Machine learning includes a neural network of impulse neural networks for predicting a time of occurrence of a clinical event

Abstract

The present disclosure relates in particular to a computer-implemented method for a machine learning neural network comprising a pulse neural network for predicting a time of occurrence of a clinical event. The method comprises providing a training dataset of training examples. Each training example comprises one or more time series of values of a respective physiological indicator of a patient over a past time period and a corresponding ground truth of a time of occurrence of a clinical event. The method further comprises training the neural network based on the training dataset. The neural network is trained to take as input one or more time series of values of a respective physiological indicator of a patient over a past time period and to output a prediction of a probability of occurrence of a clinical event within a time period. The method is an improved solution for predicting a time of occurrence of a clinical event.

Description

Technical Field

[0001] This disclosure relates to the field of computer programs and systems, and more particularly, to a method, system, and program for machine learning neural networks, including spiking neural networks for predicting the timing of clinical events. Background Technology

[0002] In the context of predicting a patient's disease progression, it is beneficial to employ mathematical models that can estimate the probability of clinical events occurring over time. These events can include disease relapse, progression to a more severe stage, complications, hospitalization, or patient death. The primary objective of using such models is to predict the occurrence of clinical events, such as death, as accurately as possible.

[0003] The mathematical models used for these predictions encompass a range of approaches, including statistical models and machine learning algorithms. The input data for these models typically consists of longitudinal patient information collected at multiple time points during consecutive visits to medical specialists. This data can include laboratory results, imaging findings, and vital signs. By utilizing such data, models can identify patterns and trends that would otherwise be difficult to discern visually, thereby improving predictive accuracy. Therefore, these models can indirectly measure medical / physiological characteristics based on the patient's already measured physiological data.

[0004] Existing survival analysis methods include statistical methods, such as joint models, as exemplified by Ferrer L., Rondeau V., Dignam J., Pickles T., Jacqmin-Gadda H., and Proust-Lima C. (2016). Joint modelling of longitudinal and multi-state processes: application to clinical progressions in prostate cancer. Statistics in medicine, 35(22), 3933-3948. However, such methods are limited by the assumptions made during the modeling process. Examples of these assumptions include a log-linear relationship between patient characteristics and the risk of clinical events, or a linear relationship between longitudinal patient characteristics and baseline patient characteristics.

[0005] Existing survival analysis methods also include the use of neural networks. However, in medical data, the time between measurements is often irregular (see, for example, discussed below). Figure 1 This can lead to a decrease in the performance of neural networks based on recurrent networks.

[0006] In this context, there is still a need for an improved solution for predicting the timing of clinical events. Summary of the Invention

[0007] Therefore, a method for a computer implementation of a machine learning neural network is provided, the neural network including a spiking neural network (SNN) for predicting the timing of clinical events. In this disclosure, this method may be referred to as a "learning method".

[0008] This learning method includes obtaining a training dataset of training examples. Each training example includes a time series of one or more values ​​of a corresponding physiological indicator of a patient over a past time period and a corresponding ground truth value for the occurrence time of a clinical event. The method also includes training a neural network based on the training dataset. This neural network is trained to take the time series of one or more values ​​of a corresponding physiological indicator of a patient over a past time period as input and output a prediction of the probability of a clinical event occurring within a given time period.

[0009] This learning method may include one or more of the following:

[0010] - Each training example also includes the patient's baseline data, and the neural network is configured to take the baseline data into account to output a prediction;

[0011] - The neural network includes an SNN and an additional module. The SNN is configured to take one or more value time series as input and output a series of spikes. The additional module is configured to take a series of spikes and baseline data as input and output a prediction.

[0012] - The additional module uses the SoftMax function to output predictions;

[0013] - The neural network includes an encoding layer that is configured to take one or more value time series as input and encode the one or more value time series into a corresponding pulse sequence;

[0014] Training involves minimizing a loss for each training example, which is penalized by the difference between the neural network's prediction for that training example and the ground truth value of that training example; and / or

[0015] - Losses can be of the following types:

[0016] ,

[0017] in, It is the number of training examples. Indicates the patient The baseline truth value for the occurrence of clinical events (1 if present, 0 otherwise). It refers to the timing of the clinical event. Indicates training example Until time Data, It is a known training example. Until time Given all the data, for the training example In time The predicted probability of a clinical event occurring, where,

[0018] ,

[0019] in, Training example The time when the final indicator is known.

[0020] And among them, The patient In time The probability of survival, and it is of the following type:

[0021] .

[0022] A neural network capable of being obtained according to a learning method is also provided, namely, a computer-implemented neural network having the weights and parameters of a neural network learned through the learning method. The provided neural network can, for example, have already been directly learned through the learning method, and its weights and parameters have been fixed through the learning steps of the learning method.

[0023] A computer-implemented method using a neural network learned according to a learning method is also provided. In this disclosure, this method may be referred to as the "method of use".

[0024] The method of use includes obtaining time series of one or more values ​​of a patient's corresponding physiological indicators over a past time period. The method of use also includes applying a neural network to the one or more value time series to predict the probability of a clinical event occurring within a time period.

[0025] A computer program including instructions for performing and / or using the method is also provided.

[0026] An apparatus is also provided, which includes a data storage medium on which computer programs and / or neural networks are recorded.

[0027] The device can form or serve as a non-transitory computer-readable medium, such as on SaaS (Software as a Service) or other servers, or cloud-based platforms. The device may alternatively include a processor coupled to the data storage medium. Therefore, the device can form, wholly or partially, a computer system (e.g., the device is a subsystem of an overall system). The system may also include a graphical user interface coupled to the processor. Attached Figure Description

[0028] A non-limiting example will now be described with reference to the accompanying drawings, in which:

[0029] - Figures 1 to 10 Examples of implementations of the method are shown; and

[0030] - Figure 11 An example of the system is shown. Detailed Implementation

[0031] A method for computer implementation of a machine learning neural network is proposed, comprising a spiking neural network (SNN) for predicting the timing of clinical events. In this disclosure, this method may be referred to as a "learning method".

[0032] This learning method includes obtaining a training dataset of training examples. Each training example includes a time series of one or more values ​​(e.g., measurements) of a corresponding physiological indicator for a patient (each of which is the same patient in one or more value time series) over a past time period (i.e., each value time series is a time series of the corresponding physiological indicator) and a corresponding ground truth value for the occurrence time of a clinical event. The method also includes training a neural network based on the training dataset. The neural network is trained to take the time series of one or more values ​​of the corresponding physiological indicator for the patient over a past time period as input and output a prediction of the probability of a clinical event occurring within a time period.

[0033] In this disclosure, a physiological indicator is any patient-related characteristic that can be obtained and / or measured, for example, by a health professional (e.g., using medical measuring devices). A physiological indicator is an indicator that changes over time, such as, by way of non-limiting example, blood pressure, heart rate, body temperature, blood test results, urine test results, radiographic results, or combinations thereof. Physiological indicators are also referred to as longitudinal indicators. Physiological indicators can be obtained, for example, using medical measuring devices and / or obtained in a laboratory setting.

