves (usually spiral waves) cause disorganized excitations. Sources may also include focal impulse areas (focal sources), from which excitations centrifugally emitted cause disorganized excitations. The present invention scheme identifies those circuit types along with additional understudied targets, such as repetitive patterns or transient rotation or focal patterns or interrupted rotation or focal patterns, which may be valuable in some individuals.

[0181] A key problem in the prior art is that some patients' AF may be closely associated with rotational or focal sites, other individuals' AF may show a less strong relationship, and still other patients' AF may have completely different mechanisms. There are also cases where AF is the "ultimate common pathway" of multiple comorbidities or diseases, which may represent some of the different types of AF. Currently, there is no way to separate these types of AF. The method of the present invention extends the prior art by adding detailed clinical data to electrical data to produce a personal digital phenotype. The detailed clinical data includes data about the mechanical and other functions of the heart, structural abnormalities of the heart from imaging such as magnetic resonance, and clinical parameters (including age, sex, demographic characteristics, family history, comorbidities (such as obesity), the presence of diabetes, genetics and metabolomics profile indicators). For example, in extremely obese patients, AF may not be driven by rotational or focal drivers, but rather rely more on structural factors. The classification method will take these clinical factors into account, generate a phenotype, and individualize the therapy accordingly. In another patient with different and / or less advanced clinical factors, AF may be driven from specific regions (focal or rotational), and these sites may be primarily located near the pulmonary veins.

[0182] Step 1229 compares the phenotype centered on the AF map with the population numerical (Bayesian) classification method 1232. Step 1238 uses the acceptance tolerance limits defined by the vector X in step 1241 using, for example, Figure 5 The comparison is performed using methods such as machine learning described in step 542 of . Step 1244 defines a therapy target through this AF phenotype based on comparison with a numerical taxonomy and data from database 1247 .

[0183] Steps 1250 to 1259 represent therapy options based on the phenotype focused on by the individual AF map. Option 1250 identifies tissue regions that subtend a clear electrical target. Depending on the patient, this electrical target may be a clear rotational or focal source / driver or other electrical feature, regardless of the structure. Option 1253 targets tissue regions that subtend structural abnormalities. Any form of electrical fingerprint may exist at this structural element, and the present invention will store electrical signals associated with these sites to update the individual and population database 1247. In both cases, therapy options 1259 include treatments such as destruction of tissue by surgery or minimally invasive ablation, modulation of tissue by electrical or mechanical pacing, gene therapy or stem cell therapy, or drug therapy. Medication may include Class I agents for reducing atrial conduction velocity or Class III agents for delaying refractory. AF ablation may be used not only to eliminate tissue, but also to target areas adjacent to fibrosis or areas with electrical vulnerability. Therapy may also be directed to the associated tissues of these areas, their nerve supply, or other regulatory biological systems.

[0184] Therapy options 1256 include other AF therapies, such as lifestyle changes (e.g., weight loss, control of ventricular afterload, and other factors). This option may be particularly effective in patients with advanced AF (e.g., patients without clear AF drivers / sources on the AF map). This option may be implemented by non-invasively determining the map from surface ECG information (step 1220). Overall, this personalized approach provides a mechanism for improving outcomes and eliminating / minimizing repeat interventions by selecting the strategy with the highest expected initial success rate for a given individual, and thereby reducing the cost and delay of treatment.

[0185] The present invention protocol can also be used to integrate anti-inflammatory agents and immunosuppression as a novel arrhythmia therapy for certain digital phenotypes at therapy option 1256. Anti-inflammatory therapy can include steroidal or non-steroidal agents, immunosuppression using agents for transplantation, and cell therapy.

[0186] FIG. 12B to FIG. 12D The utility of the inventive approach for identifying (mapping) the causes of cardiac rhythm disturbances personalized to an individual (eg, for guiding ablation therapy) is demonstrated. Fig. 12B This demonstrates an unmet need arising from the fact that clinical electrograms in complex rhythm disorders can be ambiguous and rarely provide the detail required to establish cause (mechanism) in experimental models (eg, animals). Fig. 12C An example is provided of how the present method meets this need by analyzing clinical electrograms to reconstruct signal types personalized to an individual that have previously only been seen in experimental models. Fig.12D It is shown how a customized map of arrhythmias can be generated to personalize therapy for each individual. FIG. 12B to FIG. 12D The method of the invention is demonstrated by the example of atrial fibrillation (AF), however this example is not intended to limit the scope of the invention and the example may be applied to other rhythmic disorders of the atria or ventricles or to electrical disorders in other organ systems such as the brain.

