A medical health data processing method based on deep learning

Abstract

The application discloses a kind of medical health data processing methods based on deep learning, it is related to medical data processing technical field, including the following steps: collection and pre-processing multi-source patient data, form medical observation triad sequence and diagnosis and treatment event time sequence;According to patient and index, gather observation data, generate index change form fragment sequence;Combining form fragment and diagnosis and treatment event identification stage boundary, form clinical stage flow slice set;Input PFM-STraTS model, generate patient stage flow health characterization and main change index fragment;Based on trajectory prototype memory bank retrieval history health evolution prototype, generate health status evolution representation;According to health status evolution representation, output medical health data processing result.The application utilizes PFM-STraTS model, realizes medical health status evolution analysis, with the advantages of accurate identification, clear explanation and accurate supplement.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a deep learning-based method for processing medical and health data. Background Technology

[0002] With the continuous accumulation of electronic medical records, laboratory test data, vital signs data, and clinical event data in medical institutions, the industry typically organizes patient medical data into a regularized indicator matrix through methods such as data cleaning, field standardization, missing value imputation, and fixed time window alignment. This matrix is ​​then input into recurrent neural networks, temporal convolutional networks, or Transformer models for health status prediction. While this approach can identify some of the patient's stage-specific risks, its processing is often based on fixed sampling intervals or uniform time windows. This can easily weaken the irregular observational characteristics inherent in medical data and compress the direction, rate, and fluctuation of changes between adjacent observations into ordinary numerical features, making it difficult for the model to accurately represent the changes in patient indicators.

[0003] To address these issues, existing solutions have attempted to process irregular medical observation data using methods such as sparse temporal modeling, attention mechanisms, similar case retrieval, or disease progression trajectory modeling, and have improved the interpretability of results by weighting risk factors or using historical case similarity. However, these solutions often focus on learning the associations between individual observation points or only matching similar trajectories based on the overall patient representation. They lack independent representation of the changes in adjacent observations of the same medical indicator, and fail to use indicator morphological transitions and diagnostic events together to classify patient state stages. Therefore, existing technologies struggle to simultaneously output health representations with stage-flow meaning, fragments of major changing indicators, stage transition locations, and medical indicators to be supplemented corresponding to historical health evolution processes.

[0004] Therefore, how to provide a deep learning-based method for processing medical and health data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for processing medical and health data. This invention utilizes the PFM-STraTS model to achieve evolutionary analysis of medical and health status, and has the advantages of accurate identification, clear interpretation, and precise data re-collection.

[0006] A deep learning-based medical and health data processing method according to an embodiment of the present invention includes the following steps:

[0007] Collect patients' electronic medical record data, laboratory test data, vital sign data, and medical event data, preprocess them according to patient identification, and generate medical observation ternary sequences and medical event time series;

[0008] Medical observation triplet sequences are collected according to patient identification and medical indicator type. The numerical changes between adjacent observations of the same medical indicator are read to generate a sequence of indicator change morphology fragments.

[0009] Based on the sequence of indicator change patterns and the time series of diagnosis and treatment events, the patient state stage boundaries are identified, and the medical observation triplet sequence is divided into a set of clinical stage flow slices.

[0010] The medical observation triplet sequence, the index change morphology fragment sequence, and the clinical stage flow slice set are input into the PFM-STraTS model to generate the patient stage flow health representation and the main change index fragments.

[0011] A trajectory prototype memory bank is constructed based on historical medical and health samples. Historical health evolution prototypes are retrieved based on patient stage-by-stage health representations and key change index fragments to generate health status evolution representations.

[0012] Based on the health status evolution representation, medical and health data processing results are generated, including the current health status, the health status trend in the next stage, key change indicator segments, stage transition positions, and medical indicators to be supplemented.

[0013] Optionally, the generation of the medical observation triplet sequence and the time series of diagnosis and treatment events specifically includes:

[0014] Extract the names and times of medical events from the patient's electronic medical record data and medical event data according to the patient's identifier, and generate a time series of medical events.

[0015] Extract the item name, raw value, raw unit, and collection time from the test and vital sign data according to the patient identification, and generate standardized observation data;

[0016] Based on the patient identification and observation time, standardized observation data are arranged to generate a medical observation triplet sequence for each patient, which includes the observation time, medical indicator type, and observation value.

[0017] Optionally, the generation of the indicator change pattern segment sequence specifically includes:

[0018] Based on patient identification and medical indicator type, medical observation triplet sequences are collected and arranged according to observation time to generate the same medical indicator observation sequence;

[0019] Read adjacent medical observation triplets along the same medical indicator observation sequence to generate adjacent observation change segments;

[0020] Based on adjacent observed change segments, extract the direction of change, amplitude of change, duration of change, rate of change and fluctuation state to generate index change pattern segments;

[0021] Arrange the segments of indicator change patterns according to patient identification, medical indicator type, and start observation time to generate a sequence of indicator change pattern segments.

[0022] Optionally, the generation of the clinical stage flow slice set specifically includes:

[0023] Based on patient identification and time sequence, merge the sequence of index change patterns with the time series of diagnosis and treatment events to generate a patient stage analysis time series;

[0024] The direction, speed and fluctuation of changes in adjacent indicator morphology segments are read along the patient stage analysis time sequence to generate morphological transition times.

[0025] The morphological transition time and the diagnosis and treatment event time are jointly written into the patient stage analysis timeline to generate the patient state stage boundary.

[0026] Medical observation triplet sequences are segmented based on patient status stage boundaries to generate clinical stage flow slices;

[0027] The sequence of indicator change morphology fragments is assigned to the corresponding clinical stage flow slices according to the start observation time, and the clinical stage flow slice set is generated by arranging them according to patient identification and slice time order.

[0028] Optionally, the PFM-STraTS model includes an observation point embedding branch, a morphological segment embedding branch, a stage slice embedding branch, an intra-stage attention encoding unit, and an inter-stage flow attention encoding unit.

[0029] Optionally, the generation of the patient stage flow health characterization and key change index fragments specifically includes:

[0030] The medical observation triplet sequence is sparsely encoded by the observation point embedding branch to generate the observation point embedding.

[0031] The morphological segment embedding branch is used to encode the morphological state and continuous change amount of the index change morphological segment sequence to generate morphological segment embedding.

[0032] The observation point embedding and morphological segment embedding in the clinical stage flow slice are staged and organized by the stage slice embedding branch to generate stage slice embedding;

[0033] Intra-stage fusion representations are generated by fusing observation point embeddings and morphological fragment embeddings within the same clinical stage flow slice using intra-stage attention encoding units.

[0034] Inter-stage transition representations are generated by encoding intra-stage fusion representations sequentially along slice time using inter-stage transition attention encoding units.

[0035] Based on the inter-stage flow representation and intra-stage attention weight, fragments of patient stage flow health characteristics and key change indicators are generated.

