Training method and device of time sequence application model based on medical time sequence data

Abstract

The invention discloses a training method and device of a time sequence application model based on medical time sequence data, and relates to the field of medical information and the field of artificial intelligence. The method comprises the following steps: acquiring a medical training sample; on the basis of classification judgment auxiliary information corresponding to the medical training samples, soft labels are generated for the medical training samples, and the classification judgment auxiliary information is represented in multiple samples of the same category and used for distinguishing and quantifying continuous variables of feature difference degrees of different samples; the soft label is a continuously distributed probability value, and the soft label is used for reflecting the confidence degree of the medical training sample belonging to the target category; and performing iterative training on the neural network according to the medical training sample and the soft label of the medical training sample to obtain a time sequence model. According to the method and the device, the technical problem of relatively poor robustness of a trained model due to too simple classification of training labels in the prior art is solved.

Description

Technical Field

[0001] This application relates to the fields of medical information and artificial intelligence, and more specifically, to a training method and apparatus for a time-series application model based on medical time-series data. Background Technology

[0002] In the medical field, the analysis and prediction of time-series data have significant clinical value, such as multiple physiological indicators and continuous medical imaging during a patient's treatment. A common task is to predict a patient's future health status or response to treatment by analyzing data from the previous N time points. For example, in cancer treatment, analyzing cone-beam computed tomography (CBCT) images of a patient before radiotherapy can help predict the final treatment outcome as early as possible. The clinical value of this type of prediction lies in early awareness. If the current treatment plan can be accurately determined to be ineffective early in the treatment process, doctors can adjust the treatment plan in a timely manner, such as increasing the radiation dose, thereby improving the cure rate, avoiding wasting valuable time due to ineffective treatment, and even causing unnecessary side effects to the patient.

[0003] However, existing technologies generally face a core contradiction: the predictive accuracy of a model is usually positively correlated with the length of the input data sequence. In other words, the more data, the more accurate the model prediction. But in clinical practice, high-precision predictions are most needed in the early stages of treatment when the sequence length is very short. With insufficient information in the early stages, it is difficult to balance predictive accuracy with the risk of clinical decision-making. Existing models typically pursue the highest overall accuracy without being able to optimize and balance according to specific clinical needs; for example, they cannot misclassify patients with poor treatment outcomes as having good outcomes. If the model output is overly confident, it cannot reflect the true probability. For example, traditional models often directly map good and bad treatment outcomes to hard labels 1 and 0, and use the cross-entropy loss function to train the neural network. Since hard labels only provide binary classification information and cannot express the degree of difference between samples within the same category, the model is forced to push the output probability to extreme values ​​of 0 or 1 under loss-driven conditions. This overconfident prediction makes the model more sensitive to input perturbations or data distribution shifts: when the sequence length is short and information is insufficient, the prediction results often fluctuate drastically, and robustness decreases significantly. When the model makes a mistake, it can cause serious misleading results, making it difficult to apply to clinical scenarios that require precise risk control.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a training method and apparatus for a time-series application model based on medical time-series data, which at least solves the technical problem in the prior art that the trained model has poor robustness due to the overly simple classification of training labels.

[0006] According to one aspect of the embodiments of this application, a training method for a time-series application model based on medical time-series data is provided, comprising: acquiring medical training samples, wherein the medical training samples include medical time-series data; generating soft labels for the medical training samples based on classification judgment auxiliary information corresponding to the medical training samples, wherein the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features of different samples in multiple samples of the same category; the soft label is a continuously distributed probability value, and the soft label is used to reflect the confidence level of the medical training sample belonging to the target category; and iteratively training a neural network according to the medical training samples and the soft labels of the medical training samples to obtain a time-series model.

[0007] Optionally, the classification and judgment auxiliary information includes at least one of the following indicators: tumor shrinkage rate, used to characterize the proportion of tumor volume reduction; organ-at-risk damage rate, used to characterize the proportion of damage to organs at risk during radiotherapy; polymerase chain reaction amplification rounds, used to reflect the original concentration of target nucleic acid in the nucleic acid test sample; and dose deviation, used to characterize the difference between the dose obtained by analyzing the pseudo-images after matching the first image and the second image during radiotherapy and the original planned dose. The first image is a medical image taken before each radiotherapy session, and the second image is the initial positioning medical image used to formulate the radiotherapy plan.

[0008] Optionally, based on the classification judgment auxiliary information corresponding to the medical training samples, soft labels are generated for the medical training samples, including: mapping values ​​from 0 to 1 based on the classification judgment auxiliary information; and generating soft labels based on the mapped values.

[0009] Optionally, the classification and judgment auxiliary information is mapped to obtain a value from 0 to 1, including: when the classification and judgment auxiliary information is the tumor shrinkage rate, a first value is mapped based on the actual value of the tumor shrinkage rate, wherein the value of the first value is positively correlated with the tumor shrinkage rate; when the classification and judgment auxiliary information is the organ damage severity rate, a second value is mapped based on the actual value of the organ damage severity rate, wherein the value of the second value is negatively correlated with the organ damage severity rate; when the classification and judgment auxiliary information is the polymerase chain reaction (PCR) amplification epochs, a third value is mapped based on the actual value of the PCR amplification epochs, wherein the value of the third value is negatively correlated with the PCR amplification epochs; and when the classification and judgment auxiliary information is the dose deviation, a fourth value is mapped based on the actual value of the dose deviation, wherein the value of the fourth value is negatively correlated with the dose deviation.

[0010] Optionally, the training method for the time-series application model based on medical time-series data further includes: after acquiring medical training samples, determining the result labels corresponding to the medical training samples, wherein the result labels are used to label the category to which the medical training samples belong; determining the prediction error tolerance for the category according to the category labeled by the result labels; if the prediction error tolerance is determined to be greater than a preset threshold, generating soft labels for the medical training samples based on the classification judgment auxiliary information corresponding to the medical training samples; if the prediction error tolerance is determined to be less than or equal to the preset threshold, using the result labels corresponding to the medical training samples as the soft labels of the medical training samples.

[0011] Optionally, generating soft labels based on the mapped values ​​includes: when the classification judgment auxiliary information includes multiple indicator information, and a value corresponding to each indicator information is obtained by mapping each indicator information, the following method one or method two is used to generate soft labels based on the values ​​obtained by mapping each indicator information; method one includes: performing a weighted calculation on the values ​​obtained by mapping each indicator information, and generating soft labels based on the result of the weighted calculation; method two includes: taking the logarithm of the values ​​obtained by mapping each indicator information and summing them, and generating soft labels based on the calculation result.

[0012] Optionally, after iteratively training the neural network based on medical training samples and their soft labels to obtain a time-series model, the training method for the time-series application model based on medical time-series data further includes: acquiring target medical time-series data; inputting the target medical time-series data into the time-series model, and outputting classification prediction probabilities based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase; and determining the target label corresponding to the target medical time-series data based on the classification prediction probabilities.

