Multimodal dynamic diabetes therapy evaluation method

Abstract

The application discloses a multi-modal diabetes dynamic treatment effect evaluation method, relates to the technical field of data processing, and unifies a time axis and marks reliability for multi-source data through an adaptive interpolation and alignment strategy; peak values and abnormal fluctuations within one to two weeks are detected and recognized based on a short-term window, a short-term change amount and a risk prompt are generated; a multi-layer attention or a Transformer is used to fuse data on a monthly scale, comprehensive long-term treatment effect indexes and key inflection points are extracted; short-term quantitative results and long-term indexes are brought into a multi-target reward mechanism through a continuous learning and interactive real-time feedback module, model parameters are iterated online, and clinical intervention suggestions are output; external events are quantitatively marked, and identification and response to sudden situations are strengthened through a bias injection mode. High-quality integration of multi-source heterogeneous data, short-term and long-term multi-level analysis are realized, and the precision and timeliness of diabetes dynamic monitoring and intervention are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-modal diabetes dynamic treatment evaluation method. BACKGROUND

[0002] In the clinical management of diabetes, both doctors and patients need to regularly collect and observe multi-source data such as blood glucose meters, continuous glucose monitoring devices (CGM), exercise records, electronic medical records, and diet logs. They also often pay attention to the subtle trends of patients' conditions over a period of time. For example, for type 2 diabetes patients, doctors may continue to track blood glucose stability and indicators such as weight and blood pressure for several weeks after starting an intensive insulin treatment plan to determine whether the treatment plan needs to be adjusted in a timely manner. If the annual eye examination of the patient indicates mild retinopathy, the patient's condition needs to be monitored for deterioration or improvement in the following months to take effective measures as soon as possible at an early intervention opportunity. In addition, modern diabetes management can identify abnormal blood glucose fluctuations through remote monitoring systems within a week or even a few days, automatically alert doctors or patients, and immediately analyze the potential causes of fluctuations. Due to sudden changes in patients' lifestyles, psychological states, and environmental factors, such as major surgeries, seasonal changes, or severe economic shocks, the patient's condition may also fluctuate. If the system lacks modeling and attention to these external environmental shocks, it may not be able to detect serious risks in time during a sudden situation, delaying the clinical opportunity for effective intervention.

[0003] However, most academic or commercial platforms developed based on existing information technology tools often stop at interpreting instantaneous static indicators or rough stage evaluation, lacking sufficient modeling capabilities for dynamic and continuous time series evolution, especially in real-world situations where monitoring frequency is unstable, measurement noise is large, and sampling periods of different indicators vary. Typical difficulties include: if only single test data or annual screening results are used, it is difficult to identify the potential trend of the patient's gradual deterioration over several weeks or months. Once the next outpatient or examination is awaited, the condition may have already progressed. If there is a lack of fusion and time-dependent processing capabilities for multi-modal data (such as HbA1c, blood glucose curve, exercise log, and psychological assessment) at the algorithm level, it is difficult to distinguish sudden deterioration caused by lifestyle fluctuations or external events (such as hospitalization surgery), leading to reliance on subjective experience in clinical diagnosis. If new data cannot be continuously learned and adaptively fed back, it is easy to give inaccurate or even dangerous recommendations when facing individual differences or sudden abnormal factors of patients.

[0004] Therefore, the present application provides a multi-modal diabetes dynamic treatment evaluation method. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies of the prior art, the present application provides a multi-modal diabetes dynamic efficacy evaluation method, which unifies the time axis of multi-source data and marks the reliability through an adaptive interpolation and alignment strategy; based on short-term window detection, peak values and abnormal fluctuations within one to two weeks are identified, and short-term change quantity and risk prompts are generated; multi-layer attention or Transformer technology is used to integrate data on a monthly scale, to extract comprehensive long-term efficacy indicators and key inflection points; through continuous learning and an interactive real-time feedback module, short-term quantitative results and long-term indicators are incorporated into a multi-objective reward mechanism, model parameters are iterated online, and clinical intervention suggestions are output, significantly improving the accuracy and timeliness of diabetes dynamic monitoring and intervention, thereby solving the technical problems in the background art.

[0007] (ii) Technical solutions

[0008] To achieve the above object, the present application is implemented by the following technical solutions:

[0009] The multi-modal diabetes dynamic efficacy evaluation method comprises: when real-time or batch arrival of multi-source data such as blood glucose meters, continuous glucose monitoring devices, exercise records, electronic medical records, and lifestyle logs is detected, an adaptive time alignment and interpolation algorithm is called, and the reliability Γ n (t) is calculated, noise correction is carried out, non-regular sampling points are processed in a dynamic sliding window manner, suspicious measurement values are weighted and corrected, the standardized time series matrix M is accurately constructed and significant error is suppressed, and all corrected items are recorded with time stamps, finally outputting standardized multi-modal data for calling by subsequent steps;

[0010] After receiving the standardized time series matrix M and the reliability Γ n (t), a short-term window analysis process is triggered, the short-term change quantity Δ n,k is used to represent the fluctuation amplitude within one to two weeks, a hybrid convolution-cyclic network is used to lock local peak values and mutation regions, abnormal events are labeled and mapped to the corresponding time period, noise interference is filtered out by using a weighted feature vector, and a short-term fluctuation comprehensive measure is output based on the window aggregation result;

[0011] After recording the short-term change matrix D and the abnormal label, a deep fusion algorithm is started, multi-layer attention mechanism or Transformer-based network such as HbA1c, body weight, and blood glucose fluctuation is integrated, and the short-term high-risk segment is given an additional bias, the generated long-term efficacy indicator Λ k takes into account local mutations and global trends, and identifies key inflection points based on interactive attention weights, and finally presents a comprehensive trend on a monthly dimension through multi-dimensional output;

[0012] The long-term efficacy indicator Λk with short-term variation amount Δ n,k After that, the two are fused into a multi-objective reward function through reinforcement learning or online parameter correction, and the Q function or policy network is iterated when new data arrives. If significant deterioration of the therapeutic effect is detected, an intervention suggestion is immediately pushed. If subsequent clinical verification shows that the intervention is effective, the current policy weight is enhanced;

[0013] After completing multi-modal data fusion and continuous learning, external events such as psychological stress, economic changes, seasonal changes, and surgery are collected and modeled. By adding additional dimensions to the standardized time series matrix M and assigning adaptive weights, sudden fluctuations are given more bias in the short-term detection stage, and high-intensity events are given priority focus in the long-term attention module, and the intervention strategy is dynamically adjusted;

[0014] To eliminate the uneven sampling of multi-source data and the misalignment of measurement periods, a dynamic time alignment and interpolation method based on kernel functions is constructed. For each data source A mapping Λ j is constructed to accurately project the measurement values of the data source S j onto the global time axis For any time point , an interpolation function I j (t k ) is defined

[0015] For each data source , the interpolation function I j (t k ) is calculated point by point on the global time series to obtain a new aligned sequence {I j (t1), I j (2), …, I j (t L )}; all aligned sequences of the data sources are combined into a standardized time series matrix M;

[0016] Preferably, an adaptive correction and credibility marking mechanism is introduced to form the time series data; the residual is filtered through a nonlinear function; the corrected time series value is output instead of the original interpolation I j (t k ), and the corresponding position in the standardized time series matrix M is updated: after the correction, a credibility coefficient Γ j (t k ) is assigned to each item M[j,k];