[0036] Optionally, the generation of the health state evolution representation specifically includes:

[0037] The historical medical and health samples were processed to generate medical observation triplet sequences, indicator change morphology fragment sequences, and clinical stage flow slice sets, and then input into the PFM-STraTS model to generate historical patient stage flow health representations and historical major change indicator fragments.

[0038] A historical health evolution prototype is generated based on historical patient stage-by-stage health characteristics, historical key change index segments, and historical health status results.

[0039] A trajectory prototype memory bank is constructed based on the historical health status results and the main changes of the prototype in the historical health evolution prototype.

[0040] Based on the patient's stage-based health representation and key change index fragments, retrieve the target's historical health evolution prototype from the trajectory prototype memory bank;

[0041] Generate representation offset results based on the offset between the target historical health evolution prototype and the patient stage-flow health representation;

[0042] Based on the target's historical health evolution prototype, characterization shift results, and fragments of key change indicators, a health status evolution representation is generated.

[0043] Optionally, the generation of the medical and health data processing results specifically includes:

[0044] Determine the current health status and the trend of the next stage of health status based on the health status evolution representation;

[0045] The stage transition location is determined based on the current dominant change fragments and the clinical stage flow slice set;

[0046] Generate the main change index fragment output based on the main change index fragment in the health status evolution representation;

[0047] The types of medical indicators in the subsequent observation gaps of the health status evolution representation are taken as medical indicators to be supplemented.

[0048] The data processing results are generated by combining the current health status, the trend of the health status in the next stage, the main change indicators, the stage switch position, and the medical indicators to be supplemented.

[0049] The beneficial effects of this invention are:

[0050] This invention unifies the processing of patient electronic medical record data, laboratory test data, vital sign data, and treatment event data into a medical observation triplet sequence and a treatment event time series, avoiding the information distortion caused by forcibly piecing together irregular medical observation data into a fixed time interval indicator matrix. Furthermore, this invention reads the numerical changes between adjacent observations according to patient identification and medical indicator type, generating a sequence of indicator change morphology segments containing the direction, magnitude, duration, rate of change, and fluctuation state of change. This allows the model to perceive the specific process of the same medical indicator changing from stable, rising, or falling to a turning point or plateau, rather than judging health status solely based on isolated observations, thereby improving the ability of medical and health data processing results to represent changes in the patient's true condition.

[0051] This invention identifies patient state stage boundaries by combining indicator change morphological segment sequences and treatment event time series. It divides the medical observation triplet sequence into a set of clinical stage flow slices, ensuring that stage division simultaneously reflects indicator morphological transitions and the location of treatment events, reducing stage confounding issues caused by fixed time window segmentation. The PFM-STraTS model introduces morphological segment embedding branches and stage slice embedding branches based on sparse observation point encoding. It models the interaction between observation points and morphological segments within the same clinical stage, as well as state transitions between different clinical stages, through intra-stage attention encoding units and inter-stage flow attention encoding units. The resulting patient stage flow health representation more closely reflects the patient's health evolution path, and the main changing indicator segments have clear model sources and indicator change bases.

[0052] This invention utilizes a trajectory prototype memory to create historical health evolution prototypes from historical patient stage-by-stage health representations, historical dominant change segments, and historical health status results. Based on the current patient stage-by-stage health representations and key change indicator segments, it retrieves target historical health evolution prototypes to generate a health status evolution representation. This approach provides a traceable historical evolutionary basis for the current health status, the trend of the next stage health status, and the stage transition location. Furthermore, by comparing subsequent observations from historical stages with existing medical indicator types in the current last clinical stage slice, it identifies medical indicators to be supplemented. This allows the medical and health data processing results to not only provide a health status assessment but also indicate key medical indicators that need to be supplemented, improving the interpretability, continuous tracking value, and clinical auxiliary application value of the results. Attached Figure Description

[0053] 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:

[0054] Figure 1This is a flowchart of a deep learning-based medical and health data processing method proposed in this invention;

[0055] Figure 2 This is a flowchart illustrating the stage-based flow modeling of a deep learning-based medical and health data processing method proposed in this invention.

[0056] Figure 3 This is a prototype retrieval flowchart of a deep learning-based medical and health data processing method proposed in this invention. Detailed Implementation

[0057] 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.

[0058] refer to Figures 1-3 A deep learning-based method for processing medical and health data includes the following steps:

[0059] Collect patients' electronic medical record data, laboratory test data, vital sign data, and medical event data, preprocess them according to patient identification, and generate medical observation ternary sequences and medical event time series;

[0060] Medical observation triplet sequences are collected according to patient identification and medical indicator type. The numerical changes between adjacent observations of the same medical indicator are read to generate a sequence of indicator change morphology fragments.

[0061] Based on the sequence of indicator change patterns and the time series of diagnosis and treatment events, the patient state stage boundaries are identified, and the medical observation triplet sequence is divided into a set of clinical stage flow slices.

[0062] The medical observation triplet sequence, the index change morphology fragment sequence, and the clinical stage flow slice set are input into the PFM-STraTS model to generate the patient stage flow health representation and the main change index fragments.

[0063] A trajectory prototype memory bank is constructed based on historical medical and health samples. Historical health evolution prototypes are retrieved based on patient stage-by-stage health representations and key change index fragments to generate health status evolution representations.

[0064] Based on the health status evolution representation, medical and health data processing results are generated, including the current health status, the health status trend in the next stage, key change indicator segments, stage transition positions, and medical indicators to be supplemented.

[0065] In this embodiment, the generation of the medical observation triplet sequence and the time series of diagnosis and treatment events specifically includes:

[0066] Extract the names and times of medical events from the patient's electronic medical record data and medical event data according to the patient's identifier, and generate a time series of medical events.

[0067] The time series of diagnosis and treatment events is generated by normalizing the terminology of diagnosis and treatment descriptions in the patient's electronic medical record data as the name of the diagnosis and treatment event, and taking the corresponding consultation time as the time of the diagnosis and treatment event; normalizing the event names in the diagnosis and treatment event data as the name of the diagnosis and treatment event, and taking the corresponding event occurrence time as the time of the diagnosis and treatment event; and sorting the diagnosis and treatment event names and times according to the patient identifier to generate the time series of diagnosis and treatment events.

[0068] Extract the item name, raw value, raw unit, and collection time from the test and vital sign data according to the patient identification, and generate standardized observation data;

[0069] The standardized observation data is generated by unifying the project name into medical indicator type, converting the original value into a unit based on the standard unit corresponding to the medical indicator type, using the unit-converted value as the observation value, and unifying the collection time into the observation time, thus generating standardized observation data consisting of patient identification, observation time, medical indicator type and corresponding observation value.

[0070] Based on the patient identification and observation time, standardized observation data are arranged to generate a medical observation triplet sequence for each patient, which includes the observation time, medical indicator type, and observation value.

[0071] In this embodiment, the generation of the index change pattern segment sequence specifically includes:

[0072] Based on patient identification and medical indicator type, medical observation triplet sequences are collected and arranged according to observation time to generate the same medical indicator observation sequence;

[0073] Read adjacent medical observation triplets along the same medical indicator observation sequence to generate adjacent observation change segments;

[0074] The adjacent observation change segments take the medical observation triplet with the earlier time as the starting observation and the medical observation triplet with the later time as the ending observation, and obtain the starting observation time, ending observation time, starting observation value and ending observation value from the starting observation and ending observation.