[0013] Optionally, medical time-series data includes: medical imaging time-series data and / or medical monitoring time-series data.

[0014] According to another aspect of the embodiments of this application, a method for processing medical time-series data is also provided, comprising: acquiring target medical time-series data; inputting the target medical time-series data into a time-series model, and outputting a classification prediction probability based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase. The time-series model is a neural network model trained based on medical training samples and soft labels of the medical training samples; the soft labels are generated based on classification judgment auxiliary information corresponding to the medical training samples, and the soft labels are a continuously distributed probability value used to reflect the confidence level that the medical training samples belong to the target category; wherein the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features between different samples in multiple samples of the same category; and determining the target label corresponding to the target medical time-series data based on the classification prediction probability.

[0015] Optionally, the target label corresponding to the target medical class time series data is determined based on the classification prediction probability, including: setting a target threshold based on the tolerance for prediction errors of various labels corresponding to the target medical class time series data; and determining the target label from various labels corresponding to the target medical class time series data based on the relationship between the classification prediction probability and the target threshold.

[0016] Optionally, the method for processing medical time series data further includes: when the target medical time series data is updated over time, inputting the updated target medical time series data into the time series model, outputting the updated classification prediction probability based on the data characteristics of the updated target medical time series data according to the time series model; and determining the target label corresponding to the updated target medical time series data based on the updated classification prediction probability.

[0017] According to another aspect of the embodiments of this application, a training apparatus for a time-series application model based on medical time-series data is also provided, comprising: a sample acquisition unit for acquiring medical training samples, wherein the medical training samples include medical time-series data; a soft label generation unit for generating soft labels for the medical training samples based on classification judgment auxiliary information corresponding to the medical training samples, wherein the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features of different samples among multiple samples of the same category; the soft label is a continuously distributed probability value, and the soft label is used to reflect the confidence level of the medical training sample belonging to the target category; and a model training unit for iteratively training a neural network based on the medical training samples and the soft labels of the medical training samples to obtain a time-series model.

[0018] According to another aspect of the embodiments of this application, a medical time-series data processing apparatus is also provided, comprising: a data acquisition unit for acquiring target medical time-series data; a first processing unit for inputting the target medical time-series data into a time-series model, and outputting a classification prediction probability based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase; wherein, the time-series model is a neural network model trained based on medical training samples and soft labels of the medical training samples; the soft labels are generated based on classification judgment auxiliary information corresponding to the medical training samples, and the soft labels are a continuously distributed probability value used to reflect the confidence level of the medical training samples belonging to the target category; wherein, the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features of different samples in multiple samples of the same category; and a second processing unit for determining the target label corresponding to the target medical time-series data according to the classification prediction probability.

[0019] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the above-described training method for a time-series application model based on medical time-series data.

[0020] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to execute the above-described training method for a time-series application model based on medical time-series data.

[0021] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described training method for a time-series application model based on medical time-series data.

[0022] In this embodiment, the medical data analysis system first acquires medical training samples, which include medical time-series data. Based on the classification auxiliary information corresponding to the medical training samples, soft labels are generated for the medical training samples. The classification auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features among multiple samples of the same category. The soft label is a continuously distributed probability value used to reflect the confidence level that the medical training sample belongs to the target category. Then, based on the medical training samples and their soft labels, the neural network is iteratively trained to obtain a time-series model.

[0023] As described above, obtaining medical training samples containing time-series medical data can provide a data foundation for model training and introduce classification assistance information. This classification assistance information can be a continuous variable that distinguishes and quantifies the degree of difference in features among multiple medical training samples of the same category. Based on the classification assistance information corresponding to the medical training samples, soft labels are generated for the medical training samples. The soft label can be a continuously distributed probability value, not a simple 0 or 1, but generated based on the classification assistance information. It reflects the continuous confidence level that the medical training sample belongs to the target category, thereby transforming the intra-category difference information ignored by traditional binary hard labels into gradient signals that can be used for supervised learning.

[0024] Furthermore, the neural network is iteratively trained using medical training samples and their soft labels. Since soft labels provide richer and more informative supervisory signals than hard labels, the neural network is guided during training to learn to distinguish these subtle differences in confidence levels. This training method may enable the neural network to learn more robust and discriminative feature representations, rather than simply memorizing a simple classification boundary. Therefore, the resulting time-series model exhibits better generalization ability and judgment stability when faced with data noise, inter-individual differences, or ambiguous samples at class boundaries, thus solving the technical problem in existing technologies where the training labels are too simplistic, leading to poor model robustness. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1 This is a schematic diagram of an optional training method for a time-series application model based on medical time-series data according to an embodiment of this application;

[0027] Figure 2 This is a schematic diagram of an optional medical time-series data processing method according to an embodiment of this application;

[0028] Figure 3 This is a schematic diagram of an optional training device for a time-series application model based on medical time-series data, according to an embodiment of this application.

[0029] Figure 4 This is a schematic diagram of an optional medical time-series data processing apparatus according to an embodiment of this application. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] According to an embodiment of this application, a method embodiment for training a time-series application model based on medical time-series data is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0034] According to the embodiments of this application, a medical data analysis system can be used as the execution subject of the training method of the time-series application model based on medical time-series data in the embodiments of this application. The system can be a software system or an embedded system combining software and hardware. Of course, the execution subject of the method in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the execution subject.

[0035] Figure 1 This is a training method for a time-series application model based on medical time-series data according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0036] Step S101: Obtain medical training samples, wherein the medical training samples include medical time series data.

[0037] Optionally, medical time-series data refers to medically relevant information collected continuously in chronological order within a certain time frame. Examples include dynamic data related to disease progression, such as medical images and physiological monitoring data at different time points during a patient's treatment. This dynamic data reflects the dynamic trends of a patient's changes during treatment, enabling the model to learn the characteristics and patterns of medical data changes over time, thereby improving its predictive ability for medical events or disease states.

[0038] Optionally, historical case data can be collected through a medical information system to construct a training set. For example, in the application scenario of predicting the efficacy of radiotherapy, a medical training sample can correspond to the data of a patient who has completed the entire course of radiotherapy, and the medical time-series data can be the image sequence of cone-beam computed tomography (CBCT) scans acquired before each treatment. This image sequence is arranged in order of treatment number, forming an image queue on a time axis, which can record the morphological changes of the tumor target area and surrounding organs at risk during the accumulation of radiation dose. In addition, the medical time-series data can also include time-series records of continuously monitored physiological parameters. The embodiments of this application can establish a medical training sample set consisting of medical time-series data of multiple patients.

[0039] Step S102: Based on the classification judgment auxiliary information corresponding to the medical training samples, generate soft labels for the medical training samples. The classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features between different samples in multiple samples of the same category. The soft label is a probability value of a continuous distribution, which is used to reflect the confidence level of whether the medical training sample belongs to the target category.

