Gynecological treatment efficacy evaluation method and system based on multi-modal data fusion

Abstract

The present application belongs to the technical field of medical information processing and intelligent diagnosis and treatment evaluation, and discloses a gynecological treatment efficacy evaluation method based on multi-modal data fusion. In order to solve the problems that irregular multi-time point data in a treatment course is difficult to automatically time-align, menstrual cycle phase fluctuation interferes with efficacy judgment, and missing or low-quality modalities lead to unstable conclusions, the present application acquires multi-modal data with collection time markers and treatment event data, generates modal observation features and calculates quality scores, constructs a probability state space model containing menstrual cycle phase latent variables and phase drift parameters, adopts continuous time state transition and performs jump update on efficacy state at the position triggered by the treatment event, combines temperature calibration and quality weighted product expert inference to obtain a fused posterior distribution, outputs efficacy evaluation information and confidence, and performs phase correction, thereby realizing the technical effect of stably evaluating efficacy under the conditions of irregular sampling and missing modalities.

Description

Technical Field

[0001] This invention relates to the field of medical information processing, and in particular to a method and system for evaluating the efficacy of gynecological treatments based on multimodal data fusion. Background Technology

[0002] With the development of medical informatization and intelligent diagnosis and treatment technologies, the evaluation of gynecological disease treatment efficacy is gradually shifting from single-indicator judgment to comprehensive evaluation based on multi-source data. Existing medical information systems can continuously accumulate multimodal data such as imaging data, laboratory data, electronic medical record texts, and symptom scales during the diagnosis and treatment process. Research and applications are increasingly employing machine learning models to extract and fuse features from this multimodal information to output results such as efficacy scores, efficacy grading, or disease trend changes. Simultaneously, to characterize the changes in the condition over time during treatment, some solutions introduce time series modeling approaches, such as using recurrent neural networks, attention models, or discrete-time state-space models to predict trends and infer efficacy from multiple follow-up data.

[0003] However, existing technologies still have the following shortcomings in the context of gynecological treatment efficacy evaluation. First, most methods assume approximately fixed data collection intervals or rely on post-hoc alignment and simple time coding, making it difficult to adapt to the irregular sampling and multi-time-point data in real clinical settings. They also lack alignment mechanisms that use treatment events such as medication, surgery, and physical therapy as time anchors, leading to unclear correspondences between treatment events and state changes. Second, gynecological patients exhibit significant menstrual cycle-related physiological fluctuations. Existing methods often treat the cycle merely as a general time factor or empirical grouping, lacking explicit cycle phase modeling and individualized drift correction, which can easily allow physiological fluctuations to mask the true treatment effect. Third, multimodal data often suffers from modality loss or inconsistent data quality in clinical practice. Existing fusion methods often employ simple splicing, fixed weights, or missing data imputation, lacking quality perception and uncertainty calibration mechanisms, easily leading to unstable conclusions and uncontrollable confidence levels when information is insufficient. These shortcomings limit the stability, interpretability, and clinical usability of efficacy evaluation results.

[0004] Therefore, there is a need for a method and system for evaluating the efficacy of gynecological treatments that can address the shortcomings of existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a method for evaluating the efficacy of gynecological treatments based on multimodal data fusion. Addressing the problems in existing technologies, such as irregular data collection intervals at multiple time points within a treatment course leading to difficulties in automatic temporal alignment, menstrual cycle phase fluctuations easily interfering with efficacy judgment, and unstable evaluation conclusions lacking reliability indicators under conditions of missing or low-quality multimodal data, the following technical solution is proposed: acquiring multimodal data, including images, laboratory tests, electronic medical record texts, and symptom scales with collection time stamps, as well as treatment event data within the same treatment course; generating observation features for each modality and calculating quality scores; forming a multi-time-point observation sequence according to collection time and determining treatment event triggers. The invention employs a probabilistic state-space model that incorporates cyclical physiological states and therapeutic efficacy states. Cyclic physiological states include menstrual cycle phase latent variables and phase drift parameters. A continuous-time state transition function is used for state prediction based on time intervals. At the trigger point of a treatment event, the therapeutic efficacy state is updated using a jump operator based on the event type, dosage, or duration. The fusion weights for each modality's posterior distribution are determined based on quality scores, and a temperature coefficient is introduced for calibration. A weighted product expert inference is used to generate the fused posterior distribution. Based on the fused posterior distribution, therapeutic efficacy assessment information and its confidence level are output, and phase correction is performed based on the phase latent variables to eliminate the influence of physiological fluctuations. This invention achieves stable therapeutic efficacy assessment and provides reliable output in scenarios with irregular sampling, cyclical fluctuations, and modality loss.

[0006] This invention provides a method for evaluating the efficacy of gynecological treatment based on multimodal data fusion, including:

[0007] S1. Acquire multimodal data of the target patient at multiple acquisition times within the same treatment course, record the acquisition time, and obtain treatment event data corresponding to the treatment course, including event type and event occurrence time; S2. Generate modal observation features for each modality of the multimodal data, and determine the quality score for each modal observation feature at each acquisition time; S3. Sort the modal observation features according to the acquisition time to generate a multi-time point observation sequence, calculate the time interval between adjacent acquisition times, and map the treatment event data to the multi-time point observation sequence to determine the trigger position of the treatment event; S4. Construct a probabilistic state space model based on the multi-time point observation sequence, setting potential disease states including cyclical physiological states and efficacy states, where cyclical physiological states include menstrual cycle phase latent variables and phase drift parameters corresponding to the target patient, according to... Based on time intervals, the potential disease state is predicted using a continuous-time state transition function. When a treatment event trigger location exists, an event-triggered jump update is applied to the efficacy state according to the event type. Based on the modal observation characteristics at each acquisition time, the modal posterior distribution corresponding to each modality is obtained through probabilistic inference, and the menstrual cycle phase latent variable is estimated. S5: Based on the quality score, the fusion weight is determined for each modality, and the temperature coefficient is determined to calibrate the modal posterior distribution. The calibrated modal posterior distribution is used to generate a fusion posterior distribution through weighted product expert inference. S6: Efficacy assessment information is generated based on the fusion posterior distribution, and the cyclical physiological state is phase-corrected based on the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations on efficacy assessment. The confidence level of the efficacy assessment information is generated based on the uncertainty of the fusion posterior distribution.

