A method and system for monitoring and early warning of adverse reactions of glioma radiochemotherapy

Abstract

The application discloses a glioma radiotherapy and chemotherapy adverse reaction monitoring and early warning method and system, collects blood routine data and neurological function score data for time sequence alignment, generates a toxicity response graph through channel-by-channel trend baseline correction fusion, generates toxicity response parameters and compensation depletion early warning marks; extract liver and kidney and toxicity index deviation period positioning toxicity occurrence section, generate toxicity retention aggregation marks and cross-dimension coupling with toxicity response parameters to form a toxicity feature matrix; the toxicity feature matrix is subjected to abnormal mode clustering to generate a toxicity adaptation mark, extract a toxicity gradient feature to construct an adverse reaction atlas; identify weight inversion features to determine toxicity sensitivity weights, combine the compensation depletion early warning marks to construct a hierarchical early warning rule, perform risk mapping based on the adverse reaction atlas to output an adverse reaction early warning result, realize dynamic perception and hierarchical early warning of a multi-course toxicity accumulation mode, and effectively improve the precision and timeliness of glioma radiotherapy and chemotherapy adverse reaction monitoring.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a method and system for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma. Background Technology

[0002] During radiotherapy and chemotherapy for glioma patients, while the drugs kill tumor cells, they also cause varying degrees of damage to the bone marrow hematopoietic system, neurological function, and liver and kidney metabolic function. These damages often accumulate over multiple treatment cycles, leading to complex adverse reactions. Because bone marrow suppression, neurotoxicity, and organ dysfunction overlap over time, clinicians cannot accurately determine the patient's overall toxicity load based on a single follow-up data. This often results in intervention being implemented only after toxic events have already occurred, missing the optimal treatment window.

[0003] Existing adverse reaction monitoring methods typically treat hematological indicators and neurological function assessments as independent evaluation dimensions, lacking a mechanism for joint analysis of these two types of data within a unified time reference framework. Furthermore, the identification of cumulative toxicity across multiple treatment cycles generally relies on fixed threshold determinations, lacking effective dynamic sensing methods for latent risk patterns such as declining bone marrow compensatory capacity and persistent toxicity retention. This makes it difficult to issue graded warnings before patients' reserve functions are nearly exhausted, failing to meet the individualized and refined adverse reaction management needs of long-term radiotherapy and chemotherapy for gliomas. Summary of the Invention

[0004] This invention discloses a method and system for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma. The aim is to identify patterns of toxicity retention and accumulation and bone marrow compensation exhaustion signals by integrating the dynamic characteristics of multi-course blood routine data and neurological function score data. It constructs an adverse reaction atlas covering the entire process of toxicity occurrence, adaptation and evolution, and achieves graded early warning output by combining toxicity sensitivity weights and compensation status. This provides methodological support for dynamic monitoring and individualized risk management of adverse reactions throughout the entire process of radiotherapy and chemotherapy for glioma.

[0005] The first aspect of this invention provides a method for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for gliomas, comprising the following steps:

[0006] Collect routine blood data and neurological function score data, and perform time-series alignment of the routine blood data and the neurological function score data to form a set of dynamic physiological parameters;

[0007] The physiological dynamic parameter set is fused by channel-wise trend baseline correction to generate a toxicity response map. Based on the toxicity response map, cross-index abnormality enhancement is performed to generate toxicity response parameters. The bone marrow index sudden drop and rebound characteristics are extracted from the blood routine data to generate a compensatory exhaustion early warning indicator.

[0008] The toxicity response map is used to extract the time period when the liver and kidneys deviate from the toxicity index to locate the toxicity occurrence segment. Based on the toxicity occurrence segment, the toxicity retention and accumulation period period is identified to generate a toxicity retention and accumulation identifier. Based on the toxicity retention and accumulation identifier and the toxicity response parameters, a toxicity feature matrix is ​​formed by cross-dimensional coupling.

[0009] Anomaly pattern clustering is performed on the toxicity feature matrix to determine the toxicity category distribution. A toxicity adaptation identifier is generated by identifying abrupt changes in scoring after stagnation in the toxicity category distribution. Toxicity gradient features are extracted based on the toxicity adaptation identifier to construct an adverse reaction map.

[0010] Based on the toxicity response parameters and the weight inversion feature of the toxicity feature matrix, a toxicity-sensitive weight is determined. A graded early warning rule is constructed based on the toxicity-sensitive weight and the compensation exhaustion early warning indicator. The graded early warning rule and the adverse reaction map are then risk-mapped to output the adverse reaction early warning result.

[0011] A second aspect of this invention provides a monitoring and early warning system for adverse reactions to radiotherapy and chemotherapy for gliomas, comprising:

[0012] The data acquisition module is used to collect blood routine data and neurological function score data, and to perform time-series alignment of the blood routine data and the neurological function score data to form a set of physiological dynamic parameters;

[0013] The feature extraction module is used to perform channel-by-channel trend baseline correction and fusion on the physiological dynamic parameter set to generate a toxicity response map, implement cross-index abnormal enhancement based on the toxicity response map to generate toxicity response parameters, and extract bone marrow index sudden drop and rebound features from the blood routine data to generate a compensatory exhaustion early warning indicator.

[0014] The segment positioning module is used to extract the time period when the liver and kidneys deviate from the toxicity index using the toxicity response map to locate the toxicity occurrence segment, identify the toxicity retention and aggregation time period based on the toxicity occurrence segment to generate a toxicity retention and aggregation identifier, and perform cross-dimensional coupling based on the toxicity retention and aggregation identifier and the toxicity response parameters to form a toxicity feature matrix.

[0015] The map construction module is used to perform abnormal pattern clustering on the toxicity feature matrix to determine the toxicity category distribution, generate toxicity adaptation identifiers by identifying abrupt change segments after score retention in the toxicity category distribution, and extract toxicity gradient features based on the toxicity adaptation identifiers to construct an adverse reaction map.

[0016] The early warning output module is used to determine the toxicity sensitivity weight based on the weight inversion feature of the toxicity response parameter and the toxicity feature matrix, construct a graded early warning rule based on the toxicity sensitivity weight and the compensation exhaustion early warning identifier, and perform risk mapping on the graded early warning rule and the adverse reaction map to output the adverse reaction early warning result.