[0017] If the comprehensive penalty term exceeds 1, Γ j (t k) = 0, representing that this position is basically not credible; if the penalty is small, the credibility coefficient Γ j (t k ) is close to 1, the closer to 1, the smaller the correction amplitude, the higher the interpolation weight, and the better the reliability of the original measurement; otherwise, it indicates a suspicious value, prompting the subsequent module to reduce the weight or eliminate it when determining, and generating an accompanying correction value after processing and the credibility Γ j (t k ) of the standardized time sequence matrix M;

[0018] Preferably, based on the standardized time sequence matrix M, the data in the short-term range is segmented and processed:

[0019] Let the global time index set of the standardized time sequence matrix M still be represented as {t1, t2, …, t L}, corresponding to L discrete sampling points; select a short-term window length l (2) , let each window contain l (2) consecutive sampling time points; generate a set of short-term windows according to a given step on the time axis; for each window cut off the corresponding row and column subsets in the standardized time sequence matrix M retain its corresponding credibility Γ n (t); exclude or weaken the time points with low credibility Γ n (t);

[0020] Preferably, after completing the short-term window segmentation, deep local feature extraction is performed on the data in the window to find out possible peak values, slope mutations or high-frequency fluctuation positions; for each window W k and each data source n, the network finally outputs a local feature vector f n,k to represent the main change pattern, peak and valley distribution and possible slope jump in the window with a fixed length;

[0021] Preferably, after obtaining the local feature vector f n,k , in order to map the local feature vector f n,k to the real value quantization result, a adjustable function Ω is defined to calculate the short-term change amount Δ n,k for each data source n and window , and condense the local feature vector f n,k into a one-dimensional numerical value;

[0022] A dynamic threshold curve τ (2) (f n,k ) is introduced, and if the short-term change amount Δ n ,k>τ (2) (f n,k), then mark the corresponding time period as abnormal and record its fluctuation characteristics associated with the time index of the standardized time series matrix M;

[0023] The short-term change matrix D is formed on each data source and window dimension, and the abnormal events are summarized as a short-term abnormal event list ε, where each abnormal event contains the triggering time interval, the data source identifier involved, and the fluctuation characteristic description.

[0024] Preferably, the standardized time series matrix M is segmented and aggregated, and several windows are uniformly divided on {t1,…,t L} ;

[0025] In combination with the short-term abnormal event list ε and the short-term change matrix D, they are mapped to the long window v k according to the corresponding time range;

[0026] For each long window , different indicators are aggregated to form sub-sequences; these sub-sequences are multi-modal split at the bottom layer; the short-term abnormal event list ε is injected at the middle layer, and if the comprehensive evaluation value D[n,k] of the short-term change amount shows that an indicator fluctuates violently within the time period, the corresponding abnormal mark or additional high-weight label is retained in the sequence representation of the indicator; The long window is combined with each indicator n to form a multi-modal long-term feature container H, where H[n,k] can be regarded as an aggregation of or its sequence representation at the bottom layer; at the same time, the local fluctuation data of the second step are also stored in the accompanying abnormal annotation;

[0027] Preferably, an attention mechanism is introduced to strengthen the attention to key periods and indicators. In a layer of attention mechanism, for the long window feature matrix H k corresponding to the kth long window, the attention output A k is defined;

[0028] An additional bias δ (3) >0 is applied to the vector dimension containing the abnormal mark in the feature mapping stage, that is:

[0029] If a certain indicator or time period in the long window feature matrix H k has been marked as high fluctuation, the corresponding mapping can be superimposed with a bias; a single-layer or multi-layer Transformer encoder is used for deep fusion, and a global association is established between different indicators and long time periods through multi-head attention or Transformer;

[0030] Preferably, for each long window , after multi-layer Transformer or multi-head attention encoding, the final representation is obtained. Perform global pooling or aggregation to obtain a fixed vector g representing the deep fusion features within the k-th long window, which are multimodal information. k , fix the vector g k Projected onto the real number domain, the phased efficacy score Λ of this long window is output. k Connect or visualize them in chronological order;

[0031] Constructing the difference vector ΔΛ for identifying key moments k And, combined with a pre-set difference threshold or dynamic boundary, the difference vector ΔΛ k The sign and absolute magnitude are detected by the difference vector ΔΛ. k A significant, abrupt change occurring between two consecutive windows is marked as a trend inflection point; the corresponding value of the inflection point is also recorded. Start and end times and contributions of key indicators;

[0032] After completing all long windows After processing, the final output is: a list of long-term efficacy indicators, recording the trend inflection points of the time segments in the efficacy curve where there is a significant increase or decrease.

[0033] Preferably, newly collected measurements are recorded in the standardized time series matrix M to generate new correction values ​​and corresponding confidence levels, and then inserted into the corresponding positions in the standardized time series matrix M; if a new observation point falls within the existing short-term window, the corresponding short-term change is updated.

[0034] If a new short-term time interval is reached, a new short-term window can be added, and the hybrid convolutional-recurrent network can be quickly inferred to identify whether new anomalous events have occurred.

[0035] For the defined long window on a long time scale If new data falls into it, attention or incremental inference of the Transformer can be triggered in a targeted manner to update the deep feature representation of the corresponding window;

[0036] The updated short-term change matrix D patch, the long-term efficacy indicator increments, and the list of possible new short-term anomalies ε or trend inflection points will be used. Information such as these is cached and indexed and relocated in the data dictionary;

[0037] Preferably, a reinforcement learning agent is set up to continuously interact with patient data and clinical decision-making processes;

[0038] Select the change items in the short-term change matrix D and the long-term efficacy indicator Λ k The sequence is used as the state vector s. If higher-dimensional information is needed, short-term anomaly lists and trend inflection point lists can also be used. The tags are mapped to a specific dimension;

[0039] Several intervention or correction actions can be defined, or online correction operations can be performed on model parameters. To balance short-term effects and long-term benefits, a multi-objective reward function can be defined.

[0040] The reinforcement learning agent receives a new state vector s t Perform an action a t Afterwards, receive an instant reward R. t And enter the next state s t+1 And iteratively update the parameters of the Q-function or policy network;

[0041] If certain actions are detected to repeatedly lead to negative results, the strategy update will gradually reduce the probability of selecting that action; conversely, it will increase the probability of selecting that action. This process is combined with actual clinical interventions, which can continuously optimize individualized diabetes management strategies while ensuring safety.

[0042] Preferably, based on the latest results of short-term efficacy changes and long-term efficacy indicators, if the reinforcement learning agent determines that the risk of short-term deterioration exceeds a certain warning threshold, it will generate an early warning message and push it to the clinical end: if within a long window... Intermediate-stage efficacy score Λ k A prolonged downturn or a sudden downward turning point can also trigger an alarm.

[0043] Clinically, the generated action suggestions or medication adjustment prompts can be used to make decisions. If doctors or patients actually adopt and implement them, this serves as real feedback for reinforcement learning, and the policy network or Q function can be continuously improved through iterative updates.

[0044] After doctors adjust medication, diet, or lifestyle based on early warnings, the subsequent short-term changes Δ(t+1) and long-term evolution Λ(t+1) are fed back to the management platform in real time; if several pre-validations are successful, a reward R is given. t+1 As a result, the positive learning path of the reinforcement learning agent is strengthened: if the treatment is ineffective or the situation worsens, the corresponding action is recorded as a low-return action, thereby reducing the probability of it being selected in subsequent updates.

