Multi-center medical record data decision method and system based on federated learning
By evaluating the local queue completeness and temporal truncation attributes of multi-center medical record data, an adaptive aggregation scheduling strategy is constructed to solve the temporal truncation bias problem in cross-regional medical networks, and to achieve efficient decision-making and long-term prediction of multi-center medical record data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NATIONAL HEALTH & MEDICAL BIG DATA RESEARCH INSTITUTE (SHENZHEN)
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-14
AI Technical Summary
In cross-regional medical collaboration networks, the temporal truncation bias caused by the temporal truncation of data distribution among centers causes deep neural networks to deviate from the optimization direction during feature extraction and lack the ability to predict complex evolutionary paths.
By extracting the follow-up time span and missing data points from multicenter medical records, the completeness of local cohorts is assessed, complete cohorts and truncated cohorts are divided, and an adaptive aggregation scheduling strategy is constructed. Asymmetric intervention rules are constructed using the difference between the initial gradient and the terminal gradient, and a relay fusion process of dynamic damping control is performed to generate a multicenter decision model.
It enhances the ability of multi-center collaborative networks to distinguish time-series truncation phenomena, reduces the dilution of early consensus features and late mutation features, and improves the long-term prediction capability of federated networks in time-series truncation scenarios.
Smart Images
Figure CN122392999A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data decision technology, specifically to a multi-center medical record data decision method and system based on federated learning. Background Technology
[0002] In the healthcare field, existing federated learning global aggregation algorithms (such as the traditional federated averaging algorithm FedAvg) typically use the local data sample size of each participating center as the basis for weight allocation, and perform a unified global weighted average of all network layer parameters of the local model.
[0003] In cross-regional medical collaboration networks, the data distribution among centers exhibits a physical phenomenon of temporal truncation. On the one hand, core hospitals that were established earlier often have complete follow-up cohorts, with data covering the entire process from basic signs in the early stages of the disease to rare mutations in the later stages; on the other hand, newly established tiered medical institutions often only have truncated cohorts of short-term follow-up, with data characteristics limited to common manifestations in the early stages of the disease.
[0004] Due to the inherent differences in follow-up time, a time truncation bias arises. In feature extraction, shallow layers of deep neural networks tend to extract common early consensus features, while deep layers are used to capture complex late-stage deterioration mutations. Gradient updates generated by subcenters containing only early truncation data on deep networks are lost gradients. However, truncation centers possess a large base of early outpatients, and the large sample size and weights force the global model to absorb a large amount of meaningless lost gradients during deep aggregation. This causes the global model to deviate from its optimization direction in the deep feature space, resulting in a short-sighted output model. Subcenters lack the ability to deduce and predict complex evolutionary paths in mid-to-late-stage case decisions.
[0005] Therefore, this invention provides a multi-center medical record data decision-making method and system based on federated learning. Summary of the Invention
[0006] The purpose of this invention is to provide a multi-center medical record data decision-making method and system based on federated learning to solve the aforementioned background problems.
[0007] The objective of this invention can be achieved through the following technical solutions:
[0008] A multicenter medical record data decision-making approach based on federated learning includes the following steps:
[0009] The follow-up time span and missing data points in multicenter medical record data are extracted and integrated to establish observational distribution characteristics. Based on the observational distribution characteristics, the temporal coverage integrity of multicenter medical record data in the disease evolution cycle is evaluated to obtain the local cohort completeness of the corresponding medical record data.
[0010] The local queue completeness of all participating centers is sorted out, and the relative difference distribution among the sub-centers on the follow-up time axis is extracted; the sub-centers are divided into truncated queues according to the relative difference distribution, and the temporal truncated attribute state of each sub-center is marked.
[0011] Receive local model parameters obtained from training based on local queues at each subcenter, measure the update dispersion of local model parameters according to network layers and divide them into stable layers and oscillating layers; combine the temporal truncation attribute state to compare the feature evolution differences between the complete queue and the truncation queue to obtain the initial gradient and the terminal gradient;
[0012] An asymmetric intervention rule is constructed based on the evolutionary difference between the initial gradient and the final gradient; an adaptive aggregation scheduling strategy is generated based on the asymmetric intervention rule to correct the time truncation bias.
[0013] Furthermore, the method for assessing the time coverage integrity is as follows:
[0014] Read the standard time span reference value issued by the federated learning global server, divide the average time span value in the observed distribution characteristics by the standard time span reference value to obtain the basic coverage ratio;
[0015] Based on the preset full score coefficient, a penalty adjustment coefficient is established according to the magnitude of the missing density in the observed distribution characteristics;
[0016] The time coverage integrity index is calculated by multiplying the basic coverage ratio by the penalty adjustment coefficient.
[0017] The time coverage integrity index of each center is mapped to a pre-set percentage evaluation scale to obtain the local cohort completeness.
[0018] Furthermore, the method for marking the temporal truncation attribute state is as follows:
[0019] The completeness scores of each participating center are evaluated, and each sub-center is divided into a complete queue and a truncated queue. The absolute value of the truncated deviation of each sub-center is quantified based on the score differences.
[0020] The classification labels of the corresponding queues are bound to the absolute value of the truncation deviation to generate unique time-series truncation attribute states for each sub-center.
