Medical Event Prediction Using Personalized Dual-Channel Coupling Networks

Abstract

A system and method for predicting the occurrence of a patient's medical events using a trained neural network. Historical patient data is pre-processed (902) to generate normalized training samples (904), which are sent to a deep convolutional neural network for model pre-training and model parameter update (906). The pre-trained model is stored (908) on a remote server for use in local machine personalization during preparation for treatment. A normalized fine-tuning set is generated as output (906), and the model parameters are iteratively fine-tuned (912). A personalized prediction score for the future medical event is generated (914), and the operation of a medical device is controlled in response to the prediction score (916).

Description

[Technical field]

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 170,660, filed April 5, 2021, the disclosure of which is incorporated herein in its entirety.

[0002] The present invention relates generally to predicting the occurrence of a medical event in a patient, and more particularly to predicting the occurrence of a particular medical event in a patient before, during and after treatment using a trained neural network. [Background technology]

[0003] In recent years, the extensive implementation of digital systems in hospitals and many medical institutions has generated a large amount of patient medical data. Big data has great value and allows the use of artificial intelligence (AI) to support clinical decisions in medicine. As one of the important themes in modern medicine, the number of patients with kidney disease has caused social, medical and socio-economic problems worldwide. Hemodialysis, or simply dialysis, is the process of purifying the blood of patients whose kidneys are not functioning properly and is one of the important renal replacement therapies (RRT). However, dialysis patients are at high risk of cardiovascular and other diseases, so thorough management of blood pressure, anemia, mineral metabolism, etc. is necessary. Otherwise, patients may face serious situations such as low blood pressure, leg cramps and even death during dialysis and / or other treatments. Summary of the Invention

[0004] A computer-implemented method for predicting occurrence of a patient's medical events using a trained neural network includes pre-processing received historical patient data for a plurality of patients to generate a plurality of normalized training samples. The normalized training samples are sent to a personalized deep convolutional neural network (P-DCCN) for pre-training the model and updating model parameters using the P-DCCN, the pre-trained model is stored at a remote server for use in personalization by a local machine during preparation for treatment, and a normalized fine-tuning set is generated as output from the P-DCCN by processing input personal data of the patient from the local machine. The model parameters of the P-DCCN are iteratively fine-tuned by performing multiple training iterations using the generated normalized fine-tuning set. A personalized prediction score for the patient's future medical events is generated using the P-DCCN, and the operation of the medical device is controlled according to the personalized prediction score for the future medical events.

[0005] A system for predicting occurrence of a patient's medical events using a trained neural network, comprising: pre-processing received historical patient data for a plurality of patients using a processor operatively connected to a computer-readable storage medium to generate a plurality of normalized training samples; the normalized training samples are sent to a personalized deep convolutional neural network (P-DCCN) for pre-training the model and updating model parameters using the P-DCCN; the pre-trained model is stored at a remote server for use in personalization by a local machine during preparation for treatment; and a normalized fine-tuning set is generated as output from the P-DCCN by processing input personal data of the patient from the local machine; the model parameters of the P-DCCN are iteratively fine-tuned by performing a plurality of training iterations using the generated normalized fine-tuning set; a personalized prediction score for the patient's future medical events is generated using the P-DCCN, and the operation of the medical device is controlled according to the personalized prediction score for the future medical events.

[0006] A non-transitory computer-readable recording medium including a computer-readable program for predicting occurrence of a patient's medical events using a trained neural network, the computer-readable program, when executed by a computer, causes the computer to generate a plurality of normalized training samples by pre-processing received historical patient data for a plurality of patients. The normalized training samples are sent to a personalized deep convolutional neural network (P-DCCN) for pre-training the model and updating model parameters using the P-DCCN, the pre-trained model is stored at a remote server for use in personalization by a local machine during preparation for treatment, and a normalized fine-tuning set is generated as output from the P-DCCN by processing input personal data of the patient from the local machine. The model parameters of the P-DCCN are iteratively fine-tuned by performing multiple training iterations using the generated normalized fine-tuning set. A personalized prediction score for the patient's future medical events is generated using the P-DCCN, and the operation of the medical device is controlled according to the personalized prediction score for the future medical events.

[0007] These and other features and advantages will become apparent to those skilled in the art from the following detailed description and accompanying drawings.

[0008] In the present disclosure, preferred embodiments are described in detail with reference to the following drawings, as described below. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a block / flow diagram illustrating an exemplary processing system to which the present principles may be applied, in accordance with the present principles.

[0010] [Diagram 2]FIG. 2 shows a diagram illustrating a system / method for predicting therapeutic events using a personalized dual-channel combiner network (P-DCCN), in accordance with an embodiment of the present principles.

[0011] [Diagram 3] FIG. 3 is a block / flow diagram showing exemplary time periods for treatment events, treatment preparation and prediction of treatment events, in accordance with an embodiment of the present principles.

[0012] [Figure 4] FIG. 4 shows a system / method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0013] [Figure 5A] FIG. 5A illustrates a system / method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0014] [Figure 5B] FIG. 5B illustrates a system / method for predicting therapeutic events using a Personalized Dual Channel Coupling Network (P-DCCN) computational component, in accordance with an embodiment of the present principles.

[0015] [Figure 6] FIG. 6 is a diagram illustrating a method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN) pre-processing component in accordance with an embodiment of the present principles.

[0016] [Figure 7] FIG. 7 is a high level block / flow diagram illustrating a system / method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0017] [Figure 8] FIG. 8 is a block / flow diagram illustrating a system / method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0018] [Figure 9] FIG. 9 is a block / flow diagram illustrating a method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0019] [Figure 10] FIG. 10 is a high level diagram illustrating a system for delivering therapy to a patient, predicting therapy events, and monitoring and collecting data from a patient during therapy using a Personalized Dual Channel Coupling Network (P-DCCN), in accordance with an embodiment of the present principles. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] In accordance with various embodiments of the present principles, systems and methods are provided for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN).