[0045] (III) Beneficial Effects

[0046] This invention provides a multimodal dynamic efficacy evaluation method for diabetes, which has the following beneficial effects:

[0047] The adaptive interpolation and alignment algorithm adopted can establish a unified time axis among multi-source data such as blood glucose meters, continuous glucose monitoring devices, exercise records and electronic medical records, and correct and mark the reliability of the measurement values with low quality or noise, to provide high-quality input for subsequent modeling, and the additional dimension expansion of external events is combined to make the multi-source data and external impact factors be completely characterized in the same time sequence framework.

[0048] The short-term window detection and multi-layer attention fusion can form complementarity among short-term efficacy changes: the former can sensitively capture the peak and slope mutation within one to two weeks, and the latter can integrate multi-modal trajectories such as HbA1c, blood glucose fluctuation and weight trend on the scale of several months, both of which can early warn of dramatic abnormalities and track the overall trend, organically combining fine-grained analysis and overall health management.

[0049] The continuous learning and interactive real-time feedback mechanism incorporates short-term efficacy changes and long-term efficacy indicators into the reward function through reinforcement learning or online updating, adaptively balances short-term stability and long-term optimization, and significantly improves the flexibility of personalized decision-making: when the model identifies a high-risk situation, it timely pushes intervention suggestions to doctors and patients, and corrects the strategy according to subsequent efficacy measurements, thereby realizing learning and using, and using and optimizing at the same time.

[0050] In summary, in the multi-source data preprocessing, short-term and long-term analysis, continuous learning and external event dimension expansion, from adaptive interpolation, reliability marking, deep network fusion to event bias and multi-objective reward coordination, an efficient closed loop for the whole process of diabetes efficacy management is constructed, which greatly improves the precision of perception and active intervention ability for complex conditions and changing environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 Figure 1 is a flowchart of the multi-modal diabetes dynamic efficacy evaluation method of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Please refer to Figure 1 The present application provides a multi-modal diabetes dynamic efficacy evaluation method, which comprises,

[0054] Step one, when detecting real-time or batch arrival of multi-source data such as blood glucose meter, continuous glucose monitoring device, exercise record, electronic medical record and lifestyle log, call adaptive time alignment and interpolation algorithm, and according to the credibility Γ n (t) Develop noise correction, process irregular sampling points in a dynamic sliding window manner, correct suspicious measurement values by weighting, realize accurate construction of unified time axis of standardized time series matrix M and significant error suppression, and record the time stamp of all corrected items synchronously, finally output the standardized multi-modal data for subsequent steps to call;

[0055] The step one includes the following contents:

[0056] Step 101, time alignment and adaptive interpolation

[0057] In order to eliminate the uneven sampling and measurement time period dislocation of multi-source data, a dynamic time alignment and interpolation method based on kernel function is constructed, and its main technical features are as follows:

[0058] Let be the global time series with length L; for each data source (such as blood glucose meter, exercise sensor, etc.), its original measurement time set is denoted as

[0059]

[0060] Need to construct the mapping Λ j : So that the measurement value of data source S j can be accurately projected to the global time axis So as to realize the alignment processing of subsequent multi-source data;

[0061] For any time point Define the interpolation function I j (t k ):

[0062]

[0063] Where: V j (u) represents the observation value of data source At time u; Is a composite kernel function, which is used to measure the distance between t k And u of and give corresponding weight when interpolating; α (1) ,β (1) Kernel function parameters, the values are between 0 and 1, which can be adjusted according to the data sampling characteristics and expected smoothness;

[0064] In order to enhance the ability to capture sudden abnormalities, the composite kernel function It can be composed of multiple sub-nuclei weighting, such as:

[0065]

[0066] Where: exp(-α (1) |t k -u|) is responsible for overall smoothing, φ bumF The function can highlight abnormal peaks in a local range; ω1, ω2 are constant weights, taking values between 0 and 1, used to balance the smoothness and local sensitivity;

[0067] For each data source According to the interpolation function I j (t k ) described above, it is calculated point by point on the global time series , and the new aligned sequence is obtained:

[0068] {I j (t1),I j (t2),…,I j (t L )}

[0069] Combine the aligned sequences of all data sources into a standardized time series matrix M:

[0070]

[0071] Where N is the number of all data sources, and the standardized time series matrix M will be used as the basic data structure;

[0072] When used, through the definition of mapping Λ j And the interpolation mechanism based on the composite kernel function , the time alignment and interpolation of multi-source data are realized, which can obtain smoother and more continuous data sequences without ignoring local anomalies compared with simple linear interpolation;

[0073] The introduction of φ bump In the composite kernel function can sensitively detect short-term mutations in intervention scenarios, and has important support for subsequent short-term fluctuation detection and long-term trend analysis; The output standardized time series matrix M provides a consistent data carrier for subsequent noise correction and credibility evaluation, and also provides a unified reference coordinate for the analysis of the entire scheme.

[0074] Step 102, noise correction and credibility marking

[0075] After obtaining the standardized time series matrix M, in view of the possible noise and untrusted readings in the data, an adaptive correction and credibility marking mechanism is introduced to form the most favorable time series data for subsequent detection and evaluation;

[0076] Based on the interpolation residual Δj (t k )=|I j (t k )-V j (u * )|, where u * For t k The nearest original measurement time is used to filter the residuals using a nonlinear function, as follows:

[0077]

[0078] Wherein, Θ can be defined as a piecewise saturation or exponential suppression function, used to penalize outliers that are too large or too small; γ (1) The hyperparameter used to adjust the sensitivity of outlier detection is 0 < γ. (1) ≤1;

[0079] For example, the definition:

[0080]

[0081] e i This represents the i-th type of external event (such as diet, exercise, psychological stress, etc.);

[0082] t_time represents the duration of the event.

[0083] The α saturation value represents the maximum impact of an event on the assessment of treatment efficacy.

[0084] β is the decay factor, which controls the decay rate of the event's impact;

[0085] T is the threshold time, representing the maximum duration of the event's impact. After this time, the event's impact remains unchanged at α.

[0086] Output corrected timing values Replace the original interpolation I j (t k Update the corresponding position in the normalized time series matrix M:

[0087]

[0088] After the correction is completed, a confidence coefficient Γ is assigned to each entry M[j,k]. j (t k ), obtained through the following formula:

[0089] Construct local difference vector d i,k ,make:

[0090]

[0091] The length is 2w+1, which records the difference between the corrected value and the original measurement at ±w time points near the current index k. If the difference is large overall, it indicates that there is a persistent outlier or noise in the time series;

[0092] Constructing local gradient vector Let:

[0093]

[0094] The length is 2w, and if this gradient vector is large in the norm sense, it indicates that there is a sharp change in the sequence; in the credibility assessment scenario, it can be considered as a signal of potential noise or sudden abnormal increase.