[0040] Optionally, classification auxiliary information is used to characterize a continuous variable that further distinguishes sample characteristics within a known sample category, quantifying the degree of difference between samples of the same category. The soft label generated based on the classification auxiliary information can be a continuous probability value between 0 and 1, rather than a traditional binary label. This soft label is used to characterize the confidence level that the corresponding medical training sample belongs to a certain target category, which can be, but is not limited to, efficacy or deterioration. Embodiments of this application can encode continuous prior knowledge within a category into the training label, providing the model with a more granular learning objective, thereby guiding the model to output a predicted probability that matches the true confidence level.

[0041] Optionally, classification judgment auxiliary information can be selected and defined according to different clinical scenarios. Classification judgment auxiliary information is used to characterize continuous variables that distinguish and quantify the degree of difference in characteristics among multiple samples of the same category. For example, in the scenario of predicting the efficacy of radiotherapy, the categories to which medical training samples belong can include "good efficacy" and "bad efficacy." The clinical goal in this scenario could be to identify samples belonging to the "bad efficacy" category as early as possible. This goal may be associated with asymmetric risk preferences: the tolerance for incorrectly predicting samples belonging to the "good efficacy" category as "bad efficacy" is low, as this error could lead to unnecessary dose increases and harm to patients. Therefore, the tolerance for the false negative rate (FNR) is extremely low, for example, required to be below 0.1%. Conversely, the tolerance for incorrectly predicting samples belonging to the "bad efficacy" category as "good efficacy" is relatively high. This error may temporarily maintain the original treatment plan in the early stages of treatment, but it can be continuously observed and predicted in subsequent treatments, with the opportunity to correct the judgment later. Therefore, the tolerance for the false positive rate (FPR) is relatively high. This consideration of risk preferences can serve as the background for selecting and utilizing classification judgment auxiliary information.

[0042] For example, in a scenario predicting treatment efficacy, for samples judged as having good efficacy, the classification aid information could be the tumor shrinkage rate. Generating soft labels involves mapping the actual value of this continuous variable to a soft label value. For instance, the lowest tumor shrinkage rate can be mapped to a value slightly above 0.5, such as 0.51, and the highest tumor shrinkage rate can be mapped to 1.0; values ​​in between can be mapped using a linear function. A linear function mapping ensures a positive correlation between the tumor shrinkage rate and the soft label value. For samples with a negative outcome label, their soft label can be set to 0 or a value close to 0, indicating extremely low confidence in belonging to the target category. In this way, each training sample can obtain a soft label that reflects its performance within the category.

[0043] Step S103: Based on the medical training samples and their soft labels, the neural network is iteratively trained to obtain a time-series model.

[0044] Optionally, medical training samples can be used as input features, and the soft labels of the medical training samples can be used as training targets. Multiple rounds of iterative training are performed on the selected neural network structure using the medical training samples and their soft labels. The training objective is to enable the neural network to predict probability values ​​matching the soft labels based on the input medical time-series data, ultimately achieving convergence through training to obtain a time-series model suitable for prediction. This embodiment of the application can transform medical time-series data with fine-grained labels into an AI model with predictive capabilities. The characteristic of this AI model is that its output probabilities are calibrated and can reflect different confidence levels.

[0045] Optionally, preprocessed medical time-series data is input into a neural network. The medical time-series data can be image sequence features, and the neural network can be a Long Short-Term Memory network or a Transformer model. Corresponding soft labels are used as supervision signals. The training process iteratively updates the network parameters by minimizing the difference between the model's output probability and the soft labels; for example, a mean squared error loss function can be used. After multiple rounds of iterative optimization, the neural network parameters tend to stabilize, resulting in a trained time-series model. When this time-series model receives new medical time-series data, it can output a continuously distributed probability value, which can simulate the confidence level characteristics learned during the training phase based on classification judgment auxiliary information.

[0046] After training, the model's output will no longer be a polarized 0 or 1, but a more reasonable distribution that reflects the true probability. This enhances the interpretability and credibility of the model's predictions, providing users with richer and more reliable decision-making information. For example, for a positive sample with particularly good treatment outcomes, the model can output a probability close to 1.0; while for a positive sample with less effective treatment outcomes, the model can output a probability of around 0.6. This process allows for model calibration.

[0047] In one optional embodiment, the classification and judgment auxiliary information includes at least one of the following indicators: tumor shrinkage rate, used to characterize the proportion of tumor volume reduction; organ-at-risk damage rate, used to characterize the proportion of damage to organs at risk during radiotherapy; polymerase chain reaction amplification rounds, used to reflect the original concentration of target nucleic acid in the nucleic acid test sample; and dose deviation, used to characterize the difference between the dose obtained by analyzing the pseudo-images after matching the first image and the second image during radiotherapy and the original planned dose, wherein the first image is a medical image taken before each radiotherapy session, and the second image is the initial positioning medical image used to formulate the radiotherapy plan.

[0048] Optionally, each indicator can be associated with a specific medical assessment scenario. For example, tumor shrinkage rate is suitable for assessing tumor treatment efficacy, organ-at-risk damage rate is suitable for assessing radiotherapy safety, polymerase chain reaction amplification volumes (PCR volumes), which can be characterized by Ct values, are suitable for pathogen detection, and dose deviation is suitable for decision support in adaptive radiotherapy. These indicators are all continuous variables that can further distinguish the quality of samples within the same outcome category.

[0049] Optionally, the tumor shrinkage rate directly quantifies the significance of the therapeutic effect by comparing the proportion of tumor volume reduction before and after treatment. The organ damage rate measures the proportion of damage to normal tissues caused by radiotherapy; a higher value indicates greater side effects and poorer efficacy. A higher polymerase chain reaction (PCR) amplification epoch, i.e., a higher Ct value, indicates a lower initial viral load and a weaker positivity rate in positive samples. A greater dose deviation indicates a more significant difference between the actual dose and the planned dose on that day, and a stronger negative positivity rate in negative samples.

[0050] Optionally, the corresponding indicators need to be determined based on the selected clinical task. For example, in the scenario of predicting the efficacy of radiotherapy, the tumor shrinkage rate can be used as auxiliary information for classification. This auxiliary information can be obtained by comparing medical images, such as CT (Computed Tomography), before and after the start of the treatment course, and the tumor volume change ratio can be automatically calculated by image analysis software. As another example, in the scenario of predicting whether adaptive radiotherapy needs to be triggered, the calculation of dose deviation involves deformably registering the CBCT (which can be used as the first image) taken before each treatment with the positioning CT (which can be used as the second image) based on the radiotherapy plan, generating a pseudo-CT image that reflects the anatomical structure on that day, recalculating the dose distribution on this pseudo-CT image, and finally comparing it with the original planned dose to obtain the deviation value. The introduction of these specific indicators for classification auxiliary information makes the generation of soft tags have clear medical basis and operability.