[0075] Based on adjacent observed change segments, extract the direction of change, amplitude of change, duration of change, rate of change and fluctuation state to generate index change pattern segments;

[0076] The direction of change is determined as an increase, decrease, or stagnation based on the increase or decrease of the ending observation value relative to the starting observation value; the magnitude of change is determined based on the numerical difference between the ending and starting observation values; the duration is determined based on the time interval between the ending and starting observation times; the rate of change is determined based on the degree of change of the magnitude of change per unit time within the duration, and is determined as accelerated, slowed down, or maintained by comparing the degree of change per unit time of the current adjacent observation segment with that of the previous adjacent observation segment; the fluctuation state is determined as a continuous change state, a transitional change state, or a plateau change state based on the arrangement of the change directions of adjacent observation segments in the same medical indicator observation sequence; the medical indicator type, starting observation time, ending observation time, direction of change, magnitude of change, duration, rate of change state, and fluctuation state are used to generate indicator change pattern segments.

[0077] Arrange the segments of indicator change patterns according to patient identification, medical indicator type, and start observation time to generate a sequence of indicator change pattern segments.

[0078] In this embodiment, the generation of the clinical stage flow slice set specifically includes:

[0079] Based on patient identification and time sequence, merge the sequence of index change patterns with the time series of diagnosis and treatment events to generate a patient stage analysis time series;

[0080] The direction, speed and fluctuation of changes in adjacent indicator morphology segments are read along the patient stage analysis time sequence to generate morphological transition times.

[0081] The morphological transition time is determined by the morphological changes between adjacent indicator morphological segments. The indicator morphological segment with an earlier time is taken as the preceding morphological segment, and the indicator morphological segment with a later time is taken as the following morphological segment. When the direction of change of the preceding morphological segment and the following morphological segment changes from rising to falling, from falling to rising, from level to rising, from level to falling, from rising to level, or from falling to level, the starting observation time of the following morphological segment is taken as the morphological transition time. When the change speed of the preceding morphological segment and the following morphological segment changes from accelerating to decelerating, from decelerating to accelerating, from maintaining to accelerating, from maintaining to decelerating, from accelerating to maintaining, or from decelerating to maintaining, the starting observation time of the following morphological segment is taken as the morphological transition time. When the fluctuation state of the preceding morphological segment and the following morphological segment changes from a continuous change state to a transitional change state, from a transitional change state to a continuous change state, from a continuous change state to a plateau change state, from a plateau change state to a continuous change state, from a transitional change state to a plateau change state, or from a plateau change state to a transitional change state, the starting observation time of the following morphological segment is taken as the morphological transition time.

[0082] The morphological transition time and the diagnosis and treatment event time are jointly written into the patient stage analysis timeline to generate the patient state stage boundary.

[0083] The patient state stage boundaries are obtained by sorting the execution time of the morphological transition time and the diagnosis and treatment event time. The position where the morphological transition time and the diagnosis and treatment event time exist at the same time is retained as a patient state stage boundary. The position that is before the first patient state stage boundary and has a medical observation triplet is taken as the starting boundary, and the position that is after the last patient state stage boundary and has a medical observation triplet is taken as the ending boundary. The starting boundary, patient state stage boundary and ending boundary are arranged in chronological order.

[0084] Medical observation triplet sequences are segmented based on patient status stage boundaries to generate clinical stage flow slices;

[0085] The clinical stage stream slice is formed by the time period between the boundaries of adjacent patient state stages. Medical observation triplets with observation times within the same time period are grouped into the same clinical stage stream slice, and the start observation time, end observation time, and medical indicator type are written into the same clinical stage stream slice along with the corresponding medical observation triplets.

[0086] The sequence of indicator change morphology fragments is assigned to the corresponding clinical stage flow slices according to the start observation time, and the clinical stage flow slice set is generated by arranging them according to patient identification and slice time order.

[0087] In this embodiment, the PFM-STraTS model includes an observation point embedding branch, a morphological segment embedding branch, a stage slice embedding branch, an intra-stage attention encoding unit, and an inter-stage flow attention encoding unit.

[0088] The observation point embedding branch receives a single medical observation triplet from the medical observation triplet sequence and encodes the single medical observation triplet as a sparse observation point, without padding the medical observation triplet sequence into an index matrix with fixed time intervals; the morphological segment embedding branch receives a single index change morphological segment from the index change morphological segment sequence, retaining the medical index type, start observation time, and end observation time of the index change morphological segment; the stage slice embedding branch receives a single clinical stage flow slice from the clinical stage flow slice set, retaining the slice time order of the clinical stage flow slice in the clinical stage flow slice set; the intra-stage attention encoding unit connects the observation point embedding branch, the morphological segment embedding branch, and the stage slice embedding branch, performing interactive encoding of observation points and morphological segments within the same clinical stage flow slice; the inter-stage flow attention encoding unit connects the intra-stage attention encoding unit, performing flow encoding between clinical stage flow slices according to the slice time order.

[0089] The PFM-STraTS model improves upon the STraTS sparse triplet Transformer framework. Existing STraTS primarily represents medical observations as sparse observation points composed of observation time, medical indicator type, and observation value, learning the correlation between different observation points through an attention mechanism. While this avoids forcibly filling irregular medical time sequences into a fixed-time indicator matrix, its input granularity remains primarily at the level of individual observation points, making it difficult to directly express the changes in the same medical indicator between adjacent observations, and also difficult to distinguish the patient's state transitions at different clinical stages. The PFM-STraTS model adds morphological segment embedding and stage slice embedding branches to the sparse observation point encoding of STraTS. It encodes the direction, amplitude, duration, rate of change, and fluctuation state of change between adjacent observations of the same medical indicator as morphological segment embeddings, and assigns the medical observation triplet sequence according to the patient's state stage boundaries. The segmented clinical stage flow slices are encoded as stage slice embeddings. The original attention encoding, which only interacts with observation points, is improved to include intra-stage attention encoding units and inter-stage flow attention encoding units. First, a bidirectional interaction between observation point embeddings and morphological segment embeddings is established within the same clinical stage flow slice. Then, flow encoding is performed on the intra-stage fusion representations between different clinical stage flow slices along the slice time sequence. Through this improvement, the PFM-STraTS model no longer judges the patient's status based solely on isolated observations. Instead, it simultaneously utilizes sparse observation points, indicator change patterns, and clinical stage flow information to generate a patient's stage flow health representation. It also determines the main change indicator segments based on intra-stage attention weights and stage contribution weights. This enables the trajectory prototype memory bank to retrieve historical health evolution prototypes based on the patient's stage evolution path and the dominant indicator change patterns, improving the accuracy and interpretability of current health status identification, next stage health status trend judgment, stage switching location determination, and the generation of medical indicators to be supplemented.