[0021] In particularly useful embodiments, systems and methods are provided for predicting the occurrence of certain medical events in a patient before, during, and after a treatment (e.g., a dialysis treatment) using trained neural networks.

[0022] As mentioned above, hemodialysis or simply dialysis is a process of purifying blood in patients whose kidneys are not functioning properly and is one of the important renal replacement therapies (RRT). However, dialysis patients are at high risk of cardiovascular and other diseases, and require thorough management of blood pressure, anemia, mineral metabolism, etc. Otherwise, patients may face serious events such as low blood pressure, leg cramps, and even death during dialysis and / or other treatments. Therefore, medical staff decide how to proceed with dialysis from various perspectives based on the patient's risk, potential treatment events, and various clinical factors related to dialysis events. Considering the availability of big medical data, the present invention utilizes an AI system using a dual channel coupling network (DCCN) to create a prognostic score during the pre-dialysis period for the incidence of events in future dialysis. This enhances the decision-making process of medical staff and reduces the risk of harmful and / or unwanted medical accidents.

[0023] However, (1) due to data privacy, it is usually difficult to obtain a large amount of patient data from hospitals to train an accurate model. (2) Because patients are very diverse, it is difficult to accurately fit a single pre-trained model to all new patients, who usually differ in age, gender, genetic characteristics, health conditions, etc., from the patient data in the training set. These important challenges in (1) and (2) prevent traditional AI systems from being successfully applied to accurate analysis of patient medical data. Therefore, a traditional single pre-trained model trained on a limited training dataset cannot be generalized to predictive analysis of new patient data.

[0024] According to various embodiments, the present invention harnesses the potential of dialysis patient management data by training and using a neural network (e.g., a deep neural network (e.g., DNN (Deep Neural Network), DCCN, etc.) to automatically provide high-quality, specifically personalized prognostic scores for the incidence of intradialysis events through a novel pre-training and fine-tuning technique. By personalizing the pre-trained model for each patient using a small amount of fine-tuning data, the model can generalize well to a test dataset, mitigating the two challenges mentioned above, and can be used with patients undergoing medical procedures (e.g., dialysis) to maximize the benefits of the procedure for a particular patient while minimizing the risks.

[0025] The present invention is a P-DCCN system / method that provides a systematic and data-driven solution to medical event (e.g., dialysis event) prediction that is not known in the art. The present invention is a neural network-based intelligent computing system that does not require human effort of feature engineering. The P-DCCN system / method includes a dual-channel component for integrating static features, low-frequency time-course features, and relatively high-frequency time-course features for joint representation learning and prediction of dialysis events during treatment. In various embodiments, the present invention achieves much higher processing efficiency and accuracy than non-personalized conventional models using a pre-training and fine-tuning approach that addresses and mitigates the challenges of insufficient training data and mismatched distribution of patient data. This characteristic makes P-DCCN significantly different from other models with conventional neural network training approaches.

[0026] In some embodiments, the pre-training component of the P-DCCN system can use a relatively small historical record of a patient's entire medical record data to generate a pre-trained model that is stored on a server or cloud platform for use in predictive analysis of new patients in the future. The fine-tuning component of the P-DCCN system sends the pre-trained model to a local device where new patient records are stored for fine-tuning. This component, like the pre-training component described above, uses only a relatively small number of new records. Using this small amount of personalized data allows the model to achieve significant improvements in accuracy and reduced processor requirements compared to conventional non-personalized systems and methods.

[0027] In some embodiments, the DCCN system's dialysis record data processing component converts each patient's historical record into static profile features and time series features of different frequencies that are input to the DCCN computational component for further training and / or processing. The deep neural network design of the P-DCCN computational component, according to aspects of the invention, improves prediction accuracy and significantly reduces the human effort required for feature engineering. In some embodiments, the dual channel design of the P-DCCN computational component includes a multilayer perception (MLR) and long short-term memory (LSTM) recurrent neural network that integrates both static and time series features of different frequencies for joint event prediction, according to aspects of the invention.

[0028] The embodiments described herein may be implemented entirely in hardware, entirely in software, or contain both hardware and software elements, in a preferred embodiment, the invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.

[0029] An embodiment may include a computer program product accessible from a computer usable or computer readable medium that provides program code for use by or in connection with a computer or any instruction execution system. A computer usable or computer readable medium may include any device that stores, conveys, propagates, or transfers a program for use by or in connection with an instruction execution system, device, or apparatus. The medium may be a magnetic, optical, electronic, electromagnetic, infrared, or semiconductor system (or device or apparatus), or a propagation medium. The medium may include computer readable media such as semiconductor or solid state memory, magnetic tape, removable computer diskettes, random access memory (RAM), read-only memory (ROM), rigid magnetic disk, and optical disk.

[0030] Each computer program is stored in a machine-readable storage medium or device (e.g., a program memory or a magnetic disk) that can be read by a general-purpose or special-purpose programmable computer. The computer program is read from the storage medium or device by a computer that executes the procedures described herein for configuring and controlling operations of the computer. The system of the present invention also contemplates a computer-readable storage medium containing a computer program configured to cause a computer to operate in a specific and predefined manner to perform the functions described herein.

[0031] A data processing system suitable for storing and / or executing program code may include at least one processor coupled directly or indirectly to memory elements via a system bus. The memory elements may include local memory, bulk memory devices, and cache memories for temporarily storing at least some program code used during the actual execution of the program code to reduce the number of times code is retrieved from bulk memory devices during execution of a process. Input / output or I / O devices (including but not limited to keyboards, displays, pointing devices, etc.) may be connected to the system either directly or through I / O controllers.