[0021] Furthermore, the method for extracting the sequence cliff points is as follows:
[0022] Read the completeness scores uploaded by all participating centers received by the global federated learning server;
[0023] Completeness scores are sorted in descending order of numerical value to construct a global score sequence;
[0024] Extract the global score range of the global score sequence, and calculate the numerical difference between the completeness scores of every two adjacent sorting positions in the global score sequence.
[0025] All numerical differences are concatenated according to their original sorting positions to generate adjacent decay gradient sequences. The adjacent decay gradient sequences are traversed, and the sorting node containing the largest decay gradient is located as the sequence cliff point.
[0026] Furthermore, the method for comparing the differences in feature evolution is as follows:
[0027] The historical change records of local model parameters are traced to evaluate the update dispersion of each network layer. Based on the update dispersion, a dynamic segmentation boundary is established, and the network layers are divided into stationary layers and oscillating layers.
[0028] Based on the temporal truncation attribute state of each sub-center, the local model parameters are grouped into a complete queue parameter set and a truncation queue parameter set;
[0029] In the steady layer, the similarity of parameter updates between the full queue and the truncated queue is evaluated, and the parameter update vector reflecting the early consensus characteristics is extracted as the starting gradient.
[0030] In the oscillation layer, the parameter update divergence between the complete queue and the truncated queue is evaluated, and the gradient is adjusted numerically based on the degree of truncation deviation. The enhanced feature components of the complete queue are extracted as the terminal gradient.
[0031] Furthermore, the process of extracting the complete queue feature components is as follows:
[0032] Calculate the parameter mean vector of the full queue and the parameter mean vector of the truncated queue respectively. Subtract the parameter mean vector of the truncated queue from the parameter mean vector of the full queue to obtain the update direction deviation vector. Read the absolute value of the truncated deviation of the truncated queue to construct the adaptive scaling multiplier.
[0033] The complete queue is numerically amplified and enhanced using an adaptive scaling multiplier, and the gradient components of the amplified and enhanced complete queue are extracted as the terminal gradient.
[0034] Furthermore, based on the adaptive aggregation scheduling strategy, the local model parameters are subjected to relay fusion processing with stage-by-stage damping control to obtain global shallow and deep feature matrices and perform cross-time period network alignment to output a multi-center decision model with evolution prediction capabilities.
[0035] Furthermore, the relay fusion process is performed as follows:
[0036] The local model parameters are divided into shallow parameter blocks corresponding to early features and deep parameter blocks corresponding to late features according to the network layers.
[0037] In the first fusion stage, the shallow parameter blocks are statically weighted and fused based on the basic data capacity of each sub-center to generate a global shallow feature matrix.
[0038] In the second fusion phase, a dynamic damping control mechanism related to network layer depth is activated. As the network depth increases, the aggregation weight of the truncated queue is gradually and smoothly decayed, and reverse compensation weighting is performed on the complete queue.
[0039] Based on the dynamic weights adjusted by the dynamic damping control mechanism, the deep matrices uploaded by all sub-centers are weighted and summed to output the global deep feature matrix.
[0040] Furthermore, the corresponding reverse compensation weighting method is as follows:
[0041] Obtain the sum of the base capacity weights that are reduced due to the application of the attenuation damping coefficient for all subcenters in the truncated queue;
[0042] Extract the pre-determined purification weights of each sub-center in the complete queue, calculate the proportion of the purification weight of a single sub-center in the total purification weight of the complete queue, and obtain the compensation allocation coefficient.
[0043] Multiply the sum of the basic capacity weight values by the compensation allocation coefficient of each sub-center to obtain the weight compensation increment that each sub-center in the complete queue should obtain at the current deep network level.
[0044] The basic capacity weight of each sub-center in the complete queue is added together with the corresponding weight compensation increment to obtain the dynamic weight of the complete queue.
[0045] A multi-center medical record data decision system based on federated learning includes the following modules:
[0046] Feature extraction module: used to extract the follow-up time span and missing data points in multi-center medical record data, and integrate them to establish observation distribution features; based on the observation distribution features, evaluate the time coverage integrity of multi-center medical record data, and obtain the local cohort completeness of the corresponding medical record data;
[0047] Attribute classification module: used to organize the local queue completeness of all participating centers and extract the relative difference distribution between sub-centers; divide the sub-centers into truncated queues according to the relative difference distribution and mark the temporal truncated attribute state of each sub-center;
[0048] Evolutionary analysis module: used to receive local model parameters from each sub-center, measure and update the dispersion according to the network level and divide the stable layer and oscillating layer; combined with the temporal truncation attribute state, compare the characteristic evolution differences between the complete queue and the truncation queue to obtain the initial gradient and the terminal gradient;
[0049] Strategy construction module: Constructs asymmetric intervention rules based on the evolutionary differences between the initial and final gradients; generates an adaptive aggregation scheduling strategy based on the asymmetric intervention rules;
[0050] Parameter fusion module: It is used to perform relay fusion processing on local model parameters according to the adaptive aggregation scheduling strategy, obtain global deep and shallow feature matrices, perform network alignment, and output a multi-center decision model.