[0090] The PFM-STraTS model is trained based on historical medical and health samples. First, each historical medical and health sample undergoes the same medical observation triplet sequence, indicator change morphology segment sequence, and clinical stage flow slice set generation process as the current patient. The diagnostic conclusions, follow-up outcomes, stage labels, and subsequent stage observations from the historical samples are used as training label sources. Specifically, the current historical health status is determined by the diagnostic conclusion of the corresponding clinical stage flow slice; the trend of the health status in the next historical stage is determined by the change in the diagnostic conclusions or follow-up outcomes of adjacent subsequent stages relative to the current stage; the historical stage switching position is determined by the time position corresponding to the stage label; and the historical stage subsequent observations are determined by the medical observation triplet that actually appears after the stage switching position. The historical major change indicator segments are determined by the contribution value of the indicator change morphology segment. During training, the medical observation triplet sequence, indicator change morphology segment sequence, and clinical stage flow slice set are input into the PFM-STraTS model to obtain the patient's stage flow health representation, intra-stage attention weights, inter-stage flow representation, and major change indicator segments. The loss function consists of the current health status classification loss, the next stage health status trend classification loss, the stage switching position loss, the multi-label loss of the medical indicators to be supplemented, and the major change index segments. The loss is composed of the consistency loss of the indicator segments. The current health status classification loss and the next stage health status trend classification loss are calculated by the cross-entropy between the model's predicted category and the corresponding training label. The stage switching position loss is calculated by the cross-entropy between the predicted slice time sequence and the corresponding slice time sequence of the stage label. The multi-label loss of the medical indicators to be supplemented is calculated by the binary cross-entropy between the predicted medical indicators to be supplemented and the types of medical indicators that actually appeared in subsequent observations of the historical stage. The consistency loss of the main change indicator segments is calculated by the number of consistent terms between the main change indicator segments output by the model and the historical main change indicator segments in terms of medical indicator type, change direction, change rate state, and fluctuation state. During training, the parameters of the observation point embedding branch, morphological segment embedding branch, stage slice embedding branch, intra-stage attention encoding unit, and inter-stage flow attention encoding unit are updated with backpropagation of the total loss. When the decrease in the total loss is less than the set minimum decrease in several consecutive rounds of training, or when the accuracy of current health status recognition, the accuracy of next stage health status trend recognition, and the recall rate of medical indicators to be supplemented on the validation samples no longer improve, training is stopped, and the trained PFM-STraTS model is used to generate patient stage flow health representations and main change indicator segments.

[0091] In this embodiment, the generation of the patient stage-specific health characterization and key change indicator segments specifically includes:

[0092] The medical observation triplet sequence is sparsely encoded by the observation point embedding branch to generate the observation point embedding.

[0093] The sparse observation point encoding is accomplished by processing medical observation triples one by one. The observation time of the first medical observation triple for the same patient is taken as the time starting point, and the time interval between the observation time of the current medical observation triple and the time starting point is taken as the relative observation time. The relative observation time is input into the continuous time value embedding layer to generate the observation time embedding; the observation value is input into the continuous numerical value embedding layer to generate the observation value embedding; the medical indicator type is input into the indicator embedding layer to generate the medical indicator embedding. Both the continuous time value embedding layer and the continuous numerical value embedding layer convert a single continuous value into an embedding vector through linear mapping, nonlinear activation, and re-linear mapping. The indicator embedding layer reads the corresponding indicator embedding matrix row vector according to the order position of the medical indicator type in the medical indicator dictionary as the medical indicator embedding. The observation point fusion layer connects the observation time embedding, observation value embedding, and medical indicator embedding corresponding to the same medical observation triple, performs linear mapping, normalization, and nonlinear activation on the connected embedding vector to generate the observation point fusion vector, and adds the observation point fusion vector to the medical indicator embedding by residual addition to generate the observation point embedding.

[0094] The medical indicator dictionary is established based on the types of medical indicators, and a corresponding indicator embedding matrix row vector is set for each type of medical indicator in the medical indicator dictionary. During the model training process, the parameters are updated as the loss of the PFM-STraTS model is backpropagated.

[0095] The morphological segment embedding branch is used to encode the morphological state and continuous change amount of the index change morphological segment sequence to generate morphological segment embedding.

[0096] The morphological state encoding is completed by processing each segment of the indicator change morphology, mapping the direction of change to a direction state code, the speed of change state to a speed state code, and the fluctuation state to a fluctuation state code. The direction state code is used to distinguish between rising, falling, and leveling off, the speed state code is used to distinguish between accelerating, slowing down, and maintaining, and the fluctuation state code is used to distinguish between continuous change state, turning change state, and plateau change state.

[0097] The continuous change encoding is completed through indicator change pattern segments under the same patient and the same medical indicator type. The change amplitude of the current indicator change pattern segment relative to the maximum change amplitude under the same medical indicator type is converted into a relative amplitude value, and the duration of the current indicator change pattern segment relative to the maximum duration under the same medical indicator type is converted into a relative duration value. When the maximum change amplitude under the same medical indicator type is zero, the relative amplitude value is recorded as zero, and when the maximum duration under the same medical indicator type is zero, the relative duration value is recorded as zero.

[0098] Specifically, the direction status code, velocity status code, fluctuation status code, amplitude relative value, duration relative value, and fragment medical indicator embedding are connected to form a morphological fragment input vector. Linear mapping, normalization, and nonlinear activation are performed to generate a morphological fusion vector, which is then added to the medical indicator embedding with residuals to generate a morphological fragment embedding. The start and end observation times of the corresponding indicator change morphological fragments are retained.

[0099] The observation point embedding and morphological segment embedding in the clinical stage flow slice are staged and organized by the stage slice embedding branch to generate stage slice embedding;

[0100] The phased processing is accomplished by processing each clinical phase flow slice individually. Observation points with observation times within the same clinical phase flow slice are embedded and determined as phase observation inputs, and morphological segments with start observation times within the same clinical phase flow slice are embedded and determined as phase morphological inputs. The phase duration is determined based on the start and end observation times of the clinical phase flow slices. The interval between the observation time of the phase observation input and the start observation time is converted into an observation position within the phase, and the interval between the start observation time of the phase morphological input and the start observation time is converted into a morphological position within the phase. The phase observation inputs are arranged according to their observation positions within the phase. The first phase observation input is used as the initial observation recursion result. Starting from the second phase observation input, the current phase observation input is linked with the previous observation recursion result. Following this, linear mapping and nonlinear activation are performed, and the final recursive fusion result is used as the stage observation representation. The stage morphological inputs are arranged according to their morphological positions within the stage, and recursive fusion is performed sequentially starting from the first stage morphological input after arrangement. Each recursive fusion connects the current stage morphological input with the previous recursive fusion result, followed by linear mapping and nonlinear activation, and the final recursive fusion result is used as the stage morphological representation. The stage observation representation, stage morphological representation, stage duration, and the slice time order of the clinical stage flow slice in the clinical stage flow slice set are concatenated to generate a stage slice input vector. Linear mapping, normalization, and nonlinear activation are performed on the stage slice input vector to generate a stage slice embedding, which identifies the stage observation input, stage morphological input, and slice time order within the same clinical stage flow slice.

