A dynamic monitoring and early warning ovarian cancer recurrence management method and system

Abstract

The application relates to the cross field of computer technology and medical information processing technology, in particular to a dynamic monitoring and early warning ovarian cancer recurrence management method and system, which is executed by a computer device and comprises the following steps executed by the computer device: acquiring multi-modal follow-up data of a to-be-evaluated object at different time points, analyzing and segmenting a disease course time axis of the to-be-evaluated object into multiple preset clinical stages; for any data point in the multi-modal follow-up data, a time sequence weight factor is calculated and generated; a set of time sequence enhanced features is constructed and arranged in time sequence to form a time sequence enhanced feature sequence; a dynamic risk index quantitatively representing the possibility of ovarian cancer recurrence is processed and output by a time sequence prediction model. The dynamic baseline change characteristics considering the baseline drift together constitute the time sequence enhanced features with information content far more than the original data, and dynamic and sensitive evaluation of the recurrence risk is realized.

Description

Technical Field

[0001] This application relates to the intersection of computer technology and medical information processing technology, and in particular to a method and system for dynamic monitoring and early warning management of ovarian cancer recurrence. Background Technology

[0002] Ovarian cancer is a common gynecological malignancy with a high recurrence rate and poor prognosis, posing a serious threat to women's health. Long-term dynamic follow-up monitoring of ovarian cancer patients (those under evaluation) and early warning of recurrence risk based on monitoring data are of significant clinical importance in guiding clinical interventions and improving the quality of life of those under evaluation. Existing technologies have been extensively explored in the field of ovarian cancer recurrence risk prediction.

[0003] However, analysis of existing technical solutions reveals that while existing technologies may incorporate information such as clinical staging when constructing models, they fail to effectively combine the dynamic characteristics of relapse risk evolving over time with the static contextual information of the different clinical stages of the subjects being evaluated. Whether building models based on features from a single data collection or extracting information from text, they tend to treat monitoring data at different time points as isolated or equally weighted events, ignoring the deeper meaning implied by the specific positions of data points on the long treatment and rehabilitation timeline of the subjects being evaluated. Summary of the Invention

[0004] The objective of this application is to provide a method for dynamic monitoring and early warning management of ovarian cancer recurrence. This method is executed by a computer device and includes the following steps performed by the computer device: acquiring multimodal follow-up data of a subject to be evaluated at different time points, wherein the multimodal follow-up data includes at least biomarker detection values ​​and clinical events defining the treatment cycle; based on the clinical events, parsing and segmenting the subject's disease progression timeline into multiple preset clinical stages; for any data point in the multimodal follow-up data, calculating and generating a time-series weighting factor based on the clinical stage to which the data point belongs and the time span between the data point and a preset major treatment event; constructing a set of time-series enhanced features based on the data point's data value, the time-series weighting factor, and dynamic baseline change characteristics derived from historical data, and arranging them in chronological order to form a time-series enhanced feature sequence; inputting the time-series enhanced feature sequence into a pre-trained time-series prediction model, which processes and outputs a dynamic risk index that quantifies the likelihood of ovarian cancer recurrence.

[0005] By adopting the above technical solution, two dimensions, clinical stage and time span, are introduced to evaluate the contextual significance of each data point and quantify them into time series weighting factors. Combined with the dynamic baseline change characteristics that take into account baseline drift, they together constitute a time series enhancement feature with information content far exceeding that of the original data. Finally, by using a time series prediction model designed for processing sequence data, a dynamic and sensitive assessment of relapse risk is achieved.

[0006] Optionally, before constructing the time-series enhancement feature, the method further includes: acquiring metadata associated with each data point in the multimodal follow-up data, wherein the metadata at least indicates the data detection source or data detection method; setting a data reliability score for each data point based on a preset scoring rule library and the metadata; and incorporating the data reliability score as an additional feature into the construction process of the time-series enhancement feature.

[0007] By adopting the above technical solutions, the model can distinguish and weigh data of different qualities during the learning process, giving higher trust to data with high reliability, thereby effectively suppressing the interference of noisy data on model performance and enhancing the robustness and credibility of prediction results.

[0008] Optionally, before constructing the time-series enhancement feature, the method further includes: analyzing the time series of follow-up events of the subject to be evaluated and comparing it with a standard follow-up plan to quantify the degree of abnormality of the follow-up rhythm; and incorporating the degree of abnormality of the follow-up rhythm as another additional feature into the construction process of the time-series enhancement feature.

[0009] By adopting the above technical solution, the behavioral characteristics of the subjects to be evaluated are incorporated into the risk assessment model, enabling the model to gain insight into potential health deterioration trends from a new dimension and achieve a more comprehensive and three-dimensional assessment of relapse risk.

[0010] Optionally, the metadata may also include the visit type, which is categorized as planned follow-up visit or unplanned temporary visit.