[0017] The beneficial effects of this invention are reflected in the following points: First, it aligns blood routine data and neurological function score data temporally, integrating data with inconsistent sampling densities into a unified time reference frame. Based on this, it eliminates baseline drift bias between multiple treatment courses through channel-by-channel trend baseline correction, implements cross-index abnormal enhancement to generate toxicity response parameters, and extracts bone marrow index sudden drop and rebound characteristics to generate compensatory exhaustion early warning indicators, quantifying latent risk signals such as bone marrow compensatory capacity decline, providing a complete and reliable multi-dimensional data foundation for subsequent toxicity feature extraction. Second, by identifying the divergence between liver and kidney indicators and toxicity amplitude, it isolates and locates primary toxic events of radiotherapy and chemotherapy from liver and kidney metabolic interference, then quantifies the retention and aggregation patterns in the toxicity occurrence segment, coupling them node-by-node with toxicity response parameters on the time axis, constructing a toxicity feature matrix that simultaneously carries toxicity aggregation information and intensity distribution information, achieving a structured expression of multi-treatment toxicity accumulation patterns. Finally, an adverse reaction map was constructed by abnormal pattern clustering and toxicity adaptation behavior identification, which solved the problem that high-toxicity feature samples were masked by low-toxicity categories in traditional clustering methods. On this basis, the weight reversal bias was corrected and a two-dimensional joint constraint graded early warning rule was established by combining the degree of bone marrow compensation exhaustion. This ensures that the early warning output can accurately reflect the actual risk level of patients at different stages of toxicity accumulation, and avoid false alarms and false negatives caused by a single scoring indicator. Attached Figure Description

[0018] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0019] Figure 1 This is a flowchart illustrating a method for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma according to the present invention.

[0020] Figure 2 This is a structural block diagram of a glioma radiotherapy and chemotherapy adverse reaction monitoring and early warning system according to the present invention. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0023] The technical solutions of the embodiments of this application will be described below.

[0024] like Figure 1 As shown, this embodiment of the invention provides a method for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma, including the following steps S11-S15:

[0025] Step S11: Collect blood routine data and neurological function score data, and perform time-series alignment of blood routine data and neurological function score data to form a set of physiological dynamic parameters.

[0026] Specifically, complete blood count (CBC) data and neurological function scores were collected. CBC data were collected at each follow-up stage of the radiotherapy and chemotherapy cycle, covering four bone marrow suppression-sensitive indicators: white blood cell count, absolute neutrophil count, hemoglobin concentration, and platelet count. The values ​​of these indicators showed cyclical fluctuations with each treatment cycle. The CBC data collection time was linked to the drug administration points of the treatment plan, with one collection window before and one after drug administration. Pre-administration CBC data reflected the baseline recovery status of bone marrow, while post-administration CBC data captured the magnitude of acute toxicity. The values ​​from both windows together reconstructed the dynamic changes in bone marrow suppression within a single treatment cycle. Neurological function scores were assessed by clinicians at each follow-up stage using standardized scales. The assessment results were recorded as a weighted composite score of motor function, cognitive function, and daily living activities. The collection density of neurological function scoring data and blood routine data is different. Blood routine data generates multiple measurement values ​​at multiple time points per course of treatment, while neurological function scoring data generates only one comprehensive score per follow-up. A patient may have 4 blood routine data records in a single course of treatment, but only 2 neurological function scoring data. This density difference requires the temporal alignment stage to specifically handle the node matching relationship between the two types of data.

[0027] A physiological dynamic parameter set is formed by temporal alignment of blood routine data and neurological function score data. Temporal alignment uses the natural cycle of the treatment course as the time unit, mapping the two types of data to a unified time node sequence to eliminate node misalignment caused by inconsistent sampling frequencies. During the alignment process, neurological function score data is aligned with the sampling point of blood routine data according to the nearest neighbor timestamp principle. When the time interval between the neurological function score data and the most recent blood routine data sampling point exceeds 7 days, the alignment result at the corresponding position is marked as a low-precision node. Low-precision nodes are retained in the physiological dynamic parameter set but with an added quality label. When the test value of a sampling node in the blood routine data is missing, the value of that node is interpolated by the average of adjacent time nodes. The interpolated node is marked with an interpolation label in the physiological dynamic parameter set to distinguish it from the measured value. If a patient's blood routine data could not be collected for a single session due to temporary reasons, the interpolated value is obtained by averaging the results of the two consecutive collections. The corresponding node is marked with an interpolation label in the physiological dynamic parameter set to avoid misusing the interpolated value as the measured value in subsequent analysis. After time alignment, the various indicators of the complete blood count (CBC) data and the comprehensive score of the neurological function score data correspond node by node on a unified time axis. The number of channels in the physiological dynamic parameter set is equal to the sum of the number of CBC data collection items and the number of comprehensive score channels in the neurological function score data. The CBC data collection items cover four items: white blood cell count, absolute neutrophil count, hemoglobin concentration, and platelet count. The neurological function score data contributes one comprehensive score channel. The physiological dynamic parameter set has a total of five channels, with each time node carrying both the values ​​of the CBC data and the comprehensive score of the neurological function score data. The time span covers the patient's complete radiotherapy and chemotherapy course.

[0028] Step S12: Perform channel-by-channel trend baseline correction and fusion on the physiological dynamic parameter set to generate a toxicity response map. Based on the toxicity response map, perform cross-index abnormal enhancement to generate toxicity response parameters. Extract bone marrow index sudden drop and rebound characteristics from blood routine data to generate compensatory exhaustion early warning indicators.

[0029] In some embodiments, the step of performing channel-by-channel trend baseline correction and fusion on the physiological dynamic parameter set to generate a toxicity response map includes: performing channel-by-channel trend decomposition on the physiological dynamic parameter set to generate a channel toxicity response sequence; identifying the baseline shift of adjacent period channels from the channel toxicity response sequence to generate a baseline shift factor; performing channel recalibration on the channel toxicity response sequence using the baseline shift factor to form a calibrated toxicity feature; and performing multi-scale cascade fusion based on the calibrated toxicity feature to generate a toxicity response map.