[0051] The beneficial effects of this invention are as follows:
[0052] 1. By extracting the follow-up time span and data gaps and constructing observational distribution characteristics, the completeness of local cohorts in the disease evolution cycle can be assessed, enabling the quantification of the longitudinal depth and reliability of multi-center original medical record data. Completeness scores reflect the degree of completeness of local cohort data follow-up, which helps reduce the interference of low-quality or extremely short follow-up period data on the overall disease evolution projection. The relative difference distribution of sub-centers on the follow-up time axis is extracted, thereby dividing the cohort into complete and truncated cohorts, and labeling the temporal truncation attribute state of each sub-center. By dividing the cohort with long-term observation capabilities into a truncated cohort containing only early data, the deviation of premature termination of the time series is quantified. The temporal truncation attribute state transforms clinical physical truncation phenomena into identifiable labels in federated learning, which helps enhance the ability of multi-center collaborative networks to distinguish temporal truncation phenomena.
[0053] 2. By leveraging the differences in sensitivity to feature extraction between shallow and deep layers of the network, we can locate the consensus update vector corresponding to early common features and the directional divergence vector corresponding to late complex features. By combining adaptive scaling multipliers to perform numerical augmentation and decay suppression on gradients, we transform gradient differences into cross-queue behavior comparisons. This helps identify the update misalignment phenomenon caused by the lack of late-stage data in deep networks due to truncated queues, and reduces the mutual dilution of early consensus features and late-stage mutation features during global weight aggregation.
[0054] 3. An asymmetric intervention rule is constructed based on the evolutionary differences between the initial and final gradients, and an adaptive aggregation scheduling strategy is generated based on this rule. This helps to block the ineffective interference of the truncated queue. The scheduling strategy transforms the intervention rule into weight evolution instructions for different network depths. According to the scheduling strategy, relay fusion processing is performed on local model parameters, and global parameters are integrated to complete cross-time period network alignment. Dynamic relay fusion helps to reduce the reverse dilution of short-period data on long-term disease deterioration characteristics and alleviate network structure fluctuations caused by discarding parameters. The output multi-center decision model contains massive early consensus data from the entire network, while retaining the mutation characteristics of late-stage extrapolation capabilities, which is beneficial to improving the long-term prediction capability of federated networks in time-truncated scenarios. Attached Figure Description
[0055] The invention will now be further described with reference to the accompanying drawings.
[0056] Figure 1 This is a flowchart of the multi-center medical record data decision method based on federated learning, as described in this invention;
[0057] Figure 2 This is a flowchart of the temporal truncation attribute state of each subcenter in this invention;
[0058] Figure 3 This is a functional block diagram of the multi-center medical record data decision system based on federated learning in this invention. Detailed Implementation
[0059] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0060] Example 1:
[0061] like Figure 1 As shown, the multi-center medical record data decision-making method based on federated learning includes the following steps:
[0062] S10: Extract the follow-up time span and missing data points from multicenter medical record data, and integrate them to establish observational distribution characteristics; evaluate the temporal coverage integrity of multicenter medical record data in the disease evolution cycle based on the observational distribution characteristics, and obtain the local cohort completeness of the corresponding medical record data.
[0063] The process of extracting the follow-up time span and missing data points from multi-center medical record data and integrating them to establish observational distribution characteristics is as follows:
[0064] In some embodiments, a data reading plugin for federated learning is deployed according to the interface protocol of the local medical system of each center to build a standardized local medical record data extraction interface;
[0065] By iterating through the structured follow-up record table in the local medical records through the local medical record data extraction interface, the time interval between the first diagnosis time node and the last re-examination time node of each patient is extracted as the time span of a single case follow-up.
[0066] Data missing points and abnormal patient samples were obtained by screening for missing data based on the time span of single-case follow-up.
[0067] For example, the method for data missing filtering is as follows: during the synchronous process of extracting the follow-up time span of a single case, the preset key test indicator fields in the structured follow-up record table are checked. If the value of a key test indicator in a certain follow-up record is empty, or an invalid placeholder (such as all zero characters) is entered, then the record of the corresponding key test indicator field is marked as a data missing point; cases with a single case follow-up time span of zero or negative numbers are marked as abnormal patient samples.
[0068] Remove abnormal patient samples, continuously organize the follow-up time span of single cases of valid patients and the corresponding data missing points, until all valid medical records within the local client's authorized scope have been traversed;
[0069] If the system has traversed all valid cases, the local data collection is considered complete.
[0070] Preferably, the criteria for determining a valid medical record are: the continuous follow-up records contain at least three valid follow-up time points, and the interval between a single follow-up visit does not exceed the maximum interval threshold specified in the clinical guidelines for disease monitoring; for example, for cardiovascular chronic diseases, the maximum interval threshold does not exceed 6 months.
[0071] Organize the data consisting of all valid medical records obtained from the traversal and establish a local queue;
[0072] If the system completes the construction of the local queue, it is determined that the local data acquisition is complete.
[0073] After the local data collection was completed, numerical statistical analysis was performed on the follow-up time span of each effective patient in the local cohort, and the average time span and the range of the local cohort were calculated.
[0074] The total number of missing data points is distributed across the sum of the single-instance follow-up time spans of the local queue, and the missing density per unit follow-up time is calculated.