[0187] refer to Fig. 12B, the sample clinical AF signal 1265 includes components (i.e., spurious deflections) that may not represent excitation of nearby (local) tissue because they occur faster than the cells can recover (repolarization is typically >120ms to 200ms (from Narayan et al., [Journal of the American College of Cardiology] 2008).) It should be noted that this is true even for small (e.g., 0.3mm) electrodes that are in direct contact with the human heart (Allessie et al., [Circulation] 2010) and may be even more ambiguous for larger electrodes that experts use when treating patients (2mm to 4mm). It is not clear to the expert which deflection from each AF "beat" to select as excitation (depolarization) because the recovery (repolarization) is unknown. Sub-figure 1267 shows that selecting any of three acceptable deflections within the clinical signal (based on an expert-type rule: minimum of the first derivative (dV / dt)) produces very different AF maps from the same recording (Zaman et al., Circulation Arrhythmia and Electrophysiology, 2018). Sub-figure 1269 demonstrates that some or none of these clinically labeled signals can actually coincide with the reference signal (the action potential from the monophasic action potential (MAP) recordings in these patients), which have been shown to accurately represent depolarization (MAP onset, "O") and repolarization (MAP end, "E"). Several deflections of the clinical AF electrogram do not coincide with the onset or end of the reference and are labeled "far field" (i.e., not local, "F").

[0188] Fig. 12CAn example is provided of how the inventive method reconstructs a reference (action potential, MAP output) in AF from an ambiguous clinical input signal. Multiple deflections in the clinical signal are automatically labeled by the inventive method for MAP onset 1275 (i.e., excitation, depolarization) (where the vertical solid line represents "prediction" and the vertical dashed line represents "correct") and MAP termination 1277 (i.e., recovery, repolarization). Each sub-plot shows actual data in AF (i.e., reduced to practice) in a blind test set after training this component of the inventive scheme on independently labeled training data, with the input from the clinical AF electrogram at the bottom of each graph and the MAP at the top. The labeling of MAP onset 1275 and termination 1277 is deterministic according to standard formulas. As can be seen along the lower boundary of the onset sub-plot 1275, the error in the reconstruction of the labeled depolarizations and repolarizations is very small (7.2ms to 16.5ms), thus falling well within the labeling error of a human expert. Plot 1279 shows the training accuracy of the machine learning algorithm using 16,560 EGM / MAP samples, where the algorithm converged to nearly 100% accuracy. Machine learning allows customization not only to the clinical signal, but also to personalize for the following: patient type (male / female, different ages, body mass index, comorbidities, etc.), location of signal recording (right atrium, left atrium, near pulmonary veins, left atrial appendage, left atrial posterior wall and other areas), and other tissue characteristics (e.g., the presence of fibrosis, which can change the signal characteristics relative to the action potential of the tissue).

[0189] Fig.12D An embodiment of the personalized mapping scheme of the present invention using input clinical signals for a patient's AF is illustrated. The clinical signal 1281 may be generated using a single electrode such as an ablation or mapping catheter 1282 or multiple electrodes in a circular mapping catheter 1283, a basket mapping catheter 1284, or other shaped mapping catheters. The mapping scheme applies a machine learning algorithm 1285 (or other classifier as known in the art) to reconstruct a signal of greater detail and interpretability (e.g., Fig. 12C). This processing produces arrhythmia maps 1290 that may be different from those that would be obtained using only clinical signals or analyzed by traditional expert-based rules. These maps are referred to as 'machine learning MAP maps' or 'virtual MAP maps' from clinically recorded catheters. Specifically, the resulting arrhythmia maps may show that some patients have one or more organized driver factor regions. These regions may be dominant, as evidenced by their control over large areas of the heart (1291), or may be secondary, as evidenced by their control over smaller areas of the heart (1292). Some patients may have no driver factor regions that are evident at the time and have very disorganized AF (1293). These arrhythmia maps can be used to guide therapies, such as ablation, that are personalized for that individual. This analysis can also be performed repeatedly to determine whether complex arrhythmia maps change over time or to identify secondary driver factor regions that may appear, disappear, and reappear.