[0101] Intra-stage fusion representations are generated by fusing observation point embeddings and morphological fragment embeddings within the same clinical stage flow slice using intra-stage attention encoding units.

[0102] The intra-stage attention encoding unit processes clinical stage flow slices one by one. First, it determines the objects to be encoded within the current clinical stage flow slice based on the stage observation input and stage morphological input embedded in the stage slice. For each stage observation input, it reads the medical indicator type and observation time from its corresponding medical observation triplet, and searches for stage morphological inputs within the current clinical stage flow slice that have the same medical indicator type and whose observation time is between the start and end observation times. The found stage morphological input is used as the morphological corresponding input for that stage observation input. The stage observation input is mapped to an observation query vector, and the morphological corresponding input is mapped to a morphological key vector and a morphological value vector. A dot product is performed on the observation query vector and the morphological key vector, and the dot product result is normalized to obtain the observed morphological attention weight. The morphological value vector is weighted according to the observed morphological attention weight to obtain the observation enhancement representation. When no morphological corresponding input is found, the stage observation input is used as the observation enhancement representation. For each stage morphological input, it reads the corresponding indicator change morphological segment... The process involves identifying the medical indicator type, start observation time, and end observation time within the current clinical stage stream slice. It then searches for stage observation inputs with the same medical indicator type and observation times between the start and end observation times. These found stage observation inputs are used as the corresponding observation inputs for that stage's morphological input. The stage morphological input is mapped to a morphological query vector, and the corresponding observation input is mapped to an observation key vector and an observation value vector. A dot product is calculated between the morphological query vector and the observation key vector, and the dot product result is normalized to obtain the morphological-to-observation attention weight. The observation value vector is then weighted according to this weight to obtain the morphological enhancement representation. If no corresponding observation input is found, the stage morphological input is used as the morphological enhancement representation. The observation enhancement representations within the same clinical stage stream slice are recursively fused according to the observation time, and the morphological enhancement representations within the same clinical stage stream slice are recursively fused according to the start observation time. The two recursive fusion results are then embedded and concatenated with the stage slice, followed by linear mapping, normalization, and nonlinear activation to generate an intra-stage fused representation.

[0103] Inter-stage transition representations are generated by encoding intra-stage fusion representations sequentially along slice time using inter-stage transition attention encoding units.

[0104] The inter-stage flow attention encoding unit processes the intra-stage fusion representation sequentially according to the slice time order in the clinical stage flow slice set. It concatenates the intra-stage fusion representation corresponding to the current clinical stage flow slice with the stage slice embedding to form the current stage flow input. A linear mapping is performed on the current stage flow input to obtain the current stage query vector, the current stage key vector, and the current stage value vector. When a preceding clinical stage flow slice exists before the current clinical stage flow slice, the intra-stage fusion representation corresponding to the preceding clinical stage flow slice is concatenated with the stage slice embedding to form the preceding stage flow input, and a linear mapping is performed on the preceding stage flow input to obtain the preceding stage flow input. The process involves: generating a stage key vector and a preceding stage value vector; performing a dot product between the current stage query vector and the preceding stage key vector, and normalizing the dot product result to obtain the preceding flow attention weight; weighting the preceding stage value vector according to the preceding flow attention weight to generate the preceding flow representation; when there is no preceding clinical stage flow slice before the current clinical stage flow slice, using the current stage value vector as the preceding flow representation; embedding and connecting the current stage value vector, the preceding flow representation, and the stage slice corresponding to the current clinical stage flow slice, and then performing linear mapping, normalization, and nonlinear activation to generate the inter-stage flow representation corresponding to the current clinical stage flow slice;

[0105] Based on the inter-stage flow representation and intra-stage attention weight, fragments of patient stage flow health characteristics and key change indicators are generated.

[0106] The patient staged flow health representation is obtained by performing stage contribution aggregation on the inter-stage flow representations corresponding to the clinical staged flow slice set. For each clinical staged flow slice, the average absolute value of each dimension value in the inter-stage flow representation is calculated as the stage contribution value corresponding to that clinical staged flow slice. The stage contribution value corresponding to each clinical staged flow slice is divided by the sum of all stage contribution values ​​to generate the stage contribution weight corresponding to that clinical staged flow slice. When the sum of all stage contribution values ​​is zero, the stage contribution weight corresponding to each clinical staged flow slice is set to the reciprocal of the number of clinical staged flow slices. The inter-stage flow representations corresponding to each clinical staged flow slice are weighted and aggregated according to the stage contribution weight to generate the patient staged flow health representation.

[0107] Specifically, the observed morphological attention weight and morphology-to-observation attention weight are retained in the corresponding indicator change morphological segment through the intra-stage attention encoding unit; the observed morphological attention weight and morphology-to-observation attention weight corresponding to the same indicator change morphological segment are averaged to generate the segment stage attention value, and the segment stage attention value is multiplied by the stage contribution weight of the clinical stage flow slice to which the indicator change morphological segment belongs to generate the indicator change morphological segment contribution value; the indicator change morphological segment whose contribution value is not less than the average result of the contribution values ​​of all indicator change morphological segments of the same patient is identified as the main change indicator segment.

[0108] In this embodiment, the generation of the health state evolution representation specifically includes:

[0109] The historical medical and health samples were processed to generate medical observation triplet sequences, indicator change morphology fragment sequences, and clinical stage flow slice sets, and then input into the PFM-STraTS model to generate historical patient stage flow health representations and historical major change indicator fragments.

[0110] A historical health evolution prototype is generated based on historical patient stage-by-stage health characteristics, historical key change index segments, and historical health status results.

[0111] The historical health status results are obtained from historical medical and health samples, including the historical current health status, the historical next-stage health status trend, the historical stage transition position, and subsequent observations of the historical stage. The contribution value of each indicator change pattern segment is read from the historical major change indicator segments, and the historical major change indicator segment with the largest contribution value is determined as the historical dominant change segment. The historical health evolution prototype is obtained by prototyping the health representations of historical patient stages that have the same historical current health status and the same historical next-stage health status trend, and whose historical dominant change segments have the same medical indicator type and change direction. A dimension-wise average is performed on the historical patient stage flow health representations belonging to the same historical health evolution prototype to generate prototype stage flow representations. The occurrence frequency of medical indicator type, change direction, change speed state, and fluctuation state in the historical dominant change segments belonging to the same historical health evolution prototype is counted, and the medical indicator type, change direction, change speed state, and fluctuation state with the highest occurrence frequency are retained as the prototype major change patterns. The prototype stage flow representation, prototype major change patterns, historical current health status, historical next-stage health status trend, historical stage transition position, and subsequent observations of the historical stage are collectively used to form the historical health evolution prototype.

[0112] A trajectory prototype memory bank is constructed based on the historical health status results and the main changes of the prototype in the historical health evolution prototype.