[0051] In one optional embodiment, soft labels are generated for the medical training samples based on the classification judgment auxiliary information corresponding to the medical training samples, including: mapping values ​​from 0 to 1 according to the classification judgment auxiliary information; and generating soft labels based on the mapped values.

[0052] Optionally, mapping the classification auxiliary information to obtain values ​​from 0 to 1 can normalize the classification auxiliary information, which has different physical meanings and dimensions, to a unified continuous numerical scale representing the tendency to belong to the target category. Generating soft labels based on the mapped values ​​allows the intermediate values ​​to be directly defined or further processed into the final soft labels.

[0053] For example, a mapping function can be designed from the actual value of the classification auxiliary information to a value in the interval (0,1]. Taking tumor shrinkage rate as an example, assuming the tumor shrinkage rate of the collected positive samples ranges from 10% to 90%, a linear mapping function can be used: First value = 0.51 + 0.49 ((current reduction rate - 10%) / (90% - 10%)). Thus, a reduction rate of 10% maps to 0.51, and a reduction rate of 90% maps to 1.0. The first value obtained from this mapping can be directly used as the soft label for the positive sample.

[0054] In one optional embodiment, the process of mapping a value from 0 to 1 based on classification assistance information includes: when the classification assistance information is tumor shrinkage rate, the medical data analysis system maps a first value based on the actual value of the tumor shrinkage rate, wherein the first value is positively correlated with the tumor shrinkage rate; when the classification assistance information is organ-at-risk damage rate, a second value is mapped based on the actual value of the organ-at-risk damage rate, wherein the second value is negatively correlated with the organ-at-risk damage rate; when the classification assistance information is polymerase chain reaction (PCR) amplification epochs, a third value is mapped based on the actual value of the PCR amplification epochs, wherein the third value is negatively correlated with the PCR amplification epochs; and when the classification assistance information is dose deviation, a fourth value is mapped based on the actual value of the dose deviation, wherein the fourth value is negatively correlated with the dose deviation.

[0055] Optionally, different correlation principles should be followed when mapping different indicator information. For example, for positive indicators, the better the result, the larger the value, and the mapped value should be positively correlated with the positive indicator; for negative indicators, the worse the result, the larger the value, and the mapped value should be negatively correlated with the negative indicator. For example, the first value is positively correlated with the tumor shrinkage rate, indicating that the greater the shrinkage, the higher the confidence level in the category of good efficacy. The second value is negatively correlated with the rate of organ damage at risk, indicating that the less damage, the higher the confidence level in the category of safe treatment. The third value is negatively correlated with the number of polymerase chain reaction (PCR) amplification epochs, indicating that the higher the original viral load and the lower the Ct value, the higher the confidence level in the category of positive efficacy. The fourth value is negatively correlated with the dose deviation, indicating that the more precise the dose execution and the smaller the deviation, the higher the confidence level in the category of no need for adaptive radiotherapy.

[0056] Optionally, a mapping function needs to be designed based on the nature of the indicator. For positively correlated tumor shrinkage rates, the above-mentioned linearly increasing function can be used. For negatively correlated indicators, such as dose deviation, assuming a clinically acceptable deviation range of 0% to 5%, the mapping function can be designed as: Fourth value = 1.0 - 0.49 (Current deviation / 5%), where a deviation of 0% is mapped to 1.0, and a deviation of 5% is mapped to 0.51. For the Ct value, a reasonable range can be set, such as 20 to 40, using the third value = 1.0 - 0.49. The mapping is performed using ((Ct value - 20) / (40 - 20)). This specific mapping relationship ensures that the soft label correctly encodes prior clinical knowledge. For the rate of organ-at-risk damage, the highest damage can be mapped to 0.51, and the lowest damage to 1.0, indicating that the greater the damage to normal tissue, the worse the therapeutic effect.

[0057] It should be noted that in quantitative polymerase chain reaction (PCR) detection scenarios, the classification auxiliary information can be the PCR amplification epoch, which reflects the original concentration of the target nucleic acid in the nucleic acid test sample. The specific value of the PCR amplification epoch can be called the Ct value. When generating soft tags for medical training samples, a value from 0 to 1 can be mapped based on the actual value of the PCR amplification epoch. That is, when the classification auxiliary information is the PCR amplification epoch, the third value mapped based on that amplification epoch is negatively correlated with the PCR amplification epoch. That is, the higher the PCR amplification epoch, the larger the Ct value, indicating a low original viral load, and the more closely the mapped third value can be close to 0.51; the lower the PCR amplification epoch, the smaller the Ct value, indicating a high original viral load, and the more closely the mapped third value can be close to 1.0. Based on the third value obtained from the mapping, a soft label can be generated to reflect the continuous confidence level of the medical training sample belonging to the positive category.

[0058] In scenarios predicting the necessity of adaptive radiotherapy (ART) triggering, the auxiliary information for classification judgment can be dose deviation. Dose deviation is used to characterize the difference between the dose obtained by analyzing pseudo-images obtained after matching the first and second images during radiotherapy and the originally planned dose. For example, the first image can be a medical image taken before each radiotherapy session, such as a cone-beam computed tomography (CBCT) image; the second image can be a positioning medical image initially used to formulate the radiotherapy plan, such as a computed tomography (CT) image. The process of mapping the actual value of the dose deviation to a value between 0 and 1 can follow a negative correlation: the larger the dose deviation, the closer the mapped fourth value is to 0; the smaller the dose deviation, the closer the mapped fourth value is to 1. For example, for a pre-set deviation threshold, such as 5%, the soft label of medical training samples with a dose deviation greater than this threshold can be set to 0, while medical training samples with a dose deviation less than or equal to this threshold can be mapped to a fourth value between 0.51 and 1.0 according to their deviation degree, with a deviation of 0% being mapped to 1.0. Based on the fourth value obtained from the mapping, soft labels can be generated. Soft labels generated based on dose bias can quantify the differences in dose maintenance within medical training samples that do not require adaptive radiotherapy categories.

[0059] In an optional embodiment, the training method for the time-series application model based on medical time-series data further includes: after acquiring medical training samples, the medical data analysis system determines the result labels corresponding to the medical training samples, wherein the result labels are used to label the category to which the medical training samples belong; based on the category to which the result labels belong, the system determines the prediction error tolerance for that category; if the prediction error tolerance is determined to be greater than a preset threshold, the system generates soft labels for the medical training samples based on the classification judgment auxiliary information corresponding to the medical training samples; if the prediction error tolerance is determined to be less than or equal to the preset threshold, the system uses the result labels corresponding to the medical training samples as the soft labels for the medical training samples.