[0011] By adopting the above technical solutions, behavioral patterns can be characterized more accurately.

[0012] Optionally, the step of constructing the dynamic baseline change characteristics derived from historical data includes: for a specific biomarker, calculating an independent dynamic baseline value for each clinical stage based on its historical monitoring values ​​at different clinical stages; calculating the rate of change and acceleration of change of the newly detected value of the specific biomarker relative to the dynamic baseline value of its respective clinical stage; and using the calculated rate of change and acceleration of change as components of the dynamic baseline change characteristics.

[0013] By adopting the above technical solution, it is possible to identify the danger signal of a steady to accelerated rise earlier. For example, even if the value of a certain marker is still within the normal range, but it shows an accelerated upward trend relative to the baseline of the remission period, this application can also keenly capture this potential risk and greatly advance the warning window.

[0014] Optionally, the method further includes: obtaining a subjective symptom report submitted by the subject to be evaluated in the form of unstructured natural language text; invoking a pre-trained biomedical natural language processing model to perform semantic analysis on the unstructured natural language text to extract predefined symptom entities and quantitative intensity descriptions of the symptom entities; encoding the extracted symptom entities and the quantitative intensity descriptions into a subjective symptom vector; and aligning and fusing the subjective symptom vector and the temporal enhancement feature sequence in the time dimension before inputting the temporal enhancement feature sequence into the temporal prediction model.

[0015] By adopting the above technical solution, the personal feelings of the subjects to be evaluated are transformed into quantitative characteristics that the model can understand, thus constructing a more complete and comprehensive health profile of the subjects to be evaluated. This enables the model to capture non-specific early signs of relapse that may appear earlier than changes in biomarkers, such as mild abdominal distension and fatigue, thereby further improving the comprehensiveness and sensitivity of the prediction.

[0016] Optionally, the method also includes obtaining structured health questionnaire data submitted by the subject to be evaluated, and incorporating the structured health questionnaire data when encoding the subjective symptom vector.

[0017] Optionally, the clinical event includes the surgery date, chemotherapy start date, or chemotherapy end date; the step of parsing and segmenting the disease course timeline of the subject to be evaluated into multiple preset clinical stages specifically involves segmenting the disease course timeline into at least two of the following based on the surgery date, chemotherapy start date, or chemotherapy end date: initial treatment stage, early remission stage, stable remission stage, and long-term remission stage.

[0018] By adopting the above technical solutions, the operability and clinical relevance of the solutions were ensured.

[0019] Optionally, the time-series prediction model is a long short-term memory network model or a variant thereof.

[0020] The second objective of this application is to provide a dynamic monitoring and early warning management system for ovarian cancer recurrence, comprising multiple functional modules configured to execute the aforementioned methods. Each functional module includes at least: a data acquisition module for acquiring multimodal follow-up data of a subject to be evaluated at different time points, the multimodal follow-up data including at least biomarker detection values ​​and clinical events defining the treatment cycle; and a central processing module, communicatively connected to the data acquisition module, configured to perform the following operations: based on the clinical events, analyzing and segmenting the disease progression timeline of the subject to be evaluated into multiple preset clinical stages; and targeting the... For any data point in the multimodal follow-up data, a time-series weighting factor is calculated and generated based on the clinical stage to which the data point belongs and the time span between the data point and a preset major treatment event. A set of time-series enhancement features is constructed based on the data value of the data point, the time-series weighting factor, and dynamic baseline change characteristics derived from historical data. A risk generation module, which is communicatively connected to the central processing module, is configured to input the time-series enhancement feature sequence into a pre-trained time-series prediction model, and the model processes the data to output a dynamic risk index that quantifies the possibility of ovarian cancer recurrence. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the dynamic monitoring and early warning management method for ovarian cancer recurrence in this application.

[0022] Figure 2 This is a flowchart of the ovarian cancer recurrence management system for dynamic monitoring and early warning, as described in this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0024] like Figure 1 As shown, this embodiment provides a method for dynamic monitoring and early warning management of ovarian cancer recurrence. This method is executed on a computer device, such as a software program deployed on a cloud server, a hospital local server, or a dedicated medical computing workstation. By executing a series of steps, it achieves continuous and individualized assessment of the recurrence risk of the subject to be evaluated. The method specifically includes the following steps.

[0025] S01: Obtain multimodal follow-up data of a subject to be evaluated (e.g., an ovarian cancer patient) at different time points. The multimodal follow-up data shall include at least biomarker test values ​​and clinical events that define the treatment cycle.