[0008] Optionally, S1 includes:

[0009] Obtain the patient identifier of the target patient and the treatment identifier corresponding to the same treatment course;

[0010] Multimodal data is acquired or received at multiple acquisition times during the treatment course, and the corresponding acquisition time is read or recorded for each piece of multimodal data. The multimodal data is historical data generated in the medical process and stored by the medical information system or testing equipment. The multimodal data includes at least two types of data, such as image data, test data, electronic medical record text, and symptom scale, to form a multimodal dataset with acquisition time stamps.

[0011] Acquire or receive treatment event data corresponding to the treatment course, and record the event type, event occurrence time, and dose parameter or duration parameter corresponding to the event type for each treatment event, so as to associate the treatment event data with the treatment course identifier;

[0012] Output a multimodal dataset with acquisition time stamps and treatment event data;

[0013] Furthermore, the treatment event data further includes menstrual-related events, which include the start time or end time of menstruation;

[0014] In the probabilistic inference, the menstrual-related events are used as phase-anchored observations of the menstrual cycle phase latent variables to correct the estimation results of the menstrual cycle phase latent variables.

[0015] Optionally, S2 includes:

[0016] Preprocessing is performed separately for data of different modalities in the multimodal dataset, including denoising and normalization;

[0017] The preprocessed image data is converted into image features, the preprocessed test data is converted into test features, the preprocessed symptom scale is converted into scale features, and the preprocessed electronic medical record text is structured and parsed to generate text features.

[0018] The image features, test features, scale features, and text features are respectively used as modal observation features of the corresponding modalities and organized according to the acquisition time. A quality score is calculated for each modal observation feature at each acquisition time. The quality score is determined by an integrity score and a noise score. The integrity score is calculated based on the missing proportion or the proportion of valid fields of the modal observation feature, and the noise score is calculated based on at least one of the signal-to-noise ratio, artifact intensity, or outlier proportion.

[0019] The quality score is a weighted combination of the integrity score and the noise score, and normalized to a preset range; the output is the modal observation features organized according to the acquisition time and the quality score corresponding to each modal observation feature.

[0020] Optionally, S3 includes:

[0021] Based on the modal observation features organized by acquisition time and the quality score, the modal observation features corresponding to each acquisition time are sorted according to the acquisition time to generate a multi-time point observation sequence consisting of multiple acquisition times.

[0022] For two adjacent acquisition times in the multi-timepoint observation sequence, calculate the time interval between the two adjacent acquisition times;

[0023] Based on treatment event data, the occurrence time of each treatment event is matched with the acquisition time in the multi-time point observation sequence to determine the treatment event trigger position in the multi-time point observation sequence;

[0024] The output includes a multi-timepoint observation sequence containing acquisition time, time interval, modal observation features, quality score, and treatment event trigger location.

[0025] Optionally, S4 includes:

[0026] A probabilistic state-space model is established based on multi-time point observation sequences, and a potential disease state is defined for each acquisition time. The potential disease state consists of a periodic physiological state and a therapeutic effect state. The periodic physiological state includes menstrual cycle phase latent variables and phase drift parameters corresponding to the target patient.

[0027] For two adjacent acquisition times in a multi-time point observation sequence, a predicted distribution of potential disease status is generated based on the time interval using a continuous-time state transition function. The continuous-time state transition function includes: determining a state transition rate matrix based on the time interval, obtaining a state transition matrix by performing a matrix exponentiation operation on the state transition rate matrix, and scaling the process noise covariance according to the time interval to obtain the covariance parameter of the predicted distribution.

[0028] When the multi-time-point observation sequence indicates that there is a treatment event trigger position between two adjacent acquisition times, the jump operator is invoked to update the efficacy status according to the event type and dose or duration parameter corresponding to the treatment event trigger position, so as to incorporate the efficacy change caused by the treatment event into the prediction distribution and complete the automatic time-series alignment of multiple time points. The jump operator includes applying an additive impact term or a multiplicative gain term to the efficacy status, and the impact intensity is calculated by the event representation vector corresponding to the event type and the dose or duration parameter through a mapping function.

[0029] For each modal observation feature at each acquisition time, the modal posterior distribution of the modality is calculated based on the observation model matching the corresponding modality, and the menstrual cycle phase latent variable is estimated based on the modal posterior distribution, wherein the probability inference is implemented using any one of variational inference, particle filtering, extended Kalman filtering, and unscented Kalman filtering;

[0030] Output the modal posterior distribution and the estimation results of the menstrual cycle phase latent variables;

[0031] Furthermore, the therapeutic state includes a baseline component and a therapeutic effect component. The therapeutic effect component is updated by decaying with the time interval according to the decay coefficient during the time interval when there is no therapeutic event triggering position, and when there is a therapeutic event triggering position, the jump operator applies the additive impact term or multiplicative gain term to the therapeutic effect component.

[0032] Optionally, S5 includes:

[0033] For each acquisition time, a fusion weight is determined for each modality based on the quality score, and the fusion weight is associated with the modality posterior distribution of the corresponding modality. When the observation features of a certain modality corresponding to the acquisition time are missing, the fusion weight corresponding to that modality is set to 0.

[0034] For multiple modalities acquired at the same time, the fusion weights are calculated by normalizing the quality scores of each modality, such that the sum of the fusion weights of the non-missing modalities is 1, and the fusion weight corresponding to the modality is reduced when the quality score is lower than a preset threshold. For each acquisition time, a temperature coefficient is determined for each modality based on the quality score, and the modality posterior distribution corresponding to the modality is calibrated using the temperature coefficient to suppress overconfidence in the modality posterior distribution. The calibration includes at least one of the following: normalizing the log probability density of the modality posterior distribution by multiplying it by a scaling factor corresponding to the temperature coefficient, and amplifying the variance parameter of the modality posterior distribution according to the temperature coefficient. For each acquisition time, weighted product expert inference is performed on each calibrated modality posterior distribution according to its corresponding fusion weight. By weighted fusion and normalization of each calibrated modality posterior distribution, a fused posterior distribution at that acquisition time is generated.

[0035] Optionally, S6 includes:

[0036] For each collection time, the estimated value of the therapeutic effect status is calculated based on the fusion posterior distribution, and a therapeutic effect status sequence for the corresponding collection time is generated.

[0037] For the therapeutic state sequence, the cyclical physiological state is phase-corrected based on the estimation results of the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations on the therapeutic state sequence, and a phase-corrected therapeutic state sequence is generated. The phase correction includes at least one of the following: dividing the menstrual cycle phase latent variable into multiple phase intervals, mapping the therapeutic state corresponding to different collection times to the same phase interval and then performing comparative calculations, fitting a cyclical physiological component function based on the menstrual cycle phase latent variable and subtracting the cyclical physiological component from the therapeutic state.