[0075] The average time span, the range of time span, and the missing density are organized and aggregated to obtain the observed distribution characteristics of the local cohorts of each center;
[0076] The process of assessing the temporal coverage integrity of multicenter medical record data during the disease evolution cycle based on observed distribution characteristics, and obtaining the local cohort completeness of the corresponding medical record data, is as follows:
[0077] Read the standard time span reference values for the entire natural evolution cycle of the target disease issued by the global server of federated learning;
[0078] Divide the average time span in the observed distribution characteristics by the standard time span reference value to obtain the basic coverage ratio that reflects the length of time coverage;
[0079] Using a preset full score coefficient (e.g., a constant 1) as a benchmark, subtract the corresponding weight value equal to the missing density value (i.e., directly subtract the missing density from the full score coefficient) to obtain a penalty adjustment coefficient used to characterize data reliability.
[0080] The time coverage integrity index after data quality degradation correction is calculated by multiplying the basic coverage ratio by the penalty adjustment coefficient.
[0081] Map the time coverage integrity index of each center to a preset percentage evaluation scale and assign a discrete score value to the time coverage integrity index.
[0082] The discretized score is defined as a completeness score that reflects the degree of completeness of local cohort data follow-up, and the local cohort completeness is obtained.
[0083] It is understandable that the purpose of determining the completeness of a local queue is:
[0084] Function 1: To quantify and unify the varying follow-up times and inconsistent data record quality of different centers, serving as a priori basis for weight allocation when aggregating the global federated learning model, and eliminating the interference of low-quality or extremely short follow-up period center data on the global disease evolution prediction.
[0085] Secondly, the completeness score reflects the most original vertical depth of local medical records before global federated feature interaction, providing a zero-point reference for data quality that has been measured and calibrated for subsequent evaluation of the generalization ability of federated learning models in local environments of various centers.
[0086] S20: Organize the local queue completeness of all participating centers and extract the relative difference distribution among sub-centers on the follow-up time axis; divide the sub-centers into truncated queues according to the relative difference distribution and mark the temporal truncated attribute state of each sub-center;
[0087] The process of summarizing the local cohort completeness of all participating centers and extracting the relative difference distribution among sub-centers on the follow-up time axis is as follows:
[0088] Read the completeness scores uploaded by all participating centers received by the global federated learning server;
[0089] Completeness scores are sorted in descending order of numerical value to construct a one-dimensional global score sequence.
[0090] Extract the highest and lowest score values from the global score sequence and calculate their difference to obtain the global score range.
[0091] In the global scoring sequence, calculate the numerical difference between the completeness scores of every two adjacent sorting positions one by one;
[0092] All numerical differences are concatenated according to their original sorting positions to generate an adjacent decay gradient sequence;
[0093] Traverse adjacent decay gradient sequences and retrieve the decay gradient with the largest value. Since the largest decay gradient is obtained by subtracting the score of the next position from the score of the previous position, extract the sorting node of the next position corresponding to the largest decay gradient and locate it as the cliff point of the sequence.
[0094] Establish a dataset containing the global score range, the global score sequence, and the sequence cliff point, defined as reflecting the relative difference distribution among subcenters on the follow-up time axis;
[0095] Among them, such as Figure 2 As shown, the process of dividing the subcenters into truncated queues based on the relative difference distribution and marking the temporal truncated attribute state of each subcenter is as follows:
[0096] Extract sequence cliff points from the relative differential distribution, and extract the completeness score at the sequence cliff point position in the global scoring sequence as a dynamic segmentation threshold;
[0097] Compare the completeness scores of all participating centers with the dynamic segmentation threshold one by one;
[0098] If the completeness score of a certain subcenter is greater than or equal to the dynamic segmentation threshold, the corresponding subcenter is determined to contain long-term observation features of the middle and late stages of disease evolution, and the corresponding subcenter is divided into a complete queue with long-term observation capabilities.
[0099] If the completeness score of a certain subcenter is less than the dynamic segmentation threshold, it is determined that the data chain of the corresponding subcenter is interrupted early and is mainly concentrated in the initial stage of disease evolution. The corresponding subcenter is divided into a truncated queue containing only early data.
[0100] For the subcenters that are divided into truncated queues, the completeness score of the corresponding subcenter is subtracted from the dynamic segmentation threshold to calculate the absolute value of the truncation deviation, which is used to characterize the degree of early termination of the time series.
[0101] It should be noted that the absolute value of the truncation deviation is used to quantify the specific degree of truncation for each subcenter in the truncation queue;
[0102] For the subcenters that are divided into complete queues, the absolute value of the truncation deviation is assigned to zero;
[0103] By binding the pre-determined classification labels of the complete queue or truncated queue, and the absolute value of the truncation deviation corresponding to each classification label, a unique temporal truncation attribute state is generated for each sub-center.
[0104] Example 2:
[0105] Please see Figure 1 As shown, the multi-center medical record data decision-making method based on federated learning includes the following steps:
[0106] S30: Receive the local model parameters obtained by each sub-center based on the local queue training, measure the update dispersion of the local model parameters according to the network layer and divide the stable layer and oscillating layer; combine the temporal truncation attribute state to compare the feature evolution differences between the complete queue and the truncation queue, and obtain the initial gradient and the terminal gradient;
[0107] The process of receiving local model parameters from each sub-center based on local queue training, measuring the update dispersion of the local model parameters according to network layers, and dividing the layers into stable and oscillating layers is as follows:
[0108] In the current communication round of federated learning, local model parameters uploaded by each subcenter after training based on the local local queue are received, and the local model parameters are cached in the local repository to build a historical parameter time series.