[0026] Understandably, this step is performed by the data acquisition module of the computer equipment. Biomarker detection values ​​may include, but are not limited to, the concentration values ​​of serum cancer antigen 125 (CA125), human epididymal protein 4 (HE4), and the ROMA index, an ovarian malignancy risk index calculated from these two markers. These detection values ​​usually have a timestamp accurate to the day, recording the specific time of collection. Clinical events that define the treatment cycle are the anchor points for the structured division of the entire course of the disease for the subject to be evaluated. Specifically, these may include, but are not limited to, the surgery date of the initial cytoreductive surgery, the date of the start of the first chemotherapy, the date of the end of the last chemotherapy, the date of imaging examinations such as computed tomography or magnetic resonance imaging, and the date of pathological diagnosis confirming recurrence. The data acquisition process may be to automatically pull structured data from hospital information systems, laboratory information systems, or electronic medical record systems through standardized interface protocols, or to receive manually entered data through a graphical user interface designed for the subject to be evaluated or the data entry operator. All acquired data will be uniformly stored in a time-series database associated with a unique identifier of the subject to be evaluated for subsequent processing.

[0027] S02: Based on clinical events, the disease course timeline of the subject to be evaluated is analyzed and divided into multiple pre-defined clinical stages.

[0028] Understandably, this step is executed by the central processing module of the computer device. Based on the clinical events acquired in the previous step, it analyzes and divides the seemingly continuous and complete disease course timeline of the subject to be evaluated into multiple pre-defined clinical stages in a way that has clear clinical significance. Specifically, this segmentation logic can be encoded into a set of deterministic rules. For example, if the system identifies the clinical event of "end date of last chemotherapy," it uses this as the zero point to divide the subsequent timeline. For example, clinical stages can be defined as the initial treatment stage, early remission stage, stable remission stage, and long-term remission stage. The initial treatment stage refers to the period from the date of diagnosis or surgery to the end of the last chemotherapy, during which biomarker levels usually experience a sharp decline. The early remission stage refers to the period from the first day after the end of the last chemotherapy to one year after surgery, which is an observation period with a relatively high risk of recurrence. The stable remission stage refers to the period from one year to three years after surgery, during which the condition of the subject to be evaluated tends to be stable, but close monitoring is still required. The long-term remission stage refers to the period more than three years after surgery, during which the risk of recurrence decreases year by year, but weak recurrence signals may still appear. By segmenting the data in this way, the originally simple time series data is given a rich clinical context, which allows subsequent analysis to be differentiated according to the pathophysiological characteristics of different stages. For example, a small increase in CA125 value in the long-term remission stage may have a much greater clinical significance than the same fluctuation in value at the end of the initial treatment stage.

[0029] S03: For any data point in the multimodal follow-up data, a time-series weighting factor is calculated and generated based on the clinical stage to which the data point belongs and the time span between the data point and a pre-defined major treatment event.

[0030] Understandably, this step can also be performed by the central processing module. For any isolated data point in the multimodal follow-up data, a time-series weighting factor can be customized based on two key dimensions. The first dimension is the clinical stage to which the data point belongs, reflecting its importance within the broader context of the disease course. For example, a data point in early remission might be assigned a higher base weight because this period is a critical window for assessing treatment efficacy and predicting early relapse. The second dimension is the time span between the data point and a pre-defined major treatment event. This major treatment event is typically the event most relevant to the prognosis of the subject being evaluated, such as the end date of the last chemotherapy session. The closer the time span, the more immediate and valuable the data point's reflection of the current state, and therefore, its weight should be higher. The specific calculation process can be implemented as a mathematical function; for example, the time-series weighting factor W can be expressed as W = P_weight * f(Δt), where P_weight is the basic weight value of the clinical stage (such as early remission stage, stable remission stage, etc.) obtained by looking up a table, and f(Δt) is a monotonically decreasing function with respect to the time span Δt, such as an exponential decay function f(Δt) = exp(-λ * Δt), where λ is an adjustable decay coefficient used to control the rate at which the weight decays over time. By calculating the time-series weight factor for each data point, the subsequent model can intelligently distinguish the amount of information value contained in data points at different times or different disease stages, thereby avoiding the information loss caused by treating all data points the same.

[0031] S04: Based on the data values ​​of the data points, the time-series weighting factors, and the dynamic baseline change characteristics derived from historical data, a set of time-series enhancement features are jointly constructed and arranged in chronological order to form a time-series enhancement feature sequence.

[0032] Understandably, this step can still be executed by the central processing module. For each data point, its three aspects of information are integrated and vectorized. The first aspect is the original data value of the data point itself, such as the specific detection value of CA125. The second aspect is the time-series weighting factor calculated for this data point in the previous step. The third aspect is the dynamic baseline change feature derived from historical data. It is understood that the introduction of the dynamic baseline change feature is to capture the trend of data change, not just its absolute value. By organizing these three aspects of information into a multi-dimensional feature vector, the construction of the time-series enhancement feature for a data point is completed. For example, for the CA125 detection at time point t, its original value is V_t, the calculated time-series weighting factor is W_t, and the derived dynamic baseline change feature is [R_t, A_t]. Then, the time-series enhancement feature vector at time t can be represented as [V_t, W_t, R_t, A_t]; Arranging these feature vectors of the object to be evaluated in chronological order creates a temporally enhanced feature sequence, which contains richer dynamic change information and clinical context information compared to the original detection value sequence.