[0032] Network adapters may be coupled with the data processing system to enable the system to be coupled to other data processing systems or remote printers or memory devices through private or public networks. Modems, cable modems and Ethernet cards are just a few of the currently available types of network adapters.

[0033] As used herein, the term "hardware processor subsystem" or "hardware processor" may refer to a processor, memory, software, or combination thereof that cooperate to perform one or more specific tasks. In useful embodiments, the hardware processor subsystem may include one or more data processing elements (e.g., logic circuits, processing circuits, instruction execution units, etc.). The one or more data processing elements may include a central processing unit, a graphics processing unit, and / or individual processors or computational element-based controllers (e.g., logic gates, etc.). The hardware processor subsystem may include one or more on-board memories (e.g., caches, dedicated memory arrays, read-only memories, etc.). In any embodiment, the hardware processor subsystem may include one or more memories that may be on-board or off-board, or may be dedicated for use with the hardware processor subsystem (e.g., ROM, RAM, Basic Input / Output System (BIOS), etc.).

[0034] In any embodiment, the hardware processor subsystem may include and execute one or more software elements, which may include an operating system and / or one or more applications and / or specific code to achieve a particular result.

[0035] In other embodiments, the hardware processor subsystem may include dedicated circuitry to perform one or more electronic processing functions to achieve a specified result. Such circuitry may include one or more application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), and / or programmable logic arrays (PLAs).

[0036] These and other variations of the hardware processor subsystem are also contemplated by embodiments of the present invention.

[0037]

[0023] Referring now in detail to the drawing, Figure 1, where like numerals represent the same or similar elements, there is shown an exemplary processing system 100 to which the present principles may be applied, in accordance with an embodiment of the present principles. Processing system 100 includes at least one processor (CPU) 104 operatively connected to other components via a system bus 102. Operatively connected to system bus 102 are cache 106, read only memory (ROM) 108, random access memory (RAM) 110, input / output (I / O) adapter 120, sound adapter 130, network adapter 140, user interface adapter 150, and display adapter 160.

[0038] A first storage device 122 and a second storage device 124 are operably connected to the system bus 102 by an I / O adapter 120. The storage devices 122 and 124 may be disk storage devices (e.g., magnetic or optical disk storage devices), solid state storage devices, magnetic storage devices, etc. The storage devices 122 and 124 may be the same type of storage device or different types of storage devices.

[0039] Speakers 132 are operably connected to the system bus 102 by a sound adapter 130. Transceiver 142 is operably connected to the system bus 102 by a network adapter 140. Display device 162 is operably connected to the system bus 102 by a display adapter 160.

[0040] A first user input device 152, a second user input device 154, and a third user input device 156 are operably connected to the system bus 102 by a user interface adapter 150. The user input devices 152, 154, and 156 may be any of a keyboard, mouse, keypad, image capture device, motion sensing device, microphone, or a device incorporating the functionality of at least two of these devices. Of course, other types of input devices may be used while maintaining the spirit of the present principles. The user input devices 152, 154, and 156 may be the same type of user input device or different types of user input devices. The user input devices 152, 154, and 156 are used to input information to the processing system 100 and output information from the processing system 100.

[0041] Of course, processing system 100 may include other elements (not shown) or omit certain elements, as would be readily apparent to one of ordinary skill in the art. For example, processing system 100 may include various other types of input and / or output devices, depending on the particular implementation, as would be readily apparent to one of ordinary skill in the art. For example, various wireless and / or wired input and / or output devices may be used. Furthermore, various configurations of additional processors, controllers, memory, etc. may be used, as would be readily apparent to one of ordinary skill in the art. These and other variations of processing system 100 will be readily apparent to one of ordinary skill in the art from the teachings of the present principles provided herein.

[0042] It should further be understood that systems 100, 200, 400, 500, 501, 700, 800 and 1000 described in connection with Figures 1, 2, 4, 5A, 5B, 7, 8 and 10 are systems for implementing respective embodiments of the present principles. Part or all of processing system 100 may be implemented with one or more of the elements of systems 200, 400, 500, 501, 700, 800 and 1000 according to various embodiments of the present principles.

[0043] Further, it should be understood that the processing system 100 can perform at least a portion of the methods described herein, including, for example, at least a portion of the methods 300, 400, 500, 501, 600, 700 and 900 described in Figures 3, 4, 5A, 5B, 6, 7 and 9. Similarly, some or all of the systems 200, 400, 500, 501, 700, 800 and 1000 may be used to perform at least a portion of the methods 300, 400, 500, 501, 600, 700 and 900 described in Figures 3, 4, 5A, 5B, 6, 7 and 9 in accordance with various embodiments of the present principles.

[0044] Referring now to FIG. 2, there is shown an exemplary block diagram illustrating a system / method 200 for predicting therapeutic events using a personalized dual channel combination network (P-DCCN) 218, according to an embodiment of the present principles. In one embodiment, a dual channel combination network (DCCN) 202 is used to generate a personalized DCCN (P-DCCN) 218 ​​by training the DCCN 202 using a neural network training component 206. The training component 206 includes a pre-processing component 208 (e.g., for pre-training), a computation component 210 (e.g., for processing static and temporal features), a model storage component 212 (e.g., storage), and a fine-tuning component 214. That is, a generic pre-trained model is adapted to a particular patient based on, for example, historical and / or real-time patient measurements acquired using patient measurement devices 204 and other individual patient data collected and stored on a local machine 216, according to an aspect of the present invention.