[0030] Channel-by-channel trend decomposition was performed on the physiological dynamic parameter set to generate channel toxicity response sequences. The temporal signals of each channel in the physiological dynamic parameter set consist of two components: the slow accumulation trend driven by radiotherapy and chemotherapy toxicity and normal physiological fluctuations. Trend decomposition separated the low-frequency toxicity accumulation trend from the high-frequency fluctuation components, retaining the trend component as the effective toxicity response signal for each channel. The physiological response rhythms of blood routine and neurological function scoring channels in the physiological dynamic parameter set differed significantly. Each channel independently determined its decomposition parameters based on its own acquisition frequency and response time constant. The toxicity response of blood routine indicators was relatively rapid, corresponding to a shorter trend decomposition window to capture acute inhibitory signals within a few days after drug administration. The change rhythm of neurological function scores was relatively slow, corresponding to a relatively wider window to track functional degeneration processes across treatment courses. When the trend component variance of a certain channel in the physiological dynamic parameter set was extremely low throughout the observation period, the channel was marked as a low-response channel and annotated with a low weight in the channel toxicity response sequence. If the change of a certain blood routine indicator in a patient was less than 10% of the normal fluctuation range throughout the entire treatment course, the channel was marked as a low-response channel to avoid interfering with the extraction of toxicity signals from high-response channels during the fusion phase. After the trend components of each channel are extracted, they are arranged in the order of channel identifiers and aligned one by one with the original sampling time nodes of the physiological dynamic parameter set to form a channel toxicity response sequence. Each row of the channel toxicity response sequence corresponds to the toxicity accumulation trend of a single channel. The time axis of each channel in the channel toxicity response sequence is kept consistent to ensure that the baseline shift of adjacent cycles is extracted under a consistent time reference system. The overall time span of the channel toxicity response sequence covers the patient's complete radiotherapy and chemotherapy course.

[0031] Baseline shift factors are generated by identifying baseline shifts between adjacent cycles in the channel toxicity response sequence. Accumulated damage from multiple cycles of radiotherapy and chemotherapy leads to a progressive decline in bone marrow hematopoietic capacity and neurological function as treatment progresses, manifested in the channel toxicity response sequence as a continuous unidirectional downward shift of the baseline level between adjacent cycles. The baseline shift is measured by the difference between the mean of the channel toxicity response sequence at the beginning of each cycle and the mean of the corresponding segment of the previous cycle. A positive difference indicates a baseline shift relative to the previous cycle, while a negative difference suggests local recovery. The shift is extracted independently for each channel at the boundary of each cycle to avoid interference between channels. For example, in a glioma patient, the white blood cell count channel shows a consistently positive difference at the boundary of the second to third cycles, accumulating with each cycle, indicating a continuous baseline shift across multiple cycles. Conversely, in the same patient, the hemoglobin concentration channel shows a negative difference at the boundary of the fifth cycle, suggesting short-term local recovery after blood transfusion or nutritional support intervention. The baseline shift of each channel in the toxicity response sequence is arranged into a matrix by treatment course number. A continuous increase in the value of a column along the row direction indicates that the baseline of the corresponding channel has continuously declined over multiple treatment courses. In one patient, the baseline shift of a certain channel in the fourth treatment course was more than twice the average of the previous three treatment courses. This element in the matrix was significantly larger, indicating that the degree of bone marrow suppression of that channel was significantly aggravated in that treatment course. The baseline shift matrix was subjected to global normalization across channels and treatment courses. Using the distribution of all elements as a reference, each element was mapped to a unified dimension. After normalization, the baseline shift factor was obtained. The high value of the baseline shift factor indicates that the baseline decline of the corresponding channel in the corresponding treatment course is the most prominent in the overall distribution.

[0032] The baseline shift factor is used to recalibrate the channel toxicity response sequence to form calibrated toxicity characteristics. Baseline shifts between multiple treatment courses are a global positional bias rather than a true difference in the amplitude of the toxicity response. Without eliminating this positional bias, the values ​​of the channel toxicity response sequence across different treatment courses cannot be directly compared within the same reference frame. Recalibration eliminates this bias by shifting the values ​​of each channel across each treatment course as a whole. The values ​​corresponding to each channel and treatment course in the baseline shift factor are used as the shift correction amount for that channel and treatment course. This uniformly shifts the values ​​of the corresponding channel in the channel toxicity response sequence upwards by the same amount at all time points of the corresponding treatment course, aligning the trend level of the initial segment of each treatment course to a unified reference baseline. Treatment courses corresponding to high values ​​of the baseline shift factor receive a larger shift correction amount, while low values ​​require only a smaller correction. For example, a patient with a high baseline shift factor value for a certain channel in the fifth treatment course receives a correspondingly large correction amount at all points in that course, bringing the corrected values ​​to the same baseline level as other treatment courses. In contrast, the shift amount for this channel in the first treatment course is close to zero, requiring virtually no shift correction. After recalibration, each channel in the toxic response sequence is arranged to form a calibrated toxicity feature. The amplitude of each channel in the calibrated toxicity feature can be directly compared laterally under a unified baseline reference. The difference in abnormal response intensity between different channels can be truly reflected without being affected by baseline drift between treatment courses. The amplitude structure of each channel in the calibrated toxicity feature directly determines the distribution of response intensity at each time point of the toxicity response map after multi-scale cascade fusion.