[0113] The trajectory prototype memory bank arranges historical health evolution prototypes according to their current historical health status, the trend of their health status in the next historical stage, and the types of medical indicators in the main changes of the prototype. This ensures that historical health evolution prototypes with the same current historical health status and the same types of medical indicators are located at the same search entry point. The historical health evolution prototypes within the same search entry point are then arranged according to the trend of their health status in the next historical stage and the direction of change in the main changes of the prototype. This ensures that when a patient's stage-flow health representation is searched, the medical indicator type is matched first, then the direction of change is matched, and then the differences in stage-flow representations are compared.

[0114] Based on the patient's stage-based health representation and key change index fragments, retrieve the target's historical health evolution prototype from the trajectory prototype memory bank;

[0115] The retrieval of the target historical health evolution prototype is accomplished jointly through patient stage flow health representation and major change index segments. From the major change index segments, the contribution value of each indicator change form segment is read, and the major change index segment with the largest contribution value is determined as the current dominant change segment. The medical indicator type and change direction in the current dominant change segment are read, and historical health evolution prototypes with the same medical indicator type as the prototype's major change form are searched in the trajectory prototype memory. Among the found historical health evolution prototypes, those with the same change direction as the current dominant change segment are prioritized for retention. For the retained historical health evolution prototypes, the average absolute value of the dimensionless difference between the patient stage flow health representation and the prototype stage flow representation is calculated, and the historical health evolution prototype with the smallest average absolute value of the dimensionless difference is determined as the target historical health evolution prototype. When multiple historical health evolution prototypes have the same average absolute value of the dimensionless difference, the historical health evolution prototype with the most consistent items in the change rate state and fluctuation state of the prototype's major change form with the current dominant change segment is determined as the target historical health evolution prototype.

[0116] Generate representation offset results based on the offset between the target historical health evolution prototype and the patient stage-flow health representation;

[0117] The representation offset result is obtained by comparing the patient stage flow health representation with the prototype stage flow representation in the target historical health evolution prototype dimension by dimension. The values ​​of the patient stage flow health representation and the prototype stage flow representation in the same dimension are subtracted, and the positive and negative directions and absolute magnitudes of the subtraction results are retained to generate the dimension offset result. The result is then arranged according to the dimensional order of the patient stage flow health representation to generate the representation offset result, which indicates the degree of deviation of the current patient stage flow health representation from the target historical health evolution prototype.

[0118] Based on the target's historical health evolution prototype, characterization shift results, and fragments of key change indicators, a health status evolution representation is generated;

[0119] The health status evolution representation includes the current historical health status, the trend of the next historical stage health status, the historical stage switching position, the representation offset result, and the fragments of the main change indicators in the target historical health evolution prototype; it reads the types of medical indicators included in the subsequent observations of the historical stages in the target historical health evolution prototype, and reads the types of medical indicators that already exist in the current patient's last clinical stage stream slice. The medical indicator types that exist in the subsequent observations of the historical stages but do not exist in the last clinical stage stream slice are identified as the subsequent observation gaps of the stage and written into the health status evolution representation.

[0120] In this embodiment, the generation of the medical and health data processing results specifically includes:

[0121] Determine the current health status and the trend of the next stage of health status based on the health status evolution representation;

[0122] The current health status and the trend of the next stage health status are obtained by reading the historical current health status and the historical trend of the next stage health status from the health status evolution representation;

[0123] The stage transition location is determined based on the current dominant change fragments and the clinical stage flow slice set;

[0124] The currently dominant change segment is the main change indicator segment with the largest contribution value among the identified indicator change morphology segments. The start observation time of the currently dominant change segment is read, and clinical stage flow slices with start observation times no later than the start observation time of the currently dominant change segment and end observation times no earlier than the start observation time of the currently dominant change segment are searched in the clinical stage flow slice set. The start observation time of the found clinical stage flow slice is determined as the stage switching position. When no corresponding clinical stage flow slice is found, the start observation time of the currently dominant change segment is determined as the stage switching position.

[0125] Generate the main change index fragment output based on the main change index fragment in the health status evolution representation;

[0126] The output of the main change index segments is obtained by reading the main change index segments in the health status evolution representation. When the health status evolution representation contains multiple main change index segments, they are arranged from largest to smallest according to the contribution value of the index change morphology segment corresponding to each main change index segment, and the medical indicator type, start observation time, end observation time, change direction, change amplitude, duration, change speed state and fluctuation state in each main change index segment are retained. When the contribution values ​​of the index change morphology segments of multiple main change index segments are the same, they are arranged from earliest to latest according to the start observation time.

[0127] The types of medical indicators in the subsequent observation gaps of the health status evolution representation are taken as medical indicators to be supplemented.

[0128] Among them, the medical indicators to be supplemented are arranged from earliest to latest according to the appearance time of the corresponding medical indicator type in the subsequent observations of the historical stage. When the same medical indicator type appears multiple times in the subsequent observations of the historical stage, the earliest appearance time is retained as the basis for the arrangement of the medical indicator type.

[0129] The data processing results are generated by combining the current health status, the trend of the health status in the next stage, the main change indicators, the stage switch position, and the medical indicators to be supplemented.

[0130] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous analysis scenario of medical and health data in the internal medicine ward of a tertiary hospital. In this scenario, patient laboratory tests, vital signs, and treatment events are stored in electronic medical records, laboratory platforms, bedside monitoring devices, and follow-up records. The collection frequency of different data varies greatly. For example, blood routine tests are usually collected at intervals according to doctor's orders, while body temperature and blood pressure are collected according to nursing frequency. Some treatment events are based on the time of execution of doctor's orders or the time of medical record. Traditional solutions usually supplement the above data into a regularized indicator matrix according to fixed time windows, and then input it into a recurrent neural network or a regular Transformer for risk prediction. This easily compresses the process of the same medical indicator rapidly rising, slowly recovering, or repeatedly fluctuating within a short period of time into the mean of a single window. As a result, the model can identify the existence of risk, but it is difficult to explain which indicator change segment caused the risk, and it is also difficult to indicate which medical indicators should be collected subsequently.

[0131] In this embodiment, the implementer selected inpatient data from the same period in the internal medicine ward of a tertiary hospital for validation. This included electronic medical records, laboratory test data, vital sign data, and treatment event data from 4280 patients. The training sample consisted of 3420 patients, the validation sample of 430 patients, and the test sample of 430 patients. After each patient's data was aggregated by patient identifier, a medical observation triplet sequence and a treatment event time series were generated. Subsequently, the observation data for the same patient were arranged according to medical indicator type, and the numerical changes between adjacent observations were read to generate a sequence of indicator change morphology fragments. Taking a patient with pulmonary infection and basal metabolic abnormalities as an example, within the first 36 hours after admission, the white blood cell count increased from 9.8 to 13.6, the neutrophil percentage increased from 76.1% to 84.3%, and the body temperature increased from 37.4℃ to 38.5℃. The traditional fixed 24-hour window scheme combines the above changes into the window mean and only outputs the risk of infection aggravation. After adopting the method of this invention, the PFM-STraTS model forms the indicator change morphology segments of the increase in white blood cell count, the accelerated increase in neutrophil percentage, and the turning point increase in body temperature, and generates the patient's state stage boundary by combining the anti-infective drug time. The medical observation triplet sequence is divided into three clinical stage flow slices: the admission observation stage, the abnormal aggravation stage, and the intervention response stage.