[0045] In various embodiments, after fine-tuning is completed using fine-tuning component 214, prediction component 220, according to aspects of the present invention, uses the trained P-DCCN to predict future events (e.g., treatment events, adverse patient health events, etc.) for a particular patient. It should be appreciated that the components of system 200, according to aspects of the present invention, may be connected by bus 201 or via any suitable communication means (e.g., wireless connection, remote connection via the Internet, wired Ethernet connection, other wired connection, etc.).

[0046] Reference is now made to FIG. 3, which illustratively shows a block / flow diagram illustrating an exemplary time period for treatment events, treatment preparation and prediction of treatment events 300, in accordance with one embodiment of the present principles.

[0047] Generally, a sick patient will likely have a regular routine of undergoing treatment (e.g., dialysis) for the illness at least once a week, depending on the type of illness and / or its severity. For ease of explanation, the present invention will be described with reference to a patient's dialysis treatment, however, it will be understood that the principles of the present invention, according to various embodiments of the present invention, can be applied to predict medical events for any type of illness / disease before, during, and after treatment.

[0048] Typically, a dialysis patient has regular dialysis sessions three times a week, with each session lasting 4-5 hours. An exemplary weekly schedule 301 of a dialysis patient includes a treatment day 302, non-treatment days 2 304 and 3 306, a second treatment day 308, non-treatment days 5 310 and 6 312, and a third treatment day 314. As described above, various adverse medical events may occur before, during, and / or after dialysis treatment, so the present invention uses a P-DCCN according to an embodiment of the present invention to predict the likelihood of an event occurring in each patient's upcoming dialysis session based on past record data (e.g., predicting an event before a dialysis treatment session), thereby minimizing the risk of adverse health effects before, during, and / or after dialysis treatment.

[0049] The historical record data of a dialysis patient mainly consists of four parts: the patient's static profile (e.g., age, sex, dialysis start time, etc.), dialysis measurement records (three times a week (e.g., blood pressure, weight, venous pressure, etc.)), blood test measurements (twice a month (e.g., albumin, glucose, platelet count, etc.)) and cardiothoracic ratio (CTR) (once a month). Of these, the last three parts are dynamic and change over time, so according to an embodiment of the present invention, they are modeled in a time series with different periods.

[0050] The present invention is an artificial intelligence system built on a building block architecture of a dual channel neural network called the Dual Channel Combination Network (DCCN), which integrates the different parts of the data mentioned above for model training and prognostic score prediction, as described in more detail below.

[0051] According to embodiments of the present invention, dialysis events (e.g., medical events that may occur during a dialysis treatment session 309) may be predicted during treatment preparation period 307 prior to performing the treatment session 309 on treatment day 302, treatment day 308, and treatment day 314. The treatment session is completed in block 311, and measurements and other patient data are utilized in further training of a P-DCNN for use in future dialysis treatment sessions (e.g., treatment day 314) according to embodiments of the present invention.

[0052] Reference is now made to FIG. 4, which illustrates an exemplary block diagram of a system / method 400 for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0053] In some embodiments, the present invention is an artificial intelligence system built on a building block architecture of dual channel neural networks called Dual Channel Combination Networks (DCCNs), which integrate the above-mentioned different parts of the data for model training and prognostic score prediction. In various embodiments, the present invention can generate a personalized DCCN (P-DCCN) for every individual patient's data by utilizing a pre-training and fine-tuning framework according to aspects of the present invention, which are described in more detail below.

[0054] In some embodiments, past patient records 402 for N patients (e.g., P1 401, P2 403, ..., P N405) are received as input for pre-training 404 to generate a pre-trained P-DCCN 406 according to an embodiment of the present invention. In the pre-training stage 404, the P-DCCN 406 is trained using medical history data of multiple different patients 402 obtained from a database of the patients' individual hospital records (e.g., an Electronic Health Record (EHR) database). The patient history records 402 include, for example, patient profiles, dialysis measurements, blood test measurements, CTR measurements, etc., which are used as input for training the pre-trained P-DCNN 406 according to an embodiment of the present invention.

[0055] In some embodiments, the pre-trained P-DCCN 406 is stored on a server or cloud platform (not shown) for use in further processing. In block 408, the pre-trained P-DCCN 406 is trained on a number of K new patients (e.g., P N+1 421, P N+2 423, …, P N+K 425) is sent by any suitable data transfer means (e.g., wireless, wired, remote connection, etc.) to one or more local machines 411, 413, 415 for fine-tuning. During the fine-tuning phase 408, once a certain amount of medical record data (e.g., a relatively small amount in terms of the patient's overall medical record), such as a few weeks' worth of records for a new patient, is accumulated, these records are used to fine-tune the pre-trained P-DCCN 406, which is personalized for the particular patient.

[0056] In some embodiments, the generated and trained personalized P-DCNN (e.g., P N+1 431, P N+2 433, …, P N+K435), which are each used for future predictive analysis of a particular patient to predict the probability of a medical event occurring before, during, and / or after treatment, respectively, in blocks 441, 443, and 445. Such personalization of the P-DCNN improves accuracy over conventional models that do not consider such personalization. In block 410, the predictive scores determined in blocks 441, 443, and 445 are output before, during, and / or after the patient's treatment, according to aspects of the present invention. Because many treatments (e.g., dialysis treatments) may be required over long periods of time (e.g., months, years, a lifetime, etc.), the P-DCCN 406 is continually fine-tuned in block 408 and personalized iteratively throughout the patient's treatment to improve accuracy. It should be noted that while the present invention has been described above with respect to dialysis treatments, it should be understood that the P-DCCN system and method of the present invention, according to various embodiments of the present invention, may be applied to other medical record data, diseases, and / or treatment actions.

[0057] Reference is now made to FIG. 5A, which illustrates, by way of example, a high-level diagram of a system / method 500 for predicting therapeutic events using a personalized dual-channel coupling network (P-DCCN), in accordance with one embodiment of the present principles.