[0033] Toxicity response maps are generated through multi-scale cascade fusion based on calibrated toxicity characteristics. The effects of radiotherapy and chemotherapy toxicity on patients' physiological indicators span multiple time scales. Acute toxicity exhibits rapid fluctuations within days after administration, while chronic toxicity shows a gradual trend after multiple treatment cycles. Single-scale fusion methods cannot simultaneously retain both types of features; multi-scale fusion addresses this by extracting features at each scale in a hierarchical manner. Response features are extracted at short-term, medium-term, and long-term scales based on calibrated toxicity characteristics. The short-term scale uses the sampling interval before and after a single administration as a window to capture the rapid amplitude fluctuations caused by acute toxicity. The medium-term scale covers all sampling nodes within a single treatment cycle to extract the response trend within the treatment cycle. The long-term scale spans all treatment cycles to extract the cumulative trend across multiple treatment cycles. The response features of the three scales of the toxicity profile were cascaded and merged layer by layer. Short-term and intermediate-term features were first merged to form an intermediate layer, which was then merged with the long-term features. The merging was performed using a position-weighted average to retain the feature contribution of each scale. At the end of the third course of treatment, multiple channels simultaneously showed short-term acute fluctuations and long-term cumulative upward trends. The superposition of these two at this time point formed a significant peak in the toxicity response map, indicating concentrated exposure to toxicity during this period. The temporal resolution of the toxicity response map was consistent with the sampling node density of the toxicity profile. High-amplitude positions in the toxicity response map corresponded to the time periods of concentrated toxicity in multiple channels. After normalization, the dynamic range of the toxicity response map was unified to the same dimension, and the time span of the toxicity response map covered the patient's entire course of radiotherapy and chemotherapy.

[0034] In some embodiments, the step of generating toxicity response parameters by performing cross-index anomaly enhancement based on the toxicity response map includes: performing dual-path separation of channel response and time-series response on the toxicity response map to generate an initial anomaly map; identifying the baseline-unrecovered channel after treatment based on the initial anomaly map to generate a baseline-unrecovered weight; performing weak anomaly enhancement on the initial anomaly map using the baseline-unrecovered weight to form a corrected anomaly map; and selecting high-response regions based on the corrected anomaly map to generate toxicity response parameters.

[0035] An initial anomaly map is generated by separating the channel response and time-series response in the toxicity response map. Channel-level anomalies in the toxicity response map are characterized by a sustained deviation of the response amplitude of a certain channel from the overall level of other channels. Time-series anomalies are characterized by a deviation of the response value at a certain time point from the historical mean for that period. Dual-path separation ensures that both types of anomalies are quantified independently. For the channel response path, the mean and standard deviation of all channels in the toxicity response map are used as references. Channel-level anomalies are identified when the response amplitude of any channel exceeds the mean plus or minus two times the standard deviation, and the deviation is used as the channel path anomaly value at that location. For the time-series response path, a sliding historical window is used as a reference. Time-series anomalies are identified when the absolute value of the difference between the response value at each time point and the historical sliding mean exceeds three times the standard deviation of the historical window, and the deviation is used as the time-series path anomaly value at that location. Simultaneously, the square root of the product of the two deviations is taken at the positions marked by both paths to amplify the superposition effect. This square root operation ensures that the resulting dimensions are consistent with the single-path deviation, guaranteeing comparability of the amplitude of the dual-path superposition position and the single-path marked position in the initial anomaly map. Positions marked only by a single path directly take the corresponding deviation value, while positions not marked by any path are assigned a value of zero. In a patient at the end of a treatment course, multiple channels simultaneously exhibited temporal and channel deviations. The amplitude of the initial anomaly map in this region was significantly higher than in surrounding time periods, indicating a concentrated manifestation of multidimensional toxicity abnormalities during this period. Conversely, in the early stages of the same patient's treatment, only single-path marking was triggered, resulting in a relatively lower amplitude in the initial anomaly map for that period. The difference in amplitude levels between the two periods directly reflects the shift in toxicity patterns. The spatial structure of the initial anomaly map is consistent with the toxicity response map, with high-amplitude positions in the initial anomaly map reflecting the time period of the most concentrated superposition of multidimensional toxicity abnormalities.

[0036] Based on the initial anomaly map, channels that have not recovered their baseline after treatment are identified, and baseline recovery weights are generated. The follow-up nodes at the end of each treatment cycle are the key time window for judging the recovery status of a channel. If the initial anomaly map amplitude of a certain channel at that node remains consistently high, it indicates that the toxic response of the corresponding indicator has not subsided with the end of the treatment course, and that the channel has persistent toxic residues across treatment courses. When the amplitude of each channel at the follow-up node in each treatment cycle exceeds the median of the entire map, it is determined that the baseline of that node has not recovered. The ratio of the number of nodes with baseline recovery for each channel across all follow-up nodes to the total number of follow-up nodes is calculated. A higher ratio indicates that the channel has persistent toxic residues across multiple treatment courses. In one patient, a certain channel was determined to have baseline recovery in five follow-up nodes across six treatment courses, with a ratio of 0.83, while another channel only triggered the recovery judgment in one follow-up node, with a ratio of only 0.17. The contrast is stark; the former was given a significantly higher amplification priority in the subsequent weak anomaly enhancement phase. The baseline non-recovery ratio of each channel in the initial anomaly map is normalized and mapped to the interval of 0 to 1. The set of normalized ratios of all channels constitutes the baseline non-recovery weight. The high-value channels in the baseline non-recovery weight reflect the most severe residual toxicity during the entire radiotherapy and chemotherapy course. The distribution pattern of the baseline non-recovery weight among channels serves as a reference benchmark for residual risk at each location in the regional difference analysis.

[0037] For example, the step of weakly enhancing the initial anomaly map using the baseline unrecovered weights to form a corrected anomaly map includes: performing regional difference analysis on the baseline unrecovered weights and the initial anomaly map to generate an enhancement priority map; locating adjacent periodic priority offset regions from the enhancement priority map to generate migration enhancement factors; performing adaptive difference stretching on the initial anomaly map using the migration enhancement factors to form an enhanced anomaly map; and performing response equalization processing based on the enhanced anomaly map to form a corrected anomaly map.