[0038] The efficacy assessment information is calculated based on the phase-corrected efficacy status sequence, wherein the efficacy assessment information includes at least one of efficacy score or efficacy grading information; the confidence level of the efficacy assessment information is calculated based on the uncertainty of the fusion posterior distribution, wherein the uncertainty includes at least one of the following: the entropy of the fusion posterior distribution, the variance of the fusion posterior distribution, the width of the posterior confidence interval of the efficacy status, and the posterior probability that the efficacy improvement exceeds a preset threshold; the efficacy assessment information and the confidence level are output for display or storage on the terminal for medical personnel to refer to.

[0039] On the other hand, the present invention also provides a gynecological treatment efficacy evaluation system based on multimodal data fusion, comprising:

[0040] The data acquisition module is used to acquire multimodal data and acquisition time corresponding to multiple acquisition times within the same treatment course, and to acquire treatment event data including event type and event occurrence time;

[0041] The Features and Quality module is used to generate modal observation features for each modality and determine the quality score of each modal observation feature at each acquisition time.

[0042] The sequence and event module is used to generate a multi-time point observation sequence according to the acquisition time, calculate the time interval between adjacent acquisition times, and map the treatment event data to the observation sequence to determine the trigger position of the treatment event;

[0043] The state inference module is used to construct a probabilistic state space model, setting potential disease states that include cyclical physiological states and therapeutic states. The cyclical physiological states include menstrual cycle phase latent variables and phase drift parameters. The state is predicted based on the time interval using a continuous-time state transition function. At the location of the treatment event trigger, the therapeutic state is updated by an event-triggered jump. Based on the modal observation features, probabilistic inference is performed to obtain the posterior distribution of each modality and to estimate the menstrual cycle phase latent variables.

[0044] The fusion and output module is used to determine the fusion weight and temperature coefficient based on the quality score, generate the fusion posterior distribution by weighted product expert inference after calibrating the modal posterior distribution, output the efficacy assessment information and its confidence level based on the fusion posterior distribution, and perform phase correction based on the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations.

[0045] The beneficial effects of this invention are:

[0046] 1. It can adapt to irregularly sampled multi-time point data within the treatment course and achieve automatic temporal alignment. By using a continuous-time state transition function to predict the state based on the time interval between adjacent acquisitions, and applying jump updates to the efficacy state at the treatment event trigger position according to the event type and dose or duration parameters, it establishes a correspondence between efficacy changes and treatment events on a unified time axis, thereby improving the temporal consistency and traceability of efficacy assessment.

[0047] 2. It can reduce the interference of menstrual cycle-related physiological fluctuations on the assessment of treatment efficacy. By introducing menstrual cycle phase latent variables and individualized phase drift parameters into the probabilistic state-space model, and performing phase correction or subtracting cycle physiological components based on the phase latent variables in the output stage, it achieves in-phase comparison and elimination of physiological components, thereby improving the stability and accuracy of treatment efficacy assessment results.

[0048] 3. It can still stably output efficacy conclusions and provide confidence indicators even when modalities are missing or data quality is inconsistent. By calculating quality scores for the observed features of each modality, determining the fusion weights, and setting missing modalities to zero, a temperature coefficient is introduced to calibrate the modal posterior distribution. Then, weighted product expert inference is used to generate the fusion posterior distribution, and the confidence level is output based on the uncertainty of the fusion posterior distribution, thereby suppressing overconfidence and enhancing clinical usability. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 The flowchart shows the gynecological treatment efficacy evaluation method and system based on multimodal data fusion proposed in this invention.

[0051] Figure 2 This is a flowchart of the probabilistic state space modeling and state inference process in step S4 of the present invention.

[0052] Figure 3 This is a flowchart of the multimodal posterior distribution fusion process in step S5 of the present invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 A method for evaluating the efficacy of gynecological treatments based on multimodal data fusion includes:

[0055] S1. Acquire multimodal data of the target patient at multiple acquisition times within the same treatment course, record the acquisition time, and obtain treatment event data corresponding to the treatment course, including event type and event occurrence time; S2. Generate modal observation features for each modality of the multimodal data, and determine the quality score for each modal observation feature at each acquisition time; S3. Sort the modal observation features according to the acquisition time to generate a multi-time point observation sequence, calculate the time interval between adjacent acquisition times, and map the treatment event data to the multi-time point observation sequence to determine the trigger position of the treatment event; S4. Construct a probabilistic state space model based on the multi-time point observation sequence, setting potential disease states including cyclical physiological states and efficacy states, where cyclical physiological states include menstrual cycle phase latent variables and phase drift parameters corresponding to the target patient, according to... Based on time intervals, the potential disease state is predicted using a continuous-time state transition function. When a treatment event trigger location exists, an event-triggered jump update is applied to the efficacy state according to the event type. Based on the modal observation characteristics at each acquisition time, the modal posterior distribution corresponding to each modality is obtained through probabilistic inference, and the menstrual cycle phase latent variable is estimated. S5: Based on the quality score, the fusion weight is determined for each modality, and the temperature coefficient is determined to calibrate the modal posterior distribution. The calibrated modal posterior distribution is used to generate a fusion posterior distribution through weighted product expert inference. S6: Efficacy assessment information is generated based on the fusion posterior distribution, and the cyclical physiological state is phase-corrected based on the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations on efficacy assessment. The confidence level of the efficacy assessment information is generated based on the uncertainty of the fusion posterior distribution.

[0056] In this specific embodiment, S1 includes:

[0057] The system establishes a unique patient identifier for the target patient within the hospital information system. And read the treatment course identifier corresponding to the same treatment course. Treatment course label In the treatment management table and the treatment start time and the end time of the treatment course The system is associated with... As a retrieval time window, historical data generated and stored within the treatment course were retrieved from the PACS image database, LIS laboratory database, electronic medical record system, and symptom scale system to form a multimodal dataset with acquisition time stamps. Image data was limited to gynecological ultrasound DICOM sequences, with the acquisition time read from the acquisition time field of the DICOM header and converted to a UTC epoch-second timestamp. Laboratory data was limited to structured test results of complete blood count and six hormone tests within the treatment course, with the acquisition time taken as the test report generation time and converted to a UTC epoch-second timestamp. Electronic medical record text was limited to the text content of outpatient progress notes and discharge summaries, with the acquisition time taken as the document signature time and converted to a UTC epoch-second timestamp. Symptom scales were limited to the visual analog scale for pain and the menstrual symptom scale in electronic questionnaire format, with the acquisition time taken as the questionnaire submission time and converted to a UTC epoch-second timestamp. The system added a patient identifier to each piece of multimodal data. Treatment label The original data payload is retained for subsequent feature generation, and when a certain modality does not have data at a certain acquisition time, it is not filled but the record of that modality is not generated directly to maintain the authenticity of the missing state.