[0190] The ability of the method of the present invention to reconstruct action potential mapping in the human heart is unique. This method approximates the process of optical imaging of an explanted human heart in a living patient, which is the gold standard for mapping action potentials to avoid blurring of clinical signals (Hansen, Fedorov et al., JACC Clin. Electrophysiol. [JACC Clinical Electrophysiology] 2018), but this is not applicable to living patients. The method of the present invention allows the identification of organized areas with unorganized rhythms. Tissue regions with larger sizes can be more dominant than tissue regions with smaller sizes and can be better therapeutic targets. In existing methods that do not reconstruct action potentials, the ability to identify such areas is not available.

[0191] Fig.12E The diagram shows how the inventive method can generate a "virtual signal basket catheter" 1297 from an ambiguous clinical input signal 1295. In this example, the input signal 1295 is a unipolar electrogram from an input basket probe, which is input into a machine learning algorithm 1296. In 1298, the reconstructed waveform is an action potential, which is a familiar physiological signal that the present invention has already reconstructed for multiple rhythms (AF, VT, VF; e.g., Fig. 12C) these action potentials were trained on labeled reference training data. However, other bioengineered signal types can be generated that are optimized for specific tasks. The square wave in 1298 shows local depolarization blanking until repolarization, which can be suitable for excitation mapping (minimizing stray deflections). The gradient (triangular, sawtooth) signal in 1298 shows a step-by-step gradient from onset to repolarization in amplitude and can be useful for phase mapping. Other shapes can be engineered and optimized for mapping methods using known excitation, recovery, and rate information. Each signal type can be customized by the present invention for individual characteristics, such as lower amplitudes near areas of scar or fibrosis (shown) or higher amplitudes near thick tissue areas (such as the mitral isthmus in the left atrium), etc. Several other signal types can be reconstructed for various applications and should be apparent to those skilled in the art.

[0192] The inventive approach differs from all existing mapping methods for complex arrhythmias that do not depict repolarization as well as depolarization and therefore cannot identify which signals may be far-field (false positives) because they fall within the recovery (repolarization) time (as explained by Narayan et al., Heart Rhythm 2016). Although the embodiments described and illustrated herein employ machine learning and neural networks, it will be apparent to those skilled in the art that other classifiers, including multiple parallel algorithms, may be applied as described.

[0193] Fig.13A A sample diagram of a machine learning algorithm is shown. In this case, multiple inputs 1300 are provided through an input layer 1302. Each input will vary depending on the data type. For example, the electrogram stream can take the form of a one-dimensional array of voltage-time representing multiple channels. The imaging data can be represented as a matrix representing the intensity of voxels (floating point integers) at each coordinate in three-dimensional space or pixels in two-dimensional space. An ordinal value can be used to represent a simple variable male / female for age.

[0194] Connections link each input to one or more nodes (also referred to as 'neurons' or 'elements') in hidden layer 1 (1304), hidden layer 2 (1306) to hidden layer n-1 (1308) and hidden layer n (1310). In general, hidden layers 1312 may include a variable number that may vary with the complexity of the input data to this system. Hidden layers 1312 may be any serial combination of layers, including but not limited to convolutional layers, pooling layers, locally connected layers, recursive layers, embedding layers, merging layers, activation layers, normalization layers, noise layers, dropout layers, dense layers, flattening layers, shaping layers, permutation layers, repeated vector layers, regularization layers, and masking layers. The implementation of each layer or group of layers within the network may be different from another layer or group of layers. Output 1314 may include one or more nodes that encode a classification task. For some applications, the output may be a binary (integer) value, such as "atrial fibrillation diagnosis", "yes" or "no". In other cases, the output may be represented by a spectrum or range given by a continuous variable, such as "presence of an atrial fibrillation source / driver" where 0.0 would be absent, 1.0 would be continuously present in the region, and 0.7 might be intermittent but present a certain percentage of the time.