[0038] Regional dissimilarity analysis is performed on the baseline unrecovered weights and the initial anomaly map to generate an enhancement priority map. The initial anomaly map is globally normalized and mapped to the 0-1 interval. The scalar values ​​of each channel of the baseline unrecovered weights are broadcast along the time axis to the same dimension as the normalized initial anomaly map. The difference between the broadcast weight value and the amplitude of the normalized initial anomaly map is taken at each position. Both are in the 0-1 interval, and the difference values ​​have the same dimensions. A positive difference indicates that the residual risk at that position is higher than reflected by the actual amplitude. The larger the difference, the higher the necessity of priority enhancement. A negative difference or a difference close to zero indicates that the amplitude and residual risk are basically matched. The difference values ​​at each position are normalized and mapped to the 0-1 interval to form the enhancement priority map. The high-value areas of the enhancement priority map are concentrated to identify the positions where weak anomalies are most severely underestimated. In one patient, the baseline non-recovery weight for a certain channel remained consistently high, but the initial anomalous map amplitude for the corresponding time period was low. After difference normalization, multiple nodes of this channel showed a high-priority distribution in the enhancement priority map, identifying weakly abnormal areas that urgently needed priority treatment. Conversely, another channel at the same time period had a lower baseline non-recovery weight; even with a similarly low initial anomalous map amplitude, its priority value was significantly lower after difference normalization. The priority difference between the two types of channels in the enhancement priority map truly reflects the relative levels of their respective toxicity residues. The priority level of each time node in the enhancement priority map determines the judgment criterion for adjacent cycle offset detection.

[0039] The migration enhancement factor is generated by locating the priority shift region between adjacent cycles from the enhancement priority map. Channel-by-channel difference calculations are performed on the enhancement priority maps on both sides of the boundary between adjacent treatment cycles. The difference reflects the magnitude of the change in priority level of each channel between two adjacent treatment cycles. A significant shift is defined as a difference whose absolute value exceeds 50% of the historical average of the enhancement priority map for that channel. The treatment location of the significantly shifted channel is recorded as a priority shift segment. The shift magnitude of each priority shift segment is multiplied by the priority value of the corresponding position in the enhancement priority map. The product reflects the actual migration intensity of the shift segment under the overall priority background. Since the baseline non-recovery weight for residual risk assessment of each channel has been incorporated into the enhancement priority map, the product at high priority positions naturally amplifies the migration signal of channels with high toxicity residue. For example, in a glioma patient, the absolute neutrophil count channel showed a significant priority shift at the boundary between the third and fourth treatment cycles. Furthermore, the enhancement priority map value at this location remained persistently high due to the failure to recover the baseline after multiple treatment cycles, resulting in a significantly larger product of these two values. This suggests active weak abnormal migration in this channel at the treatment cycle transition point. Conversely, in the same patient, the hemoglobin concentration channel showed similar shifts between adjacent treatment cycles, but the corresponding enhancement priority map value was close to zero, resulting in a very small product. This meant that the channel produced almost no effective stretching signal in the migration enhancement factor. The difference in the product results of the two channels accurately reflects the modulation effect of their respective residual toxicity background on migration intensity. The migration intensity values ​​of all priority shift segments were arranged by channel and time node position, then normalized to the maximum value and mapped to the 0-1 interval to form the migration enhancement factor. This ensured that the subsequent adaptive differential stretching amplification factor at each position remained within a controllable range. The non-zero positions of the migration enhancement factor identified the spatiotemporal regions where weak abnormal migration was most active.

[0040] An enhanced anomaly map is formed by adaptive differential stretching of the initial anomaly map using a migration enhancement factor. The values ​​of the migration enhancement factor are unevenly distributed across different channels and time points, with high values ​​concentrated in the most active regions of weak anomaly migration. Differential stretching utilizes this non-uniform distribution to apply position-specific amplitude amplification to the initial anomaly map, ensuring targeted enhancement of signals in weak anomaly migration regions rather than uniform amplification across the entire map. Let the amplitude at time point t in channel c of the initial anomaly map be I(c,t), and the value of the migration enhancement factor at the corresponding position be M(c,t). Differential stretching calculates the stretched amplitude E(c,t) position by position using the formula E(c,t) = I(c,t) × (1 + γ × M(c,t)), where γ is the global stretching coefficient, determined by the formula γ = γ_min + (γ_max − γ_min) × SNR_norm, where SNR_norm = clip((μ_I / σ_I)). γ_min = (SNR_min) / (SNR_max−SNR_min), 0, 1), μ_I is the mean amplitude of all positions in the initial anomaly map, σ_I is the corresponding standard deviation, SNR_min is 0.5, SNR_max is 3.0, γ_min is 0.5, γ_max is 2.0, and SNR_norm is truncated to the [0,1] interval. When the signal-to-noise ratio is low, γ approaches 0.5 to prevent excessive amplification of noise; when the signal-to-noise ratio is high, γ approaches 2.0 to fully amplify the weak anomaly signal. The E(c,t) values ​​of all channels and time nodes are arranged in the order of channel identifier and time node to form the enhanced anomaly map. The number of rows in the enhanced anomaly map is equal to the number of channels in the initial anomaly map, and the number of columns is equal to the number of time nodes. Even if M(c,t) is non-zero, the E(c,t) result remains zero at locations where I(c,t) is zero. This prevents the introduction of false amplitudes at locations without abnormal evidence. For example, in a certain channel of a patient, I(c,t) may be close to zero during a certain treatment course, but M(c,t) may have a higher value at that location. After stretching, E(c,t) at that location remains close to zero and does not trigger an abnormality marker. However, at another location within the same treatment course, although I(c,t) is low, it is not zero. After stretching, E(c,t) is effectively increased, forming a visible weak abnormal signal in the enhanced abnormality map. The amplitude change in the enhanced abnormality map relative to the initial abnormality map is concentrated in the high value region of M(c,t), while the other locations remain highly consistent with the initial abnormality map.