[0034] In this disclosure, a clinical event refers to any observable and / or measurable event that is of significant importance to the patient's health. A clinical event is associated with a time of occurrence, which is the point in time when the clinical event first occurs (e.g., an hour, day, or month). A clinical event can be any measurable or observable change in the patient's overall health status, including the patient's death. Additionally or alternatively, a clinical event may be associated with a specific disease and may include, but is not limited to, the onset, recurrence, or subsequent recovery of the disease.

[0035] In this disclosure, the prediction of the probability of a clinical event occurring within a time period can be a probability sequence within that time period (e.g., the probability for each day or month within that time period, such as the probability for 300 days or a future year), indicating the likelihood of the clinical event occurring at each specified point in time within that time period. These point in time can be regularly distributed within that time period (e.g., daily / monthly within that time period). The probability can indicate the level of confidence the method has in predicting the occurrence of a clinical event. The aforementioned time period can, for example, coincide with the time period of the training data, or it can be a future time period (e.g., the time period can be a time span starting immediately after the method is used). Therefore, the method of this disclosure can be used both to predict future clinical events and to obtain information about the first occurrence of a clinical event that has already occurred (e.g., the first occurrence of cancer recurrence in the past). The time span can be days, months, or years, such as 300 days or 3 years.

[0036] In embodiments of the methods disclosed herein, clinical events may be associated with cancer, and examples of clinical events may be cancer recurrence, characterized by the reappearance or progression of malignant cells (e.g., after a remission period). The prediction of the timing of recurrence may be the expected moment of cancer recurrence, and the associated probability may be the confidence level of the method in the prediction.

[0037] In this disclosure, the time series of values ​​for a (corresponding) physiological indicator consists of one or more values ​​(e.g., measurements) of a patient's physiological indicator, each value associated with a time point (also called a time step). Each time point may represent a specific moment (e.g., an hour, a day, or a month) that indicates when the value was obtained (e.g., the moment the measurement was performed). In the example, each of the time series of one or more values ​​of a patient's corresponding physiological indicator may include one or more time series, each time series comprising entries, each entry including the value and / or measurement of the patient's corresponding physiological indicator and the date (e.g., an hour, a day, or a month) on which the measurement was performed. In the example, each time series may be represented as a vector, with the indices of the vector corresponding to time points, and the value at each index corresponding to the value and / or measurement of the patient's corresponding physiological indicator at the corresponding time point. In the formula, each time series may be represented as a sequence. ,index Representing a point in time (e.g., each index is associated with a different hour, day, and / or month), and the value... Indicates at a point in time The values ​​of the corresponding physiological indicators.

[0038] The method disclosed herein may include obtaining a time series of one or more values ​​of a patient's corresponding physiological indicator by collecting measurements of the patient's physiological indicator at different time points, for example, one or more values ​​at each visit to a healthcare professional. Thus, one or more values ​​of the physiological indicator may be obtained, for example, by a healthcare professional, such as by using medical measurement equipment. For example, a patient may visit a healthcare professional regularly and undergo blood / urine tests during each appointment. Therefore, each time series of a patient's corresponding physiological indicator may consist of values ​​of that physiological indicator obtained at irregular intervals (e.g., corresponding to the patient's visits to a healthcare professional). The irregularity of these intervals poses a significant challenge to training neural networks and deep learning algorithms. In fact, many models used in prior art methods do not and are unable to account for these irregularities.

[0039] The method disclosed herein may include normalizing the time series values ​​of relevant physiological indicators of the obtained patient. In the formula, the value... Normalization can be performed using any method known in the art (e.g., considering the mean and / or standard deviation).

[0040] Figure 1 This example illustrates two patients who visited the clinic at irregular intervals.

[0041] Furthermore, training examples that include time series of one or more physiological indicators of the patient over a past period and corresponding ground truth values ​​of clinical events may be incomplete. This could be due to patient data becoming unavailable after a period of time, for example, due to the patient being transferred to another hospital (e.g., another city) or the completion of a clinical study / trial. This poses additional difficulties when training deep learning models based on such data.

[0042] Figure 2 Examples of the medical history and clinical events of four different patients are shown.

[0043] Time series analysis of relevant physiological indicators (e.g., analysis of longitudinal feature evolution) can be important for understanding disease progression, assessing response to medical interventions, and evaluating the overall impact of pathology on patients. For example, some physiological indicators or their changes / evolutions over time may be associated with specific diseases. In a specific example, the evolution of the presence and / or concentration of certain proteins may be important for understanding cancer evolution. In a specific example, prostate-specific antigen (PSA) and its evolution over time may be important for prostate cancer because it suggests it can serve as an indicator of recurrence in this type of cancer.

[0044] However, in most cases, the correlation between the time series values ​​of corresponding physiological indicators and (e.g., disease-related) clinical events can be complex. In the example, a combination of the evolution of several different physiological indicators (including different weights and complex mathematical functions) could be the optimal predictor for a given (e.g., disease-related) clinical event. Because such combinations can be very complex and disease-specific, it may be infeasible to equip each clinical event associated with each disease with (e.g., different or specific) mathematical models. Health professionals often make these predictions by mentally comparing a patient's specific situation with other cases they know, relying on their (biased) personal experience. Such predictions are often imprecise and may be heavily dependent on the health professional.

[0045] The learning method is an improved solution that learns neural networks to predict the timing of clinical events.

[0046] In fact, this learning method provides a framework applicable to any possible clinical event, such as disease-related clinical events. A neural network trained using this method can automatically learn from a training dataset to predict the occurrence of clinical events. As a non-limiting example, the dataset could be a clinical study targeting a given clinical event and / or disease. Therefore, the learning method can train a neural network that, starting from data obtained from clinical studies, outputs predictions of the probability of a clinical event occurring over a given time period. Thus, the learning method eliminates the need for a specific mathematical model for each different clinical event, providing a unified framework.

[0047] In implementations of the learning method, clinical events can be cancer-related, such as cancer recurrence or patient death from cancer. Each physiological indicator can be an indicator with a specific relevance to cancer. As a non-limiting example, a physiological indicator can be (e.g., the concentration of a specific molecule (such as a protein) in the body, as detected in a blood test or urine test). Alternatively, a physiological indicator can be the size of a tumor (e.g., diagonal, volume), as measured by radiographic imaging. Data can be obtained from clinical studies of cancer. Thus, the learning method can, based on the aforementioned measurements, machine learn / train a neural network to predict cancer recurrence and / or patient death from cancer.

[0048] Health professionals can use neural networks trained using learning methods to machine learning. They can feed longitudinal data of patients accumulated over time (i.e., time series of one or more values ​​of the patient's corresponding physiological indicators) as input to the trained neural network. Therefore, using the neural network, health professionals can measure and / or predict the (e.g., future) occurrence of clinical events (such as death or illness due to cancer) with an accuracy unattainable, for example, by guessing based on professional experience. This measurement can be used to obtain indirect indicators of treatment efficacy or disease progression. Furthermore, as illustrated in the following embodiments, the method can be used to obtain indirect indicators of tumor size.