[0195] exist Fig. 13BIn FIG. 12 , steps 1320 to 1347 show another sample embodiment modified from AlexNet (convolutional neural network) that can input the step of image input (a preferred embodiment for the AF map analysis in FIG. 12 ). The steps cycle from image input, convolution, rectified linear unit (ReLU), cross-channel normalization, maximum pooling, maximum and average pooling layers (shown), fully connected layers, and dropout layers, which set the input elements of the layer to zero with a given probability. For classification problems, a Softmax layer or a binary cross entropy layer must be located after the final fully connected layer. Step 1320 represents the input to the network (e.g., an image). Steps 1323 to 1326 represent two consecutive blocks of convolution (Conv), activation (ReLU), batch normalization (BN), and maximum pooling (MP) CNN layers. Steps 1329 to 1335 represent three consecutive Conv and ReLU blocks. Steps 1338 to 1341 represent two consecutive blocks of fully connected (FC), ReLU, and dropout (DP) blocks. 1344 represents an FC block, and 1347 represents a binary cross entropy output classification layer. The convolutional neural network (CNN) above is an illustration of an AlexNet implementation in Python. Other networks can be used on different computing platforms, including but not limited to Inception-v3, ResNet-50, ResNet-101, GoogLeNet, VGG-16, VGG-19, and DenseNet. A specific network design can be implemented to merge all desired data. Other network architectures can be used, and the selection of these network architectures will be obvious to those skilled in the art.

[0196] FIG. 14A to FIG. 14E Provides sample results processed using the method of the present invention, such as simplified use Fig.13A and Fig. 13B The embodiment of the present invention is used to practice that. Fig.14A An AF map generated by phase analysis is shown, segmented into sub-regions exhibiting rotation and sub-regions without rotation. Fig. 14B is a plot of the trained behavior of a CNN with very high accuracy for detecting critical regions of AF, where localized ablation terminated persistent AF prior to pulmonary vein isolation. These data show results with >90% accuracy on test data (i.e., data not used to train the convolutional neural network).

[0197] Fig. 14C Unique aspects of the inventive approach for interpreting machine learning in arrhythmia diseases are demonstrated. Since lack of interpretability is considered a limitation of most existing machine learning applications, the inventive approach has been designed to provide interpretability for its learned patterns. Fig. 14CIt is illustrated how the trained network achieves its classification. Several methods can be used and will be familiar to those skilled in the art. The process shown is the gradient-weighted class activation mapping (Grad-CAM) designed by RR. Selvaraju et al. merged the class-conditional properties of CAM with existing pixel-space gradient visualization techniques to highlight fine details on the image. Grad-CAM can be applied to various CNN model families (including CNNs with fully connected layers (e.g., VGG19), CNNs for structured outputs (e.g., subtitle shorthand), CNNs for tasks with multimodal inputs (e.g., VQA) or reinforcement learning) without architecture changes or retraining.

[0198] exist Fig. 14C , the input map 1405 is an image showing rotational excitation in AF, which was acquired using a standard mapping method. The data from the input map 1405 is transformed for machine learning, and in step 1407, a forward propagation of the input vector is performed through a trained CNN (Conv 5+ReLU) to generate a predicted label (1 / 0), such as "rotation exists". In step 1409, a backpropagation algorithm is performed to calculate the gradient of the output label (1 / 0) relative to each convolutional layer, each of these gradients representing a neuron importance weight (m,n,k) of size equal to the convolutional layer. In step 1411, the values ​​are averaged along m and n to obtain a vector of size (k). In step 1413, the dot product between the average gradient vector and the output of the convolutional layer along the third dimension (the k dimension) results in an m×n matrix 1415, where higher values ​​1417 represent more important areas than lower values ​​1418. (For ease of illustration, only the convolutional layers are shown.) In heat map 1415, the warmer colors (at the upper end of the color spectrum, labeled "red," "yellow," "green," "blue" in color scale 1419) indicate the regions in the trained network that are most responsible for the classification. In other words, these warmer colors indicate how the network makes its decisions.