[0109] For example, the local model parameters specifically include the weight matrix and bias vector used to construct the disease evolution prediction neural network;
[0110] Retrieve the local repository to trace the local model parameter changes of each sub-center in multiple historical communication rounds;
[0111] It should be noted that the local model parameter change record refers to the historical sequence of the numerical difference of the local model parameters between two adjacent communication rounds for the same subcenter.
[0112] The numerical fluctuation variance of local model parameters in consecutive historical communication rounds is calculated layer by layer according to the network hierarchy, and the numerical fluctuation variance is defined as the update dispersion of the corresponding network hierarchy.
[0113] Calculate the global arithmetic mean of the updated discreteness across all network layers, and use the global arithmetic mean as the dynamic discreteness split boundary.
[0114] Network layers with update discreteness strictly less than the discreteness split boundary are classified as stationary layers, while network layers with update discreteness greater than or equal to the discreteness split boundary are classified as oscillating layers.
[0115] The process of obtaining the initial gradient and the final gradient by combining the temporal truncation attribute states with the feature evolution differences between the complete queue and the truncated queue is as follows:
[0116] Based on the temporal truncation attribute state of each sub-center, the local model parameters uploaded by each sub-center are classified and grouped to obtain the parameter set belonging to the complete queue and the parameter set belonging to the truncation queue.
[0117] In the stationary layer, the update vectors corresponding to the parameter sets of the complete queue and the parameter sets of the truncated queue are extracted, and the gradient direction cosine similarity between the update vectors of the complete queue and the truncated queue is calculated.
[0118] The cosine similarity of all calculated gradient directions within the stationary layer is statistically analyzed, and the median is extracted as the dynamic similarity benchmark.
[0119] Since both the complete queue and the truncated queue contain medical record data from the initial stage of disease evolution, the update vector with a gradient direction cosine similarity greater than the similarity benchmark value is extracted and used as the consensus update vector with consensus features.
[0120] Extract the consensus update vector as the initial gradient representing early signs;
[0121] In the oscillation layer, the mean vector of parameters for the parameter set of the complete queue and the mean vector of parameters for the parameter set of the truncated queue are calculated respectively.
[0122] Calculate the parameter mean vector of the complete queue and the parameter mean vector of the truncated queue respectively. Subtract the parameter mean vector of the truncated queue from the parameter mean vector of the complete queue to obtain the update direction divergence vector between the complete queue and the truncated queue.
[0123] Read the absolute value of the truncation deviation calculated for the truncation queue, and establish an adaptive scaling multiplier based on the absolute value of the truncation deviation;
[0124] The complete queue is numerically amplified and enhanced based on an adaptive scaling multiplier, while the truncated queue is attenuated and suppressed.
[0125] Preferably, the amplification and enhancement process is as follows: the absolute value of the truncation deviation is added to the basic constant one to obtain the adaptive scaling multiplier used to adjust the gradient; the gradient components of the complete queue in the oscillation layer are multiplied by the adaptive scaling multiplier along the direction of the update vector to complete the numerical amplification and enhancement process of the complete queue.
[0126] The gradient components of the truncated queue in the oscillation layer are divided by the adaptive scaling multiplier to complete the numerical decay suppression processing of the truncated queue.
[0127] Extract the complete queue gradient components after numerical amplification and enhancement processing, and use them as the terminal gradients to characterize late-stage features;
[0128] It is understandable that the purpose of queue alignment and gradient stripping is:
[0129] Function 1: By utilizing the time-series truncation attribute state, which has natural control properties, mathematical gradient values are transformed into cross-cohort behavioral difference comparisons with clinical physical significance. By identifying the update misalignment phenomenon in deep networks caused by the lack of late-stage data in the truncation cohort, key parameters representing the evolution of late-stage diseases can be accurately located.
[0130] Function 2: It distinguishes the initial gradient from the final gradient in physical and network space, and adjusts the numerical multiplication and division differently through adaptive scaling multipliers, which helps to reduce the mutual dilution of early common features and late rare mutations during global model weight aggregation and update.
[0131] S40: Construct asymmetric intervention rules based on the evolutionary differences between the initial gradient and the final gradient; generate an adaptive aggregation scheduling strategy to correct time truncation bias based on the asymmetric intervention rules.
[0132] The process of constructing asymmetric intervention rules based on the evolutionary differences between the initial and final gradients is as follows:
[0133] Tracing back to the location of the initial gradient in the stable layer network, the network weight parameters corresponding to the stable layer are divided into early parameter blocks for processing consensus features;
[0134] By tracing the location of the oscillating layer network where the terminal gradient is located, the network weight parameters corresponding to the oscillating layer are divided into late parameter blocks for handling abrupt changes.
[0135] For early parameter blocks, establish basic allocation rules based on data abundance;
[0136] Preferably, the basic allocation rule is set as follows: extract the total number of medical record samples contained in the local queues of all sub-centers, divide the number of medical record samples in a single sub-center by the total number of medical record samples, and obtain the basic capacity weight corresponding to each sub-center.
[0137] For late-stage parameter blocks, a deep purification rule is set to cut off noise pollution.
[0138] Preferably, the method for setting the deep purification rules is as follows: extract the completeness scores of each sub-center in the complete queue and sum them to obtain the total completeness benchmark; divide the completeness score of a single sub-center in the complete queue by the total completeness benchmark to obtain the purification weight corresponding to the sub-center in the complete queue.
[0139] At the same time, the purification weight of all sub-centers in the truncated queue will be forcibly set to zero.