[0041] A corrected anomaly map is generated by performing response equalization processing based on the enhanced anomaly map. Adaptive differential stretching applies varying degrees of amplitude amplification to each channel, resulting in a general increase in the amplitude of channels with high mobility enhancement factors. Without equalization, this overall amplitude advantage could bias cross-channel comparisons during the weak anomaly screening stage. Equalization is performed independently for each channel. The maximum amplitude of each channel across all time points in the enhanced anomaly map is used as the equalization benchmark for that channel. The amplitude of each time point of that channel is then divided sequentially by the corresponding equalization benchmark to achieve intra-channel normalization, ensuring that the peak values ​​of each channel are uniformly raised to the same level. After equalization, the amplitudes of different channels can be directly compared on the same scale. When the maximum amplitude of a channel is zero, the equalization results for all nodes of that channel are directly set to zero to avoid numerical anomalies. The amplitude values ​​of all channels after equalization are arranged in order of channel identifier and time point to obtain the corrected anomaly map. The number of rows in the corrected anomaly map equals the number of channels in the enhanced anomaly map, and the number of columns equals the number of time points. The structure is identical to that of the enhanced anomaly map. Figure 1 This leads to a uniformity in the dynamic range of each channel, ensuring a fair and comparable benchmark for each channel when the 75th percentile of the amplitude of all positions in the corrected anomaly map is used as the threshold during the high-response region screening stage. For example, a patient's weak anomaly node in a certain channel had an amplitude of 0.41 after stretching, but the maximum response value of the same channel's enhancement anomaly map was 0.85. After equalization, the node's amplitude was normalized to 0.48, placing it at a similar level to a high-response node in another channel with an equalized amplitude of 0.51. This establishes the comparability of the two types of nodes in the corrected anomaly map. Without equalization, the former would still be at a disadvantage in high-response region screening due to its lower overall channel gain, and the weak anomaly signal would still face the risk of being suppressed. The coverage of weak anomaly regions in the corrected anomaly map is more complete than that in the enhancement anomaly map.

[0042] Toxicity response parameters are generated by screening high-response regions based on the calibration anomaly map. High-amplitude locations in the calibration anomaly map clinically correspond to time periods of dense superposition of multiple toxicity abnormalities. In glioma patients undergoing radiotherapy and chemotherapy, the bone marrow suppression trough and the simultaneous decline in neurological function scores after temozolomide administration typically form high-amplitude regions with multiple channels superimposed on the calibration anomaly map. The core objective of the screening is to extract toxic events from the continuous amplitude field into a discrete set of feature descriptions. The screening uses the 75th percentile of the amplitude across all locations in the calibration anomaly map as the threshold. Locations exceeding the threshold constitute a candidate set. Isolated candidate points with a time span less than one acquisition interval are considered noise and discarded. The remaining continuous high-response segments are confirmed as valid target regions. For each effective high-response segment, three features—peak amplitude, duration, and number of channels involved—were extracted from the corrected anomaly map. These three features constitute a toxicity description triplicate for that segment. Peak amplitude reflects the superposition intensity of the lowest point of myelosuppression and the decline in neurological function score; duration reflects the extent of interference of the toxic event with the radiotherapy and chemotherapy rhythm; and the number of channels involved reflects the combined impact of this toxic event on the hematopoietic system and neurological function. For a glioma patient, the white blood cell count, absolute neutrophil count, and neurological function score were analyzed in the corrected anomaly map from days 14 to 21 after the third course of temozolomide administration. The high response triggered by the simultaneous activation of one channel, with a high peak amplitude and a duration spanning a complete dosing window, indicates that the overall weight of this ternary group is at the highest level in the toxicity response parameters. The corresponding toxic event is given priority in cross-dimensional coupling, and the ternary group is assigned a high priority weight in the toxicity response parameters. In contrast, another patient only had a high amplitude in the platelet count channel in a certain segment, and the duration was less than one treatment course. The overall weight of the ternary group was lower, reflecting that this toxic event was limited to a single hematopoietic branch and did not form multi-system combined inhibition. In the toxicity response parameters, it only participated in subsequent coupling as a secondary feature.

[0043] This study extracts the characteristics of sudden drops and rebounds in bone marrow indicators from complete blood count (CBC) data to generate early warning indicators for compensatory depletion. After chemotherapy, bone marrow indicators typically undergo a two-stage process: a rapid decline to a trough, followed by a rebound to a post-trough peak driven by the bone marrow's compensatory mechanism. When the post-trough peak continues to decrease across multiple consecutive treatment cycles, it indicates a gradual decline in bone marrow compensatory capacity with cumulative damage over treatment. This progressive decline is a typical signal of bone marrow reserve depletion. The study extracts sudden drops and rebounds in bone marrow indicators from CBC data for three indicators: white blood cell count, absolute neutrophil count, and platelet count. These three indicators cover different hematopoietic branches, and the combined assessment of multiple indicators reduces the probability of misjudgment caused by fluctuations in a single indicator. For each indicator, the study identifies the trough time point and the post-trough peak time point for each treatment cycle. The trough value is the lowest value at each sampling point within the treatment cycle, and the rebound amplitude is the difference between the post-trough peak and the trough value. If the rebound amplitude of a certain indicator shows a monotonically decreasing trend over three or more consecutive treatment cycles, and the rebound amplitude in the current treatment cycle is less than 70% of the average of the previous three cycles, it is determined that the indicator has triggered compensatory attenuation in the current treatment cycle. When at least two of the three indicators simultaneously meet the compensatory attenuation condition, the current treatment course is marked as a compensatory exhaustion warning node. The compensatory exhaustion warning marker records the triggering treatment course number, the names of the indicators that meet the conditions, and the percentage decrease in their respective recovery rates. The number of indicators that meet the conditions and their respective decrease rates jointly determine the severity level—when all three indicators are met and the decrease rates are all significant, it is judged as high severity; when only two indicators are met, it is judged as medium severity; and when the decrease rate is close to the 70% threshold and only two indicators are critically met, it is judged as low severity. The triggering treatment course number serves as an index for aligning the compensatory exhaustion warning marker with the adverse reaction profile timeline during the risk mapping stage. The severity level provides a quantitative basis for judging the degree of bone marrow reserve depletion in the graded warning rules.

[0044] Step S13: Use the toxicity response map to extract the time period when the liver and kidneys deviate from the toxicity index to locate the toxicity occurrence segment. Based on the toxicity occurrence segment, identify the time period of toxicity retention and aggregation to generate toxicity retention and aggregation markers. Based on the toxicity retention and aggregation markers and toxicity response parameters, perform cross-dimensional coupling to form a toxicity feature matrix.

[0045] In some embodiments, the step of using the toxicity response map to extract the time period of divergence between liver and kidney and toxicity indicators to locate the toxicity occurrence segment includes: mapping the toxicity response map to the clinical indicator dimension space to generate an indicator distribution prediction map; identifying normal areas of liver and kidney indicators based on the indicator distribution prediction map to obtain divergence candidate time periods; using the divergence candidate time periods to perform indicator change rate constraint screening to form rate correction time periods; and performing local peak screening on the rate correction time periods to locate the toxicity occurrence segment.