[0095] To integrate the above vector information into the integrated measure, define:

[0096]

[0097] ||d j,k || p represents the local difference vector d j,k The size in the p-norm measures the overall deviation between the corrected value and the original observation within the same window;

[0098] represents the size of the gradient vector in the g-norm, which is used to quantify the local time series fluctuation amplitude; α (2) , β (2) : the penalty coefficients of the difference part and the gradient part respectively, both taking values (0, +∞);

[0099] When α (2) or β (2) ) becomes larger, the local difference or fluctuation will lower the credibility more strongly; when both values are small, a certain degree of noise or rapid change is allowed without significantly reducing Γ j (t k ); w is the local window radius, typically taking values in the range 1≤w≤5; p, q are vector norm types, which can be selected from {1, 2, ∞};

[0100] The final credibility coefficient Γ j (t k ) is forced to be limited in the range [0, 1]:

[0101] If the comprehensive penalty term exceeds 1, Γ j (t k ) = 0, which represents that this position is basically not credible; if the penalty is small, the credibility coefficient Γ j (t k) approaches 1, the closer to 1 indicates that the correction amplitude is smaller, the interpolation weight is higher, and the original measurement is more reliable; otherwise, it indicates a suspicious value, prompting the subsequent module to reduce the weight or eliminate it when judging,

[0102] After processing, a correction value is generated and a reliability Γ j (t k ) standardized time sequence matrix M;

[0103] In use, the measurement reliability is converted into a numerical reliability coefficient Γ j (t k ), which provides data quality prior for subsequent methods, can be flexibly weighted or selectively ignored when model inference and risk detection, directly uses interpolation residual information, can quickly locate noise concentration area, and enhances the pertinence of correction; lays a solid data foundation for subsequent dynamic detection and deep evaluation of efficacy; step 101 realizes the alignment and fine interpolation of multi-source data on a unified time coordinate, and retains the sensitivity to abnormal values; step 102 further eliminates noise and digitizes the data quality, so that the results not only have formal consistency, but also achieve reliability in quality.

[0104] Step two, receive standardized time sequence matrix M and reliability Γ n (t) after, trigger short-term window analysis process, short-term change Δ n,k represent the fluctuation amplitude in the past one to two weeks, and use the hybrid convolution-cyclic network to lock the local peak value and mutation area, add labels to the captured abnormal events and map them to the corresponding time period, combine the weighted feature vector to exclude noise interference, and output the short-term fluctuation comprehensive measure based on the window aggregation result;

[0105] The step two includes the following contents:

[0106] Step 201, short-term window segmentation and data preparation

[0107] Based on the standardized time sequence matrix M, the data in the short-term range is segmented for local feature extraction and anomaly detection in the subsequent steps. The specific logic is as follows:

[0108] Let the global time index set of the standardized time sequence matrix M be represented as {t1,t2,…,t L}, corresponding to L discrete sampling points; select a short-term window length l (2) (Units can be days or hours), let each window contain l (2) consecutive sampling time; in the common diabetes management scenario, l (2) can take a value of about 7, 14 days, to capture the fluctuations in the past one to two weeks, and the time axis is divided according to the given step (which can be the same as l(2) The same or partially overlapping) to generate a set of short-term windows Wherein These windows will be used for subsequent local operations;

[0109] For each window Cut out the corresponding row and column subsets in the standardized timing matrix M

[0110]

[0111] Retain its corresponding credibility Γ n (t) in order to dynamically weight or filter in subsequent models; filter out moments or weaken credibility Γ n (t) is too low, which can effectively reduce the interference of noise points on local model determination;

[0112] When in use, after the data is segmented into short-term windows for easy management, the model can focus more on capturing fluctuations and mutations in a small range, which is conducive to quickly identifying abnormalities; if the clinical demand pays more attention to subtle changes, a shorter step size can be selected; if stability is emphasized, it can be relaxed moderately.

[0113] Step 202, local feature extraction based on hybrid convolution-cyclic network

[0114] After completing the short-term window segmentation, deep local feature extraction is performed on the data in the window to find the possible peak, slope mutation or high-frequency fluctuation position, and an adjustable hyperparameter θ (2) and ζ (2) , respectively, control the convolution kernel size and GRU hidden state dimension;

[0115] For a short-term window W k extracted from the timing vector First stack in the channel direction to form a ζ (2) dimensional local representation; the convolution kernel size is controlled by θ (2) , and convolution operations are performed on consecutive observation points to extract important local patterns and feature slopes between adjacent time points;

[0116] This convolution process reduces the impact of noise and provides more concentrated feature vectors for subsequent GRU processing, and in addition, credibility Γ n (t) does mask or weight reduction operation on low credibility positions in the channel dimension;

[0117] Input the sequence output by convolution into the GRU layer to capture timing dependence within the window range;

[0118] GRU has the mechanism of update gate and reset gate, which can retain historical information while weakening irrelevant past readings, and improve the sensitivity of identifying trend mutations in short periods;

[0119] Adjustable hyperparameter ζ is added when GRU hidden state is transmitted (2) to determine the size of the hidden state dimension to adapt to the complexity of multi-modal input and distinguish different indicators and window positions; for each window W k and each data source n, the network finally outputs a local feature vector f n,k , denoted as:

[0120]

[0121] The vector represents the main change pattern, peak and valley distribution, and possible slope jump in the window with a fixed length;

[0122] When used, the convolutional network has strong capturing ability for local patterns and abnormal peaks, and GRU can further integrate the sequential relationship in the time dimension. The coupling of the two can improve the comprehensive representation ability in the short-term window. In the convolution and GRU stage, the mask operation can be used to weaken the data with very low credibility Γ n (t) and thus reduce the destructive effect of abnormal noise points;

[0123] The combination of credibility and deep network not only uses convolution-recurrent network to capture features, but also further combines the credibility Γ n (t) information output in the first step to weight or mask the network input, which has higher robustness and creativity in medical noise scenarios;

[0124] Each indicator data is taken as a different input channel in the same window to fully utilize the local co-occurrence features across indicators, such as whether blood glucose and exercise volume mutate at the same time in the same period.

[0125] Step 203, calculation of short-term efficacy change and marking of abnormal events

[0126] After obtaining the local feature vector f n,k , in order to map the local feature vector f n,k to the real value quantization result, an adjustable function Ω is defined for each data source n and window to calculate and obtain the short-term change Δ n,k :

[0127] Δ n,k =Ω(f n,k ;λ (2) )

[0128] where λ (2)An additional hyperparameter is introduced, with a value between 0 and 1, to control the nonlinearity and bias term of the adjustable function Ω;

[0129] The local eigenvector f can be transformed using an adjustable function Ω. n,k Condensed into a one-dimensional numerical value, it intuitively measures the extent of improvement or deterioration in efficacy within this window; the larger the value, the more significant the fluctuation in the short term.

[0130] Introducing a dynamic threshold curve τ (2) (f n,k ), by local eigenvector f n,k Together with global reference information (such as historical average change levels), it determines whether to distinguish between normal fluctuations and suspected abnormal fluctuations;

[0131] If the short-term change Δ n ,k>τ (2) (f n,k If an anomaly is detected, an anomaly marker will be added to the corresponding time period and its fluctuation characteristics (such as peak time, maximum slope, etc.) will be recorded; these markers will be associated with the time index of the normalized time series matrix M.

[0132] Through the above process, a short-term change matrix D can be formed across various data sources and window dimensions, with its element D[n,k] = Δ n,k That is, the comprehensive evaluation value of the nth data source in the kth window, and at the same time, the abnormal events are summarized into a short-term abnormal list ε, where each abnormal event includes the trigger time interval, the identifier of the data source involved, and a description of the fluctuation characteristics.