[0058] The system simultaneously retrieves treatment event data corresponding to the treatment course from the medical orders and treatment record tables, and records the event type, event occurrence time, and event parameters for each treatment event. The event type is limited to oral medication, intravenous administration, surgical procedures, physical therapy procedures, menstruation start, and menstruation end, and the event occurrence time is uniformly converted to a UTC epoch second timestamp. Event parameters are characterized by both dosage and duration parameters. For oral and intravenous medications, the dosage parameter is the prescribed dosage, and the duration parameter is 0. For surgical procedures, the dosage parameter is 0, and the duration parameter is the duration from the start to the end of the surgery. For physical therapy procedures, the dosage parameter is 0, and the duration parameter is the duration from the start to the end of the physical therapy. For menstruation start and menstruation end, both the dosage and duration parameters are set to 0, and additional phase anchoring observations are written for subsequent anchoring correction of the menstrual cycle phase latent variable. Specifically, the phase anchoring observation for menstruation start is 0, and the phase anchoring observation for menstruation end is... and will Fixed setting The phase span corresponding to the menstrual period being 5 days in a standard 28-day cycle;

[0059] The above output is written to the database using a unified data structure and submitted to subsequent steps, denoted as:

[0060] ;

[0061] in Represents a multimodal dataset with time stamps for data collection and Indicates the number of its records. Indicates the first The timestamp of the collection time of the multimodal data. Indicates the first The modality labels for each multimodal data point are limited to one of the following: ultrasound images, test results, electronic medical record text, or symptom scales. Indicates and The corresponding raw data payload, Represents treatment event data and Indicates the number of its events. Indicates the first The timestamp of the occurrence time of each treatment event. Indicates the first Each treatment event is labeled with an event type, and the values ​​are limited to one of the following: oral medication, intravenous administration, surgical procedure, physical therapy procedure, start of menstruation, and end of menstruation. Indicates the relationship with the first The dosage parameters for each treatment event are consistent with the units of the prescribed dosage. Indicates the relationship with the first The duration parameter for each treatment event is in seconds. Indicates the first Phase anchoring observations of each event and when Take 0 or 0 for the start or end of menstruation, respectively. And when For other treatment events, take -1 to indicate that no phase anchoring observation is provided.

[0062] In this specific embodiment, S2 includes:

[0063] The system uses a multimodal dataset with acquisition time stamps. As input and processed one by one, Indicates the first The data collection time and timestamp Indicates the first Modal labels for each data item Indicates and For the corresponding raw data payload, the system first performs mode-matching preprocessing on each data point to complete denoising and normalization, and generate modal observation features. ;

[0064] when When generating ultrasound images, the system from The grayscale frame sequence is decoded and the pixel values ​​are restored to a uniform grayscale scale according to the recalibration parameters in the DICOM header. Then, each frame is processed. Median filtering removes salt-and-pepper noise and spatially crops the area to a centered size before scaling. To eliminate resolution differences, pixel-based intensity normalization is performed on each frame's pixels. The average of the normalized multi-frame images is then calculated over time to obtain a single-frame grayscale image. Finally, statistical and textural features of a fixed set are calculated on this single-frame grayscale image and sequentially concatenated to form an image feature vector. The statistical feature set consists of pixel mean, pixel standard deviation, skewness, kurtosis, energy, and entropy, while the texture feature set consists of the gray-level co-occurrence matrix at an angle. , The contrast, correlation, homogeneity, and energy are calculated at a pixel distance of 1 and averaged over the angle to form a composition. It is a vector of fixed length, and each dimension of it has a definite physical meaning;

[0065] when When verifying the results, the system will The parsing process generates a structured item table, and the set of test items is limited to 10 items: 4 blood routine tests and 6 hormone tests. These items are arranged in a fixed order to form the test feature vector. For each test item, the raw values ​​are linearly normalized according to the lower and upper reference limits recorded in the LIS and then clipped to the nearest integer. To eliminate dimensional differences and address missing items in Enter 0 in the blank, but simultaneously count it as missing in the quality score calculation to maintain traceability of the missing status;

[0066] when When converting to electronic medical record text, the system... Denoising is achieved by unifying full-width and half-width characters, compressing consecutive whitespace, and removing non-Chinese and non-numeric characters. Then, character tuples are extracted using a fixed window, and a hash mapping modulo 256 is performed on each tuple. The count of each hash bucket is accumulated and processed. Normalization yields a text feature vector of length 256. The hash mapping rule and modulo radix remain constant to ensure that the same text is processed consistently in different operating environments. ;

[0067] when When using a symptom scale, the system starts from... Read the scores from the Visual Analogue Scale for Pain and the total score from the Menstrual Symptom Scale, and normalize them to their maximum scores. The resulting feature vector is of length 2. ;

[0068] In generation At the same time, the system calculates the relationship between each data point and the given information. One-to-one corresponding quality score And decompose it into integrity score With noise score Two parts, both normalized to , among which when Integrity score for ultrasound images The value is determined by the number of valid frames and the required DICOM field integrity rate. If the number of valid frames is less than 30, the value is taken in proportion to the number of valid frames and 30. The cropping and missing fields directly cause Set to 0 for noise score. The signal-to-noise ratio (SNR) of the central region is mapped and set to 0 when the SNR is less than 5 and 1 when the SNR is greater than 20, and determined by linear interpolation within the interval.

[0069] when Integrity score when verifying results The noise score is determined by the effective reporting rate of 10 test items. Determined by the proportion of valid projects falling within the reference range;

[0070] when Integrity scoring for electronic medical record text The ratio of the number of valid characters after cleaning to 500 is determined and in Cutting and noise scoring The percentage of non-Chinese and non-numeric characters in the text before cleaning is determined, and when this percentage is greater than 0.1, then... Take 0 and perform linear interpolation within the interval;

[0071] when Completeness score for symptom scale The score is determined by the proportion of valid entries for both scores and the noise score. The score is determined by whether it falls within the defined range of the scale, and is set to 0 if it exceeds the range.