[0199] Fig.14D Provided Fig. 14C1420 shows a single region of interest (rotational activity) in the AF of a 64 year old male. The trained network is able to detect this region, and the resulting heat map 1422 accurately shows the location 1423 in the heart that would be coded by the expert as a region of rotational activity. From the input image 1425 showing AF in a 74 year old male, the heat map 1426 indicates two concurrent regions of interest 1427 that were automatically identified by the Grad-CAM processing, showing excellent agreement with the region that the expert would select. Input image 1430 shows AF in a 65 year old male: three regions of interest are shown and detected by Grad-CAM, with a heat map 1431 accurately indicating each region 1432 that would be similarly labeled by the expert.

[0200] Previous "interpretability" analysis of the network showed that the trained network, without being explicitly trained on the pathophysiology (causes) of AF, could make its successful classification decisions by empirically determining the features that an expert would visually select. Overall, the method identified features that an expert would select with >90% sensitivity and >90% specificity in a series of 35 patients and 175,000 images.

[0201] Different interpretability analyses can be used, and these interpretability analyses will be obvious to those skilled in the art. Grad-CAM does not require modification of the trained CNN and can be automated, however, other methods can be used. One method is to extract features from different layers, visualize these features as images, and visually analyze specific features of these images. Sensitivity analysis has been proposed, in which certain parts of the input are systematically omitted or grayed out, and the importance of these parts is inferred by changes in classification. Such methods may be limited because it is difficult to detect multiple concurrent areas at the same time, because omitting certain parts of the input may produce artifact features for classification, and changing the representation may fundamentally change the confidence of the network prediction. In different interpretability analyses, the network can identify features that are different from those identified by the expert. This can be an advantage or disadvantage for each embodiment of the present invention.

[0202] Fig.14EA sample user interface display of data on a specially designed smartphone app is shown. Display panel 1440 shows the termination site of personal data digitally acquired from the imaging / mapping system. This example display shows the left atrium with the left superior pulmonary vein (LSPV), left inferior pulmonary vein (LIPV), right superior pulmonary vein (RSPV), and right inferior pulmonary vein (RIPV) labeled. The termination site corresponds to the lesion within the circled area 1442 for this particular patient. Personal digital data 1445 shows the electrogram when AF terminates to sinus rhythm. Display panel 1450 shows the personal streaming data of the AF map generated by the freely available online method played in the smartphone app. Display panel 1455 shows interactive input with the physician to identify key sites for personalized therapy.

[0203] Fig.15 A computer system that can be used to implement the method of the present invention is illustrated, such as can be incorporated into various devices, such as a personal computer (PC), a tablet PC, a personal digital assistant (PDA), a mobile device, a palmtop computer, a laptop computer, a desktop computer, a communication device, a control system, a web appliance, or any other machine capable of executing (sequentially or otherwise) a set of instructions that specify actions to be performed by the machine. Further, although a single computer system 2300 is shown, the term "system" should also be construed to include any collection of systems or subsystems that can execute one or more sets of instructions, either individually or in combination, to perform one or more computing functions.

[0204] like Fig.15 As shown, the computer system 2300 may include a processor 2302 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), or both). The computer system may include a main memory 2304 and a static memory 2306 that may communicate with each other via a bus 2326. As shown, the computer system 2300 may further include a video display unit 2310 (such as a liquid crystal display (LCD), an organic light emitting diode (OLED), a flat panel display, a solid state display, or a cathode ray tube (CRT)). Additionally, the computer system 2300 may include an input device 2312 such as a keyboard and a cursor control device 2314 such as a mouse. The computer system 2300 may also include a drive unit 2316, a signal generating device 2322 (such as a speaker or a remote control), and a network interface device 2308.

[0205] In a specific embodiment, if Fig.15As depicted, the drive unit 2316 may include a computer-readable medium 2318 having one or more sets of instructions 2320 (e.g., software) stored therein. The drive unit 2316 may be a disk drive, a thumb drive (USB flash drive), or other storage device. Further, the instructions 2320 may embody one or more of the methods or logic described herein. In a particular embodiment, the instructions 2320 may reside completely or at least partially within the main memory 2304, the static memory 2306, and / or the processor 2302 during execution by the computer system 2300. The main memory 2304 and the processor 2302 may also include computer-readable media.