[0133] When using it, the local feature vector f is adjusted through the adjustable function Ω. n,k Mapped to an intuitively interpretable numerical value Δ n,k It can provide clear risk warnings or improvement indicators; it introduces a dynamic threshold curve τ. (2) (f n,k This approach, combining local features with historical benchmarks, avoids the limitations of a single fixed threshold while effectively distinguishing between normal fluctuations and truly anomalous sharp jumps. The dynamic threshold curve τ (2) (f n,k Compared to common static numerical thresholds, dynamic threshold curves τ generated based on feature vector adaptation or the ARIMA time series model offer advantages. (2) (f n,k It is more sensitive to local context and can flexibly adjust multidimensional anomaly records: it not only simply outputs anomaly markers, but also saves feature information such as peak time and maximum slope, which can provide data support for more refined intervention strategies in the future.

[0134] Step three, after recording the short-term change matrix D and the abnormality label, start the deep fusion algorithm to integrate long-term sequences such as HbA1c, weight, blood glucose fluctuation, etc. with multi-layer attention mechanism or Transformer-based network, and give additional bias to the short-term high-risk segment, and generate the long-term therapeutic effect index Λ k The step three includes the following contents:

[0135] The step three includes the following contents:

[0136] Step 301, hierarchical feature splitting of multi-modal long-term sequence

[0137] The standardized time sequence matrix M contains multiple modal indicators such as blood glucose, HbA1c, exercise data, and weight. The corrected value on the unified time axis {t1,…,t L} is The credibility Γ n (t) is

[0138] To realize the evaluation of a longer time scale (such as months), the standardized time sequence matrix M needs to be segmented and aggregated. Denote T (3) as the length of the long-time window (the unit can be weeks or months), and divide the {t1,…,t L} into several windows Each window has a length of T (3) discrete points (or moderately overlap according to clinical habits) for subsequent extraction of long-term features;

[0139] In addition, the short-term abnormal list ε and the short-term change matrix D also need to be mapped to the long window according to the corresponding time range, so as to give higher attention to these local abnormalities in the subsequent attention or Transformer model;

[0140] For each long window Different indicators are aggregated respectively to form sub-sequences

[0141] At the bottom, these sub-sequences are multi-modal split: such as glycated hemoglobin (HbA1c) has a low sampling frequency, updated every 3 months: while blood glucose or exercise amount may be measured multiple times a day: need to be adaptively processed (such as repeated mapping or interpolation) according to their original sampling characteristics;

[0142] Inject short-term anomaly list ε in the middle layer: if the comprehensive evaluation value D[n, k] of the short-term change amount shows that a certain index fluctuates violently in this time period, then keep the corresponding abnormal mark or add a high weight label in the sequence list corresponding to this index, so that the next step can identify the potential impact on the long-term trend.

[0143] Split each long window into a multi-modal long-term feature container H combined with each index n, where H[n, k] can be regarded as the aggregation of or its sequence representation at the bottom layer; at the same time, the local fluctuation data of the second step is also stored in the accompanying abnormal annotation, and the multi-modal long-term feature container H is taken as the output for deep fusion and attention analysis.

[0144] Where, after processing the multi-modal time series data (such as blood glucose value, exercise record, etc.) by the short-term window analysis method, the numerical information reflecting the data change in this time window is extracted. This includes the calculated short-term change amount, as well as the feature data of local peak value, fluctuation slope and mutation captured by the hybrid convolution-cyclic network, which are used to identify whether there is abnormal fluctuation in this window, thereby providing a basis for subsequent abnormal detection and efficacy evaluation. These data are local fluctuation data.

[0145] When used, the original standardized time series matrix MI is split along the long time window , and the short-term anomaly list ε and the short-term change matrix D are embedded, providing a unified data framework for subsequent construction of overall trend; different sampling characteristics are adapted and treated differently for different layers, ensuring the accuracy and flexibility of long-term trend analysis. Step 301 maps local anomalies to a larger time scale, enabling step 302 to focus more accurately on these short-term events that may affect the long-term trajectory.

[0146] Step 302, multi-layer attention mechanism or deep fusion based on Transformer

[0147] When processing multi-modal long-term sequences, attention mechanism is introduced to strengthen the attention to key periods and indexes, where Q, K, V represent query, key and value vectors or matrices respectively: when multi-head attention is used, there are multiple sets of Q h ,K h ,V h ; in a layer of attention mechanism, for the long window feature matrix H k (which can be regarded as a vector or sequence spliced with all indexes n) corresponding to the kth long window, the attention output A k is defined as:

[0148] ​

[0149] where μ (3) > 0 is a scale hyper-parameter to balance the dot product size; Q, K, V are then derived from H k by a learnable linear projection, e.g. where are the parameter matrices to be learned in this step.

[0150] To effectively utilize the short-term anomaly list ε, in the calculation of QK T , an additional bias δ (3) > 0 is imposed on the dimension of the vector containing the anomaly label in the feature mapping stage, i.e.

[0151] If a certain index or period in the long window feature matrix H k has been labeled as high volatility, it can add a bias when mapping in Q and K, so that it occupies a higher weight after Softmax normalization;

[0152] Specifically, an anomaly bias function Ψ(ε k ; δ (3) ) can be defined to describe:

[0153] Q' = Q + Ψ(ε k ; δ (3) ), K' = K + Ψ(ε k ; δ (3) )

[0154] Then calculate Q'K' T to obtain enhanced attention, so that more accurate focus can be achieved on long-term sequence fragments involving abnormal fluctuations;

[0155] Single or multi-layer Transformer encoder is used for deep fusion: multi-layer Transformer encoder can transfer attention weights across time periods and across indicators, thereby capturing complex patterns such as: HbA1c continuously rising within six months but blood glucose daily fluctuations abnormally increasing in a specific segment;

[0156] If multi-layer stacking is performed, the r-th layer can be represented as

[0157]

[0158] where: is the input initial value, is the output result of the R-th layer, and the final output is

[0159] In use, for the unified processing of multi-modal and long time series, through multi-head attention or Transformer, global correlation can be established between different indicators and long time periods, so that the model has the ability to identify cross-modal dependence and long time lag effect;

[0160] The marked short-term abnormal information is incorporated into the attention mechanism in a bias δ (3) way, so that key local events have a more significant impact on long-term assessment, avoiding being obscured by a large amount of smooth data in a long time range; the multi-layer stacking structure can refine features layer by layer, gradually focusing on long-term patterns of different complexity, thereby outputting more hierarchical and forward-looking feature representations;

[0161] Abnormal bias is incorporated into attention, which, compared with traditional attention mechanism based on sequence autocorrelation score calculation, integrates local fluctuation markers to dynamically bias the attention to short-term abnormalities and improve the attention to long-term impact; projection + attention processing of multiple indicators with different sampling frequencies and attributes under the Transformer architecture is beneficial to capture the interaction and potential coupling relationship across indicators.