[0046] The toxicity response map is mapped to the clinical indicator dimension space to generate an indicator distribution prediction map. The clinical indicator response model is pre-calibrated using historical treatment data of similar patients. The correspondence between the overall toxicity magnitude and the expected abnormality of each clinical indicator is encoded as a dimension projection weight matrix. The magnitude vector of each time node in the toxicity response map is projected onto this matrix to form predicted intensity values ​​in each clinical indicator dimension. Liver and kidney indicators and blood toxicity indicators present their prediction results independently in their respective dimensions. The predicted values ​​of each dimension together constitute the clinical indicator distribution vector at that time node. The projection results of all time nodes in the toxicity response map are arranged in chronological order to form the indicator distribution prediction map. The rows of the indicator distribution prediction map correspond to the time nodes, and the columns correspond to each clinical indicator dimension. The numerical sequences of the liver and kidney indicator dimensions and the numerical sequences of the toxicity indicator dimensions are expanded synchronously on the time axis. A patient's toxicity response map showed a significant increase in amplitude during a certain period. The indicator distribution prediction map showed high predictive strength in both the liver / kidney and toxicity dimensions during this period, indicating synchronous abnormalities in both indicators. Subsequent discrepancy assessment will rely on the presence of an inverse distribution between the two dimensions to identify the true toxicity occurrence window. Conversely, the same patient's toxicity response map also showed elevated amplitude during another period, but the liver / kidney dimension values ​​in the indicator distribution prediction map were significantly lower than the toxicity dimension values, showing an inverse distribution. This suggests that this period occurred against a backdrop of normal liver and kidney function, representing a primary toxicity event. The values ​​in each dimension of the indicator distribution prediction map were normalized to unify dimensions, and the time span of the indicator distribution prediction map remained completely consistent with that of the toxicity response map.

[0047] Based on the indicator distribution prediction map, candidate time periods for deviation are obtained by identifying normal areas of liver and kidney indicators. Both normal liver and kidney function and abnormal toxicity indicators must be simultaneously met to trigger a deviation judgment; neither condition can be omitted. The combined judgment of these two conditions effectively eliminates false positive interference when either condition is met alone. The upper and lower limits of the clinical reference range and the indicator distribution prediction map are mapped to the 0-1 interval using the same normalization transformation, ensuring that subsequent judgments are performed under a unified dimension. Time nodes in the indicator distribution prediction map where the liver and kidney indicator dimensions fall within the normalized clinical reference range are identified as normal liver and kidney nodes. Based on this, the toxicity indicator dimensions of the same node are examined in the indicator distribution prediction map to see if they exceed the upper limit of the normalized clinical reference range. Nodes that simultaneously meet both conditions are marked as deviation nodes. The determination of deviation nodes relies entirely on the normalized values ​​of each dimension of the indicator distribution prediction map. Continuously distributed deviation nodes on the time axis are merged into a single candidate time period. The start and end nodes of the candidate time period determine the time range of that period. Single-node deviations with a time span less than one acquisition interval are removed as isolated points during merging to eliminate occasional disturbances. In a certain patient's indicator distribution prediction chart, the liver and kidney dimension values ​​were consistently within the reference range at multiple consecutive time points during a certain treatment course, while the toxicity indicator dimension values ​​consistently exceeded the upper limit. After merging, a divergence candidate period spanning multiple nodes was formed. However, in another treatment course of the same patient, the liver and kidney dimension values ​​exceeded the reference range at a certain time point. Even if the toxicity indicator dimension was also abnormal during that time point, the divergence judgment was not triggered. The difference in the judgment results of the two situations clearly distinguishes the temporal distribution of primary toxic events and secondary toxicity caused by liver and kidney stress.

[0048] Rate correction periods are formed by constraining the rate of change of indicators through deviation from candidate periods. The rate of change is obtained by taking the difference between the values ​​of adjacent nodes in the toxicity indicator dimension from the indicator distribution prediction map and dividing it by the time interval. Nodes whose rate exceeds three times the standard deviation of the overall mean of the current deviation candidate period are identified as rate anomalous nodes. The rate anomalous node and one node before and after it are removed from the deviation candidate period, and the remaining consecutive nodes are merged to form the rate correction period. The threshold is set at three times the standard deviation of the mean of the time period, rather than a fixed value. This ensures that the screening criteria are adaptively adjusted when there are significant differences in the rhythm of change among different patients and different treatment courses. For example, a glioma patient experienced a sudden drop in white blood cell count on days 7 to 14 after temozolomide administration due to a temporary infection. The overall change within the candidate time period was already relatively drastic. If a fixed threshold were used, a large number of real bone marrow suppression nodes, in addition to the transient disturbance node caused by the infection, would be mistakenly eliminated. The adaptive threshold, based on the distribution of the time period, identifies the sudden drop node and one node before and after it as a rate abnormality node and eliminates it, while reasonably retaining the remaining continuous bone marrow suppression nodes as valid signals. On the other hand, another patient's blood routine changes were stable during the same period, and the mean and standard deviation of the rate throughout the candidate time period were low. The adaptive threshold was tightened accordingly, eliminating only a few extreme rate nodes, effectively avoiding the problem of missing real toxic signals during stable periods. When all nodes within a candidate time period are determined to be rate abnormalities, the entire candidate time period is removed and does not enter the peak screening stage. The time range of the rate correction period is usually less than or equal to the corresponding deviation candidate period. The rate of change of toxicity indicators at each node is within a reasonable range, and the amplitude distribution tends to be stable after removing instantaneous noise.