[0049] In this implementation, the learning method was tested on a dataset containing time series values ​​of relevant physiological indicators from patients with metastatic prostate cancer. Metastatic prostate cancer is an advanced stage of the disease. In this case, the tumor has spread to other parts of the body besides the tissues surrounding the prostate. The five-year survival rate for patients with metastatic prostate cancer is approximately 50%. The data used in this implementation came from four different studies on patients with metastatic prostate cancer. The data included longitudinal variables (such as PSA) and general patient information variables (such as BMI and medical history). Prostate-specific antigen (PSA) is a protein produced by prostate cells. Changes in blood PSA concentration are an important marker of tumor size evolution. The dataset used included data from over 1000 prostate cancer patients. For each patient, the data consisted of discrete longitudinal data, obtained over several visits to a healthcare professional, with the first visit corresponding to the start of the clinical study. The clinical event studied was patient death over time. This implementation was able to accurately predict the probability of patient death over time.

[0050] Furthermore, unlike existing algorithms that do not consider the irregularities between patient visit times in prediction tasks, the neural network of the above-described learning method takes into account the temporal irregularities in the time series of each value of the corresponding physiological indicator. This is thanks to spiking neural networks. As will be described in detail below in this disclosure, spiking neural networks possess a temporal dimension and are capable of correctly processing data obtained at irregular time intervals.

[0051] Therefore, compared with existing technical methods, neural networks learned through learning methods can predict clinical events more accurately.

[0052] A computer implementation method for using a neural network learned according to this learning method is also proposed. In this disclosure, this method may be referred to as the "method of use".

[0053] The method of use includes obtaining time series of one or more values ​​of a patient's corresponding physiological indicators over a past time period. The method of use also includes applying a neural network to the one or more value time series to predict the probability of a clinical event occurring within a time period.

[0054] In implementations, the learning method and the usage method can be combined. In implementations, any method may include steps of other methods (e.g., complete other methods).

[0055] For the reasons given above, the method constitutes an improved solution for predicting patient clinical events. The method can be used to accurately predict the timing of clinical events.

[0056] In one implementation, the clinical event can be a recurrence of a disease (such as cancer). Cancer recurrence can be characterized and measured by the reappearance or progression of malignant cells, for example, after a period of remission. Therefore, the method can be used to measure the progression of malignant cells. In other implementations, the clinical event can be the death of a patient.

[0057] This implementation method can be used, for example, by healthcare professionals. Therefore, healthcare professionals may be able to measure the efficacy of a treatment by providing predictive analyses of its impact on cancer recurrence and / or patient mortality.

[0058] In this disclosure, and as is known from machine learning itself, a neural network can be defined by its architecture, parameters, and hyperparameters. The architecture consists of layers, starting with an input layer whose neuron count equals the dimension of the input data. Following the input layer can be several hidden layers with a given number of neurons and activation functions. These layers and neurons define the depth and width of the network. The neural network may include, for example, inter-layer activation functions that can introduce non-linearity into the model. The output layer may have as many neurons as there are variables in the output data. In an implementation, the output layer may include neurons for each considered time point within a time period. The interconnections between these layers define the topology of the neural network. The parameters of the neural network are learnable weights and biases that are determined / configured / modified during training. In contrast, hyperparameters are predefined values / settings that are not learned from the training data. These include the number of hidden layers, the number of neurons per layer, and so on.

[0059] The main difference between spiking neural networks (SNNs) and traditional neural networks lies in the structure of each artificial neuron unit. SNNs are composed of spiking neurons. Spiking neurons more closely resemble biological neuronal units and can process temporal data. Learning and application methods leverage this ability to process temporal data to account for the irregularities in the time series of each value of a patient's corresponding physiological indicator. Compared to existing technology models, this ability to process temporal data makes the learned neural network (depending on the learning method) more accurate in predictions. Spiking neurons take signals as input and output signals, each signal consisting of a series of pulses. Spiking neurons utilize a membrane potential system, heavily inspired by mathematical models of biological neurons. Spiking neurons include thresholds used to determine when and / or whether they should fire pulses.

[0060] In this disclosure, a spiking neural network (SNN) can be defined by its architecture, parameters, and hyperparameters. The architecture consists of (hidden) layers of spiking neurons that communicate through discrete events (spiks) rather than continuous activation. The number of (hidden) layers and the number of neurons per layer can vary depending on each implementation and can be determined after several benchmark tests (e.g., depending on the training dataset and / or depending on clinical events). The first layer of the spiking neural network can be an input layer, where each neuron can correspond to a dimension or feature of the input data. Following the input layer, one or more hidden layers of the spiking neurons process the incoming signals.

[0061] In this disclosure, each spiking neural network and each spiking neuron constituting the spiking neural network has a time dimension. The time dimension can be discretized and can depend on several time points. The difference between consecutive time points (the time interval herein) can be constant, such that the time points can be uniformly distributed over a period of time. At each time point, each spiking neuron can receive pulses as input from one or more connected spiking neurons and / or emit pulses as output to these spiking neurons. The discretization, i.e., the length of the time interval, can depend on the specific implementation.

[0062] In implementations, each method of this disclosure may include: normalizing, in the temporal direction, the time series of each value of the corresponding physiological indicator of the patient over a past time period before inputting each time series into the SNN. Normalization means that the time intervals (i.e., the differences between consecutive time points) can all be multiples of a base time interval. Normalization may include approximating each time point as the closest time point consisting of multiples of the base time interval. The base time interval may be one day or more days, or one week or more weeks, for example, one day or one month. In such implementations, the time series may be provided as input to a neural network learned using methods and / or learning methods, and the base time interval may be the same as the time interval used in the SNN, as described above. Alternatively, in other implementations, the neural network may include normalizing, in the temporal direction, the time series of each value of the corresponding physiological indicator of the patient over a past time period provided as input.

[0063] In this disclosure, each given spiking neuron may have an internal memory (e.g., membrane potential) that is time-dependent. This memory may be represented by a real number and may include information about the number of past pulses received by the spiking neuron. The memory of past pulses may decrease over time, such that pulses received in the more distant past contribute less to the memory.

[0064] In embodiments of the method disclosed herein, each given spiking neuron may have a time-dependent variable. membrane potential And can at every time Output value (For example, if) Then a pulse is emitted, and if (Then no pulse is emitted). Membrane potential and output Updates can be made at each of the following time steps:

[0065]

[0066]

[0067] in:

[0068] - At a certain point in time The sum of all received pulses (or, equivalently, all quantities of the corresponding spiking neurons). The sum of all spiking neurons connected (forward) to a given spiking neuron;

[0069] - It is the attenuation factor; and

[0070] - It is a threshold.

[0071] Figure 3 A diagram of a spiking neuron is shown.

[0072] In embodiments of the method disclosed herein, each pulse can contribute a value of 1. In other words, each pulse can be encoded as an output value of 1. For example, if at time... If a spiking neuron outputs a pulse, then the function defined above... It can be 1, otherwise It can be 0.

[0073] In other embodiments of the method disclosed herein, each pulse can be a positive pulse or a negative pulse. A positive pulse can contribute a value of 1, while a negative pulse can contribute a value of -1. In other words, each positive pulse can be encoded as an output value of 1, and each negative pulse can be encoded as an output value of -1. In the formula, if at time... If a spiking neuron outputs a positive pulse, then the function defined above... It can be 1, if in time If the pulsatile nerve outputs a negative pulse, then It can be -1, otherwise It can be 0. In such implementations, each given spiking neuron may include a positive threshold. and negative threshold This allows for the determination of membrane potential whenever (e.g., memory). When the voltage is above a positive threshold, a positive pulse is output / emitted, and whenever the membrane potential (e.g., memory) is above a positive threshold, a positive pulse is output / emitted. When the value is below the negative threshold, a negative pulse is output / issued. In an implementation, the negative threshold may be the opposite of the positive threshold. .