[0206] In alternative embodiments, dedicated hardware implementations such as application specific integrated circuits (ASICs), programmable logic arrays (PLAs) and other hardware devices can be constructed to implement one or more of the methods described herein. The applications of the apparatus and systems that may include various embodiments may include a wide range of electronic and computer systems. One or more embodiments described herein may use two or more specific interconnected hardware modules or devices with related control and data signals that may be transmitted between modules and through modules to implement functions or to implement functions as part of an application specific integrated circuit. Therefore, the system of the present invention includes software, firmware and hardware implementations.

[0207] According to various embodiments, the methods described herein may be implemented by a software program tangibly embodied in a processor-readable medium and may be executed by a processor. Further, in an exemplary non-limiting embodiment, implementations may include distributed processing, component / object distributed processing, and parallel processing. Alternatively, a virtual computer system process may be configured to implement one or more of the methods or functions as described herein.

[0208] It is also contemplated that the computer-readable medium includes instructions 2320 or receives instructions 2320 and executes these instructions in response to the propagated signal so that devices connected to the network 2324 can communicate voice, video, or data over the network 2324. Further, the instructions 2320 can be transmitted or received over the network 2324 through the network interface device 2308.

[0209] Although the computer-readable medium is illustrated as a single medium, the term "computer-readable medium" includes a single medium or multiple media, such as a centralized or distributed database and / or associated caches and servers storing one or more sets of instructions. The term "computer-readable medium" should also include any medium that can store, encode, or carry a set of instructions for execution by a processor or cause a computer system to perform any one or more of the methods or operations disclosed herein.

[0210] Embodiments of systems and methods for generating personalized digital phenotypes of diseases, which are compared with digital classifications to personalize therapies, are described above. Although specific example embodiments have been described, it is apparent that various modifications and changes may be made to these embodiments without departing from the broader scope of the invention. Therefore, this specification and the accompanying drawings should be viewed in an illustrative rather than a restrictive sense. The accompanying drawings forming part of this specification illustrate specific embodiments in which the subject matter can be practiced by way of illustration and not limitation. The embodiments shown are described in sufficient detail to enable those skilled in the art to practice the teachings disclosed herein. Other embodiments may be utilized and derived therefrom, so that structural and logical substitutions and changes may be made without departing from the scope of the present disclosure. Therefore, this detailed description should not be viewed in a restrictive sense, and the scope of each embodiment is defined only by the appended claims together with the full range of equivalents to which such claims are entitled.

[0211] Such embodiments of the inventive subject matter may be referred to herein, individually and / or collectively, by the term "invention", which is merely for convenience and is not intended to limit the scope of the present application to any single invention or inventive concept when more than one invention or inventive concept is actually disclosed. Thus, although specific embodiments have been shown and described herein, it should be understood that any arrangement intended to achieve the same purpose may be substituted for the specific embodiments shown.

[0212] References (Incorporated herein by reference)

[0213] 1. Barbato E et al., Review and Updates in Regenerative and Personalized Medicine, Preclinical Animal Models, and Clinical Care in Cardiovascular Medicine. Journal of Cardiovascular Translational Research, 2015;8:466-74.

[0214] 2. Darbar D, Genetics of atrial fibrillation: rare mutations, common polymorphisms, and clinical relevance. Heart Rhythm 2008;5:483-6.

[0215] 3. Van Driest SL et al., Association of Arrhythmia-Related Genetic Variants With Phenotypes Documented in Electronic Medical Records. JAMA, 2016;315:47-57.

[0216] 4. Lahrouchi N et al., Utility of Post-Mortem Genetic Testing in Cases of Sudden Arrhythmic Death Syndrome. J Am Coll Cardiol. 2017;69:2134-2145.

[0217] 5. Parvez B et al., A common beta1-adrenergic receptor polymorphism predicts favorable response to rate-control therapy in atrial fibrillation. [Journal of the American College of Cardiology] 2012;59:49-56.