[0140] Logically bind the basic allocation rules based on data abundance with the deep purification rules, and define them uniformly as asymmetric intervention rules.
[0141] The process of generating an adaptive aggregation scheduling strategy to correct time truncation bias based on asymmetric intervention rules is as follows:
[0142] At the global server of federated learning, asymmetric intervention rules are transformed into weight evolution instructions applicable to different network depths;
[0143] Specifically, for the shallow layers of the network corresponding to the early parameter blocks, basic capacity weights are assigned as the aggregation benchmark.
[0144] For the network depth corresponding to the late parameter block, purification weights are assigned as the aggregation endpoint;
[0145] The weight evolution instructions for different network depths are encapsulated into a complete computation instruction chain, and the complete computation instruction chain is defined as an adaptive aggregation scheduling strategy.
[0146] It is understandable that the purpose of constructing asymmetric intervention rules and adaptive aggregation scheduling strategies is to:
[0147] Breaking away from the traditional federated learning mindset of treating a single model as a whole and aggregating the mean, this approach establishes a parameter decoupling strategy based on the time dimension. This provides a rule-based approach calibrated by objective data for the subsequent differentiated blocking and truncation of contamination gradients at the physical network level.
[0148] S50: Based on the adaptive aggregation scheduling strategy, the local model parameters are processed by relay fusion with stage-by-stage damping control to obtain the global shallow and deep feature matrices and perform cross-time period network alignment to output a multi-center decision model with evolution prediction capabilities.
[0149] The relay fusion process, which involves performing stage-by-stage damping control on local model parameters based on an adaptive aggregation scheduling strategy, is as follows:
[0150] On the global server of federated learning, local model parameters uploaded by all sub-centers are received and the local model parameters are decomposed into shallow matrices corresponding to early parameter blocks and deep matrices corresponding to late parameter blocks.
[0151] In the first fusion stage of processing the shallow matrix, the basic capacity weight in the adaptive aggregation scheduling strategy is applied. The shallow matrices uploaded by all sub-centers are multiplied by their corresponding basic capacity weights and then summed to obtain the global shallow feature matrix that fuses the full early medical record features of multiple centers.
[0152] In the second fusion stage of processing deep matrices, the stage-by-stage damping control mechanism is activated.
[0153] Preferably, the activation method of the phased damping control mechanism is as follows: starting from the global shallow feature matrix, the deep matrix is traversed layer by layer in the order of deepening of the network layers;
[0154] Calculate the deep network layer number, divide it by the total number of layers in the deep network to obtain the depth evolution factor, and subtract the depth evolution factor from the constant to obtain the actual attenuation damping coefficient.
[0155] As the network depth increases, a decreasing attenuation damping coefficient is applied to the truncated queue layer by layer. The basic capacity weight of the truncated queue in the corresponding deep network is multiplied by the attenuation damping coefficient to continuously weaken its aggregate weight, until the aggregate weight of the truncated queue is attenuated to the value of zero specified by the purification weight in the deepest network that represents the late mutation characteristics.
[0156] For the complete queue, reverse compensation weighting with the same magnitude is applied according to the total attenuation of the damping coefficient until the aggregation weight of the complete queue is increased to the corresponding purification weight in the deepest network.
[0157] Based on the dynamic weights adjusted by attenuation damping and reverse compensation, the deep matrices uploaded by all sub-centers are weighted and summed to complete the relay fusion and output the global deep feature matrix.
[0158] The preferred method for applying anti-compensation weighting is as follows:
[0159] S501. Calculate the total reduction of basic capacity weight values of all sub-centers in the truncated queue due to the application of attenuation damping coefficient at the current deep network level, and use the total reduction of basic capacity weight values as the weight pool to be compensated.
[0160] S502. Extract the pre-determined purification weight of each sub-center in the complete queue, calculate the proportion of the purification weight of a single sub-center in the total purification weight of the complete queue, and obtain the compensation allocation coefficient for a single sub-center.
[0161] S503. Multiply the value of the weight pool to be compensated by the compensation allocation coefficient of each sub-center to obtain the weight compensation increment that each sub-center in the complete queue should obtain under the current deep network level.
[0162] S504. Add the basic capacity weight of each sub-center in the complete queue to the corresponding weight compensation increment to complete the reverse compensation weighting operation for the complete queue and obtain the dynamic weight of the complete queue at the current deep network level.
[0163] The weighted summation method for the deep matrices uploaded by all sub-centers is as follows:
[0164] S511. The basic capacity weight of each sub-center in the truncated queue after being weakened by the damping coefficient s is used as the dynamic weight of the truncated queue at the current deep network level.
[0165] S512. Based on the dynamic weights of all sub-centers, multiply the deep matrix uploaded by the truncated queue by the dynamic weights corresponding to the truncated queue to generate a truncated feature weighted matrix, and multiply the deep matrix uploaded by the complete queue by the dynamic weights corresponding to the complete queue to generate a complete feature weighted matrix.
[0166] S513. Perform element-level numerical summation on the truncated feature weighted matrix and the complete feature weighted matrix to complete the weighted summation operation. The matrix after numerical summation is used as the global deep feature matrix representing the current deep network layer relay fusion result.