[0058] According to an embodiment of the present invention, historical medical records (e.g., historical electronic medical records (EMR)) are input as a neural network training set in block 502 for use in the P-DCNN data pre-processing component 504. Dialysis patient medical history records are stored in any suitable format (e.g., CVS, excel, etc.). Each patient may have a file containing medical information regarding their static profile (e.g., age, sex, dialysis initiation, etc.), which may be input as static input X in block 506. The file may also contain medical information regarding their longitudinal profile (e.g., dialysis measurements, blood test measurements, event incidence, albumin, glucose, platelet count, etc.), which may be input as longitudinal inputs X1, X2, ..., X in block 508, according to an aspect of the present invention. TEach row represents a specific date that a patient visited the hospital. In some embodiments, the static inputs 506 and the temporal inputs 508 are sent to a P-DCCN computation component 518 for further training and / or processing using the static channel 512 and the temporal channel 514, in accordance with aspects of the invention.

[0059] In some embodiments, each column represents a particular feature, such as some index metric (e.g., blood pressure, weight, venous pressure, etc.) in the dialysis measurements. Since different parts have different frequencies, some entries in the form may be left blank to indicate that the feature is not measured on a particular day, according to an aspect of the invention. In block 510, training labels y are generated and sent as inputs to the P-DCCN computation component 518, including the training loss in block 516, which is described in more detail below with reference to FIG. 5B.

[0060] In various embodiments, the P-DCCN pre-processing component 504, according to aspects of the invention, can extract different parts of the data from the file, remove noisy information, and fill missing values ​​by using the average value of corresponding features in the historical data and / or by using values ​​from adjacent previous time steps.

[0061] In some embodiments, the pre-processing component 504 can set a time window of width w to segment the time series data, which is described in more detail below with reference to FIG. 6. In one embodiment, each time window can generate a sample X from time step Tw to time step T, which can be associated with an event label Y at time step T+1. This focuses the samples on features that are relatively close in date to the future event. All dialysis measurements within the time window can be included, since different parts have different frequencies, but blood test measurements with dates closest to the time window can be included. The time window can then be slid from the start of a date in the record to the end of a date to generate multiple samples.

[0062] In one embodiment, after the samples are generated, the pre-processing component 504 normalizes all samples using Gaussian normalization so that the features of the training samples have a mean value of 0 and a variance of 1. This aids in the stability of the computation component algorithm in block 518. When testing the samples, the samples are normalized using the mean value and variance obtained from the training data in block 518 according to an embodiment of the present invention, and the normalized samples are sent to the model storage component 520 for further model training, testing, and / or storage. In various embodiments, the P-DCCN computation component 518 may include two channels, a static channel for processing static and relatively low frequency temporal features, and a temporal channel for processing relatively high frequency temporal features, which will be described in more detail below with reference to FIG. 5B according to an embodiment of the present invention.

[0063] In some embodiments, the P-DCCN model storage component 520 receives as input a pre-trained P-DCCN model from the P-DCCN computation component 518, where the pre-trained model is trained using a certain (e.g., threshold) amount of input historical patient data 502 provided as a training set. This process updates the parameters of the P-DCCN to match the data in the training set, so that sufficient knowledge is extracted for further fine-tuning and personalization in block 526 using data from the local machine 524 and the P-DCCN personalization component 522, in accordance with aspects of the present invention.

[0064] In various embodiments, the P-DCCN model is pre-trained in block 518 using an optimizer with the following regression loss function:

number

number

[0065] According to an embodiment of the present invention, after pre-training is performed in block 518, the pre-trained P-DCCN (with all parameters updated and fixed) can be sent to a server or cloud platform and stored in block 520 so that it can be easily distributed to one or more local machines 524 for further fine-tuning and personalization in block 526 using a relatively small number of recordings from new patients collected on the local machine.

[0066] In fact, if a new patient undergoes dialysis treatment for several weeks, the local machine 524 collects multiple different types of records (e.g., static and longitudinal measurements) for that patient during that time. Although the amount of records collected by the local machine 524 is much smaller than the data size in the pre-training dataset, these records are specific to a particular patient and therefore help to adapt the generic pre-trained model to the patient's condition. This personalization process using the P-DCCN personalization component of block 522 via a relatively small amount of fine-tuning data according to an embodiment of the present invention takes advantage of few-shot learning.

[0067] In some embodiments, the pre-trained P-DCCN stored in the model storage component 520 is sent to a local machine 524 where a fine-tuning dataset is collected and stored locally. The fine-tuning dataset is again pre-processed during fine-tuning in block 526, similar to the pre-processing described above with respect to the pre-processing component 504 to generate training samples.

[0068] In some embodiments, the regression loss function described in Section 3

number

number

[0069] In some embodiments, performing the fine-tuning in block 526 may result in future time steps being calculated in block 502 based on an analysis of the patient history data received as input from the local machine 524 and the entered historical EMR data.

number

[0070] Reference is now made to FIG. 5B, which illustratively illustrates a diagram 501 of a personalized dual channel coupling network (P-DCCN) computation component 518 for predicting therapeutic events, in accordance with an embodiment of the present principles.

[0071] In various embodiments, in accordance with aspects of the present invention, the P-DCCN computation component 518 includes two channels: a static channel 512 for processing static and relatively infrequent temporal features, and a temporal channel 514 for processing relatively frequent temporal features.