[0218] 6. Ritchie MD et al. Chromosome 4q25 variants are genetic modifiers of rare ion channel mutations associated with familial atrial fibrillation. J Am Cardiol 2012;60:1173-81.

[0219] 7. Kuck KH et al., Cryoballoon or Radiofrequency Ablation for Paroxysmal Atrial Fibrillation. N Engl J Med. 2016;74:2235-45.

[0220] 8. Verma A et al., Approaches to catheter ablation for persistent atrial fibrillation. [New England Journal of Medicine] 2015;372:1812-22.

[0221] 9. Narayan SM et al., Panoramic Electrophysiological Mapping But Not Individual Electrogram Morphology Identifies Sustaining Sites for Human Atrial Fibrillation (AF Rotors and Focal Sources Relate Poorly to Fractionated Electrograms). Circ Arrhythm Electrophysiol. 2013;6:58-67.

[0222] 10. Sahli Costabal F et al., Interpreting Activation Mapping of Atrial Fibrillation: A Hybrid Computational / Physiological Study. Ann Biomed Eng. 2018; 46(2): 257-269.

[0223] 11. Zaman JAB et al., Identification and Characterization of Sites Where Persistent Atrial Fibrillation is Terminated by Localized Ablation. [Circulation: Arrhythmias / Electrophysiology] 2018;11:e005258; in press.

[0224] 12. Sahadevan J et al., Epicardial Mapping of Chronic Atrial Fibrillation in Patients: Preliminary Observations. Circulation, 2004; 110: 3293-3299.

[0225] 13. Sanders P et al., Localization of maximal dominant frequency sources correlates with the termination of atrial fibrillation during catheter ablation (abstract). [Heart Rhythm] 2004;1:S12.

[0226] 14. Nademanee K et al., A new approach for catheter ablation of atrial fibrillation: mapping of the electrophysiologic substrate. [Journal of the American College of Cardiology] 2004a; 43: 2044-2053.

[0227] 15. Cuculich PS et al., Noninvasive Characterization of Epicardial Activation in Humans With Diverse Atrial Fibrillation Patterns. [Circulation], 2010;122:1364-72.

[0228] 16. Wong KC et al., No Benefit of Complex Fractionated Atrial Electrogram (CFAE) Ablation in Addition to Circumferential Pulmonary Vein Ablation and Linear Ablation: BOCA Study. Circulation: Arrhythmias and Electrophysiology, 2015;8:1316-24.

[0229] 17. Narayan SM et al., Treatment of Atrial Fibrillation by the Ablation of Localized Sources: The Conventional Ablation for Atrial Fibrillation With or Without Focal Impulse and Rotor Modulation: CONFIRM Trial. J Am Cardiol 2012;60:628-636.

[0230] 18. Haissaguerre M et al., Driver Domains in Persistent Atrial Fibrillation. Circulation, 2014;130:530-8.

[0231] 19. Krummen DE et al., Rotor Stability Separates Sustained Ventricular Fibrillation From Self-Terminating Episodes in Humans. J Am Cardiol 2014;63:2712-21.

[0232] 20. Alhusseini M et al., Two Independent Mapping Techniques Identify Rotational Activity Patterns at Sites of Local Termination during Persistent Atrial Fibrillation. J Cardiovasc Electrophys. 2017;28:615-622.

[0233] 21. Zaman JAB et al., Rotational Drivers in Atrial Fibrillation: Are Multiple Techniques Circling Similar Mechanisms? [Circulation: Arrhythmias and Electrophysiology], 2017;10.

[0234] 22. Esteva A et al., Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017; 542: 115-118.

[0235] 23. Rajpurkar P et al., Cardiologist level arrhythmia detection with convolutional neural networks. 2017; arXiv:1707.01836 [cs.CV].

[0236] 24. Abed HS et al., Effect of weight reduction and cardiometabolic riskfactor management on symptom burden and severity in patients with atrial fibrillation: a randomized clinical trial. JAMA. 2013;310:2050-60.

[0237] 25. Selvaraju, RR et al., Grad=CAM: Visal Explanation from Deep Networks via Gradient-based Localization, arXiv:1610.02391v3, March 21, 2017.