[0033] S05: Input the time-enhanced feature sequence into the pre-trained time-series prediction model, which processes and outputs a dynamic risk index that quantifies the likelihood of ovarian cancer recurrence.

[0034] Understandably, this step is performed by the risk generation module, which receives the complete time-series augmented feature sequence output by the central processing module and sends it as input to a pre-trained time-series prediction model. This time-series prediction model is adept at handling and learning dependencies in long-term time-series data; for example, it could employ LSTM (Long Short-Term Memory). Long Short-Term Memory (LSTM) networks or variants thereof, such as gated recurrent units (GRUs), utilize special structures such as forget gates, input gates, and output gates within their models. These structures enable LTMs to effectively learn which historical information needs to be remembered long-term and which can be forgotten. This is crucial for capturing weak but persistent recurrence signals in ovarian cancer after years of remission. The time-series prediction model, after processing the entire input feature sequence, outputs a continuous numerical value at its output layer, defined as a dynamic risk index. This dynamic risk index is typically normalized to a range of 0 to 1; a higher value indicates a greater likelihood of ovarian cancer recurrence in the currently evaluated subject. Since the entire time-series enhanced feature sequence is updated each time new follow-up data points are added and can be re-inputted into the model for calculation, this method generates a dynamically updated risk curve over time, providing clinicians with an intuitive risk monitoring tool. This enables a quantitative characterization and dynamic early warning of the likelihood of ovarian cancer recurrence.

[0035] Understandably, this embodiment adds a preprocessing step of data quality assessment and quantification before the step of constructing time-series enhanced features. Specifically, when performing the data acquisition step, in addition to acquiring biomarker detection values ​​and clinical events, the method further acquires metadata associated with each data point in the multimodal follow-up data. This metadata describes the source and background of the data itself, and at least indicates the detection source or detection method of the data. For example, for a CA125 detection value, its metadata may include the name of the hospital that performed the test, the qualification level of the laboratory, the specific detection instrument model used, and the detection technology used, such as chemiluminescence immunoassay or enzyme-linked immunosorbent assay. After acquiring this metadata, the system will assign a score to each data point based on a preset scoring rule base. A data reliability score is established. This score rule base can be jointly developed by medical experts and data scientists, containing rules that map different metadata to specific scores. For example, the rules can define: CA125 values ​​detected by a nationally certified laboratory using a mainstream brand fully automated chemiluminescence analyzer have a data reliability score of 0.98; values ​​detected by a community health service center using a semi-automated enzyme-linked immunosorbent assay (ELISA) may have a reliability score of 0.75; manually entered data from unknown sources may have a reliability score as low as 0.4. After calculating this data reliability score between 0 and 1 for each data point, this score will be incorporated as a new additional feature into the construction process of time-series enhancement features. For example, it can be included in the feature vector [V_t, ...] described in the above embodiment. Based on [W_t, R_t, A_t], a data reliability score S_t is added to form a higher-dimensional feature vector [V_t, W_t, R_t, A_t, S_t]. In this way, the time series prediction model can learn the relationship between the data reliability score and the prediction target during training and prediction. It will automatically give higher trust and weight to data with high scores and good quality, while suppressing the interference of data with low scores and possible noise or bias on the model's judgment. This enhances the stability of the entire method when facing complex and ever-changing real-world data and the reliability of the final output dynamic risk index.

[0036] Understandably, as a further enrichment and utilization of data metadata, the metadata in this embodiment may also include visit type, which is divided into planned follow-up visits or unplanned temporary visits. Planned follow-up visits refer to the patient coming for a follow-up examination according to a pre-established regular schedule by the doctor, such as a check-up every three months in the first year after surgery. Unplanned temporary visits, on the other hand, refer to the patient coming to the hospital voluntarily or temporarily due to the appearance of new discomfort symptoms or a perceived change in their condition. These two different visit types imply drastically different clinical significance; an unplanned visit itself is a very strong signal, potentially indicating that the patient... The health condition of the person being assessed is potentially deteriorating; therefore, when processing the data, the system encodes the type of medical visit as well. For example, planned follow-up is encoded as 0, and unplanned medical visits are encoded as 1. This encoded value is also incorporated as a feature into the time-series augmentation feature vector, or used to adjust the calculation of the time-series weight factor. For example, for a data point from an unplanned medical visit, even if its value is still within the normal range, the system can amplify its influence in the model by assigning it a higher time-series weight or by using its unique feature encoding, so that the model can more sensitively capture the early risk signals reflected by changes in the behavior patterns of the person being assessed.