[0072] In some embodiments, the static channel 512 receives static features (e.g., fluid temperature, hourly water removal rate, target water removal, dry weight, weight after last dialysis treatment, weight measurement before dialysis, gain over this time period, etc.) as static input x 506. The static features and relatively infrequent temporal features are stored in a vector x s and the static channel 512 can be represented as x s Information in

number

number

number

[0073] In various embodiments, the temporal channel 514 includes multiple Long Short Term Memory (LSTM) layers 515A, 515B, 515C for processing the temporal feature inputs 508A, 508B, 508C, which are represented as vectors x1, x2, ..., x T The LSTM layer is represented as a sequence of

number

number

[0074] In some embodiments, in addition to the LSTM layers 515A, 515B, 515C, the compact representations 513, 517, 512, 525 h0, h1, h2, ..., h T may be sent to the attention layers 519, 523, 527 for combination. The attention layers 519, 523, 527 calculate the following temporal importance scores, or attention weights, for each time step in blocks 519, 523, 527:

number

number

number

number

number

[0075] In various embodiments, the static representation h s and chronological expression d After the vectors , , are obtained from the static channel 512 and the temporal channel 514, the prediction layer 509 concatenates them and

number

number

[0076] With continuing reference to FIG. 5A and with reference to FIG. 6, there is shown an exemplary diagram 600 of a method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN) pre-processing component 504, in accordance with an embodiment of the present principles.

[0077] According to an embodiment of the invention, diagram 600 illustrates a segmentation process over a time window 604 for dialysis treatment measurements and data. Historical dialysis patient measurement data can be input in block 602, and longitudinal dialysis measurement data is measured and / or analyzed over time during dialysis treatment in blocks 606 and 608. According to an aspect of the invention, static blood test data 610 and static patient historical medical record data 612, in addition to static real-time patient measurement data 614 and / or other dialysis measurement data 616, are utilized as inputs to predict future medical events 618.

[0078] In some embodiments, each time window 604 generates a sample X from time step T-w 620 to time step T 622 and associates it with an event label Y at time step T+1 624 in block 618. According to aspects of the invention, this association can generate samples that focus on features of the date closest to the future event, and can include all dialysis measurements in the time window 604, while including blood test measurements 610 in the time window for the date closest to the future event, since different parts have different frequencies. The time window may slide from the beginning of a date in the record to the end of a date (e.g., a period of dialysis treatment, a full day, etc.) to generate multiple samples.

[0079] After the samples are generated, the pre-processing component 504 normalizes all samples using Gaussian normalization so that the features of the training samples have a mean value of 0 and a variance of 1. This improves the accuracy and stability of the computation algorithm of the P-DCNN computation component 518 according to an embodiment of the present invention. When testing samples, they can be normalized by using the mean value and variance obtained from the training data, and the normalized samples are sent to the next component (e.g., the model storage component 520) according to an embodiment of the present invention for further model training and testing. In fact, some of the dialysis measurements can be obtained on the same day that the event is predicted. These measurements (e.g., fluid temperature, hourly water removal rate, target water removal, dry weight, weight after the last dialysis treatment, time to dialysis weight measurement, increase during this period, etc.) can be obtained just before the start of dialysis, and can therefore be included as static features 614 for further processing according to an embodiment of the present invention.

[0080] Reference is now made to FIG. 7, which illustrates, by way of example, a high level block / flow diagram of a system / method 700 for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0081] According to various embodiments, in block 702, past pre-training data of a patient is input to a pre-training module 704. According to aspects of the invention, the pre-training data 702 (e.g., historical record data of multiple patients) can be input to a P-DCCN data pre-processing component 706, which can output normalized samples as a pre-training set. According to aspects of the invention, the normalized samples from block 706 are sent to a P-DCCN computation component 708 to update the P-DCCN parameters, which outputs a pre-trained P-DCCN, and the pre-trained P-DCCN from the P-DCCN computation component 708 is sent to a model storage component 710 for future deployment and / or personalization from a local machine.

[0082] In one embodiment, a relatively small amount of historical medical data for a particular patient (e.g., compared to the patient's complete medical history record) is input from the local machine 712 to the P-DCCN pre-processing component 716 using the personalization module 712, which can output normalized samples as a fine-tuning set and send to the P-DCCN data collection component 718, according to aspects of the present invention. The pre-trained P-DCCN from the model storage component 710 is sent to the P-DCCN personalization module 712 and utilized by the P-DCCN data collection component 718 to generate personalized prediction scores, output at block 720, according to aspects of the present invention. In some embodiments, fine-tuning of the model parameters of the pre-trained P-DCCN can be performed by multiple training iterations using the fine-tuning dataset, and the fine-tuned P-DCCN can be used to generate personalized prediction scores, output at block 720, according to aspects of the present invention.

number

[0083] Reference is now made to FIG. 8, which illustrates, by way of example, a block / flow diagram of a system / method 800 for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0084] According to an embodiment of the present invention, a system architecture of a P-DCCN system 802 is provided. A P-DCNN pre-training module 804 includes a P-DCCN pre-processing component 806, a P-DCNN computation component 814, and a model storage component 822. In some embodiments, according to aspects of the present invention, the P-DCNN pre-training module 804 is configured to perform data cleaning and data imputation to improve the quality of historical data at block 808, segmentation of recorded data and generation of time series samples at block 810, and / or Gaussian normalization of data samples for computational stability at block 812.

[0085] In some embodiments, the P-DCNN computation component 814 may include a dual channel neural network (DCNN) for simultaneously processing static and temporal features of different frequencies in block 816, an attention mechanism for the temporal channel to learn the relative importance of different time steps during integration for performance improvement and interpretation in block 818, and / or a combination layer for integrating the static and temporal features to calculate an event prediction score in block 820. In some embodiments, the P-DCCN model storage component 822, according to aspects of the present invention, is configured for platform support for executing the pre-processing component 806 and the computation component 814 in block 824, for pre-training, collecting and storing models in block 826, and / or for efficient communication with local machines for sharing pre-trained models received as input in block 828.