[0167] The process of integrating various global parameters to achieve cross-time period network alignment and outputting an evolution prediction capability multicenter decision model is as follows:
[0168] Extract the global shallow feature matrix and the global deep feature matrix after the relay fusion is completed;
[0169] The global shallow feature matrix and the global deep feature matrix are concatenated in terms of network depth and then aligned in the network.
[0170] To eliminate numerical gaps between shallow and deep layers caused by dynamic weight allocation, a weight smoothing operation is performed on the concatenated network layers to achieve cross-time period network alignment. This smoothing operation involves introducing a fixed small transition distance parameter (e.g., 0.2) at the junction of the shallow network end and the deep network beginning. Within this junction region, the global weights applied to adjacent network layers are numerically smoothed using linear interpolation. This ensures a smooth transition from the static basic capacity weights of the shallow layers to the dynamic decay weights of the deep layers, preventing feature gaps caused by abrupt changes in weights between adjacent feature extraction layers within the same multi-center model.
[0171] The complete network parameters after network alignment are overlaid with the historical parameters of the global server of federated learning to output the final multi-center decision model, and the multi-center decision model is distributed to each sub-center for decision prediction.
[0172] It is understandable that the role of damping relay fusion and network alignment is:
[0173] Function 1: Through dynamically increasing damping control, the influence of the truncated queue (new hospitals lacking late-stage data) gradually diminishes as the network deepens, while the influence of the complete queue (old hospitals containing late-stage data) smoothly takes over, reducing network structure damage caused by discarding parameters.
[0174] Function 2: Reduces the adverse dilution effect of short-cycle truncated data on long-term disease deterioration characteristic parameters, outputs a high-performance multi-center decision model that includes early consensus across the entire network and has pure late-stage prediction capabilities, and achieves anti-biased federated learning.
[0175] Example 3:
[0176] Please see Figure 3 As shown, the multi-center medical record data decision system based on federated learning includes the following modules:
[0177] Feature extraction module: used to extract the follow-up time span and missing data points in multicenter medical record data, and integrate them to establish observational distribution features; based on the observational distribution features, evaluate the temporal coverage integrity of multicenter medical record data in the disease evolution cycle, and obtain the local cohort completeness of the corresponding medical record data;
[0178] Attribute classification module: used to organize the local queue completeness of all participating centers and extract the relative difference distribution between sub-centers on the follow-up time axis; divide the sub-centers into truncated queues according to the relative difference distribution and mark the temporal truncated attribute state of each sub-center;
[0179] Evolutionary analysis module: Used to receive local model parameters obtained by each sub-center based on local queue training, measure the update dispersion of local model parameters according to network level and divide the stable layer and oscillating layer; combine the temporal truncation attribute state to compare the feature evolution differences between the complete queue and the truncation queue, and obtain the initial gradient and terminal gradient;
[0180] Strategy construction module: Constructs asymmetric intervention rules based on the evolutionary differences between the initial gradient and the final gradient; Generates an adaptive aggregation scheduling strategy to correct time truncation bias based on the asymmetric intervention rules;
[0181] Parameter fusion module: It is used to perform relay fusion processing of local model parameters by stage-by-stage damping control according to the adaptive aggregation scheduling strategy, to obtain global shallow and deep feature matrices and perform cross-time period network alignment, and output a multi-center decision model with evolution prediction capabilities.
[0182] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.
Claims
1. A multi-center medical record data decision-making method based on federated learning, characterized in that, Includes the following steps: The follow-up time span and missing data points in multicenter medical record data are extracted and integrated to establish observational distribution characteristics. Based on the observational distribution characteristics, the temporal coverage integrity of multicenter medical record data in the disease evolution cycle is evaluated to obtain the local cohort completeness of the corresponding medical record data. The local queue completeness of all participating centers is sorted out, and the relative difference distribution between sub-centers on the follow-up time axis is extracted; the sub-centers are divided into truncated queues according to the relative difference distribution, and the temporal truncation attribute status of each sub-center is marked. Receive local model parameters obtained from training based on local queues at each subcenter, measure the update dispersion of local model parameters according to network layers and divide them into stable layers and oscillating layers; combine the temporal truncation attribute state to compare the feature evolution differences between the complete queue and the truncation queue to obtain the initial gradient and the terminal gradient; An asymmetric intervention rule is constructed based on the evolutionary difference between the initial gradient and the final gradient; an adaptive aggregation scheduling strategy is generated based on the asymmetric intervention rule to correct the time truncation bias.
2. The multi-center medical record data decision method based on federated learning according to claim 1, characterized in that: The method for assessing the time coverage integrity is as follows: Read the standard time span reference value issued by the federated learning global server, divide the average time span value in the observed distribution characteristics by the standard time span reference value to obtain the basic coverage ratio; Based on the preset full score coefficient, a penalty adjustment coefficient is established according to the magnitude of the missing density in the observed distribution characteristics; The time coverage integrity index is calculated by multiplying the basic coverage ratio by the penalty adjustment coefficient. The time coverage integrity index of each center is mapped to a pre-set percentage evaluation scale to obtain the local cohort completeness.
3. The multi-center medical record data decision method based on federated learning according to claim 1, characterized in that: The method for marking the timing truncation attribute state is as follows: The completeness scores of each participating center are evaluated, and each sub-center is divided into a complete queue and a truncated queue. The absolute value of the truncated deviation of each sub-center is quantified based on the score differences. The classification labels of the corresponding queues are bound to the absolute value of the truncation deviation to generate unique time-series truncation attribute states for each sub-center.