[0162] Step 303, generating long-term efficacy indicators and marking key trend inflection points

[0163] For each long window After multi-layer Transformer or multi-head attention encoding, the final representation is obtained

[0164] The final representation is globally pooled or aggregated (such as taking the hidden state of the last position in the sequence or weighted sum with learnable weights) to obtain a fixed vector g representing the deep fusion features of multi-modal information in the kth long window k ;

[0165] Since each window is sequentially connected in time, a list of g1, g2, …, g p can be formed to identify the overall trend;

[0166] Let Φ be a learnable or adjustable mapping function, which projects the fixed vector g k to the real number domain and outputs the phase efficacy score Λ of the long window k :

[0167] Λ k = Φ(g k ; v (3) )

[0168] Where v (3) is an adjustable parameter or network structure, taking a value between 0 and 1, used to control the complexity and bias of the mapping layer;

[0169] After obtaining {Λ1, Λ2, …, Λ p}, it can be connected or visualized in time sequence, which can reflect the evolution trend of efficacy in the dimension of months;

[0170] To automatically identify the critical moment when the treatment changes from improvement to deterioration or vice versa, a differential vector ΔΛ k is constructed to identify the critical moment

[0171] ΔΛ k = Λ k+1 - Λ k

[0172] And in combination with the pre-set differential threshold or dynamic boundary, the positive and negative and absolute amplitude of the differential vector ΔΛ k is detected, and if the differential vector ΔΛ k has a significant cross change between two consecutive windows, it can be marked as a trend inflection point; At the same time, the start and end time corresponding to the inflection point and the main index contribution (which can be based on the attention weight in step 302 to find out which modal has the maximum contribution) are recorded;

[0173] After completing the processing of all long windows , the final output is:

[0174] Long-term efficacy indicators {Λ1,Λ2,…,Λ p} representing the overall efficacy level in each long window; Record the trend inflection point list of the time segment with significant rise or fall in the efficacy curve

[0175] In use, through the periodic efficacy score Λ k , the complex multi-modal long-term data is intuitively quantified, and the key turning point of the efficacy curve can be quickly located. Combined with attention weight and abnormal bias information, the cause of each inflection point can be explained in multiple dimensions (such as blood glucose, body weight, HbA1c, etc. Who dominates the trend change), providing transparent basis for clinical decision-making; The high-dimensional vector refined by attention or Transformer is mapped to a real efficacy indicator in this step, forming a complete chain from deep representation to visualized value, balancing the flexibility of deep learning and the ease of clinical application, not only giving the overall score, but also identifying significant turning points, making long-term evaluation have dynamic tracking and warning functions. k

[0176] Step four, update the long-term efficacy indicator Λ k and the short-term change Δ n,k , then fuse them into a multi-objective reward function through reinforcement learning or online parameter correction, and iterate the Q function or policy network when new data comes, if the efficacy is significantly deteriorated, the intervention suggestion is immediately pushed, if the subsequent clinical verification shows that the intervention is effective, the current strategy weight is enhanced;

[0177] The step four includes the following contents:​​

[0178] Step 401, real-time data access and model input update

[0179] Newly collected measurement values are recorded in the standardized time series matrix M with uniform timestamps and correction methods. When new readings of existing indicators (such as blood glucose, HbA1c, weight, exercise log) appear, interpolation and credibility marking are performed in a timely manner to generate new correction values And the corresponding credibility Γ n (t new ), and insert them into the corresponding positions of the standardized time series matrix M to ensure dynamic expansion of the data structure in the time dimension;

[0180] When new data arrives, incremental update can be triggered:

[0181] If the new observation point falls within the existing short-term window, update the corresponding short-term change;

[0182] If it touches a new short-term time segment, a new short-term window can also be supplemented, and a fast inference is made on the hybrid convolution-cyclic network to identify whether a new abnormal event has occurred;

[0183] On the long-term scale, for the defined long window If new data falls within it, attention or Transformer micro-incremental reasoning can also be triggered to update the corresponding window's deep feature representation;

[0184] The short-term change matrix D patch, long-term efficacy indicator increment, and possibly new short-term abnormal list ε or trend inflection point list and other information obtained after updating are cached and indexed and relocated in the data dictionary;

[0185] Thus, the latest short-term and long-term evaluation results can be directly read, providing traceable time series context for subsequent online updating and decision optimization; through the incremental interpolation correction and credibility marking mechanism, new clinical data can be immediately integrated into the standardized time series matrix M, and the subsequent short-term and long-term evaluation results and the overall management platform are kept synchronized; the newly collected data is quickly spliced into the existing short-term and long-term analysis window, avoiding model input discontinuity or inconsistency due to irregular data arrival times. Only by forming a unified time series database with the latest observation data and existing results can reinforcement learning or online updating modules be called and utilized in a timely manner.

[0186] Step 402, reinforcement learning or online updating mechanism

[0187] Set up a reinforcement learning (RL) agent that continuously interacts with the patient's data information and clinical decision-making process;

[0188] State space: can be selected by combining the change items in the short-term change matrix D with the long-term efficacy indicator Λ k The sequence is used as the state vector s. If higher-dimensional information is needed, short-term anomaly lists and trend inflection point lists can also be used. The tags are mapped to a specific dimension;

[0189] Several interventions or correction actions can be defined, such as fine-tuning insulin dosage, suggesting an increase in exercise, prompting further examination, or online correction operations on model parameters (such as certain hidden layer weights or thresholds);

[0190] To balance short-term effects and long-term benefits, a multi-objective reward function Υ can be defined:

[0191] R t =Υ(Δ (t) ,Λ (t) ;α ( 4 ) ,β (4) ,η (4) )

[0192] Among them, R t For multi-objective reward values, Δ (t) This represents the overall score of the short-term change observed at time t (which can be derived from the short-term change matrix D and its local statistics), Λ (t) For current long-term or comprehensive efficacy indicators; α (4) ,β (4) ,η (4) For the introduced hyperparameter, Δ (4) ,β (4) The value of η is between 0 and 1. (4) The value ranges from 1 to 3, used to control the trade-off between short-term volatility penalties and long-term indicator optimization; where Δ(t) represents the overall score of short-term changes observed at time t, and its acquisition process is as follows: First, calculate the local change Δ for each data source n within each short-term window. n,k (For example, using the maximum-minimum difference or average rate of change), and then based on the weight of each data source and the corresponding time confidence level Γ n (t) for Δ in the time interval t n,k Weighted aggregation yields a comprehensive score, namely:

[0193] Δ(t)=f(Δ 1,k ,Δ 2,k ,…,Δ N,k ),

[0194] where f denotes a mapping function for weighted normalization, which ensures that Δ(t) can fully reflect the fluctuations of each data source in the short term, providing accurate basis for subsequent anomaly detection and dynamic efficacy evaluation;

[0195] The multi-objective reward function Y can be designed as:

[0196]

[0197] where φ is a positive incentive function for long-term indicators Λ, for suppressing large short-term deterioration; α (4) ,β (4) ,η (4) can be optimized in implementation according to clinical concerns;

[0198] The reinforcement learning agent receives a new state vector s t (generated by the latest data after step 401 merging), performs an action a t (such as making drug adjustment suggestions to doctors / patients), obtains immediate reward R t and enters the next state s t+1 ;

[0199] Through Q-learning, policy gradient or other RL algorithms, the parameters of the Q function or policy network are iteratively updated;

[0200] If some actions repeatedly lead to negative results (such as short-term deterioration or serious deviation of long-term indicators), the policy update will gradually reduce the selection probability of that action, and vice versa. The combination of this process and clinical actual intervention can continuously optimize individualized diabetes management strategies on the basis of safety;

[0201] When used, short-term and long-term goals are compatible: using the multi-objective reward function Y, the generated key indicators are considered simultaneously, so that both short-term abnormalities and long-term improvements can be focused on; through online updating or reinforcement learning agent, the model and decision scheme can be continuously optimized in the scenario of constantly arriving data and dynamic changes of patient status; model suggestions or actions not only exist in the algorithm layer, but also can be directly mapped to whether to notify the doctor to adjust the scheme or to push the reminder to the patient, etc. actual operation, greatly improving the initiative of dynamic management of efficacy; in the strong noise and strong heterogeneity of medical scene, through online reinforcement learning to adapt to the characteristics of different patients, realize the truly personalized and dynamic scheme.