[0037] It is understood that, before constructing the time-series enhancement features, this application embodiment can add a step of analyzing the follow-up behavior pattern of the subject to be evaluated. Specifically, after the data acquisition module collects all follow-up events, the central processing module can execute a subroutine for follow-up rhythm analysis. This subroutine first reads the standard follow-up plan for the current clinical stage of the subject to be evaluated from a configuration library. This standard follow-up plan is usually defined by clinical guidelines or the attending physician. For example, for the subject to be evaluated in the stable remission stage, the standard follow-up plan may be a follow-up examination every six months. Then, the subroutine accurately compares the time series of all follow-up events actually occurring in the subject to be evaluated with this standard follow-up plan. The purpose of the comparison is to quantify the degree of abnormality in the follow-up rhythm. There are various quantification methods. The system can calculate the statistical measure of the deviation between the actual follow-up interval and the standard interval, such as calculating its variance or standard deviation. A large variance means that the follow-up behavior of the subject to be evaluated is very irregular. The system can also statistically analyze the number or frequency of unplanned medical visits within a specific time window, such as the past year, and use this frequency as a direct measure of the degree of abnormality. The calculated abnormality value of this follow-up rhythm, whether variance or frequency, will be used as a new behaviorally significant additional feature and will be used together to generate the final dynamic risk index. Specifically, it can be directly concatenated into the time-series enhanced feature vector and passed to the downstream time-series prediction model. The advantage of doing so is that the model no longer relies solely on physiological and biochemical indicators, but begins to understand the behavioral patterns of the subject being evaluated. For example, if a subject who was previously followed up very regularly suddenly experiences multiple unplanned intensive medical visits recently, even if their biomarkers have not yet exceeded the warning line, this drastic change in behavioral pattern may itself be an important early warning signal, indicating that the subject may be experiencing some discomfort. The method of this application can capture and quantify this signal, thereby improving the sensitivity and foresight of the warning.

[0038] Understandably, the steps for constructing dynamic baseline change characteristics can be performed by the central processing module, specifically including the following steps.

[0039] The first step involves calculating an independent dynamic baseline value for each clinical stage based on historical monitoring values ​​of a specific biomarker at different clinical phases. Specifically, for a given biomarker, such as CA125, the system first filters out all historical monitoring values ​​of the subject within a specific clinical phase, such as stable remission. Then, these selected values ​​are statistically calculated to obtain a baseline representing the normal level for the subject at this stage. This calculation can be the simplest mean or median, or a more robust statistic, such as a truncated mean, to exclude the influence of extreme outliers. This dynamic baseline value is calculated for a specific subject at a specific stage. For example, subject A's CA125 dynamic baseline during stable remission might be 8.5 U / mL, while subject B's baseline during the same stage might be 12.0 U / mL, fully reflecting individual differences. Furthermore, when the subject enters the next clinical stage, such as long-term remission, the system will recalculate a completely new dynamic baseline value using data from the new stage.

[0040] The second step, after obtaining the dynamic baseline values ​​for each stage, involves calculating the rate of change and acceleration of a new test value relative to the dynamic baseline value of its corresponding clinical stage when a new test value is generated. The rate of change calculation directly describes how quickly the value rises or falls. For example, assuming a subject in stable remission has a dynamic baseline CA125 value of 10 U / mL, a previous test value of 11 U / mL, and a new test value of 13 U / mL, with a 30-day interval between the two tests, the absolute change in this test value is 13-10=3 U / mL, and the rate of change can be calculated as (13-11) / 30 days, or 0.067. U / mL per day; this rate of change is itself an important feature, reflecting the short-term upward trend of the biomarker level; furthermore, in order to capture the changing trend, i.e. whether the trend is accelerating or decelerating, the system also calculates the acceleration of change. The calculation of the acceleration of change requires at least three consecutive detection data, which is obtained by calculating the difference in the rate of change between two consecutive time periods and then dividing by the time interval; a positive acceleration value means that the upward trend of the biomarker level is accelerating, which is an extremely dangerous signal. Even if the current value is still within the normal range, this accelerated upward trend strongly indicates the potential risk of relapse.

[0041] The third step involves incorporating the calculated rate of change and acceleration from the previous step into the dynamic baseline change characteristics. In practice, this dynamic baseline change characteristic can be a vector containing multiple elements, such as the difference between the current value and the dynamic baseline value, the ratio between the current value and the dynamic baseline value, the first derivative (rate of change), and the second derivative (acceleration). These calculated trend characteristics are then integrated into the time-enhanced feature vector of the above embodiment. In this way, the method of this application can greatly improve the sensitivity of early warning, and can identify this dangerous upward trend months or even longer in advance by analyzing the rate of change and acceleration relative to the individual baseline during the process of a subject's CA125 value slowly rising from 8 U / mL to 15 U / mL, thereby gaining a valuable time window for clinical intervention.