4. The multi-center medical record data decision method based on federated learning according to claim 3, characterized in that: The method for extracting the cliff points of the sequence is as follows: Read the completeness scores uploaded by all participating centers received by the global federated learning server; Completeness scores are sorted in descending order of numerical value to construct a global score sequence; Extract the global score range of the global score sequence, and calculate the numerical difference between the completeness scores of every two adjacent sorting positions in the global score sequence. All numerical differences are concatenated according to their original sorting positions to generate adjacent decay gradient sequences. The adjacent decay gradient sequences are traversed, and the sorting node containing the largest decay gradient is located as the sequence cliff point.
5. The multi-center medical record data decision method based on federated learning according to claim 1, characterized in that: The method for comparing the differences in feature evolution is as follows: The historical change records of local model parameters are traced to evaluate the update dispersion of each network layer. Based on the update dispersion, a dynamic segmentation boundary is established, and the network layers are divided into stationary layers and oscillating layers. Based on the temporal truncation attribute state of each sub-center, the local model parameters are grouped into a complete queue parameter set and a truncation queue parameter set; In the steady layer, the similarity of parameter updates between the full queue and the truncated queue is evaluated, and the parameter update vector reflecting the early consensus characteristics is extracted as the starting gradient. In the oscillation layer, the parameter update divergence between the full queue and the truncated queue is evaluated, and the gradient is adjusted numerically based on the degree of truncation deviation. The enhanced full queue feature components are extracted as the terminal gradient.
6. The multi-center medical record data decision method based on federated learning according to claim 5, characterized in that: The process of extracting the complete queue feature components is as follows: Calculate the parameter mean vector of the full queue and the parameter mean vector of the truncated queue respectively. Subtract the parameter mean vector of the truncated queue from the parameter mean vector of the full queue to obtain the update direction deviation vector. Read the absolute value of the truncated deviation of the truncated queue to construct the adaptive scaling multiplier. The complete queue is numerically amplified and enhanced using an adaptive scaling multiplier, and the gradient components of the amplified and enhanced complete queue are extracted as the terminal gradient.
7. The multi-center medical record data decision method based on federated learning according to claim 1, characterized in that: Based on the adaptive aggregation scheduling strategy, the local model parameters are processed by relay fusion with stage-by-stage damping control to obtain the global shallow and deep feature matrices. The network is then aligned across time periods to output a multi-center decision model with evolution prediction capabilities.
8. The multi-center medical record data decision method based on federated learning according to claim 7, characterized in that: The relay fusion process is performed as follows: The local model parameters are divided into shallow parameter blocks corresponding to early features and deep parameter blocks corresponding to late features according to the network layers. In the first fusion stage, the shallow parameter blocks are statically weighted and fused based on the basic data capacity of each sub-center to generate a global shallow feature matrix. In the second fusion phase, a dynamic damping control mechanism related to network layer depth is activated. As the network depth increases, the aggregation weight of the truncated queue is gradually and smoothly decayed, and reverse compensation weighting is performed on the complete queue. Based on the dynamic weights adjusted by the dynamic damping control mechanism, the deep matrices uploaded by all sub-centers are weighted and summed to output the global deep feature matrix.
9. The multi-center medical record data decision method based on federated learning according to claim 8, characterized in that: The corresponding reverse compensation weighting method is as follows: Obtain the sum of the base capacity weights that are reduced due to the application of the attenuation damping coefficient for all subcenters in the truncated queue; Extract the pre-determined purification weights of each sub-center in the complete queue, calculate the proportion of the purification weight of a single sub-center in the total purification weight of the complete queue, and obtain the compensation allocation coefficient. Multiply the sum of the basic capacity weight values by the compensation allocation coefficient of each sub-center to obtain the weight compensation increment that each sub-center in the complete queue should obtain at the current deep network level. The basic capacity weight of each sub-center in the complete queue is added together with the corresponding weight compensation increment to obtain the dynamic weight of the complete queue.
10. A multi-center medical record data decision system based on federated learning, used to implement the multi-center medical record data decision method based on federated learning as described in any one of claims 1-9, characterized in that, Includes the following modules: Feature extraction module: used to extract the follow-up time span and missing data points in multi-center medical record data, and integrate them to establish observation distribution features; based on the observation distribution features, evaluate the time coverage integrity of multi-center medical record data, and obtain the local cohort completeness of the corresponding medical record data; Attribute classification module: used to organize the local queue completeness of all participating centers and extract the relative difference distribution between sub-centers; divide the sub-centers into truncated queues according to the relative difference distribution and mark the temporal truncated attribute state of each sub-center; Evolutionary analysis module: used to receive local model parameters from each sub-center, measure and update the dispersion according to the network level and divide the stable layer and oscillating layer; combined with the temporal truncation attribute state, compare the characteristic evolution differences between the complete queue and the truncation queue to obtain the initial gradient and the terminal gradient; Strategy construction module: Constructs asymmetric intervention rules based on the evolutionary differences between the initial and final gradients; generates an adaptive aggregation scheduling strategy based on the asymmetric intervention rules; Parameter fusion module: It is used to perform relay fusion processing on local model parameters according to the adaptive aggregation scheduling strategy, obtain global deep and shallow feature matrices, perform network alignment, and output a multi-center decision model.