[0132] During PFM-STraTS model training, the training data consists of medical observation triplet sequences, indicator change pattern fragment sequences, and clinical stage flow slice sets generated from the same batch of historical medical and health samples. Labels are derived from diagnostic conclusions, stage annotations, follow-up outcomes, and subsequent stage observations. The current health status label is determined by the diagnostic conclusion of the corresponding clinical stage; the next stage health status trend label is determined by the change in the subsequent stage diagnostic conclusion or follow-up outcome relative to the current stage; the stage switch position label is determined by the time position in the stage annotation; the label for medical indicators to be supplemented is determined by the actual medical observation triplet that appears after the stage switch position; and the historical key change indicator fragment labels are determined by the key indicator change fragments reviewed by physicians. The training loss consists of the current health status classification loss, the next stage health status trend classification loss, the stage switch position loss, the multi-label loss for medical indicators to be supplemented, and the consistency loss of key change indicator fragments. The classification loss is calculated using cross-entropy, the multi-label loss is calculated using binary cross-entropy, and the consistency loss of key change indicator fragments is calculated based on the number of inconsistencies between the model output fragments and the manually reviewed fragments in terms of medical indicator type, direction of change, rate of change, and fluctuation state. During training, training is stopped when the total loss of validation samples decreases by less than 0.003 for eight consecutive rounds, and the accuracy of current health status identification, the accuracy of next stage health status trend identification, and the recall rate of medical indicators to be supplemented all show no improvement.

[0133] Test results show that, compared to the scheme using fixed time window completion and inputting into LSTM, the accuracy rate for identifying the current health status is 82.4%, the accuracy rate for identifying the trend of the next stage health status is 75.9%, the hit rate of the main change indicator segments is 68.7%, and the recall rate of medical indicators to be supplemented is 64.5%. When using the ordinary StraTS model to directly process the medical observation triplet sequence, the above four results are 85.6%, 78.8%, 72.3%, and 68.1%, respectively. After using the method of this invention, the above four results are improved to 89.7%, 83.5%, 81.6%, and 76.9%, respectively. In 430 test patients, the median deviation between the stage switching position output by the method of this invention and the physician review time is 3.8 hours, while that of the fixed time window scheme is 9.6 hours. The above results demonstrate that the present invention, through the combined processing of indicator change morphology fragments, clinical stage flow slice sets, and trajectory prototype memory, can more accurately express the evolution process of a patient's health status, while providing key change indicator fragments and medical indicators to be supplemented, thereby improving the accuracy, interpretability, and continuous tracking value of medical and health data processing results.

[0134] Table 1. Performance Comparison of Medical and Health Data Processing Based on Stage-Based Health Representation

[0135] Fixed Time Window - LSTM 82.4 75.9 68.7 9.6 64.5 70.8 GRU-D 83.9 76.6 69.8 8.4 66.9 71.5 IP-Net 84.7 77.4 70.6 7.9 67.3 72.6 STraTS 85.6 78.8 72.3 7.2 68.1 74.9 PFM-STraTS 89.7 83.5 81.6 3.8 76.9 82.4

[0136] The above experiments used the same test sample for comparison, which included electronic medical record data, laboratory test data, vital sign data, and treatment event data of 430 hospitalized patients. The accuracy rate of current health status identification was used to evaluate the ability of different methods to judge the current health status of patients; the accuracy rate of the next stage health status trend was used to evaluate the model's ability to predict the direction of subsequent health evolution; the hit rate of key change indicator segments was used to evaluate the consistency between the output key change indicator segments and the key indicator change segments reviewed by physicians; the median deviation of the stage switching position was used to evaluate the time deviation of the model in locating the patient's state stage boundary; the smaller this value, the closer the stage switching position is to the review result; the recall rate of medical indicators to be supplemented was used to evaluate whether the model can identify the types of medical indicators that are missing from subsequent observations of the current patient relative to the historical stage; and the consistency rate of historical health evolution prototype matching was used to evaluate whether the retrieved historical health evolution prototypes are consistent with the health evolution categories reviewed by physicians.

[0137] As shown in Table 1, the fixed-time-window LSTM achieved an accuracy of 82.4% in identifying the current health status, but only 75.9% in identifying the trend of the next health status. The median bias of the stage switching position was 9.6 hours. This indicates that while compressing irregular medical observation data into a fixed-time window allows for status judgment using the overall numerical distribution within the window, it fails to adequately express the changes in indicators between adjacent observations. This is particularly true when the patient's status changes from stable to abnormal, or from abnormal to intervention response, leading to lag in stage boundary localization. GRU-D and IP-Net addressed missing and irregular sampling, improving the accuracy of current health status identification to 83.9% and 84.7%, respectively, and reducing the median bias of the stage switching position to 8.4 hours and 7.9 hours, respectively. However, both still primarily model observations, missing states, and time intervals, failing to independently form indicator change pattern segments based on the direction of change, rate of change, and fluctuation. Therefore, the hit rates of the main change indicator segments were only 69.8% and 70.6%, respectively, showing limited improvement.

[0138] STraTS directly processes medical observation triple sequences, avoiding some information loss caused by fixed window completion. This results in an accuracy rate of 85.6% for current health status identification and 78.8% for the next stage health status trend, representing improvements of 3.2 and 2.9 percentage points respectively compared to fixed-time-window LSTM. These results demonstrate that sparse triple representations that preserve observation time, medical indicator type, and observation value have a positive effect on irregular medical time series analysis. However, STraTS still uses a single sparse observation point as the primary modeling object, failing to explicitly generate sequence sequences of indicator change morphology fragments or generate a set of clinical stage flow slices based on indicator morphology transitions and treatment event times. Therefore, its median bias for stage switching positions remains 7.2 hours, and the hit rate for major change indicator fragments is only 72.3%, indicating insufficient ability to locate key indicator change processes and stage boundaries.

[0139] PFM-STraTS outperformed PFM-STraTS across all metrics, with an accuracy rate of 89.7% for current health status identification, 83.5% for the next stage health status trend, 81.6% for the hit rate of major change indicator segments, a median deviation of the stage switching position reduced to 3.8 hours, a recall rate of 76.9% for medical indicators to be supplemented, and a historical health evolution prototype matching consistency rate of 82.4%. This performance improvement primarily stems from three aspects: First, the indicator change morphology fragment sequence explicitly expresses the direction, magnitude, duration, rate of change, and fluctuation state of adjacent observations of the same medical indicator, enabling the model to identify clinically significant changes such as a sustained increase in white blood cell count, a turning point in body temperature, and a rapid increase in the proportion of neutrophils. Second, the clinical stage flow slice set is jointly generated by the indicator change morphology fragment sequence and the time series of diagnosis and treatment events, ensuring that stage boundaries simultaneously reflect both indicator morphology turning points and the location of diagnosis and treatment interventions, reducing stage confounding caused by fixed window segmentation. Third, the trajectory prototype memory bank retrieves historical health evolution prototypes based on the patient's stage flow health representation and major change indicator fragments, providing a historical evolutionary path basis for the generation of current health status, the trend of the next stage health status, and medical indicators to be supplemented, rather than relying solely on the single model classification output.