[0060] Optionally, embodiments of this application introduce a conditional judgment process, which adaptively selects between using soft labels and traditional outcome labels based on the clinical tolerance for different types of prediction errors. Outcome labels refer to the final, definitive category classification of a sample, such as good or bad efficacy. By introducing a conditional judgment process, embodiments of this application can implement a risk-sensitive labeling strategy during model training: for low-tolerance error types, such as misclassifying good as bad, soft labels are used for meticulous calibration to minimize the risk; for high-tolerance error types, outcome labels can be used directly to simplify the process, enhancing the flexibility and practicality of the method in complex clinical decision-making environments.

[0061] For example, firstly, a corresponding result label needs to be determined for each training sample. Then, based on the category to which the result label belongs, the prediction error tolerance for that category is determined. The prediction error tolerance can be preset by the user or set based on historical data. If the prediction error tolerance for a certain category is determined to be greater than the preset threshold, it indicates that such prediction errors are more acceptable, triggering a soft label generation process based on classification judgment auxiliary information to fine-tune the model's behavior near the category boundary. For example, for a sample with a "good" efficacy (result label is positive), the soft label for that sample can be set to a continuous value between 0 and 1 based on the "tumor shrinkage rate." For example, the lowest shrinkage rate is mapped to 0.51, and the highest shrinkage rate is mapped to 1.0. That is, the lower the "tumor shrinkage rate," the worse the efficacy may be. If the prediction error tolerance for a certain category is determined to be less than or equal to the preset threshold, it indicates that such prediction errors have serious consequences and low tolerance, and the result label can be directly used as a soft label. For example, for a sample with a "bad" efficacy (result label is negative, negative is represented by 0), its label is still set to 0.

[0062] In one optional embodiment, generating soft labels based on the mapped values ​​includes: when the classification judgment auxiliary information includes multiple indicator information, and a value corresponding to each indicator information is obtained by mapping each indicator information, the following method one or method two is used to generate soft labels based on the values ​​obtained by mapping each indicator information; method one includes: performing a weighted calculation on the values ​​obtained by mapping each indicator information, and generating soft labels based on the result of the weighted calculation; method two includes: taking the logarithm of the values ​​obtained by mapping each indicator information and summing them, and generating soft labels based on the calculation result.

[0063] Optionally, when multiple indicator information exists, multiple mapping values ​​can be integrated to generate a single soft label. This integration operation assumes that each indicator information has been independently mapped to a value between 0 and 1. The purpose of the integration operation is to comprehensively consider clinical information from multiple dimensions to generate a more comprehensive and robust soft label. This application provides two integration methods: Method 1 is implemented through weighted calculation, allowing different weights to be assigned based on the clinical importance of different indicator information, reflecting the configurability and emphasis among indicator information. Method 2 is implemented by taking the logarithm and then summing the results. Method 2 tends to give higher overall confidence to samples that perform well in all indicator information and significantly lower confidence to samples that perform poorly in any indicator information, having a veto effect.

[0064] In an optional embodiment, after iteratively training the neural network based on medical training samples and their soft labels to obtain a time series model, the training method for the time series application model based on medical time series data further includes: the medical data analysis system can acquire target medical time series data and input the target medical time series data into the time series model, output classification prediction probabilities based on the data features of the target medical time series data according to the prior knowledge learned by the time series model during the training phase, and then determine the target label corresponding to the target medical time series data based on the classification prediction probabilities.

[0065] Optionally, the application phase of the time series model may include three consecutive actions: acquiring new input data corresponding to the target medical time series data; inputting the target medical time series data into the time series model, and outputting classification prediction probabilities based on the data characteristics of the target medical time series data according to the prior knowledge learned by the time series model during the training phase; and determining the target label corresponding to the target medical time series data based on the classification prediction probabilities and making probability-based decisions. By applying the time series model, a complete closed-loop technical process from training to application can be described, clarifying the specific purpose and output form of the trained model. This application embodiment emphasizes that the core function of the model is to output a classification prediction probability with calibration significance, rather than directly outputting a hard category.

[0066] For example, after the time series model is trained, it can be deployed in a medical data analysis system. When a prediction is needed for a new patient, the medical data analysis system obtains the patient's target medical time series data up to the current moment, such as the CBCT sequences corresponding to the first five completed treatments. This target medical time series data is input into the loaded time series model. Based on prior knowledge learned from soft labels during training, the time series model extracts features from the input data and outputs a classification prediction probability, for example, 0.75. This probability value represents the model's confidence that the patient's final treatment outcome is good. Subsequently, the medical data analysis system can determine the final target label based on this probability value through simple threshold comparison or other decision rules, demonstrating the direct application of the training method. Threshold comparison can refer to classifying a patient as having a good outcome if the probability is greater than 0.5.

[0067] In one alternative embodiment, the medical time-series data includes: medical image time-series data and / or medical monitoring time-series data.

[0068] Optionally, medical time-series data can include two main categories: medical imaging time-series data and medical monitoring time-series data. This classification can cover the main sources of time-series data in medicine. Medical imaging time-series data emphasizes the changes in spatial information over time, while medical monitoring time-series data emphasizes the changes in physiological parameters over time. The embodiments of this application are applicable to scenarios where medical time-series data includes only one type of time-series data, and also to fusion scenarios that include both types of time-series data.

[0069] Optionally, medical imaging time-series data can be a series of medical images of the same patient collected at different time points, such as a series of CBCT images during radiotherapy or a series of CT or MRI images used to assess tumor shrinkage. Medical monitoring time-series data can be patient vital signs data recorded continuously over a period of time (such as heart rate and blood pressure time-series curves), continuous blood glucose monitoring data, or multiple collections of laboratory test indicators (such as white blood cell count sequences). When constructing medical training samples, only image time-series data or only monitoring time-series data can be used. Alternatively, image features and monitoring indicator features on the same time axis can be fused together as input features for the neural network.

[0070] See Figure 2 According to another aspect of the embodiments of this application, a method for processing medical time-series data is also provided, including:

[0071] Step S201: Obtain the target medical time series data;

[0072] Step S202: Input the target medical time-series data into the time-series model, and output the classification prediction probability based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase. The time-series model is a neural network model trained based on medical training samples and their soft labels; the soft labels are generated based on the classification judgment auxiliary information corresponding to the medical training samples, and are continuously distributed probability values ​​used to reflect the confidence level of whether the medical training samples belong to the target category; the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features between different samples in the same category.

[0073] Step S203: Determine the target label corresponding to the target medical time series data based on the classification prediction probability.