[0074] In further embodiments of the method disclosed herein, each pulse may be associated with a number (e.g., an integer or any other subset of real numbers) that can be directed to the sum described above. The contribution is equal to the value of the number it is associated with, and the spiking neural network can include a corresponding threshold for each different number that can be associated with the spiking.

[0075] Therefore, in this disclosure, a spiking neuron can take a signal as input and output a signal. The signal can be generated from time points. The index is a sequence, so its length can be the number of time points, and at each time point it records whether the associated spiking neuron has output a pulse (and if so, what kind of pulse it was).

[0076] Figure 4 The signal shown consists of several pulses.

[0077] In one implementation, the signal can be represented by a sequence of binary values, for example, each value in the sequence being either "0" or "1". In such implementations, a value of 0 can correspond to a pulse that is absent, and 1 to a pulse that is present. In other implementations, the signal can be represented by a sequence of values, each value being one of a list including "0", "-1", and "1". These values ​​can correspond to pulses that are absent, negative, and positive, respectively. In yet another implementation, the signal can be a sequence of numbers that identify associated pulses.

[0078] The learning method involves obtaining a training dataset of training examples. As explained below in this disclosure, the training dataset will be used for machine learning / training neural networks. Each training example includes a time series of one or more values ​​of a corresponding physiological indicator of a patient over a past time period. Each training example is associated with a patient and may be associated with one or more physiological indicators, the training example including a time series of values ​​for each of the patient's physiological indicators. As explained above, this time series of values ​​can be obtained by repeatedly measuring the patient's physiological indicators over time, and the time series also includes the time points at which each value was obtained. These values ​​were obtained over a past time period; that is, the time points are within a past time period. These values ​​may have already been obtained by a healthcare professional, for example, by using medical measurement equipment. Any training dataset in this document may include a large number of training examples, for example, greater than 10,000 or 50,000.

[0079] Each training example also includes a corresponding ground truth value for the occurrence time of the clinical event. That is, each training example includes information on whether a clinical event (e.g., cancer recurrence or patient death) has occurred in the patient associated with the training example. If the clinical event has occurred, the training example includes the occurrence time of the clinical event. The ground truth value is important for training the neural network because during training, the results obtained after inputting one or more time series into the neural network are compared to the ground truth value.

[0080] In implementation, the training dataset for obtaining training examples may include data preparation and / or data preprocessing. In reality, different clinical studies may involve different physiological indicators, and there may be significant data gaps.

[0081] In this implementation, the past time period can be a time window (e.g., the duration of a clinical trial / study) and can be fixed. A time window can include two dates (a start date and an end date) separated by a time span. This time span can be one year or more, such as greater than one year or five years.

[0082] Obtaining the training dataset for training examples may include retrieving at least a portion of the training dataset from (e.g., local or remote) storage, or receiving (e.g., from a remote system) at least a portion of a training dataset already acquired therefrom. Acquisition may include obtaining at least a portion of the training dataset by measuring patients' physiological indicators and recording the occurrence of patients' clinical events. Acquisition may include obtaining at least a portion of the training dataset from clinical trials / research.

[0083] The training dataset obtained in the learning method includes training examples, each consisting of time series values ​​of one or more corresponding physiological indicators of a patient. Each physiological indicator may be correlated with a clinical event (e.g., each physiological indicator may be correlated with a clinical event to be predicted). That is, the values ​​of physiological indicators (e.g., measurements) may be correlated with clinical events, but they may not be independent. It may be important to use only physiological indicators that are correlated with clinical events because a neural network trained using the training dataset can learn the correlations. Therefore, predictions made by the machine learning-trained neural network can be based on these correlations. In practice, using irrelevant indicators may bias the neural network in an undesirable way, causing it to learn correlations that are purely accidental.

[0084] In this implementation, for each training example, the training dataset obtained in the learning method may include an ID corresponding to a patient. This allows the neural network to correlate values ​​of different physiological indicators of the same patient to predict the same clinical event. Each training example of the learning method may also include baseline data for the patient. Baseline data may include any patient-dependent data (e.g., the patient's age) that can be used to predict clinical events and remains relatively constant over time or independent of changes in a particular patient over time. In the example, the baseline data may vary in the same way for each patient (e.g., linearly for each patient, such as the patient's age). As a non-limiting example, baseline data may include age, sex, BMI, height, genetic characteristics, medical history, and / or other (Boolean) values ​​related to the patient's state before the clinical study. The baseline data depends only on the patient, not on a specific point in time. Baseline data enables the neural network to better predict clinical events.

[0085] The method was implemented, and a training dataset was obtained. The training dataset can be visualized in tabular form and includes the table below.

[0086]

[0087] The learning method involves training a neural network based on a training dataset. Training the neural network (or equivalently, machine learning) involves providing entries (e.g., training examples) from the training dataset (obtained in the learning method) as input to the neural network. Each entry in the training dataset comprises one or more time series values ​​of a corresponding physiological indicator of a patient over a past time period (e.g., each time series is a portion of the column corresponding to the patient in the table above). Each value time series is processed by the neural network by multiplying the values ​​by weights, applying an activation function, and combining the results across layers. The processing may include a time dimension, as the neural network includes an SNN. In an implementation, each entry may include baseline data of the patient, and the baseline data may also be processed by the network as explained above. The input layer for the baseline data may differ from the input layer for the time series data. Training may then include comparing the output of the neural network with a baseline ground truth of the occurrence time of clinical events included in the training examples. This comparison may include / compute an error. This comparison may include outputting the error based on a loss function. The loss function may quantify the difference between the prediction and the expected outcome. The error (e.g., the error of the output) may then be used for backpropagation through the network, resulting in adjustments to the weights to minimize the error. During training, various optimization solvers can be employed, each with its own set of hyperparameters, to efficiently converge to the optimal parameter set. Therefore, as is known in the field of machine learning itself, training a neural network (e.g., a neural network based on learning methods) involves adjusting the network's weights, biases, and other parameters so that the processing of inputs can accurately predict the occurrence of clinical events.

[0088] The neural network for the learning method is trained to use the same type of input as during training (i.e., time series of one or more values ​​of the corresponding physiological indicators of the patient over a past time period). If baseline data is used during training, the trained neural network can also use baseline data as input.

[0089] The neural network of the learning method is trained to output predictions of the same type as the baseline ground truth used during training, i.e., predictions of the occurrence of a clinical event within a time period (e.g., in the future). This prediction includes the probability (e.g., confidence level) of the clinical event occurring within that time period. This time period can be equivalent to the time period used in each time series (e.g., a time window, the length of a clinical trial / study) (e.g., it can be the same time period), or it can be shorter. In implementations, training the neural network may include setting a time period (e.g., in days or years, e.g., less than 3 or 5 years). This time period may correspond to or be shorter than the time window associated with the time series. In such implementations, the trained neural network can output the probability of a clinical event occurring at each time point within the time period (e.g., daily or monthly). The sum of all probabilities at all time points can be 1 (indicating that the clinical event is expected to occur within that time period).