[0086] According to various embodiments, the P-DCCN personalization module 830 includes a P-DCCN local data collection component 832 and a P-DCCN fine-tuning component 840. According to aspects of the invention, the local data collection component 832 is configured to efficiently communicate with the model storage component 822 for data exchange, which includes platform support for timely collection and recording of new data from a treatment (e.g., dialysis) session in block 834, receiving pre-trained models from the model storage component 822, and aligning the operation of the fine-tuning component 840 to the data collected from the local data collection component 832.

[0087] In some embodiments, in accordance with aspects of the invention, the fine-tuning component 840 is configured to collect a pre-trained model and fine-tuning data in block 842, relatively quickly fit the pre-trained model to the fine-tuning data using a few-shot learning approach in block 844, fine-tune the model using a regression objective function and / or a gradient optimization algorithm in block 846, and generate a personalized prediction score based on new input data from the local machine in block 846.

[0088] Reference is now made to FIG. 9, which illustrates an exemplary block / flow diagram 900 of a method for predicting therapeutic events using a personalized dual channel coupling network (P-DCCN), in accordance with an embodiment of the present principles.

[0089] According to various embodiments, historical patient data is input at block 902, and a pre-training set is generated by pre-processing the historical patient data at block 904. The normalized samples are output as a pre-training set to a P-DCCN computation component at block 906, according to aspects of the present invention, and the pre-trained P-DCCN is stored at a model storage component at block 908. At block 910, measurements (e.g., blood pressure, heart rate, etc.) of the patient before, during, and / or after treatment (e.g., dialysis) are obtained, and iterative fine-tuning and personalization of the P-DCNN is performed based on the measurements from block 910 and other data stored on the local machine at block 912. At block 914, a personalized prediction score for future medical events for a particular patient is generated using the fine-tuned P-DCCN, according to aspects of the present invention. In block 916, in accordance with an embodiment of the present invention, a controller (e.g., automatic or manual) is used to control the operation of a medical device (e.g., dialysis device) and / or a number of measurement devices (e.g., blood pressure monitor, heart rate monitor, etc.) in response to the personalized prediction score generated in block 914.

[0090] Reference is now made to FIG. 10, which illustrates, by way of example, a high level diagram of a system 1000 for administering patient therapy, predicting therapy events, and monitoring and collecting data from a patient during therapy using a Personalized Dual Channel Coupling Network (P-DCCN), in accordance with one embodiment of the present principles.

[0091] According to various embodiments of the present invention, a patient 1001 is connected to a treatment and / or measurement device 1002 (e.g., dialysis machine, electrocardiogram (EKG) machine, blood pressure monitor, etc.) to receive treatment by a medical professional. According to aspects of the present invention, prior to treatment, a P-DCCN system 1006 can be used to predict potential medical events that may occur during treatment using the treatment and / or measurement device 1002. For repeated prediction of potential medical events in real time by the P-DCCN system 1006, the P-DCCN system 106 can be integrated (e.g., built-in) in the treatment and / or measurement device 1002 or can be connected to the treatment and / or measurement device via a port (e.g., USB, Ethernet, etc.) so that real-time measurements of the patient 1001 can be performed before as well as during treatment. A medical professional 1003 can use a controller (e.g., wired, remote, etc.) to control the operation of the medical and / or measurement device 1002 in response to event predictions output in real time by the P-DCCN system according to aspects of the present invention.

[0092] The foregoing should be considered in all respects as illustrative and exemplary, and not restrictive, and the scope of the invention disclosed herein should be determined not from the detailed description, but from the claims to be interpreted in accordance with the broadest possible interpretation permitted by the patent laws. It should be understood that the embodiments shown and described herein are merely illustrative of the principles of the invention, and that various modifications may be made by those skilled in the art without departing from the scope and spirit of the invention. Those skilled in the art may implement various other combinations of features without departing from the scope and spirit of the invention. Although aspects of the invention have been described above with the fine detail and particularity required by the patent laws, the scope of the claims which are sought to be protected by letters patent are set forth in the appended claims.

Claims

1. 1. A computer-implemented method for predicting an occurrence of a medical event in a patient using a trained neural network, comprising: generating 904 a plurality of normalized training samples by pre-processing the received historical patient data for a plurality of patients; Sending the normalized training samples to a personalized deep convolutional neural network (P-DCCN) and initiating model pre-training and model parameter updating using the P-DCCN with the normalized training samples (906); storing 908 the pre-trained model on a remote server for use in local machine personalization during preparation for treatment; generating a normalized fine-tuning set as output from the P-DCCN by processing the input personal data of the patient from the local machine (906); iteratively fine-tune the model parameters of the P-DCCN by performing multiple training iterations using the generated normalized fine-tuning set (912); generating 914 a personalized prediction score for a future medical event for the patient using the P-DCCN; Controlling (916) an operation of a medical device as a function of the personalized prediction score for a future medical event.

2. y i is the true indicator of the occurrence rate of the event of the i-th sample in the training sample, [0010] Let i be the predicted score of the i-th sample, N be the total number of training samples, θ represent the model parameters, and λ be a hyperparameter that controls regularization of the model parameters to avoid overfitting during the training. The P-DCCN is [0025] The method of claim 1 , wherein the regression loss function is fine-tuned by optimizing with

3. The method of claim 1 , wherein the pre-processing of the received historical patient data for a plurality of patients further comprises segmenting the recorded data to generate time series samples.