[0202] Step 403, closed-loop feedback and clinical interaction execution

[0203] According to the latest results of short-term efficacy change and long-term efficacy indicators, if the reinforcement learning agent judges that the short-term deterioration risk exceeds a certain warning line, it will generate a warning information and push it to the clinical end: if the long-term efficacy score Λ k continues to decline or a downward inflection point appears, an alarm will also be triggered;

[0204] The clinical end can make decisions with the generated action suggestions or drug adjustment prompts, and if the doctor or patient actually adopts and executes them, they will be used as real feedback for reinforcement learning to continuously improve the policy network or Q function through iteration;

[0205] When the doctor adjusts the medication, diet or lifestyle according to the warning, the subsequent short-term change Δ(t+1) and long-term evolution Λ(t+1) will be fed back to the management platform in real time.

[0206] If the intervention is verified to be effective, the reward R t+1 followed by an increase, reinforcing the positive learning path of the reinforcement learning agent:

[0207] If the treatment is ineffective or worsens, the corresponding action is recorded as a low-yield action, thereby reducing its probability of being selected in subsequent updates;

[0208] The management platform can also interact directly with the patient through a human-machine interface, such as customized push messages, health education, reminders for disease review, etc., to achieve more precise self-management

[0209] At each data update cycle, the management platform automatically completes the interpolation correction of new data, short-term or long-term analysis, and policy iteration of reinforcement learning, and the effectiveness of clinical intervention is also continuously promoted through the RL feedback loop to optimize the model, ultimately achieving dynamic, personalized and intelligent management of diabetes efficacy;

[0210] In use, a complete management closed loop is formed, integrating data collection, analysis, decision-making suggestions, intervention execution and feedback evaluation into the same management platform, ensuring smooth information flow and timely response; improving the efficiency of doctor-patient collaboration, enabling proactive intervention when abnormalities are detected early, and no longer relying on traditional low-frequency follow-up mode. The operations of the clinical or patient in turn affect the reinforcement learning strategy, improving the dynamic learning ability.

[0211] Step five, after completing multi-modal data fusion and continuous learning, external events such as psychological stress, economic changes, seasonal changes and surgery are collected and modeled, additional dimensions are added to the standardized time series matrix M and adaptive weights are assigned to increase the bias processing of sudden fluctuations in the short-term detection stage, and to add priority focus to high-intensity events in the long-term attention module, and dynamically adjust the intervention strategy;

[0212] The step five includes the following contents:​

[0213] Step 501, collection and labeling of external event information

[0214] Based on the standardized time sequence matrix M and its credibility Γ n (t), set a dimension to store non-physiological or non-regular clinical measurement information, such as patient psychological stress (PSS scale self-evaluation), sudden economic difficulties, seasonal change nodes, surgical records, etc.

[0215] Define the above events as ε' = {e1, e2, …}, and determine the effective interval of each event e on the time axis and the event type (psychological, economic, seasonal, medical, etc.), so that it has a unified mapping relationship with {t1, …, t L};

[0216] If the event has the characteristics of continuous or intermittent occurrence, then according to the characteristics of the event, the time is explicitly marked in the standardized time sequence matrix M in the form of Boolean or weight, to ensure time alignment with the existing indicators;

[0217] In order to deal with the diversity of external events and the difference in their impact on patients, an adaptive weight function Θ is introduced, which gives each event e a strength value ω i ∈(0, +∞), for example, a major surgery can be given a higher weight to represent a possible profound impact on blood glucose control, a short-term economic fluctuation can be given a medium weight, and a mild emotional fluctuation can be given a lower weight;

[0218] Specifically, it can be defined as:

[0219] ω i = Θ(e i ; α(5), β (5) )

[0220] Wherein, α (5) and β (5) are adjustment parameters specific to this step, both of which are between 0 and 1, and can be calculated based on event categories, clinical experience, patient self-evaluation results, etc., to produce quantifiable event strength;

[0221] After this processing, one or more events and their strength values ω i are associated with each time t in the extended structure of the standardized time sequence matrix M, denoted as the external event label matrix X;

[0222] If there is no external event at a certain time, it can be set to 0 or default for that dimension;

[0223] The external event marker matrix X is updated in time, and interpolation or deletion operations are performed according to the event type and intensity when new external events enter or old events expire, to ensure that the latest and accurate external information can be called;

[0224] In use, the multi-dimensional external factor data is digitized: the triggering events originally scattered in lifestyle, economic environment or concurrent conditions are included in a unified time sequence matrix, facilitating consistency with the physiological indicators in the first step; the influence degree of the events can be distinguished: the adaptive weight function Θ can be differentiated for different types or sizes of external events, more truly reflecting the potential impact of events on blood glucose or concurrent risk; through time indexing and interpolation strategy, the external event information is aligned with the existing short-term and long-term indicators, without damaging the existing data structure and credibility mechanism.

[0225] Step 502, joint modeling of external events and short-term-long-term evaluation

[0226] Backtracking short-term window analysis method, which uses convolution-recurrent network (or one-dimensional convolution + gated recurrent unit) to capture local peaks or mutations; now in the convolution-recurrent network input, the external event entries of the external event marker matrix X are additionally added to construct the fusion vector

[0227] To enhance event sensitivity, event bias terms δ can be defined in network convolution kernels or recurrent units (5) > 0, if the internal and external event weight ω i (t) of a certain period t is significantly greater than the average level, the bias is strengthened for the input features of this period, so as to more accurately identify and quantify the instantaneous fluctuations caused by external factors;

[0228] If the short-term change amount Δ n,k is significantly correlated with the external event ω i (t), the external trigger annotation is added to the abnormal event marker ε;

[0229] In a multi-layer attention or Transformer network, the event intensity ω i can be used as an attention bias, similar to the existing short-term abnormal bias mechanism:

[0230] Q' = Q + Ω(ω i ; ζ (5) ), K' = K + Ω(ω i ; ζ (5) )

[0231] Where Ω is a mapping function for external events, ζ (5) is a learnable parameter, taking a value between 0 and 1, used to control the weight relationship between external events and long-term attention mechanism;

[0232] If the intensity value ω i is at a high level, Q' and K' will increase the contribution of the relevant period and indicators in the attention score calculation, thereby improving the long-term efficacy indicator Λ k The generation takes into full consideration the impact of external environment;

[0233] When the external event intensity value ω i (t) exceeds a certain threshold, it automatically switches to a more conservative or more sensitive decision-making strategy, such as adjusting the penalty factor in the reward function, and pays more attention to short-term stability when facing high-intensity external interference, in order to avoid sudden deterioration of the disease;

[0234] At the same time, if the external event ends and is smoothly passed, the strategy center of gravity will be returned to long-term optimization again according to historical learning experience, so as not to intervene too much;

[0235] After completing the integration of external events at short-term and long-term levels, external event information and weights will be automatically added to the abnormal event marker ε, trend inflection point list and continuous learning log; the output to the clinical or patient end can include visual reports such as event type, event time window, event intensity, and disease fluctuation correlation;

[0236] When it is passed to the established closed-loop system, it can ensure that the next data collection cycle (back to the first step) or even the next model inference has complete external event context information;

[0237] By adding external event bias in the short-term feature extraction stage, it can quickly discover sharp fluctuations caused by major psychological or surgical factors and reduce the blind area of traditional models that ignore external triggers;

[0238] With the help of attention or Transformer mechanism, external event information is introduced into cross-week or cross-month evaluation, which can more accurately divide controllable fluctuations and abnormalities caused by external interference, thereby improving the credibility of the long-term efficacy indicator Λ k ; In the reinforcement learning reward function, according to the external event ω i (t), it is dynamically optimized to make necessary precautions when facing high-intensity external shocks, and timely return to the regular strategy after the disturbance subsides, achieving flexible and continuous learning;

[0239] External event bias or weight mapping is introduced in both short-term detection and long-term fusion, covering multi-time scale models. This unified event fusion framework significantly improves the sensitivity and interpretability of the management platform to sudden situations. By adjusting the multi-objective reward function in real time through the external event intensity, the management platform can adaptively switch between short-term urgent teaching and long-term stability, far exceeding the flexibility and safety of traditional single reward strategies.