[0090] Training a learning method may include minimizing a loss (e.g., a loss function). For each training example, the loss may penalize the difference between the prediction output by the neural network for the training example and the ground truth value of the training example. The loss function may be a distance, or any function used to compute the difference and / or comparison between the output and the ground truth value, such as based on mean squared error and / or mean absolute error. In implementations, the loss (e.g., the loss function) that can be minimized in the learning method may be of the following type (i.e., given by the following formula):

[0091] ,

[0092] in, It is the number of training examples. Indicates the patient (For example, with index) The baseline truth value for the occurrence of a medical event (if the associated patient) is 1, otherwise 0. It is the time of the event. Indicates training example Until time Data, It is a known training example. Until time Given all the data, for the training example In time The predicted probability of a medical event occurring, where,

[0093] ,

[0094] in, Training example Finally, the time for the indicator is known, and among them, The patient In time The probability of survival and is of the following type:

[0095] .

[0096] These probabilities can form what is known as a "survival function." This loss function fits well into neural networks, allowing for efficient and improved training of them.

[0097] In implementations of the learning and / or usage methods, the neural network may include (e.g., composed of) several modules (e.g., a spiking neural network, additional modules, and / or additional layers of a spiking neural network). In such implementations, training the neural network may include training each of the several modules, for example, adjusting the weights, biases, and other parameters of each of the several modules. These modules may be trained together during neural network training.

[0098] Figure 5 An example of an implementation of a neural network trained in a learning method and the method of use is shown. The neural network of the implementation includes several modules, including an encoding layer 50, a longitudinal data processing module 51 (which may be an SNN), a baseline data insertion module 52, and a final processing module 53.

[0099] In implementations of the learning and / or usage methods, in addition to the SNN, the neural network may also include one or more functions and / or one or more layers that can, for example, form additional modules. The SNN can be configured to output one or more signals, each signal comprising a series of pulses. The additional modules can be configured to take the signals comprising a series of pulses output by the SNN as input. The additional modules can output predictions, i.e., predictions of the probability of a clinical event occurring over a time period. The additional modules can be trained simultaneously with the SNN.

[0100] The add-on module may include a transformation function for each spiking neuron in the final layer of the SNN, each transformation function taking the output of the corresponding spiking neuron as input and giving a scalar as output. The transformation function may be important because the output of each spiking neuron (i.e., a signal comprising a series of pulses) may not be directly processed by subsequent layers of the neural network (e.g., the add-on module). Subsequent layers may be expected to take a scalar (e.g., a vector of scalars) as input. In an implementation, the transformation function may be a frequency function. The frequency function can transform the signal (which includes a series of pulses) into a scalar vector. Within a certain time point The input is the next pulse, and the return value is the frequency of the pulse. For example, if the signal includes 20 time points and 15 pulses occur within those time points, the conversion function / frequency function can give 15 / 20 = 0.75 as the output. Figure 4 An example of a signal with 10 pulses at 20 time points is shown. In this case, the transfer function / frequency function can give 10 / 20 = 0.5 as the output. Therefore, the value of the frequency function can be... Within the range, however, variations of the conversion function / frequency function can also be used in other implementations.

[0101] Additional modules may include fully connected layers that take the output of a transformation function as input. Additionally or alternatively, additional modules may include survival neural networks (e.g., the survival neural network described in EP4057297A1, which is incorporated herein by reference).

[0102] Additional modules, which may be included in the neural network used in the method and / or learned in the learning method, may apply the SoftMax function to output a prediction. That is, the additional module may include the SoftMax function, which takes the output of the previous layer as input and outputs a prediction. In some implementations, the SoftMax function may be applied in the final step of the additional module (e.g., in the last layer). The output of the SoftMax function is a probability distribution. In some implementations, this probability distribution may be the output of a neural network; that is, the probability distribution may be the probability of a clinical event occurring over a time period. In some implementations, the output of the SoftMax function may indicate the probability of a clinical event occurring at each time point. For example, the additional module may include fully connected layers and / or a survival network, and the output of the fully connected layers and / or the survival network may be used as input to the SoftMax function.

[0103] In one example, the learning method can be used to machine learn / train a neural network to predict the occurrence of a clinical event (e.g., death) within 300 days. The SNN can have two hidden layers, for example, with 100 neurons. An add-on module can have a final layer with 300 neurons, each neuron corresponding to one day out of the 300 days. A SoftMax function, which can be included in the add-on module, can output the probability of a clinical event occurring on a given date for each day out of the 300 days. In an implementation, the hyperparameters and training parameters of the neural network can be as follows:

[0104]

[0105] This implementation method was used in the above implementation of the dataset, which includes time series values ​​of physiological indicators of patients with metastatic prostate cancer.

[0106] Therefore, the add-on module can transform the signals output by spiking neurons (i.e., by the SNN), converting them into scalars, processing these scalars, and applying the SoftMax function to generate a time-varying probability distribution. Thus, the add-on module can interpret the SNN's output and convert it into a user-understandable output.

[0107] The add-on module can also be configured to take baseline data as input. The baseline data can be taken from the last layer of the add-on module, such as... Figure 5 As shown in module 52. In an implementation, the additional module may include fully connected layers and / or a survival neural network that, in addition to receiving the output of the SNN (which may have been processed by the transformation function described above), can also take baseline data as input. The additional module can be beneficial because it can combine the output of the SNN, consisting of spikes, with the baseline data to obtain a practical prediction of clinical events. The ability to simultaneously consider longitudinal and baseline data in the same algorithm is an improvement over existing technology models. Most existing technology models do not and cannot simultaneously consider longitudinal and baseline data.

[0108] The neural network used in the method and / or the neural network obtained by machine learning / training in the learning method may include an encoding layer configured to take as input one or more time series of values ​​of a corresponding physiological indicator of the patient over a past time period, and encode each time series into a corresponding signal (i.e., a series of pulses). The encoding layer may be important because the time series of values ​​of the corresponding physiological indicators included in the training dataset and / or provided as input to the trained neural network may include real continuous values, and the SNN included in the neural network can be expected to take signals as input, that is, it can be expected to take pulse sequences as input (e.g., which can be encoded as binary / ternary signals).

[0109] The coding layer has a significant advantage: it breaks the convention that "spiking neurons always take pulse sequences as input" and enables the use of spiking neurons to encode data for other spiking neurons.

[0110] In implementations of the learning method and / or usage method, the encoding layer may consist of layers trained simultaneously with the neural network. That is, training the neural network included in the learning method may also include training the encoding layer (i.e., adjusting weights and other parameters). The advantage of training the encoding layer simultaneously with the rest of the neural network is that the weights and parameters are automatically determined during training, eliminating the need for manual configuration. The training process can automatically find the optimal weights and parameters for the encoding layer. During training, the weights of the encoding layer can be adjusted to improve the encoding.

[0111] Alternatively, in other implementations of the learning and / or usage methods, the coding layer may consist of pre-trained modules trained independently of (and prior to) the neural network. The coding layer may be trained using a training dataset that includes time series values ​​of the patient's corresponding physiological indicators and ground truth values ​​of encoded impulses. Training the coding layer independently can have the additional advantage that the same coding layer can be used with different neural networks.