[0074] Optionally, this application defines a method for processing medical time-series data corresponding to the training method of a time-series application model based on medical time-series data, i.e., a method for applying the trained model. This method for processing medical time-series data can be independent of the training process, forming a complete workflow. The training method of a time-series application model based on medical time-series data focuses on the offline training stage of the model, with the core being the use of soft labels generated based on classification judgment auxiliary information to train the neural network to obtain the time-series model. In contrast, the method for processing medical time-series data in this application focuses on the online application stage after the model training is completed. Its input is the target medical time-series data to be analyzed, and its output is the classification prediction probability based on the time-series model, ultimately determining the target label. The two methods, the training method of the time-series application model based on medical time-series data and the method for processing medical time-series data, can be interconnected by sharing a time-series model. This time-series model is the output of the training method of the time-series application model based on medical time-series data and also the key execution unit of the method for processing medical time-series data.

[0075] Optionally, the target medical time-series data can be a sequence of multiple cone-beam computed tomography (CBCT) images acquired sequentially from the start of radiotherapy to the current time point. Then, the step of inputting the target medical time-series data into the time-series model is performed. This time-series model is a neural network model trained using a training method for time-series application models based on medical time-series data. During the training phase, this model has learned through soft labeling how to output continuous classification prediction probabilities based on the data characteristics of the medical time-series data.

[0076] Optionally, based on the prior knowledge learned by the time series model during the training phase, a classification prediction probability is output based on the data features of the target medical time series data. The time series model applies the patterns it has learned internally to the input target medical time series data. For example, in the scenario of predicting the efficacy of radiotherapy, the time series model can analyze the feature change trends in CBCT image sequences and output a classification prediction probability between 0 and 1. This classification prediction probability can be interpreted as the confidence level that the final efficacy of the case represented by the current time series data belongs to the target category. Finally, the target label corresponding to the target medical time series data can be determined based on the classification prediction probability, and the classification prediction probability can be compared with one or more preset thresholds. For example, it can be set that when the classification prediction probability is lower than 0.3, the target label is determined to be "poor efficacy"; when the classification prediction probability is higher than 0.3, the target label is determined to be "good efficacy". This probability-based decision-making method can provide users with quantitative, non-absolute reference information.

[0077] Optionally, the entity executing the medical time-series data processing method can be a pre-deployed medical data processing software module or system. When it is necessary to predict new cases, the system first acquires the target medical time-series data for that case, such as the patient's vital signs sequence collected hourly over the past week. Then, it calls a pre-stored time-series model trained using a training method for time-series application models based on medical time-series data. After processing the input data, the time-series model outputs a classification prediction probability, for example, the probability of the patient developing septic shock is 0.3. Finally, the processing logic determines the target label as high-risk based on this probability value (e.g., combined with a threshold of 0.2), to ensure that the prediction model used is calibrated with soft labels.

[0078] In one optional embodiment, determining the target label corresponding to the target medical time series data based on the classification prediction probability includes: setting a target threshold based on the tolerance for prediction errors of various labels corresponding to the target medical time series data; and determining the target label from various labels corresponding to the target medical time series data based on the relationship between the classification prediction probability and the target threshold.

[0079] Optionally, embodiments of this application introduce a tolerance-based dynamic target threshold setting mechanism to quantify and integrate the asymmetric risk preference for different error types such as false positives and false negatives in clinical decision-making into the automated decision-making process. Instead of using a fixed threshold, the decision threshold is dynamically set based on the label type targeted by the current decision and the severity of the consequences of misjudging the label, i.e., the tolerance, so that the final classification decision is closely linked to the clinical risk control objective.

[0080] For example, suppose the target label in a treatment efficacy prediction scenario is "poor efficacy". Clinically, the tolerance for misclassifying a poor efficacy result as a good one (false negative) is extremely low, denoted as L; the tolerance for misclassifying a good efficacy result as a poor one (false positive) is relatively high, denoted as H. The system sets a target threshold based on the tolerance L for prediction errors (false negatives) for the "poor efficacy" label. For example, a very strict threshold T (e.g., 0.3) can be set. During decision-making, the model's output classification prediction probability (here, the probability that the model outputs "good efficacy") is compared with T. If the probability is lower than T (i.e., the model's confidence in "good efficacy" is very low), the target label is determined to be "poor efficacy"; otherwise, it is "good efficacy". In this way, only when the model is highly confident will it make a high-risk (low-tolerance error) judgment, thus controlling risk.

[0081] In an optional embodiment, the method for processing medical time series data further includes: when the target medical time series data is updated over time, inputting the updated target medical time series data into a time series model, outputting an updated classification prediction probability based on the data features of the updated target medical time series data according to the time series model; and determining the target label corresponding to the updated target medical time series data based on the updated classification prediction probability.

[0082] Optionally, this application embodiment, by iteratively re-predicting time-series data that changes over time, can simulate the real-world scenario of continuous observation and dynamic assessment in clinical practice. It can utilize longer data sequences accumulated over time to improve prediction accuracy and correct early, uncertain predictions. This application embodiment demonstrates the continuity and self-updating capability of time-series prediction, and is particularly suitable for the management of chronic diseases or treatment processes with long treatment courses and multiple assessments.

[0083] For example, when acquiring a time-series model based on medical time-series data processing methods and applying it to clinical prediction, an iterative processing strategy can be implemented. For instance, the steps of setting a target threshold based on tolerance can be combined with re-predicting using updated data. When applying this time-series model for early prediction, a low target threshold, such as 0.3, is first set based on the tolerance for the error of predicting a target medical time-series data belonging to the "good treatment" category as "bad treatment." When the target medical time-series data (such as a sequence of cone-beam computed tomography images from the five sessions prior to radiotherapy) is input into the time-series model, a decision is made by comparing the classification prediction probability output by the time-series model with the target threshold.

[0084] If the classification prediction probability is lower than the target threshold, the target label for the target medical time series data can be determined as "bad efficacy" based on the classification prediction probability. If the classification prediction probability is higher than the target threshold, the target label can be temporarily determined as "good efficacy". In this case, no immediate intervention is necessary; instead, data collection can continue.

[0085] In subsequent processes, the target medical time-series data will be updated over time; for example, the number of treatments may increase to 15, lengthening the sequence. At this point, the updated target medical time-series data can be input back into the time-series model to obtain updated classification prediction probabilities. For samples initially predicted as having good efficacy but whose updated classification prediction probabilities significantly decrease (e.g., far below 0.3), the target label can be corrected to "bad efficacy." This iterative process of dynamically updating prediction results based on data sequence growth helps to re-evaluate samples with high uncertainty using richer information, while ensuring the reliability of early decisions.

[0086] See Figure 3 According to another aspect of the embodiments of this application, a training device for a time-series application model based on medical time-series data is also provided, including: a sample acquisition unit 301, a soft label generation unit 302, and a model training unit 303.

[0087] The sample acquisition unit 301 is used to acquire medical training samples, which include medical time-series data. The soft label generation unit 302 is used to generate soft labels for the medical training samples based on the classification judgment auxiliary information corresponding to the medical training samples. The classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features between different samples in the same category. The soft label is a probability value of a continuous distribution, which is used to reflect the confidence level of the medical training sample belonging to the target category. The model training unit 303 is used to iteratively train the neural network based on the medical training samples and the soft labels of the medical training samples to obtain a time-series model.