[0240] Those skilled in the art can clearly understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0241] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0242] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0243] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0244] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-modal dynamic diabetes therapy evaluation method, characterized in that: comprising, When the multimodal data is detected to arrive, an adaptive alignment and interpolation algorithm is called, and outlier correction is performed according to the reliability, so that irregular sampling points are smoothly integrated into the standardized matrix, and each source heterogeneous data is preprocessed; Based on the standardized time series matrix, the data in the short-term range is segmented, when the standardized matrix and the reliability are ready, the short-term window detection is triggered, the fusion convolution and the cyclic network are used to perform local peak and mutation recognition on the recent core indicators, and the short-term change amount and the abnormal marker are written in the detected abnormal section, and a high-sensitivity fluctuation feature is realized; after the short-term window segmentation is completed, deep local feature extraction is performed on the data in the window, and the existing peak, slope mutation or high-frequency fluctuation position is found out; the local feature vector is finally output for each window and data source, and the main change mode, peak and valley distribution and possible slope jump in the window are represented in a fixed length; A multi-layer attention is called to deeply fuse the detection indicator sequence in the long-term span, and an additional bias is given to the local high-risk section to generate a comprehensive long-term efficacy indicator and a key inflection point; when processing the multimodal long-term sequence, an attention mechanism is introduced to strengthen the attention to the key period and the indicator, wherein: in a layer of attention mechanism, for the long window feature matrix corresponding to the selected long window, an attention output is defined, an additional bias is applied to the vector dimension containing the abnormal marker in the feature mapping stage, and a global correlation is established between different indicators and long periods; After synchronously updating the long-term efficacy indicator and the short-term change amount, a reinforcement learning or online updating mechanism is started, the two are included in a multi-objective reward function, the model parameters are corrected in a policy iteration manner, if a major deterioration is identified, an intervention suggestion is pushed, otherwise the current decision strategy is enhanced.

2. The multimodal dynamic diabetes efficacy evaluation method according to claim 1, characterized in that: A dynamic time alignment and interpolation method based on a kernel function is constructed to eliminate uneven sampling and measurement time period misalignment of the multi-source data, wherein: a mapping is constructed for each data source, so that the measurement values of the data source are projected onto a global time axis; For any time point, an interpolation function is defined, each data source is calculated point by point on the global time sequence according to the interpolation function, a new aligned sequence is obtained, and the aligned sequences of all data sources are combined into a standardized time series matrix.

3. The multimodal dynamic diabetes efficacy evaluation method according to claim 2, characterized in that: An adaptive correction and reliability marking mechanism is introduced to form a time series data, wherein a nonlinear function is used to filter the residual error, the corrected time series value is output to replace the original interpolation, and the corresponding position in the standardized time series matrix is updated, and after the correction is completed, a reliability coefficient is assigned to each item; If the reliability coefficient is not close to 1, the suspicious value is identified, the weight is reduced or removed during the judgment, and the standardized time series matrix with the corrected value and the reliability is generated after processing.

4. The multimodal dynamic diabetes efficacy evaluation method according to claim 3, characterized in that: Selecting short-term windows, each containing multiple consecutive sampling times, and generating a set of short-term windows according to a given step size on the time axis; for each window, extracting a subset of rows and columns from the standardized time series matrix and retaining the corresponding confidence; excluding or weakening time points with low confidence.

5. The multi-modal diabetes dynamic therapeutic effect evaluation method of claim 4, wherein: Mapping the local feature vector to a real-valued quantization result, defining an adjustable function for each data source and window, calculating the short-term change, and condensing the local feature vector into a one-dimensional value; Introducing a dynamic threshold curve, if the short-term change > dynamic threshold curve, marking the corresponding time period as abnormal and recording its fluctuation characteristics associated with the time index of the standardized time series matrix; forming a short-term change matrix in each data source and window dimension, and summarizing the abnormal events into a short-term anomaly list.

6. The multi-modal diabetes dynamic therapeutic effect evaluation method of claim 5, wherein: Segmenting and aggregating the standardized time series matrix, evenly dividing into several windows, combining the short-term anomaly list and the short-term change matrix, mapping them into long windows according to the corresponding time range, and aggregating different indicators for each long window to form a subsequence; If a certain indicator fluctuates sharply, the corresponding abnormal mark or additional high-weight label is retained in the sequence representation corresponding to the indicator, and the combination of each long window and each indicator forms a multi-modal long-term feature container, and the local fluctuation data is stored in the attached abnormal annotation.

7. The multi-modal diabetes dynamic therapeutic effect evaluation method of claim 6, wherein: After each long window is encoded by multiple layers of Transformer or multi-head attention, the final representation is obtained; Global pooling or aggregation is performed on the final representation to obtain a fixed vector representing the deep fusion features of the multi-modal information in the long window; The fixed vector is projected to the real number domain, and the phase therapeutic effect score of the long window is output, and it is connected or visualized in time sequence; a key time difference vector is constructed, and the positive and negative and absolute amplitude of the difference vector are detected in combination with the pre-set dynamic boundary: If the difference vector shows a significant cross-over change between two consecutive windows, it is marked as a trend inflection point, and the start and end time corresponding to the inflection point and the main indicator contribution are recorded; After processing all long windows, the final output is: long-term therapeutic effect indicators, recording the trend inflection point list of the time segment with significant rising or falling in the therapeutic effect curve.

8. The multi-modal diabetes dynamic therapeutic effect evaluation method of claim 7, wherein: New measurement values are recorded in the standardized time series matrix to generate new corrected values and corresponding confidence, and they are inserted into the corresponding position of the standardized time series matrix; if the new observation point falls within the existing short-term window, the corresponding short-term change is updated; If a new short-term time segment is touched, a new short-term window is supplemented and a mixed convolution-cyclic network is quickly inferred to identify whether a new abnormal event occurs; on the long-term scale, for the defined long window, if the new data falls into it, the deep feature representation of the corresponding window is updated.

9. The multi-modal dynamic diabetes therapy evaluation method of claim 8, wherein: a reinforcement learning agent is set to interact with the patient's data information and clinical decision-making process continuously; a sequence of the change amount entries in the short-term change matrix and the long-term efficacy indicators are selected as the state vector, and a number of intervention or correction actions or online correction operations on the model parameters are defined; if some actions repeatedly lead to negative results, the strategy update will gradually reduce the selection probability of the action, and vice versa, which combines the process with the actual clinical intervention; according to the latest results of the short-term efficacy change amount and the long-term efficacy indicators, if the reinforcement learning agent judges that the short-term deterioration risk exceeds a certain warning line, it will generate a warning information and push it to the clinical end: if the stage efficacy score is continuously low or a downward inflection point suddenly appears in a long window, an alarm is triggered.

Images