[0042] It is understood that this application embodiment adds a process for processing unstructured text information; firstly, through the data acquisition module, the subjective symptom report submitted by the subject to be evaluated in the form of unstructured natural language text is acquired; such reports can come from various sources, such as daily or weekly health diaries entered by the subject to be evaluated through a dedicated mobile application or portal website, chat logs of online communication with medical staff, or the subject's chief complaint text recorded by the doctor in the electronic medical record during the outpatient visit; these texts usually contain a lot of valuable information describing the subject's self-feelings, such as "I have felt a slight heaviness in my lower abdomen for the past week and I feel more tired than usual."

[0043] After acquiring the text, a pre-trained biomedical natural language processing model is invoked to perform deep semantic analysis on these unstructured natural language texts. This biomedical natural language processing model has been pre-trained on massive amounts of medical literature, clinical notes, and medical records, enabling it to recognize and understand medical terms and symptom descriptions. The model processes the input text to extract predefined symptom entities and quantified intensity descriptions of these symptom entities. Here, symptom entities refer to a list of specific symptoms related to ovarian cancer recurrence that the model can identify, such as abdominal distension, abdominal pain, fatigue, loss of appetite, and weight loss. Quantified intensity descriptions refer to words extracted from the text by the model to modify the severity of symptoms, such as "mild," "moderate," "severe," "persistent," and "intermittent," and mapped to a standardized numerical scale, such as floating-point numbers from 0 to 1.

[0044] After extracting the symptom entities and their quantitative intensity descriptions, the system encodes this information into standardized subjective symptom vectors. These vectors are multi-dimensional, with each dimension corresponding to a predefined symptom entity, and the value in that dimension representing the quantitative intensity of the symptom. For example, a subjective symptom vector might be in the form of [bloating intensity: 0.3, fatigue intensity: 0.6, abdominal pain intensity: 0.0, ... ], where 0.3 corresponds to "mild bloating" in the text, 0.6 corresponds to "more prone to fatigue," and 0.0 indicates that abdominal pain is not mentioned. This vector is generated for each time point, thus transforming the originally unstructured text into structured and quantitative data that the model can process.

[0045] Finally, before inputting the time-series augmented feature sequence into the time-series prediction model, a fusion step can be performed to align and fuse the generated subjective symptom vector with the previously constructed biomarker-based time-series augmented feature sequence in the time dimension. Alignment means ensuring that the subjective symptom vector of a certain date is associated with the biomarker feature vector of the same day or within a nearby time window. The fusion method can be simple vector concatenation, that is, appending the subjective symptom vector to the end of the time-series augmented feature vector to form a longer and more comprehensive feature vector. Through this process, the personal feelings of the subject being evaluated are successfully transformed into quantitative features that the model can understand. This allows the model to understand both objective detection data and subjective symptom reports when making risk predictions. This is of irreplaceable value for identifying occult relapses that present with nonspecific symptoms as the initial manifestation, thereby significantly improving the comprehensiveness and sensitivity of the prediction.

[0046] Understandably, to further enhance the completeness and reliability of data sources, embodiments of this application can also acquire structured health questionnaire data submitted by the subject to be evaluated while obtaining subjective symptom reports. These questionnaires can be in the form of scales, such as the European Organization for Research and Treatment of Cancer's Quality of Life Core Questionnaire QLQ-C30, which requires the subject to be evaluated to rate their fatigue, pain, and other conditions over the past week on a scale of 1 to 4. This type of structured data does not require complex natural language processing and can be directly and accurately mapped to the corresponding dimensions of the subjective symptom vector. When encoding the subjective symptom vector, the system will include this data from the structured health questionnaire. For example, if the subject to be evaluated scores fatigue at 3 out of 4 in the questionnaire, it can be standardized to 0.75 and used to update or supplement the fatigue intensity extracted from the text.

[0047] It is understood that the embodiments of this application specify the key concepts in the disease course segmentation step (that is, analyzing and dividing the disease course timeline of the subject to be evaluated into multiple preset clinical stages); among them, clinical events include, but are not limited to, the surgery date, the chemotherapy start date, or the chemotherapy end date; these are all milestone clinical events in the diagnosis and treatment process of the ovarian cancer subject to be evaluated and will be accurately recorded in the medical record; based on these specific date anchors, the step of analyzing and dividing the disease course timeline into multiple preset clinical stages can be implemented as follows: the system takes the surgery date or diagnosis date as the starting point of the entire timeline, and the end date of the last chemotherapy as the end point of the initial treatment stage and the starting point of the remission stage. Then, based on the end date of the last chemotherapy, the remission stage can be further subdivided into the early remission stage, the stable remission stage, and the long-term remission stage. Through such a clear and clinically consistent definition, the accuracy and consistency of the key step of disease course segmentation are ensured, making the disease course stages of different subjects to be evaluated comparable.