[0072] The system obtains a quality score using a weighted combination method and then performs normalization. The calculation formula is as follows:

[0073] ;

[0074] in Indicates the first Bar modal observation features Quality rating Indicates the first Completeness score of each data item Indicates the first Noise score of each data point The weights for the integrity score are fixed. Indicates the noise score weight and is fixed. This means cropping the value within the parentheses to... The normalization operator is used to ensure the stability of the quality score value;

[0075] The system will process each piece of data Write to the feature table and sort by acquisition time An index is created so that step S3 can sort and serialize the data collected at different times.

[0076] In this specific embodiment, S3 includes:

[0077] The system reads the identifiers belonging to the same treatment course from the feature table. All records And read treatment event data:

[0078] ;

[0079] in Indicates the first The timestamps of the collection timestamps for each multimodal data point are in seconds. Indicates the first Modality labels for multimodal data. Indicates and Corresponding modal observation features, Indicates and One-to-one corresponding quality score Indicates the first The timestamp of each treatment event, in seconds. Indicates the first The event type of each treatment event, Indicates the first Dosage parameters for each treatment event, Indicates the first The duration parameter for each treatment event, in seconds. Indicates the first Phase anchoring observations for each treatment event;

[0080] The system first sorts all records by collection time. Sort in ascending order and extract the duplicate collection time set ,in Indicates the first Each collection time and meets the requirements , Indicates the number of collection times within this treatment course;

[0081] The system at each acquisition time Construct observation containers organized by modality With mass container ,in This is used to store the modal observation features of each modality, with the modal label as the key. Used for storage and The system will satisfy the one-to-one correspondence quality score of the same key. All records by modal label Merge Write and When the same collection time Same modality tag When multiple records exist, the system retains only one record and writes it to the container according to a deterministic deduplication rule, which compares the quality scores sequentially. The largest value is retained. When there are the same maximum quality score, the original data payload byte length is compared and the largest value is retained. If they are still the same, the lexicographical order of the record identifier is compared and the smallest value is retained. This ensures that the same modality at the same acquisition time corresponds to only one modal observation feature and one quality score.

[0082] The system then calculates the time interval between adjacent acquisition times and generates a time interval sequence. ,in Fixed setting to 0 and for According to the formula Calculate, where Indicates the first The collection time and the first The time interval between each data collection point is in seconds. Indicates the first Each collection time timestamp Indicates the first Each collection time timestamp;

[0083] The system will categorize treatment events by the time they occur. The system sorts the data in ascending order and maps treatment events to a set of acquisition times to determine the trigger location of each treatment event. Each acquisition time is indexed... Initialize event list Empty and process treatment events one by one. ,when Write the treatment event to the trigger location. Corresponding event list When there is satisfy Write the treatment event to the trigger location. Corresponding event list and in each Internal event holding Arrange in ascending order to ensure that the new execution order is determined when subsequent events trigger jump changes, and when The treatment event will be added to the list of unmapped events. It is marked as exceeding the observation sequence range to avoid ambiguity in subsequent trigger updates based on adjacent acquisition time intervals;

[0084] The system's final output includes a multi-timepoint observation sequence containing acquisition time, time interval, modal observation features, quality score, and treatment event trigger location. And It is stored as an ancillary output for adjustment of the treatment scope or subsequent expansion processing.

[0085] In this specific embodiment, S4 includes:

[0086] The system reads observation sequences at multiple time points. And at each collection time A probabilistic state-space model is established to recursively estimate potential disease states, where potential disease states are represented by state vectors. This represents constant components, and the meaning of each component is fixed. Menstrual cycle phase latent variables Phase drift parameters corresponding to the target patient Baseline components of therapeutic status Therapeutic effect component relative to therapeutic status And define the periodic physiological state as And define the therapeutic state as The overall efficacy level is defined as For use by observation models;

[0087] The system in Time Assign a Gaussian prior distribution and fix the prior mean as And the prior covariance is fixed as a diagonal matrix:

[0088] ;

[0089] in Represents the initial state mean vector and its component order is the same as... Consistent, This represents the initial state covariance matrix, with the first dimension having a variance of 0, used to constrain the constant components. It is always 1;

[0090] against to Adjacent acquisition time interval The system uses a continuous-time state transition function. Make predictions and construct a fixed state transition rate matrix as follows:

[0091] ;

[0092] in The phase propulsion angular velocity is represented by a standard 28-day cycle and is kept constant. , The continuous-time decay coefficient of the therapeutic effect component is represented and fixed. To correspond to a 7-day half-life, thereby quantifying the therapeutic effect. During the time interval at the location where no treatment event is triggered Exponential decay update and phase drift parameters occur Maintain a constant value over time;

[0093] The system is based on time intervals The state transition matrix is ​​obtained through matrix exponentiation and the formula is:

[0094] ;

[0095] in Indicates the first State transition matrix for each acquisition time, Represents matrix exponentiation. Represents the state transition rate matrix. Indicates the first The collection time and the first The time interval between each data collection time is in seconds;

[0096] The system uses the fused posterior distribution parameters of the previous time step as the prediction input and employs... The predicted distribution is obtained by linear propagation of the state mean and covariance, while the process noise baseline covariance is fixed at a certain value. and according to Scaling is performed to obtain the covariance increment of the prediction distribution, thereby ensuring that the prediction uncertainty is consistent with the time scale under different sampling intervals;

[0097] When multi-time-point observation sequences indicate that There is a location where treatment events are triggered, i.e., an event list. When not empty, the system classifies events according to their occurrence time. In the process, each treatment event is executed with an event-triggered jump update, and the jump operator is fixed to the treatment effect component. Apply an additive shock term and apply it to the baseline component. No jump update is performed, where for each treatment event Only when the event type The therapeutic effect jump was triggered by oral medication, intravenous administration, surgical procedures, or physical therapy procedures, and the event intensity was normalized to [missing information]. And stipulate when For oral or intravenous administration And when For surgical or physical therapy procedures ,in This represents a clipping operator used to limit the intensity of events. The system further classifies event types Mapped to a fixed impact coefficient And stipulate This updates the mean of the therapeutic effect component to... The increase in variance of the treatment effect component was fixed at 1. To reflect individual differences in treatment response;

[0098] when The event type includes the start or end of menstruation. When a period begins or ends, the system treats this event as a latent phase variable of the menstrual cycle. The phase anchoring observations are performed and phase anchoring updates are executed, where the anchoring observation values ​​are taken from the event carried by the event. And the observation noise variance is fixed at 1. The phase observation function is defined as a... conduct Surround normalization to ensure phase falls within Thus correct Estimate and suppress phase drift accumulation error;