[0112] Alternatively, in other implementations of the learning method and / or usage method, the encoding layer may be a deterministic layer, i.e., it is not learned during the training process.

[0113] In implementations of the learning and / or usage methods, the coding layer may include delta modulation. Delta modulation may be inspired by the way the retina perceives changes in its field of view and sends signals only when it detects a change. When no change is detected, no signal is transmitted. Delta modulation may be based on this premise and may be designed to analyze changes in a time series to generate a binary signal.

[0114] In implementations of the learning method and / or the method of use, incremental modulation may include: given a time series of values ​​of a patient's physiological indicators over a past time period, for each time point, calculating the difference between the value at that time point and the value at the previous time point. If the difference (or the absolute value of the difference) is greater than a predetermined threshold, a pulse is output at that time point. Otherwise, no pulse is output. In implementations, each pulse may be a positive pulse or a negative pulse. In such implementations, a positive pulse is output at that time point if the difference is greater than a predetermined positive threshold, and a negative pulse is output at that time point if the difference is less than a predetermined negative threshold. In implementations, the predetermined threshold may be determined in different ways; for example, the predetermined threshold may be defined using the mean and standard deviation of the difference.

[0115] The coding layer, including delta modulation, can be implemented as follows, and this coding layer can be included in both the neural network learned by this method and the neural network using the method. The time series can be represented as a sequence of real values. Here, the index corresponds to a point in time. That is, each index... It can correspond to a date (e.g., a specific hour, day, or month). Value It is the patient at a certain point in time. The values ​​of physiological indicators. In an implementation, incremental modulation may include calculating a difference sequence. In the formula, the difference sequence may be determined as follows:

[0116]

[0117] .

[0118] The coding layer can utilize delta modulation to transmit signals. For the sequence Encode the given time series, where the signal Defined as:

[0119] .

[0120] In implementations where the pulse can be either a positive or negative pulse, the coding layer can utilize delta modulation to transmit the signal. For the sequence Encode the given time series, where the signal Defined as:

[0121] .

[0122] In implementation, the above threshold It can be defined as ,in It can be bad The average value, It can be bad The standard deviation of . Value It can be a parameter (e.g., a real number), which can be adjusted within the method. The larger the value, the fewer pulses are recorded in the output signal. For example, the parameter... It can be adjusted according to the noise level of the time series.

[0123] Figure 6 This paper illustrates delta modulation methods applied to time series composed of noisy Gaussian functions (e.g., the sum of a Gaussian function and a white noise function) and time series composed of white noise functions. It can be seen that in the case of Gaussian functions, delta modulation successfully extracts the trends of positive and negative pulses, with very few pulses caused by noise. However, the delta modulation method is less effective when the time series is a white noise function. In the former case, the time series is globally stationary, yet several pulses are encoded. The presence of these pulses is due to a deviation in the threshold. In practice, the threshold is determined using the mean and standard deviation of the difference sequence.

[0124] Alternatively, the encoding layer can be a deterministic layer, i.e., no learning occurs during the training process. The encoding layer can be obtained using the encoding methods described in this invention. That is, in embodiments, the learning method and / or usage method can include an encoder employing this encoding method.

[0125] A computer-implemented method for encoding time series values ​​of a patient's physiological indicators is also provided. In this disclosure, this method may be referred to as an "encoding method".

[0126] The encoding method includes obtaining a time series of values ​​for a patient's physiological indicators. The encoding method further includes: for each time point in the value time series, obtaining a slope coefficient of the value time series at that time point. The slope coefficient is obtained based on (e.g., using calculations such as the following) that time point (e.g., based on the value of the physiological indicator associated with that time point), a predetermined number of previous time points (e.g., based on the values ​​of the physiological indicator associated with a predetermined number of previous time points), and a predetermined number of subsequent time points (e.g., based on the values ​​of the physiological indicator associated with a predetermined number of subsequent time points). The encoding method further includes: for each time point in the value time series, comparing the slope coefficient with a reference slope coefficient. The comparison of the slope coefficients includes: if the absolute value of the calculated slope coefficient is greater than the reference slope coefficient, then encoding that time point as a pulse. The comparison of the slope coefficients also includes: if the absolute value of the calculated slope coefficient is less than the reference slope coefficient, then encoding that time point as no pulse.

[0127] Therefore, the encoding method is a method of taking a time series of values ​​of a patient's physiological indicators as input and outputting an encoded sequence (e.g., a signal), which can be represented, for example, as a real-number vector with indices corresponding to time points, and the encoded sequence includes a series of pulses. Therefore, the encoding method can be used to implement an encoding layer in a neural network, such as the encoding layer discussed above. Thus, the encoding method can also include providing the neural network with encoded time series of values ​​of the patient's physiological indicators (e.g., encoding one or more time series of values ​​of the corresponding physiological indicators of the patient, and using each encoded time series as input to the neural network). The neural network can include a spiking neural network that predicts the timing of clinical / medical events. The neural network can be a neural network learned in a learning method and / or a neural network using a method. Therefore, in embodiments, the encoding method can be combined with the learning method and / or the using method (or with steps of either method), such that the encoding method can be included in or incorporated into either method. In such embodiments, each time series of the corresponding physiological indicator of the learning method and / or the using method can be preprocessed using the encoding method. In other embodiments, the encoding method can be used to implement an encoding layer in another neural network that requires a series of impulses as input, which may or may not include an SNN.

[0128] This encoding method is an improved solution for encoding time series values ​​of patients' physiological indicators.

[0129] In fact, unlike existing coding methods (e.g., the incremental modulation method described above), this coding method mitigates the impact of unwanted noise in the data. This unwanted noise can originate from the medical measurement equipment used to obtain values ​​of a patient's physiological indicators or from the measurement process itself. This mitigation cannot be achieved using existing methods (e.g., incremental modulation methods). Instead of simply examining the difference between the values ​​of a time series between two consecutive time points, the coding method involves obtaining a slope coefficient based on the time point, a predetermined number of previous time points, and a predetermined number of subsequent time points. Therefore, the coding method can analyze the evolution of a time series by using the slopes of temporally close data.

[0130] As discussed below Figure 8 , Figure 9 and Figure 10 As shown, this mitigation is particularly effective. These figures illustrate the effectiveness of encoding methods that encode time series without encoding unwanted fluctuations in noise.

[0131] The slope coefficient can be an optimal indicator of data variability, capturing local trends in the data—for example, whether the data changes significantly or remains roughly constant over time. The slope coefficient is calculated based on a given time point, a predetermined number of previous time points, and a predetermined number of subsequent time points. Therefore, the slope coefficient encodes the trend at both previous and subsequent time points.

[0132] Furthermore, the flexibility in the selection of method parameters (selection of references, selection of time windows to be considered) can enable a better interpretation of our data encoding.

[0133] Obtaining time series values ​​of a patient's physiological indicators through encoding methods may include retrieving at least a portion of the time series from (e.g., local or remote) memory, or receiving (e.g., from a remote system) at least a portion of the time series thus acquired. Acquisition may include directly obtaining at least a portion of the time series by measuring the patient's physiological indicators and recording the occurrence of the patient's clinical events. Acquisition may include obtaining at least a portion of the time series from clinical trials / research.