Claims

1. A system for identifying and treating a cardiac rhythm disorder in a patient, the system comprising: an electrode catheter including an electrical sensor configured to detect electrical excitation generated within a patient's heart by recording electrical signals generated within the patient's heart over time, wherein the electrode catheter is configured for delivery from within a chamber of the patient's heart; A computing device configured to: collecting electrical signals recorded by electrical sensors on a region of the patient's heart; collect patient data elements that include one or more of the patient's demographic, clinical, laboratory, pathology, chemistry, imaging, historical, genetic, and activity data; reconstructing an action potential of the patient's heart by filtering noise from the electrical signal; generating a map of the electrical activity indicative of a location of the patient's heart using the reconstructed action potentials; identifying one or more sources of a complex rhythm disorder based on the map of the locations of the electrical activation and the patient data elements; comparing the identified one or more source and patient data elements to a population digital classification that includes stored prior data about other patients having the one or more quantitative disease classifications to classify the patient into one or more quantitative disease classifications; and A personalized intervention for treating the patient is determined based on the classified one or more quantitative disease classifications for the patient.

2. The system of claim 1, wherein: The electrical signal is transmitted via one of wired or wireless communications.

3. The system of claim 1, wherein: The computing device is also configured to collect data streams from one or more of electrodes, optical sensors, piezoelectric sensors, acoustic sensors, resistive sensors, thermal sensors, accelerometers, pressure sensors, flow sensors, and electrochemical sensors.

4. The system of claim 3, wherein: The data stream includes one or more of mechanical heart signals, heart rate, heart sounds, respiratory sounds, respiratory rate, respiratory volume, neural activity, and immunological signals.

5. The system of claim 1, wherein: These patient data elements include one or more of hemodynamics, cardiac structure from imaging, clinical factors associated with cardiac or pulmonary conduction, neural signals, genetic profiles, metabolic state biomarkers, and patient movement.

6. The system of claim 1, wherein: The computing device is further configured to apply a time stamp to the electrical signal and each of the patient data elements.

7. The system of claim 1, wherein: The computing device is also configured to segment the electrical signal to generate a personalized digital phenotype, and the segmentation algorithm includes one or more of supervised machine learning, neural networks, correlation analysis, logistic regression analysis, decision trees, time domain analysis, frequency domain analysis, triangular transformation, logarithmic transformation, cluster analysis and unsupervised machine learning.

8. The system of claim 1, wherein: The digital taxonomy further includes previous data of the patient.

9. The system of claim 1, wherein: The one or more quantitative disease classifications include one or more of a rotational or focal activation pattern, an intermittent rotational or focal activation pattern, an incomplete activation pattern, and a location to a specific cardiac structure or a specific anatomical location of the patient.

10. The system of claim 1, wherein: The heart rhythm disorder includes one or more of atrial fibrillation, ventricular fibrillation, atrial tachycardia, atrial flutter, polymorphic or monomorphic ventricular tachycardia, ventricular flutter, or other electrical disturbances within the heart.

11. The system of claim 10, wherein: The computing device is also configured to identify locations of relatively higher excitement in the graph.

12. The system of claim 10, wherein: The computing device is also configured to generate a reconstructed electrogram using a machine learning algorithm trained on one or more reference signals associated with different heart rhythms.

13. The system of claim 1, wherein: Personalized treatment includes modifying at least a portion of the patient's heart by performing one or more of ablation via a contact device, energy delivery via a non-contact device, electrical therapy, thermal therapy, mechanical therapy, delivery of drug therapy, delivery of immunosuppression, delivery of stem cell therapy, and delivery of gene therapy.

14. The system of claim 7, wherein the computing device is configured to generate updated personal history data of the personalized digital phenotype, the classified one or more quantitative disease classifications, the personalized treatment, and personal profile results.

15. The system of claim 1, further comprising a display device configured to display the classified one or more quantitative disease classifications.

16. The system of claim 1, wherein: The computing device is further configured to iteratively collect additional electrical signals from the electrical sensor.

17. The system of claim 1, wherein: The electrode catheter includes a plurality of electrical sensors configured to detect cardiac signals.

Images