[0099] After obtaining the state prediction distribution including event transitions and phase anchoring, the system performs a specific task for each acquisition time. For each modality of observation features present, an observation model matching that modality is constructed, and an unscented Kalman filter is used to perform probability inference to obtain the modality posterior distribution corresponding to each modality. The unscented transform parameter of the unscented Kalman filter is fixed at [value missing]. The observation model will use the state vector Mapped to a three-dimensional interpretation vector And through offline training and solidified storage of linear mapping parameters. Generate the observation mean for the corresponding mode, assuming the observation noise is a zero-mean Gaussian distribution and fixing its covariance as a diagonal matrix. The feature dimension of the ultrasound image modality is set to 10 and fixed. All diagonal elements are The feature dimension of the test result modality is set to 10 and fixed. All diagonal elements are The feature dimension of the electronic medical record text modality is 256 and fixed. All diagonal elements are The feature dimension 2 of the symptom scale modality is taken and fixed. All diagonal elements are The system for each mode Each observation update is performed independently using the same event-corrected predicted distribution as a priori, and the updated Gaussian distribution parameters are then updated. Record this mode at the acquisition time The modal posterior distribution at the location, and simultaneously the event-corrected prior or any modal posterior distribution. The mean and variance of the components serve as latent phase variables in the menstrual cycle. The estimated results are output along with the data and used in step S5 to perform multimodal fusion with mass weighting and temperature calibration.

[0100] In this specific embodiment, S5 includes:

[0101] The system uses each collection time Modal posterior distributions corresponding to each mode To fuse the input and read the mass container at the same acquisition time in step S3 To obtain a quality rating ,in Represents the collection time index and satisfies This represents a modal label, and its values ​​are limited to one of the following: ultrasound images, test results, electronic medical record text, or symptom scales. Representing modes During the collection time State vector The posterior mean vector with a fixed dimension of 5 and is coupled with The order of the components is consistent. Indicates and The paired posterior covariance matrix is Symmetric positive definite matrix This represents a quality score that corresponds one-to-one with each modal observation feature and has a value range of [value range missing]. The system first determines the collection time. Available mode set at the location And Defined as in The system contains modal observation features and generates a complete set of modal labels corresponding to the modal posterior distribution in step S4. Therefore, when a modal observation feature is missing, the modality will not be included. And its fusion weight is considered to be 0;

[0102] The system then processed each Constructing unnormalized weights And a deterministic value is assigned based solely on the quality score, where when season To eliminate low-quality modes, when season To reduce the contribution of this mode, when season The system is based on The normalization factor is used to calculate the fusion weight. And ensure that the sum of the fusion weights of all non-removed modes is 1 and the fusion weights of the removed modes are 0, when The system will adjust the predicted distribution parameters after event correction. This is directly used as the fusion posterior distribution at that acquisition time to ensure stable results can still be output even under low quality conditions across all modes;

[0103] The system targets each Further determine the temperature coefficient based on the quality score. And the mapping is fixed as follows:

[0104] ;

[0105] in Representing modes During the collection time The temperature coefficient at the location and the range of values ​​is To suppress overconfidence in the modal posterior distribution, the system employs variance amplification for calibration, and defines the calibrated covariance matrix as follows: And keep the mean vector unchanged. Thus in At lower levels, through larger Reduce the contribution of this mode to the accuracy of the fusion results;

[0106] After obtaining the fusion weights and calibrated modal posterior distributions, the system performs weighted product expert inference on multiple modalities acquired at the same time and expresses the fusion posterior distribution under the Gaussian distribution assumption as follows: Its parameters are determined according to the following formula: ;

[0107] in Indicates the collection time The covariance matrix of the fused posterior distribution. Indicates the collection time The mean vector of the fused posterior distribution. Represents the calibrated covariance matrix The inverse matrix is ​​the precision matrix. This represents the fusion weights obtained by normalizing the quality scores. Indicates the collection time The available modal set will be used by the system. Write it into the fusion table and use it as the sole input for generating efficacy assessment information and confidence levels in step S6.

[0108] In this specific embodiment, S6 includes:

[0109] The system reads each acquisition time from the fusion table. Corresponding fusion posterior distribution parameters And generate the therapeutic state sequence and the phase-corrected therapeutic state sequence, wherein Indicates the collection time State vector The fused posterior mean vector, with its 4th and 5th dimensions corresponding to the baseline components of the therapeutic status, respectively. Therapeutic effect component Indicates and The fusion posterior covariance matrix of the pair and its elements They represent and covariance and covariance and The system defines the point estimate of the therapeutic status as the overall therapeutic level based on the covariance. The posterior mean was taken as the estimated value of the therapeutic state to form a therapeutic state sequence, and the data were read simultaneously. The second dimension serves as a latent phase variable in the menstrual cycle. The estimated value is obtained and it is ensured that the estimated value always falls within the phase wrap-around normalization action in step S4. ;

[0110] To eliminate the influence of periodic physiological fluctuations on the assessment of treatment efficacy, the system consistently employs a periodic physiological component subtraction method based on phase latent variables to perform phase correction and constructs the periodic physiological component function into a single harmonic form:

[0111] ;

[0112] in To solidify the constant parameters stored in the system configuration and fix their values ​​respectively , and Representing the cosine and sine functions respectively, with the input angle unit in radians, thus utilizing at each acquisition time... The corresponding periodic physiological components are calculated and subtracted from the estimated therapeutic state to obtain the phase-corrected therapeutic state sequence;

[0113] The system then generates efficacy assessment information based on the phase-corrected efficacy status sequence and outputs it in the form of efficacy scores and efficacy grades, where the efficacy scores take a fixed range. And based on the first collection time within the treatment course The phase-corrected therapeutic status is used as a baseline for linear mapping. The therapeutic grade is determined by the therapeutic score threshold and is fixedly divided into three levels: significant effect, effective and ineffective, with thresholds of 80 and 60, respectively.