[0134] The encoding method includes: for each time point of the value time series, obtaining the slope coefficient of the value time series at that time point. In other words, the encoding method includes calculating the slope coefficient, such as the slope coefficient of a line, at each time point of the time series. The slope coefficient is based on the time point, a predetermined number of previous time points, and a predetermined number of subsequent time points.

[0135] The predetermined number of previous / subsequent time points can be the predetermined number of adjacent previous / subsequent time points, that is, the set of time points immediately before / after that time point, whose base is equal to the predetermined number (e.g., a positive integer).

[0136] A predetermined number of previous time points and a predetermined number of subsequent time points can each be a non-negative integer. The predetermined number of previous time points and the predetermined number of subsequent time points can be the same or different. In one implementation, the predetermined number of previous / subsequent time points may not depend on a base point; for example, it may always be the same integer (for all time points, this integer may depend only on "previous" or "subsequent," or be independent of "previous" or "subsequent"). In other implementations, the predetermined number of previous / subsequent time points may depend on the time point. In such implementations, this can be done whenever possible (i.e., whenever at least...). (a predetermined number of previous time points can be taken as numerical values) (For example, It can be a non-negative integer; otherwise, it is the number at a previous time point. Similarly, in such implementations, it is possible (i.e., as long as at least...) (a number of subsequent time points), the predetermined number of subsequent time points can be taken as a numerical value. (For example, (Can be a non-negative integer), otherwise it represents the number of subsequent time points. In the formula, as mentioned above, a time series can be represented as a sequence. Non-negative integers and It can be fixed. At a specific point in time. At this point, the previous time point for the predetermined quantity can be the quantity. The subsequent time point for the pre-ordered quantity can be .

[0137] Therefore, in the implementation, the slope coefficient can be calculated based on / according to a set of time points including a given time point, the immediately preceding time point, and the immediately following time point. In the formula, as described above, the time series can be represented as a sequence. Based on time point The slope coefficients for a predetermined number of previous time points and a predetermined number of subsequent time points can be based on the following elements:

[0138] .

[0139] In an implementation, the slope coefficient can be the slope coefficient of a line. This line can be an optimal line describing a set of time points, which includes (e.g., composed of) time points, a predetermined number of previous time points, and a predetermined number of subsequent time points. Therefore, in an implementation, the encoding method can include determining the set of time points to describe the time points. The optimal line. This line can be a line in 2D space, where one dimension is given by time (i.e., time points), and the other dimension is given by the values ​​of the time series. Provided.

[0140] Figure 7 The diagram shows a 2D space containing points (shown with crosses) indicating the values ​​of a time series, with the coordinates of each point being the time point and the value. Figure 7 The diagram also shows line segments at each time point, which represent the aforementioned line based on the corresponding time point and indicate the slope coefficient.

[0141] In the encoding method, obtaining the slope coefficient of the time series at a given time point can include calculating a linear regression on the time series values, where the time series values ​​include that time point, a predetermined number of previous time points, and a predetermined number of subsequent time points. The linear regression can be a linear least squares regression, that is, a line that optimally approximates the data with respect to the mean square error. Therefore, in an implementation, for each time point... Encoding methods may include time-point-based methods. (That is, calculate linear regression (e.g., linear least squares regression) based on the correlation values ​​of time points and physiological indicators.)

[0142] Calculating linear regression at each time point can be advantageous because linear regression can be the best possible line to describe the data under consideration (and the metric / error under consideration), and therefore the slope of linear regression can be the best descriptor of the local trend of the time series.

[0143] The encoding method also includes comparing the slope coefficient with a reference slope coefficient for each time point of the value time series. The comparison includes:

[0144] If the absolute value of the calculated slope coefficient is greater than the reference slope coefficient, then the time point is encoded as a pulse; and

[0145] If the absolute value of the calculated slope coefficient is less than the reference slope coefficient, then the time point is encoded as pulseless.

[0146] In other words, the encoding method may include: for each time point of the value time series, comparing the trend of change given by the slope coefficient with a reference trend of change. The encoding method may also include: if the trend of change is greater than the reference change, then encoding it as an impulse.

[0147] In the formula, as mentioned above, a time series can be represented as a sequence. Time point The slope coefficient at point can be expressed as And the reference slope coefficient can be expressed as The encoding method can use signals defined as follows. Encoded sequence :

[0148] .

[0149] Each pulse in the encoding method can be a positive pulse or a negative pulse. The step of comparing the slope coefficient with the reference coefficient in the encoding method may also include:

[0150] If the calculated slope coefficient is greater than the positive reference slope coefficient (e.g., the absolute value form of the reference slope coefficient discussed above), then that time point is encoded as a positive pulse; and

[0151] If the calculated slope coefficient is less than the negative reference slope coefficient (e.g., the opposite of the absolute value of the reference slope coefficient discussed above), then that time point is encoded as a negative pulse.

[0152] In other words, the encoding method may also include: for each time point of the value time series, comparing the trend of change given by the slope coefficient with a reference change. This comparison may consider whether the trend is increasing or decreasing. The encoding method may also include: if the trend of change is greater than the reference change, then encoding it as an impulse.

[0153] In the formula, as mentioned above, a time series can be represented as a sequence. Time point The slope coefficient at point can be expressed as And the reference slope coefficient can be expressed as We can assume a reference slope coefficient. If positive, a similar formula can be used in other cases. The encoding method can use a signal defined as follows. Encoded sequence :

[0154] .

[0155] The reference slope of the encoding method (e.g., referred to above) The reference slope coefficient can be 1, or the quotient of the maximum amplitude of the time series value and the maximum time amplitude. In other words, the reference slope coefficient can adapt to the changes in the time series of the data. In implementations, the time series of data may not be normalized, and in such implementations, it may be advantageous to define the reference slope coefficient considering the maximum change in the time series.

[0156] In the formula, the reference slope coefficient can be... Alternatively, by The time series of a patient's physiological indicators can be represented by the following reference slope:

[0157]

[0158] in It is a sequence The maximum value, It is a sequence The minimum value.

[0159] Figure 8 The results of implementing the encoding method are shown. Figure 8 In the middle, consider with Figure 6 The same value in the time series. Figure 8 The coded pulse generated by an implementation of the encoding method is shown. Figure 8 The coded pulse ratio shown Figure 6 The ones shown better illustrate the changes in the curve. For example, in Figure 8 In the noisy Gaussian function, the rising and falling trends have been correctly encoded. Furthermore, with... Figure 6 The difference lies in the fact that it differs from existing technological methods, and more specifically, from incremental modulation methods. Figure 8 In this method, no pulse is encoded due to noise. This is an advantage over the delta modulation method, which may fail to correctly encode pulse data due to data instability caused by noise.

[0160] Furthermore, even with frequent local variations in relatively high amplitude, this encoding method can accurately encode the trend of changes in the time series values. (Reference) Figure 9 Provide an example.

[0161] Figure 9 Another example of a comparison between the encoding method and existing methods (specifically, the delta modulation method) is shown. Figure 9 A time series including several oscillations is shown. The trend of the time series is well captured in the pulses generated by the encoding method, where a positive pulse indicating an upward trend is followed by a negative pulse indicating a negative trend. In contrast, the delta modulation method generates an output consisting of alternating positive and negative pulses, and a clear trend cannot be obtained from the output.