[0140] As shown in Table 1, PFM-STraTS, through joint modeling of medical observation triplet sequences, indicator change morphology fragment sequences, and clinical stage flow slice sets, simultaneously improves the capabilities of health status identification, trend prediction, stage boundary localization, key indicator interpretation, and supplementary indicator identification in irregular medical observation scenarios. Compared with fixed time window-LSTM, PFM-STraTS improves the accuracy of current health status identification by 7.3 percentage points, the accuracy of the next stage health status trend by 7.6 percentage points, the hit rate of major change indicator fragments by 12.9 percentage points, reduces the median bias of stage switching positions by 5.8 hours, and improves the recall rate of medical indicators to be supplemented by 12.4 percentage points. These results indicate that the processing method based on stage-flow health representation and trajectory prototype memory can more completely preserve the patient's indicator change process and clinical stage evolution path, giving the medical and health data processing results higher accuracy, interpretability, and continuous tracking value.

[0141] The above are merely preferred embodiments 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.

Claims

1. A deep learning-based method for processing medical and health data, characterized in that, Includes the following steps: Collect patients' electronic medical record data, laboratory test data, vital sign data, and medical event data, preprocess them according to patient identification, and generate medical observation ternary sequences and medical event time series; Medical observation triplet sequences are collected according to patient identification and medical indicator type. The numerical changes between adjacent observations of the same medical indicator are read to generate a sequence of indicator change morphology fragments. Based on the sequence of indicator change patterns and the time series of diagnosis and treatment events, the patient state stage boundaries are identified, and the medical observation triplet sequence is divided into a set of clinical stage flow slices. The medical observation triplet sequence, the index change morphology fragment sequence, and the clinical stage flow slice set are input into the PFM-STraTS model to generate the patient stage flow health representation and the main change index fragments. A trajectory prototype memory bank is constructed based on historical medical and health samples. Historical health evolution prototypes are retrieved based on patient stage-by-stage health representations and key change index fragments to generate health status evolution representations. Based on the health status evolution representation, medical and health data processing results are generated, including the current health status, the health status trend in the next stage, key change indicator segments, stage transition positions, and medical indicators to be supplemented.

2. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the medical observation triplet sequence and the time series of diagnosis and treatment events specifically includes: Extract the names and times of medical events from the patient's electronic medical record data and medical event data according to the patient's identifier, and generate a time series of medical events. Extract the item name, raw value, raw unit, and collection time from the test and vital sign data according to the patient identification, and generate standardized observation data; Based on the patient identification and observation time, standardized observation data are arranged to generate a medical observation triplet sequence for each patient, which includes the observation time, medical indicator type, and observation value.

3. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the indicator change pattern segment sequence specifically includes: Based on patient identification and medical indicator type, medical observation triplet sequences are collected and arranged according to observation time to generate the same medical indicator observation sequence; Read adjacent medical observation triplets along the same medical indicator observation sequence to generate adjacent observation change segments; Based on adjacent observed change segments, extract the direction of change, amplitude of change, duration of change, rate of change and fluctuation state to generate index change pattern segments; Arrange the segments of indicator change patterns according to patient identification, medical indicator type, and start observation time to generate a sequence of indicator change pattern segments.

4. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the clinical stage flow slice set specifically includes: Based on patient identification and time sequence, merge the sequence of index change patterns with the time series of diagnosis and treatment events to generate a patient stage analysis time series; The direction, speed and fluctuation of changes in adjacent indicator morphology segments are read along the patient stage analysis time sequence to generate morphological transition times. The morphological transition time and the diagnosis and treatment event time are jointly written into the patient stage analysis timeline to generate the patient state stage boundary. Medical observation triplet sequences are segmented based on patient status stage boundaries to generate clinical stage flow slices; The sequence of indicator change morphology fragments is assigned to the corresponding clinical stage flow slices according to the start observation time, and the clinical stage flow slice set is generated by arranging them according to patient identification and slice time order.

5. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The PFM-STraTS model includes an observation point embedding branch, a morphological segment embedding branch, a stage slice embedding branch, an intra-stage attention coding unit, and an inter-stage flow attention coding unit.

6. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the patient stage-specific health characterization and key change index segments specifically includes: The medical observation triplet sequence is sparsely encoded by the observation point embedding branch to generate the observation point embedding. The morphological segment embedding branch is used to encode the morphological state and continuous change amount of the index change morphological segment sequence to generate morphological segment embedding. The observation point embedding and morphological segment embedding in the clinical stage flow slice are staged and organized by the stage slice embedding branch to generate stage slice embedding; Intra-stage fusion representations are generated by fusing observation point embeddings and morphological fragment embeddings within the same clinical stage flow slice using intra-stage attention encoding units. Inter-stage transition representations are generated by encoding intra-stage fusion representations sequentially along slice time using inter-stage transition attention encoding units. Based on the inter-stage flow representation and intra-stage attention weight, fragments of patient stage flow health characteristics and key change indicators are generated.

7. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the health state evolution representation specifically includes: The historical medical and health samples were processed to generate medical observation triplet sequences, indicator change morphology fragment sequences, and clinical stage flow slice sets, and then input into the PFM-STraTS model to generate historical patient stage flow health representations and historical major change indicator fragments. A historical health evolution prototype is generated based on historical patient stage-by-stage health characteristics, historical key change index segments, and historical health status results. A trajectory prototype memory bank is constructed based on the historical health status results and the main changes of the prototype in the historical health evolution prototype. Based on the patient's stage-based health representation and key change index fragments, retrieve the target's historical health evolution prototype from the trajectory prototype memory bank; Generate representation offset results based on the offset between the target historical health evolution prototype and the patient stage-flow health representation; Based on the target's historical health evolution prototype, characterization shift results, and fragments of key change indicators, a health status evolution representation is generated.

8. The medical and health data processing method based on deep learning according to claim 1, characterized in that, The generation of the medical and health data processing results specifically includes: Determine the current health status and the trend of the next stage of health status based on the health status evolution representation; The stage transition location is determined based on the current dominant change fragments and the clinical stage flow slice set; Generate the main change index fragment output based on the main change index fragment in the health status evolution representation; The types of medical indicators in the subsequent observation gaps of the health status evolution representation are taken as medical indicators to be supplemented. The data processing results are generated by combining the current health status, the trend of the health status in the next stage, the main change indicators, the stage switch position, and the medical indicators to be supplemented.

Images