[0088] Optionally, the soft label generation unit 302 includes at least one of the following determining subunits: a first determining subunit for determining the tumor shrinkage rate, which characterizes the proportion of tumor volume reduction; a second determining subunit for determining the organ-at-risk damage rate, which characterizes the proportion of damage to organs at risk during radiotherapy; a third determining subunit for determining the polymerase chain reaction (PCR) amplification rounds, which reflects the original concentration of the target nucleic acid in the nucleic acid test sample; and a fourth determining subunit for determining the dose deviation, which characterizes the difference between the dose obtained by analyzing the pseudo-images after matching the first and second images during radiotherapy and the original planned dose, wherein the first image is a medical image taken before each radiotherapy session, and the second image is a positioning medical image initially used to formulate the radiotherapy plan.

[0089] Optionally, the soft tag generation unit 302 includes: an information mapping subunit, used to obtain a value from 0 to 1 by mapping auxiliary information based on classification judgment; and a soft tag generation subunit, used to generate a soft tag based on the value obtained by mapping.

[0090] Optionally, the information mapping subunit includes: a first value determination module, used to map a first value based on the actual value of the tumor shrinkage rate when the auxiliary information for classification is tumor shrinkage rate, wherein the value of the first value is positively correlated with the tumor shrinkage rate; a second value determination module, used to map a second value based on the actual value of the organ-at-risk damage rate when the auxiliary information for classification is organ-at-risk damage rate, wherein the value of the second value is negatively correlated with the organ-at-risk damage rate; a third value determination module, used to map a third value based on the actual value of the polymerase chain reaction (PCR) amplification epochs when the auxiliary information for classification is polymerase chain reaction (PCR) amplification epochs, wherein the value of the third value is negatively correlated with the PCR amplification epochs; and a fourth value determination module, used to map a fourth value based on the actual value of the dose deviation when the auxiliary information for classification is dose deviation, wherein the value of the fourth value is negatively correlated with the dose deviation.

[0091] Optionally, the training device for the time-series application model based on medical time-series data further includes: a result label determination unit, used to determine the result label corresponding to the medical training sample after acquiring the medical training sample, wherein the result label is used to mark the category to which the medical training sample belongs; a tolerance determination unit, used to determine the prediction error tolerance for the category marked by the result label; a first determination unit, used to generate a soft label for the medical training sample based on the classification judgment auxiliary information corresponding to the medical training sample when it is determined that the prediction error tolerance is greater than a preset threshold; and a second determination unit, used to use the result label corresponding to the medical training sample as the soft label of the medical training sample when it is determined that the prediction error tolerance is less than or equal to the preset threshold.

[0092] Optionally, the soft tag generation subunit includes: a first soft tag generation module, used to generate soft tags based on the values ​​obtained from mapping each indicator information when the classification judgment auxiliary information includes multiple indicator information and a corresponding value is obtained based on mapping each indicator information, using method one; method one includes: performing a weighted calculation on the values ​​obtained from mapping each indicator information and generating soft tags based on the result of the weighted calculation; and a second soft tag generation module, used to generate soft tags based on the values ​​obtained from mapping each indicator information when the classification judgment auxiliary information includes multiple indicator information and a corresponding value is obtained based on mapping each indicator information, using method two; method two includes: taking the logarithm of the values ​​obtained from mapping each indicator information and summing them, generating soft tags based on the calculated result.

[0093] Optionally, the training device for the time-series application model based on medical time-series data further includes: a time-series data acquisition unit for acquiring target medical time-series data; a data processing unit for inputting the target medical time-series data into the time-series model and outputting classification prediction probabilities based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase; and a target label determination unit for determining the target label corresponding to the target medical time-series data according to the classification prediction probabilities.

[0094] Optionally, the sample acquisition unit 301 includes a medical image time series data acquisition subunit and / or a medical monitoring time series data acquisition subunit.

[0095] See Figure 4 According to another aspect of the embodiments of this application, a medical time-series data processing device is also provided, including: a data acquisition unit 401, a first processing unit 402, and a second processing unit 403.

[0096] The system includes a data acquisition unit 401 for acquiring target medical time-series data; a first processing unit 402 for inputting the target medical time-series data into a time-series model and outputting classification prediction probabilities based on the data features of the target medical time-series data according to the prior knowledge learned by the time-series model during the training phase; wherein the time-series model is a neural network model trained based on medical training samples and their soft labels; the soft labels are generated based on the classification judgment auxiliary information corresponding to the medical training samples, and the soft labels are a continuously distributed probability value used to reflect the confidence level of the medical training samples belonging to the target category; wherein the classification judgment auxiliary information represents a continuous variable used to distinguish and quantify the degree of difference in features between different samples in the same category; and a second processing unit 403 for determining the target label corresponding to the target medical time-series data based on the classification prediction probabilities.

[0097] Optionally, the second processing unit 403 includes: a target threshold setting subunit, used to set a target threshold based on the tolerance for prediction errors of various labels corresponding to the target medical class time series data; and a target label determination subunit, used to determine a target label from various labels corresponding to the target medical class time series data based on the relationship between the classification prediction probability and the target threshold.

[0098] Optionally, the medical time-series data processing device further includes: a data update processing unit, used to input the updated target medical time-series data into the time-series model when the target medical time-series data is updated over time, and output the updated classification prediction probability based on the data characteristics of the updated target medical time-series data according to the time-series model; and an update information determination unit, used to determine the target label corresponding to the updated target medical time-series data according to the updated classification prediction probability.

[0099] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the above-described training method for a time-series application model based on medical time-series data.

[0100] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to execute the above-described training method for a time-series application model based on medical time-series data.

[0101] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described training method for a time-series application model based on medical time-series data.

[0102] In addition, according to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described method for processing medical time-series data.

[0103] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the above-described method for processing medical time-series data.

[0104] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described method for processing medical time-series data.