[0162] Figure 10Results of an implementation of this method are shown. In this implementation, publicly available data from patients with primary biliary cirrhosis were used. This data was obtained by the Mayo Clinic (a consortium of university hospitals and research institutions) between January 1974 and May 1984. The Mayo Clinic conducted a randomized, double-blind clinical study comparing the effects of D-penicillamine versus placebo in primary biliary cirrhosis (PBC). PBC is an autoimmune-originating, progressive disease, and the ensuing inflammatory process ultimately leads to cirrhosis and destruction of the liver and bile. The data obtained included longitudinal data from 312 patients (1945 rows). Some patients had multiple hospital visits. Figure 10 The code for albumin variables is shown, corresponding to the concentration of albumin in the blood, in mg / dL. Albumin is the most abundant plasma protein. In patients with advanced cirrhosis, albumin levels are low due to the reduced number of hepatocytes caused by the disease itself. Figure 10 In this encoding method, a predetermined number of previous time points and a predetermined number of subsequent time points are both used, each with a value of 4. The (positive) reference slope coefficient is 1, and the negative reference slope coefficient is -1. For example... Figure 10 As shown, compared to existing incremental modulation methods, the encoding method is able to better describe the overall change (i.e., decrease) of this variable in the patient. In fact, the incremental modulation method cannot accurately encode the trend of change in this variable.

[0163] These methods (e.g., learning methods, usage methods, and coding methods) are implemented by computers. This means that the steps (or essentially all steps) of each method are performed by at least one computer or any similar system. Therefore, the steps of each method can be performed fully or semi-automatically by a computer. In the example, the triggering of at least some steps of any method can be performed through user-computer interaction. The required level of user-computer interaction can depend on the expected level of automation and be balanced with the need to fulfill user expectations. In the example, this level can be user-defined and / or predefined.

[0164] A typical example of a computer implementation of a method is to execute the method using a system adapted for this purpose. This system may include a processor coupled to memory and a graphical user interface (GUI); the memory stores a computer program containing instructions for executing the method. The memory may also store a database. The memory is any hardware suitable for such storage and may comprise several physically distinct parts (e.g., a portion for the program and possibly a portion for the database).

[0165] Figure 11An example of the system is shown, where the system is a client computer system, such as a user's workstation.

[0166] The client computer in this example includes a central processing unit (CPU) 1010 connected to an internal communication bus 1000, and random access memory (RAM) 1070 also connected to the bus. The client computer also includes a graphics processing unit (GPU) 1110 associated with video random access memory 1100 connected to the bus. The video RAM 1100 is also referred to in the art as a frame buffer. A mass storage device controller 1020 manages access to mass storage devices, such as hard disk drives 1030. Mass storage devices suitable for tangibly representing computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor storage devices such as EPROM, EEPROM, and flash memory devices; disks such as internal hard disks and removable disks; and magneto-optical disks. Any of the above may be supplemented or incorporated by a specially designed application-specific integrated circuit (ASIC). A network adapter 1050 manages access to a network 1060. The client computer may also include a tactile device 1090, such as a cursor control device, a keyboard, etc. The cursor control device is used in the client computer to allow the user to selectively position the cursor at any desired location on the display 1080. Furthermore, the cursor control device allows the user to select various commands and input control signals. This cursor control device includes multiple signal generating devices for inputting control signals to the system. Typically, the cursor control device can be a mouse, with mouse buttons used to generate signals. Alternatively or additionally, the client computer system may include a sensitive pad and / or a sensitive screen.

[0167] A computer program may include computer-executable instructions, which include means for causing the system to perform any of the methods. The program may be recorded on any data storage medium, including the system's memory. The program may be implemented, for example, in digital electronic circuits, or in computer hardware, firmware, software, or a combination thereof. The program may be implemented as means, for example, tangibly embodied in a machine-readable storage device for use in a product executed by a programmable processor. The method steps may be executed by a programmable processor that executes the program of instructions to perform the function of the method by manipulating input data and generating output. Thus, the processor may be programmable and coupled to receive data and instructions from the data storage system, at least one input device, and at least one output device, and to send data and instructions to the data storage system, at least one input device, and at least one output device. The application program may be implemented in a high-level procedural programming language or an object-oriented programming language, or, as desired, in assembly language or machine language. In any case, the language may be a compiled or interpreted language. The program may be a complete installation program or update program. In either case, applying the program to the system generates instructions for performing the method. Alternatively, the computer program can be stored and executed on a server in a cloud computing environment that communicates with one or more clients via a network. In this case, the processing unit executes the instructions included in the program, thereby enabling the method to execute on the cloud computing environment.

Claims

1. A computer-implemented method for a machine learning neural network, said neural network including a spiking neural network (SNN) for predicting the timing of clinical events, the method comprising: - Obtain a training dataset for training examples, each training example including a time series of one or more values ​​of the corresponding physiological indicators of the patient over a past time period and a corresponding baseline ground truth value of the occurrence time of the clinical event; as well as - The neural network is trained based on the training dataset. The neural network is trained to take as input a time series of one or more values ​​of the corresponding physiological indicators of the patient in the past time period and output a prediction of the probability of the clinical event occurring in a time period.

2. The method according to claim 1, wherein, Each training example also includes the patient's baseline data, and the neural network is configured to take the baseline data into account to output the prediction.

3. The method according to claim 2, wherein, The neural network includes the SNN and an additional module. The SNN is configured to take the one or more value time series as input and output a series of pulses. The additional module is configured to take the series of pulses and the baseline data as input and output the prediction.

4. The method according to claim 3, wherein, The additional module uses the SoftMax function to output the prediction.

5. The method according to any one of claims 1 to 4, wherein, The neural network includes an encoding layer configured to take the one or more value time series as input and encode the one or more value time series into corresponding pulse sequences.

6. The method according to any one of claims 1 to 5, wherein, The training includes minimizing a loss, for each training example, where the loss penalty is the difference between the prediction output by the neural network for that training example and the baseline ground truth value of that training example.

7. The method according to claim 6, wherein, The loss is of the following type: ， in, It is the number of training examples. Indicates the patient The baseline truth value for the occurrence of a clinical event; if a medical event exists, then... =1, otherwise =0, This refers to the time of the clinical event. Indicates training example Until time Data, It is a known training example. Until time Given all the data, for the training example In time The predicted probability of a clinical event occurring, where, ， in, Training example The time when the final indicator is known. And among them, The patient is In time The probability of survival, and it is of the following type: 。 8. A neural network that can be obtained by the method according to any one of claims 1 to 7.

9. A computer-implemented method using the neural network according to claim 8, the method comprising: - Obtain time series values ​​of one or more physiological indicators of the patient over a past period; as well as - Apply the neural network to the one or more value time series to predict the probability of the clinical event occurring within a time period.

10. A computer program product comprising instructions, wherein the instructions, when executed by a computer system, cause the computer system to perform the method according to any one of claims 1 to 7 and / or the method according to claim 9.

11. A computer-readable data storage medium having a computer program included in a computer program product according to claim 10 and / or a neural network according to claim 8 recorded thereon.

12. A computer system comprising a processor coupled to a memory, the memory storing a computer program included in a computer program product according to claim 10 and / or a neural network according to claim 8.

Images