[0048] It is understood that, in the embodiments of this application, the time series prediction model can be a Long Short-Term Memory (LSTM) network model or a variant thereof; an LSM network is a special type of recurrent neural network designed to solve the gradient vanishing or gradient exploding problems encountered by traditional recurrent neural networks when processing long sequences; a standard LSM network unit contains a cell state and three gating structures, namely a forget gate, an input gate, and an output gate; the cell state spans the entire time series and is responsible for transmitting long-term memory; the forget gate is responsible for reviewing the output of the previous time step and the input of the current time step, deciding which old information to discard from the cell state; the input gate is responsible for deciding which new information to update the cell state; and the output gate, based on the current cell state, decides what information to output; this This gating mechanism enables Long Short-Term Memory (LSTM) networks to selectively remember and forget information, making them particularly adept at capturing long-term dependencies in time series. In the scenario of ovarian cancer recurrence monitoring, the remission period of the subjects being evaluated may last for several years, and the recurrence signal may be composed of a series of weak but persistent abnormal changes after a long period of stable data. The long-term memory capability of LTM networks is well-suited to capturing such signals. Its variants, such as gated recurrent units, simplify the model structure by merging the forget gate and input gate into an update gate and fusing cell states and hidden states. They can achieve performance comparable to LTM networks in many tasks while being more computationally efficient, thus serving as an effective alternative.

[0049] like Figure 2As shown, this application also discloses a dynamic monitoring and early warning management system for ovarian cancer recurrence, including a data acquisition module, a central processing module, and a risk generation module. These modules can be independent software services or hardware units, and they communicate and collaborate through predefined interfaces to jointly complete the task of dynamic monitoring and early warning.

[0050] Specifically, the data acquisition module is configured to execute the data acquisition steps in the aforementioned method. This module may contain multiple sub-components, such as a database adapter for connecting to the hospital database, capable of extracting data from relational databases in hospital or laboratory information systems using languages ​​like SQL; an API client for interacting with modern electronic medical record systems or third-party health platforms via RESTful APIs to obtain data in standard formats such as rapid medical interoperability resources; and a web frontend or mobile application backend interface for receiving data manually entered by the assessed individuals or doctors, such as subjective symptom texts and quality-of-life questionnaires. This data acquisition module is responsible for data collection, cleaning, format standardization, and preliminary validation. It then transmits the processed multimodal time-series data to the central processing module via a message queue or direct function calls.

[0051] The central processing module communicates with the data acquisition module, receiving raw data streams from the latter. This central processing module is configured to perform the computational tasks described above, specifically including: First, performing disease course segmentation, i.e., based on the received clinical event data, calling the built-in rule engine to segment the timeline of the subject's disease course into multiple preset clinical stages such as the initial treatment stage and the early remission stage; Second, performing time-series weight generation, i.e., for each data point, calculating its time-series weight factor using a preset mathematical model based on its clinical stage and the time difference between the data point and a major treatment event; Third, performing feature... The enhancement operation involves calculating time-series weighting factors based on the original data values ​​of the data points, and calling the dynamic baseline analysis subroutine to derive dynamic baseline change features. Finally, all this information is integrated into a high-dimensional time-series enhanced feature vector and organized into a feature sequence in chronological order. If the system implements optional optimization functions, this module is also responsible for tasks such as calculating data reliability scores, quantifying the degree of abnormality in follow-up rhythms, and calling biomedical natural language processing models to process subjective symptom reports and generate subjective symptom vectors. After completing all feature engineering, the central processing module outputs the final constructed time-series enhanced feature sequence to the risk generation module.

[0052] The risk generation module communicates with the central processing module and receives the time-enhanced feature sequences constructed by the latter. Internally, this module encapsulates one or more pre-trained time-series prediction models, such as long short-term memory networks. Its configuration allows it to take the received feature sequences as input and feed them into the model for forward propagation calculations. After a series of complex nonlinear transformations, the model ultimately outputs a dynamic risk index that quantifies the likelihood of ovarian cancer recurrence. The risk generation module may also include a post-processing unit for interpreting and visualizing the raw values ​​output by the model, such as generating a curve showing the risk over time or triggering an early warning event based on whether the risk index exceeds a preset threshold. The final risk index, chart, or early warning signal can be pushed to the dashboard of the clinical decision support system through the system's output interface, displayed to doctors, or alerted to the assessed individuals and health managers via mobile application notifications, thus closing the entire loop from data collection to clinical decision support.

[0053] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0055] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0056] 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 network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] 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.