[0105] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0106] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0107] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0108] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0109] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0110] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0111] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for training a time series application model based on medical time series data, characterized in that, The method comprises: obtaining a medical training sample, wherein the medical training sample comprises medical time series data; generating a soft label for the medical training sample based on classification judgment auxiliary information corresponding to the medical training sample, wherein the classification judgment auxiliary information represents a continuous variable for distinguishing and quantifying the difference between different sample characteristics in the same category; the soft label is a continuous distribution probability value, and the soft label is used to reflect the confidence degree of the medical training sample belonging to a target category; iteratively training a neural network according to the medical training sample and the soft label of the medical training sample to obtain a time series model. 2.The method of claim 1, wherein, The classification judgment auxiliary information comprises at least one of the following index information: a tumor shrinkage rate for representing a reduction ratio of a tumor volume; an organ-at-risk damage degree rate for representing a damage ratio of an organ-at-risk in a radiotherapy process; a polymerase chain reaction amplification round for reflecting an original concentration of a target nucleic acid in a nucleic acid test sample; a dose deviation for representing a difference between a dose obtained by analyzing a pseudo image matched by a first image and a second image and a planned dose, the first image being a medical image taken before each radiotherapy, and the second image being a positioning medical image used for initially formulating a radiotherapy plan. 3.The method of claim 2, wherein, The method for generating a soft label for the medical training sample based on classification judgment auxiliary information corresponding to the medical training sample comprises: mapping the classification judgment auxiliary information to obtain a value between 0 and 1; generating the soft label according to the obtained value. 4.The method of claim 3, wherein, The method for mapping the classification judgment auxiliary information to obtain a value between 0 and 1 comprises: when the classification judgment auxiliary information is the tumor shrinkage rate, mapping an actual value of the tumor shrinkage rate to obtain a first value, wherein the first value is positively correlated with the tumor shrinkage rate; when the classification judgment auxiliary information is the organ-at-risk damage degree rate, mapping an actual value of the organ-at-risk damage degree rate to obtain a second value, wherein the second value is negatively correlated with the organ-at-risk damage degree rate; when the classification judgment auxiliary information is the polymerase chain reaction amplification round, mapping an actual value of the polymerase chain reaction amplification round to obtain a third value, wherein the third value is negatively correlated with the polymerase chain reaction amplification round; when the classification judgment auxiliary information is the dose deviation, mapping an actual value of the dose deviation to obtain a fourth value, wherein the fourth value is negatively correlated with the dose deviation. 5.The method of claim 3, wherein, The method for training a time series application model based on medical time series data further comprises: after obtaining the medical training sample, determining a result label corresponding to the medical training sample, wherein the result label is used to mark a category to which the medical training sample belongs; determining a prediction error tolerance for the category according to the category marked by the result label; In a case where it is determined that the prediction error tolerance is greater than a preset threshold, a soft label is generated for the medical training sample based on classification judgment auxiliary information corresponding to the medical training sample; In a case where it is determined that the prediction error tolerance is less than or equal to the preset threshold, a result label corresponding to the medical training sample is taken as the soft label of the medical training sample. 6.The method of claim 3, wherein, The soft label is generated according to the values obtained through the mapping, including: In a case where the classification judgment auxiliary information includes multiple index information, and a value corresponding to each index information is obtained through the mapping of the index information, the soft label is generated according to the values obtained through the mapping of each index information in the following manner one or manner two: Manner one, including: performing weighted calculation on the values obtained through the mapping of each index information, and generating the soft label according to a result of the weighted calculation; Manner two, including: taking logarithms of the values obtained through the mapping of each index information and adding them, and generating the soft label according to a calculation result obtained. 7.The method of claim 1, wherein, After the neural network is iteratively trained according to the medical training sample and the soft label of the medical training sample to obtain a time sequence model, the training method of the time sequence application model based on medical time sequence data further includes: Obtaining target medical time sequence data; Inputting the target medical time sequence data into the time sequence model, outputting a classification prediction probability based on data characteristics of the target medical time sequence data according to prior knowledge learned by the time sequence model in a training stage; and Determining a target label corresponding to the target medical time sequence data according to the classification prediction probability. 8.The method of claim 1 to 7, wherein, The medical time sequence data includes medical image time sequence data and / or medical monitoring time sequence data.

9. A method of processing medical time series data, the method comprising: Including: Obtaining target medical time sequence data; Inputting the target medical time sequence data into the time sequence model, outputting a classification prediction probability based on data characteristics of the target medical time sequence data according to prior knowledge learned by the time sequence model in a training stage; and The time sequence model is a neural network model trained according to a medical training sample and a soft label of the medical training sample; the soft label is generated based on classification judgment auxiliary information corresponding to the medical training sample, and the soft label is a continuous distribution probability value used to reflect a confidence degree of the medical training sample belonging to a target category; the classification judgment auxiliary information represents continuous variables used to distinguish and quantify differences in characteristics of different samples in a same category; Determining a target label corresponding to the target medical time sequence data according to the classification prediction probability.

10. The method of claim 9, wherein, Determining a target label corresponding to the target medical time sequence data according to the classification prediction probability, including: Setting a target threshold according to a tolerance of prediction errors of various labels corresponding to the target medical time sequence data; Determining the target label from the various labels corresponding to the target medical time sequence data according to a size relationship between the classification prediction probability and the target threshold.

11. The method of claim 9, wherein, The processing method of the medical time sequence data further includes: In a case where the target medical class time series data is updated over time, the updated target medical class time series data is input into the time series model, and the time series model outputs updated classification prediction probability based on data features of the updated target medical class time series data; The target label corresponding to the updated target medical class time series data is determined according to the updated classification prediction probability. 12.A model training apparatus based on medical time-series data, characterized by comprising: The method comprises the following steps: An acquisition unit is configured to acquire medical training samples, wherein the medical training samples comprise medical time series data; A soft label generation unit is configured to generate soft labels for the medical training samples based on classification judgment auxiliary information corresponding to the medical training samples, wherein the classification judgment auxiliary information represents continuous variables for distinguishing and quantifying differences in features of different samples in the same class; the soft label is a continuous distribution probability value, and the soft label is used to reflect a confidence degree of the medical training sample belonging to a target class; A model training unit is configured to iteratively train a neural network according to the medical training samples and the soft labels of the medical training samples, to obtain a time series model.

13. An apparatus for processing medical time series data, characterized by The method comprises the following steps: A data acquisition unit is configured to acquire target medical class time series data; A first processing unit is configured to input the target medical class time series data into a time series model, and output classification prediction probability based on data features of the target medical class time series data according to prior knowledge learned by the time series model in a training stage. The time series model is a neural network model trained according to medical training samples and soft labels of the medical training samples; the soft label is generated based on classification judgment auxiliary information corresponding to the medical training samples, and the soft label is a continuous distribution probability value used to reflect a confidence degree of the medical training sample belonging to a target class; the classification judgment auxiliary information represents continuous variables for distinguishing and quantifying differences in features of different samples in the same class; A second processing unit is configured to determine a target label corresponding to the target medical class time series data according to the classification prediction probability.

14. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and when the computer program runs, the device in which the computer readable storage medium is located executes the training method of the time series application model based on medical time series data according to any one of claims 1 to 8.

15. An electronic device, comprising: The device comprises one or more processors and a memory, and the memory is configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the training method of the time series application model based on medical time series data according to any one of claims 1 to 8.

16. A computer program product, characterised in that, The computer program or instructions are executed by the processor to implement the training method of the time series application model based on medical time series data according to any one of claims 1 to 8.

Images