4. The method of claim 1 , wherein the static channel is used to process static and relatively infrequent temporal features, and the temporal channel is used to process relatively frequent temporal features. 【Request 5】 【Number 3】 but [0045] represents multiple layers of a fully connected network in the form [0050] are the model parameters to be trained, the static channel is s Information in [006] The method of claim 4 , further comprising: a multi-layer perceptron (MLP) configured to encode the signal h s into a compact static representation h s according to

6. The time-dependent channel is a sequence of vectors x 1 , . . . , x T Processing the time-series feature represented by [0070] When contains multiple layers of LSTM units, each containing trainable model parameters, [0080] A set of compact representations h 1 ,...,h T 5. The method of claim 4, further comprising a plurality of long short-term memory (LSTM) layers for outputting:

7. The static representations received from the static and temporal channels s and time-dependent expression d The probability of occurrence of a medical event is calculated by concatenating [0090] is a score indicating the occurrence probability of the medical event, [0010] The method of claim 1 , further comprising: calculating the occurrence probability using a multi-layer perceptron (MLP) according to

8. 1. A system for predicting an occurrence of a medical event in a patient using a trained neural network, comprising: A processor (104) operably connected to a computer readable storage medium (122), generating 904 a plurality of normalized training samples by pre-processing the received historical patient data for a plurality of patients; Sending the normalized training samples to a personalized deep convolutional neural network (P-DCCN) and initiating model pre-training and model parameter updating using the P-DCCN with the normalized training samples (906); storing 908 the pre-trained model on a remote server for use in local machine personalization during preparation for treatment; generating a normalized fine-tuning set as output from the P-DCCN by processing the input personal data of the patient from the local machine (906); iteratively fine-tune the model parameters of the P-DCCN by performing multiple training iterations using the generated normalized fine-tuning set (912); generating 914 a personalized prediction score for a future medical event for the patient using the P-DCCN; The system is configured to control (916) an operation of a medical device as a function of the personalized prediction score for a future medical event.

9. y i is the true indicator of the occurrence rate of the event of the i-th sample in the training sample, ##EQU00011## Let i be the predicted score of the i-th sample, N be the total number of training samples, θ represent the model parameters, and λ be a hyperparameter that controls regularization of the model parameters to avoid overfitting during the training. The P-DCCN is ##EQU00012## The system of claim 8, wherein the regression loss function is fine-tuned by optimizing with

10. The system of claim 8 , wherein the pre-processing of the received historical patient data for a plurality of patients further comprises segmenting the recorded data and generating time series samples.

11. The system of claim 8 , wherein the static channel is used to process static and relatively infrequent temporal features, and the temporal channel is used to process relatively frequent temporal features. 【Request 12】 【Number 13】 but ##EQU14## represents multiple layers of a fully connected network in the form ##EQU00015## are the model parameters to be trained, the static channel is s Information in ##EQU00016## 12. The system of claim 11, further comprising a multi-layer perceptron (MLP) configured to encode the signal hs into a compact static representation hs according to

13. The time-dependent channel is a sequence of vectors x 1 , . . . , x T Processing the time-series feature represented by ##EQU00017## When contains multiple layers of LSTM units, each containing trainable model parameters, [0018] A set of compact representations h 1 ,...,h T 12. The system of claim 11, comprising a plurality of long short-term memory (LSTM) layers for outputting:

14. The processor receives the static representations h from the static channel and the temporal channel. s and time-dependent expression d The probability of occurrence of a medical event is calculated by concatenating [0019] is a score indicating the probability of occurrence of a medical event, [0020] The system of claim 8 , further configured to calculate the occurrence probability using a multi-layer perceptron (MLP) according to

15. 1. A non-transitory computer readable recording medium comprising a computer readable program for predicting the occurrence of a medical event in a patient using a trained neural network, the non-transitory computer readable recording medium comprising: When the computer readable program is executed on the computer, the computer Pre-processing the received historical patient data for a plurality of patients to generate a plurality of normalized training samples (904); Sending the normalized training samples to a personalized deep convolutional neural network (P-DCCN) and initiating model pre-training and model parameter updating using the normalized training samples (906); Storing (908) the pre-trained model on a remote server for use in local machine personalization during preparation for treatment; Processing the input personal data of the patient from the local machine to generate a normalized fine-tuning set as output from the P-DCCN (906); iteratively fine-tune the model parameters of the P-DCCN by performing multiple training iterations using the generated normalized fine-tuning set (912); generating 914 a personalized prediction score for a future medical event for the patient using the P-DCCN; A computer readable medium for controlling (916) an operation of a medical device as a function of the personalized prediction score for a future medical event.

16. y i is the true indicator of the occurrence rate of the event of the i-th sample in the training sample, ##EQU00021## Let i be the predicted score of the i-th sample, N be the total number of training samples, θ represent the model parameters, and λ be a hyperparameter that controls regularization of the model parameters to avoid overfitting during training. The P-DCCN is [0022] The computer-readable storage medium of claim 15, wherein the regression loss function is fine-tuned by optimizing with

17. 16. The computer-readable medium of claim 15, wherein the static channel is used to process static and relatively infrequent temporal features, and the temporal channel is used to process relatively frequent temporal features. 【Request 18】 【Number 23】 but [0024] represents multiple layers of a fully connected network in the form [0025] are the model parameters to be trained, the static channel is s Information in [0026] 20. The computer readable medium of claim 17, further comprising a multi-layer perceptron (MLP) configured to encode the signal hs into a compact static representation hs according to

19. The time-dependent channel is a sequence of vectors x 1 , . . . , x T Processing the time-series feature represented by [0027] When contains multiple layers of LSTM units, each containing trainable model parameters, [0028] A set of compact representations h 1 ,...,h T 20. The computer readable medium of claim 17, further comprising a plurality of long short term memory (LSTM) layers for outputting:

20. The static representations received from the static and temporal channels s and time-dependent expression d The probability of occurrence of a medical event is calculated by concatenating [0029] is a score indicating the probability of occurrence of a medical event, [0030] 20. The computer readable storage medium of claim 19, wherein the occurrence probability is calculated using a multi-layer perceptron (MLP) according to

Images