[0114] The system further generates confidence levels for efficacy assessment information based on the uncertainty of the fused posterior distribution and fixes the overall efficacy level posterior variance before phase correction as the uncertainty measure to ensure repeatability of the calculation and consistency with the fused posterior distribution. The above calculations are determined in one step by the following formula:

[0115] , ;

[0116] In the formula Indicates the collection time Estimate the posterior mean at the overall efficacy level express Chinese correspondence The component values, express Chinese correspondence The component values, express Chinese correspondence The component values, Indicates the collection time Estimation of periodic physiological components at the location, Indicates the collection time The estimated therapeutic status after phase correction. Indicates the overall therapeutic level The posterior variance, They represent middle and and and The corresponding covariance element, Indicates the collection time Treatment efficacy score Indicates baseline acquisition time The estimated value of the phase-corrected therapeutic status at that location. This means cropping the value within the parentheses to... The clipping operator, Indicates the collection time The confidence level of the treatment efficacy score and the range of values ​​are as follows: And it decreases as uncertainty increases;

[0117] The system outputs records for each acquisition time. And classify the efficacy of these treatments According to the threshold rule Determined and when When it is effective, When it is effective, when If the time data is invalid, the system will sort all output records by collection time and write them into the results table, and simultaneously display the efficacy score curve, efficacy grade and confidence level on the terminal interface for medical staff to refer to.

[0118] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0119] This invention addresses the complex technical challenges of irregular data acquisition intervals at multiple time points during treatment, interference from menstrual cycle phase fluctuations, and the need for stable output despite modal gaps. It maps multimodal observation sequences and treatment event data onto a probabilistic state-space model for joint inference. Firstly, it uses a continuous-time state transition function based on adjacent acquisition time intervals to predict potential disease states, ensuring a consistent time scale for state evolution under irregular sampling conditions. Secondly, it performs event-triggered jump updates on the therapeutic state at treatment event trigger locations, establishing a clear correlation between therapeutic changes and event type, dosage, or duration. Simultaneously, it characterizes individualized cyclical changes using menstrual cycle phase latent variables and phase drift parameters, and performs phase correction or subtraction of cyclical physiological components during the output stage, thereby reducing the interference of physiological fluctuations on therapeutic conclusions. Finally, it combines quality-score-driven fusion weights and temperature coefficient calibration to perform quality-perceived fusion of the posterior distributions of each modality and outputs the confidence level corresponding to the uncertainty, thus achieving stable output and credibility indication of therapeutic evaluation results.

[0120] Compared to solutions that use time as discrete step-length encoding or alignment in post-processing, this invention makes targeted structural improvements to address the aforementioned technical problems: First, it changes state transitions from discrete updates to continuous time updates and introduces jump operators and impact intensity mapping mechanisms bound to treatment events, making treatment events endogenous anchors for alignment and efficacy change modeling, thus improving adaptability to irregular sampling and complex treatment courses; Second, it elevates the menstrual cycle from a general time factor to a cyclic phase latent variable and incorporates individualized phase drift, while using menstrual-related events as phase anchoring observations to improve the reliability of phase estimation and phase correction; Third, it integrates modal missingness and low modal quality into a quality-perceived weighted product expert inference framework and suppresses posterior overconfidence through temperature calibration, enabling stable efficacy conclusions and interpretable confidence levels even with incomplete information, thereby better achieving the expected technical effects.

Claims

1. A method for evaluating the efficacy of gynecological treatment based on multimodal data fusion, characterized in that, include: S1. Acquire multimodal data of the target patient at multiple collection times within the same treatment course, record the collection time, and acquire treatment event data corresponding to the treatment course, including event type and event occurrence time; S2. Generate modal observation features for each modality of the multimodal data, and determine the quality score for each modal observation feature at each acquisition time. S3. Sort the modal observation features according to the acquisition time to generate a multi-time-point observation sequence, calculate the time interval between adjacent acquisition times, and map the treatment event data to the multi-time-point observation sequence to determine the trigger position of the treatment event. S4. Construct a probabilistic state-space model based on the multi-time-point observation sequence, setting potential disease states including cyclical physiological states and therapeutic efficacy states. The cyclical physiological states include menstrual cycle phase latent variables and phase drift parameters corresponding to the target patient. Based on the time interval, use a continuous-time state transition function to predict the potential disease states. When a treatment event trigger location exists, an event-triggered jump update is applied to the efficacy status based on the event type. The modal posterior distribution corresponding to each modality is obtained through probabilistic inference based on the modal observation characteristics at each acquisition time, and the menstrual cycle phase latent variable is estimated. S5: The fusion weight is determined for each modality based on the quality score, and the temperature coefficient is determined to calibrate the modal posterior distribution. The fused posterior distribution is generated by weighted product expert inference on the calibrated modal posterior distribution. S6: Efficacy assessment information is generated based on the fused posterior distribution, and the phase correction of the cyclical physiological state is performed based on the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations on efficacy assessment. The confidence level of the efficacy assessment information is generated based on the uncertainty of the fused posterior distribution.

2. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S1 includes: Obtain the patient identifier of the target patient and the treatment identifier corresponding to the same treatment course; Multimodal data is acquired or received at multiple acquisition times during the treatment course, and the corresponding acquisition time is read or recorded for each piece of multimodal data. The multimodal data is historical data generated in the medical process and stored by the medical information system or testing equipment. The multimodal data includes at least two types of data, such as image data, test data, electronic medical record text, and symptom scale, to form a multimodal dataset with acquisition time stamps. Acquire or receive treatment event data corresponding to the treatment course, and record the event type, event occurrence time, and dose parameter or duration parameter corresponding to the event type for each treatment event, so as to associate the treatment event data with the treatment course identifier; Output a multimodal dataset with acquisition time stamps and treatment event data.

3. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S2 include: Preprocessing is performed separately for data of different modalities in the multimodal dataset, including denoising and normalization; The preprocessed image data is converted into image features, the preprocessed test data is converted into test features, the preprocessed symptom scale is converted into scale features, and the preprocessed electronic medical record text is structured and parsed to generate text features. The image features, test features, scale features, and text features are respectively used as modal observation features of the corresponding modalities and organized according to the acquisition time. A quality score is calculated for each modal observation feature at each acquisition time. The quality score is determined by an integrity score and a noise score. The integrity score is calculated based on the missing proportion or the proportion of valid fields of the modal observation feature, and the noise score is calculated based on at least one of the signal-to-noise ratio, artifact intensity, or outlier proportion. The quality score is a weighted combination of the integrity score and the noise score, and normalized to a preset range; the output is the modal observation features organized according to the acquisition time and the quality score corresponding to each modal observation feature.

4. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S3 includes: Based on the modal observation features organized by acquisition time and the quality score, the modal observation features corresponding to each acquisition time are sorted according to the acquisition time to generate a multi-time point observation sequence consisting of multiple acquisition times. For two adjacent acquisition times in the multi-timepoint observation sequence, calculate the time interval between the two adjacent acquisition times; Based on treatment event data, the occurrence time of each treatment event is matched with the acquisition time in the multi-time point observation sequence to determine the treatment event trigger position in the multi-time point observation sequence; The output includes a multi-timepoint observation sequence containing acquisition time, time interval, modal observation features, quality score, and treatment event trigger location.

5. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S4 include: A probabilistic state-space model is established based on multi-time point observation sequences, and a potential disease state is defined for each acquisition time. The potential disease state consists of a periodic physiological state and a therapeutic effect state. The periodic physiological state includes menstrual cycle phase latent variables and phase drift parameters corresponding to the target patient. For two adjacent acquisition times in a multi-time point observation sequence, a predicted distribution of potential disease status is generated based on the time interval using a continuous-time state transition function. The continuous-time state transition function includes: determining a state transition rate matrix based on the time interval, obtaining a state transition matrix by performing a matrix exponentiation operation on the state transition rate matrix, and scaling the process noise covariance according to the time interval to obtain the covariance parameter of the predicted distribution. When the multi-time-point observation sequence indicates that there is a treatment event trigger position between two adjacent acquisition times, the jump operator is invoked to update the efficacy status according to the event type and dose or duration parameter corresponding to the treatment event trigger position, so as to incorporate the efficacy change caused by the treatment event into the prediction distribution and complete the automatic time-series alignment of multiple time points. The jump operator includes applying an additive impact term or a multiplicative gain term to the efficacy status, and the impact intensity is calculated by the event representation vector corresponding to the event type and the dose or duration parameter through a mapping function. For each modal observation feature at each acquisition time, the modal posterior distribution of the modality is calculated based on the observation model matching the corresponding modality, and the menstrual cycle phase latent variable is estimated based on the modal posterior distribution, wherein the probability inference is implemented using any one of variational inference, particle filtering, extended Kalman filtering, and unscented Kalman filtering; Output the modal posterior distribution and the estimation results of the menstrual cycle phase latent variable.

6. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S5 include: For each acquisition time, a fusion weight is determined for each modality based on the quality score, and the fusion weight is associated with the modality posterior distribution of the corresponding modality. When the observation features of a certain modality corresponding to the acquisition time are missing, the fusion weight corresponding to that modality is set to 0. For multiple modalities acquired at the same time, the fusion weights are calculated by normalizing the quality scores of each modality, such that the sum of the fusion weights of the non-missing modalities is 1, and the fusion weight corresponding to the modality is reduced when the quality score is lower than a preset threshold. For each acquisition time, a temperature coefficient is determined for each modality based on the quality score, and the modality posterior distribution corresponding to the modality is calibrated using the temperature coefficient to suppress overconfidence in the modality posterior distribution. The calibration includes at least one of the following: normalizing the log probability density of the modality posterior distribution by multiplying it by a scaling factor corresponding to the temperature coefficient, and amplifying the variance parameter of the modality posterior distribution according to the temperature coefficient. For each acquisition time, weighted product expert inference is performed on each calibrated modality posterior distribution according to its corresponding fusion weight. By weighted fusion and normalization of each calibrated modality posterior distribution, a fused posterior distribution at that acquisition time is generated.

7. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 1, characterized in that, S6 include: For each collection time, the estimated value of the therapeutic effect status is calculated based on the fusion posterior distribution, and a therapeutic effect status sequence for the corresponding collection time is generated. For the therapeutic state sequence, the cyclical physiological state is phase-corrected based on the estimation results of the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations on the therapeutic state sequence, and a phase-corrected therapeutic state sequence is generated. The phase correction includes at least one of the following: dividing the menstrual cycle phase latent variable into multiple phase intervals, mapping the therapeutic state corresponding to different collection times to the same phase interval and then performing comparative calculations, fitting a cyclical physiological component function based on the menstrual cycle phase latent variable and subtracting the cyclical physiological component from the therapeutic state. The efficacy assessment information is calculated based on the phase-corrected efficacy status sequence, wherein the efficacy assessment information includes at least one of efficacy score or efficacy grading information; the confidence level of the efficacy assessment information is calculated based on the uncertainty of the fusion posterior distribution, wherein the uncertainty includes at least one of the following: the entropy of the fusion posterior distribution, the variance of the fusion posterior distribution, the width of the posterior confidence interval of the efficacy status, and the posterior probability that the efficacy improvement exceeds a preset threshold; the efficacy assessment information and the confidence level are output for display or storage on the terminal for medical personnel to refer to.

8. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 2, characterized in that, The treatment event data further includes menstrual-related events, which include the start time or end time of menstruation; In the probabilistic inference, the menstrual-related events are used as phase-anchored observations of the menstrual cycle phase latent variables to correct the estimation results of the menstrual cycle phase latent variables.

9. The method for evaluating the efficacy of gynecological treatment based on multimodal data fusion according to claim 5, characterized in that, The therapeutic effect state includes a baseline component and a therapeutic effect component. The therapeutic effect component is updated by decaying with the time interval according to the decay coefficient during the time interval when there is no therapeutic event triggering position. When there is a therapeutic event triggering position, the jump operator applies the additive impact term or multiplicative gain term to the therapeutic effect component.

10. A gynecological treatment efficacy evaluation system based on multimodal data fusion, used to execute the gynecological treatment efficacy evaluation method based on multimodal data fusion as described in any one of claims 1 to 9, comprising: The data acquisition module is used to acquire multimodal data and acquisition time corresponding to multiple acquisition times within the same treatment course, and to acquire treatment event data including event type and event occurrence time; The Features and Quality module is used to generate modal observation features for each modality and determine the quality score of each modal observation feature at each acquisition time. The sequence and event module is used to generate a multi-time point observation sequence according to the acquisition time, calculate the time interval between adjacent acquisition times, and map the treatment event data to the observation sequence to determine the trigger position of the treatment event; The state inference module is used to construct a probabilistic state space model, setting potential disease states that include cyclical physiological states and therapeutic states. The cyclical physiological states include menstrual cycle phase latent variables and phase drift parameters. The state is predicted based on the time interval using a continuous-time state transition function. At the location of the treatment event trigger, the therapeutic state is updated by an event-triggered jump. Based on the modal observation features, probabilistic inference is performed to obtain the posterior distribution of each modality and to estimate the menstrual cycle phase latent variables. The fusion and output module is used to determine the fusion weight and temperature coefficient based on the quality score, generate the fusion posterior distribution by weighted product expert inference after calibrating the modal posterior distribution, output the efficacy assessment information and its confidence level based on the fusion posterior distribution, and perform phase correction based on the menstrual cycle phase latent variable to eliminate the influence of physiological fluctuations.

Images