[0058] 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.) or processor 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 USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0059] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for managing ovarian cancer recurrence through dynamic monitoring and early warning, characterized in that, The method is performed by a computer device and includes the following steps performed by the computer device: Obtain multimodal follow-up data of a subject to be evaluated at different time points, wherein the multimodal follow-up data includes at least biomarker detection values ​​and clinical events that define the treatment cycle; Based on the clinical events, the disease course timeline of the subject to be evaluated is analyzed and divided into multiple preset clinical stages; For any data point in the multimodal follow-up data, a time-series weighting factor is calculated and generated based on the clinical stage to which the data point belongs and the time span between the data point and a preset major treatment event. Based on the data values ​​of the data points, the time-series weighting factors, and the dynamic baseline change characteristics derived from historical data, a set of time-series enhancement features are jointly constructed and arranged in chronological order to form a time-series enhancement feature sequence. The time-enhanced feature sequence is input into a pre-trained time-series prediction model, which processes and outputs a dynamic risk index that quantifies the likelihood of ovarian cancer recurrence.

2. The method according to claim 1, characterized in that, Before constructing the temporal enhancement features, the following is also included: Further, obtain metadata associated with each data point in the multimodal follow-up data, wherein the metadata at least indicates the data detection source or data detection method; Based on a preset scoring rule base and the metadata, a data reliability score is set for each data point; The data reliability score is incorporated as an additional feature into the construction process of the time-series enhancement feature.

3. The method according to claim 2, characterized in that, Before constructing the temporal enhancement features, the following is also included: The time series of follow-up events of the subject under evaluation were analyzed and compared with a standard follow-up plan to quantify the degree of abnormality of the follow-up rhythm. The degree of abnormality of the follow-up rhythm is incorporated as another additional feature into the construction process of the time-series enhancement feature.

4. The method according to claim 2, characterized in that, The metadata also includes the type of visit, which is categorized as planned follow-up visits or unplanned temporary visits.

5. The method according to claim 1, characterized in that, The steps for constructing the dynamic baseline change features derived from historical data include: For a specific biomarker, an independent dynamic baseline value is calculated for each clinical stage based on its historical monitoring values ​​at different clinical stages. Calculate the rate of change and acceleration of change of the newly detected value of the specific biomarker relative to the dynamic baseline value of its clinical stage; The calculated rate of change and acceleration of change are used as components of the dynamic baseline change characteristics.

6. The method according to claim 1, characterized in that, Also includes: Obtain the subjective symptom reports submitted by the subjects to be evaluated in the form of unstructured natural language text; A pre-trained biomedical natural language processing model is invoked to perform semantic analysis on the unstructured natural language text in order to extract predefined symptom entities and quantitative intensity descriptions of the symptom entities. The extracted symptom entities and the quantified intensity description are encoded into a subjective symptom vector; Before inputting the time-series enhanced feature sequence into the time-series prediction model, the subjective symptom vector and the time-series enhanced feature sequence are aligned and fused in the time dimension.

7. The method according to claim 6, characterized in that, It also includes obtaining structured health questionnaire data submitted by the subject to be evaluated, and incorporating the structured health questionnaire data when encoding the subjective symptom vector.

8. The method according to claim 1, characterized in that, The clinical events include the surgery date, chemotherapy start date, or chemotherapy end date; the step of parsing and segmenting the disease course timeline of the subject to be evaluated into multiple preset clinical stages specifically involves segmenting the disease course timeline into at least two of the following based on the surgery date, chemotherapy start date, or chemotherapy end date: initial treatment stage, early remission stage, stable remission stage, and long-term remission stage.

9. The method according to claim 1, characterized in that, The time-series prediction model is a long short-term memory network model or a variant thereof.

10. A dynamic monitoring and early warning management system for ovarian cancer recurrence, characterized in that, This includes a plurality of functional modules configured to perform the method as described in any one of claims 1 to 9, said functional modules comprising at least: A data acquisition module is used to acquire multimodal follow-up data of a subject to be evaluated at different time points. The multimodal follow-up data includes at least biomarker detection values ​​and clinical events that define the treatment cycle. A central processing module, communicatively connected to the data acquisition module, is configured to perform the following operations: based on the clinical events, analyze and segment the disease timeline of the subject to be evaluated into multiple preset clinical stages; for any data point in the multimodal follow-up data, calculate and generate a time-series weighting factor based on the clinical stage to which the data point belongs and the time span between the data point and a preset major treatment event; and construct a set of time-series enhancement features based on the data value of the data point, the time-series weighting factor, and dynamic baseline change features derived from historical data. A risk generation module, which is communicatively connected to the central processing module, is configured to input the time-series enhanced feature sequence into a pre-trained time-series prediction model, and output a dynamic risk index that quantifies the likelihood of ovarian cancer recurrence after processing by the model.

Images