[0049] Local peak screening is used to locate toxicity-affected segments during the rate correction period. A local peak node is identified when its toxicity index magnitude is simultaneously higher than the magnitudes of the two nodes preceding and following it within the rate correction period. Nodes meeting this maximum condition are confirmed as local peak nodes. The segment boundary is determined by extending the amplitude to both sides until it drops to 50% of the peak value. This 50% threshold strikes a balance between sufficient segment coverage and clear boundaries between adjacent peak segments. Each local peak node is used to determine the start and end boundaries of its corresponding segment according to the above rules. When the boundaries between two adjacent local peaks overlap, the time point where the amplitude between the two peaks equals the average amplitude of the two peaks is taken as the dividing point to prevent adjacent toxic segments from merging on the time axis, which would distort the segment boundaries. When no local peak nodes are detected within the rate correction period, the entire rate correction period is retained as a toxicity occurrence segment to capture plateau-type toxicity patterns with sustained high amplitude but no obvious single-peak structure. For example, in a glioma patient, after the third course of temozolomide administration, bone marrow suppression remained in the severe range for 14 days, with white blood cell and platelet counts simultaneously remaining low without significant rebound peaks. The entire rate correction period is retained to fully record this plateau-type bone marrow suppression accumulation range. In another patient, two obvious local peaks appeared within the rate correction period of the same course of treatment, corresponding to two bone marrow suppression troughs on days 7 and 14 after administration, respectively. The amplitude between the two peaks dropped to 45% of the peak value. Based on the boundary judgment rules, it was divided into two independent toxicity occurrence segments, and their start and end points and peak intensities were recorded separately. All segments confirmed after local peak screening within the entire rate correction period constitute toxicity occurrence segments, and the start and end time points of each segment define the detection window for identifying the toxicity retention and accumulation period.

[0050] Toxicity retention clustering markers are generated based on the identification of toxicity occurrence segments and the generation of toxicity retention clustering periods. The determination of toxicity retention clustering relies on the joint verification of two independent conditions: first, whether the recovery period between adjacent segments is sufficient; and second, whether the toxicity amplitude has returned to the background level by the end of the previous segment. Meeting only one condition is insufficient for clustering. The interval between two adjacent toxicity occurrence segments is measured by the time span from the end of the previous segment to the beginning of the next segment. An interval shorter than 50% of a complete treatment course is considered too short. Given a short interval, the marker is examined to see if the amplitude at the end of the previous toxicity occurrence segment is still higher than the background mean of the corresponding time period in the indicator distribution prediction map. If the amplitude remains high, toxicity retention is determined. Adjacent toxicity occurrence segments that simultaneously meet the criteria of both short interval and toxicity retention are marked as toxicity retention clustering periods. Toxicity retention clustering markers are generated based on the coverage of the toxicity retention clustering period. When multiple consecutive segments are merged, the toxicity retention clustering markers cover the complete time range from the beginning of the first segment to the end of the last segment. One patient had two adjacent toxicity episodes separated by only three days, and the amplitude was still significantly high at the end of the first episode, indicating toxicity retention and accumulation. Another patient also had two adjacent episodes with a short interval, but the amplitude had completely returned to background levels by the end of the first episode, thus not meeting the toxicity retention criterion. This difference in criterion demonstrates the ability of dual-condition joint validation to distinguish different toxicity dissipation patterns. When multiple consecutive toxicity episodes meet the toxicity retention criterion in every pair, these episodes and their intervals are merged into a single consecutive toxicity retention accumulation period, covering the complete time range from the beginning of the first episode to the end of the last episode. The toxicity retention accumulation indicator is determined by three factors: the start and end times of the toxicity retention accumulation period, the number of toxicity episodes involved, and the peak amplitude within the period. It is graded according to severity; the more episodes involved and the higher the peak amplitude, the higher the severity.

[0051] A toxicity feature matrix is ​​formed by cross-dimensional coupling based on toxicity retention clustering identifiers and toxicity response parameters. Cross-dimensional coupling integrates the clustering dimension information of toxicity retention clustering identifiers and the intensity dimension information of toxicity response parameters into the same feature vector node by node on a unified time axis. The clustering dimension value comes from the clustering intensity S of each clustering period in the toxicity retention clustering identifier, which is determined by the formula S=N×P / P_max, where N is the number of toxicity occurrence segments involved in the clustering period, P is the peak amplitude within the clustering period, and P_max is the highest peak value among all toxicity occurrence segments. P and P_max are compared with the same reference benchmark to ensure that the ratio is in the range of [0,1]. The value of S ranges from 0 to N, and S is assigned zero for nodes outside the clustering period. The intensity dimension value comes from the expanded value of the peak amplitude, duration, and number of channels involved in each toxicity description triple in the toxicity response parameters at the corresponding node. When a node is covered by multiple triples, the average value is taken as the representative value of the intensity dimension. The values ​​of each dimension are independently normalized and mapped to the range of 0 to 1. For example, in a glioma patient at the end of the fourth treatment cycle, bone marrow suppression and neurological function scores declined simultaneously, with a very short interval between the previous toxic event. The corresponding node showed high levels of both aggregation and intensity, forming a typical high-toxicity row vector in the toxicity feature matrix. Conversely, in the first treatment cycle of the same patient, a single node showed only a single fluctuation in platelet count, with zero aggregation and only one channel involved. The structural differences in the feature vectors of these two types of nodes clearly distinguish between isolated single-system toxic events and multi-cycle, sustained aggregation toxic events in the toxicity feature matrix. The feature vectors of all time nodes are arranged chronologically to form the toxicity feature matrix. Rows correspond to time nodes, and columns correspond to four feature dimensions: aggregation intensity, peak amplitude, duration, and number of channels involved. The number of rows equals the total number of sampling nodes within the radiotherapy and chemotherapy cycle.

[0052] Step S14: Perform abnormal pattern clustering on the toxicity feature matrix to determine the toxicity category distribution. Generate toxicity adaptation identifiers by identifying abrupt changes in the scoring after stagnation in the toxicity category distribution. Extract toxicity gradient features based on the toxicity adaptation identifiers to construct an adverse reaction map.

[0053] In some embodiments, the step of determining the toxicity category distribution by performing abnormal pattern clustering on the toxicity feature matrix includes: generating a sample distance metric based on the toxicity feature matrix; detecting high-toxicity feature samples within low-toxicity categories from the sample distance metric to generate a masking correction factor; performing adaptive initialization of centroids on the sample distance metric and the masking correction factor to form dynamic cluster centers; and determining the toxicity category distribution by classifying samples according to the dynamic cluster centers.

[0054] A sample distance metric is generated based on the toxicity feature matrix. The feature vectors of any two time points in the toxicity feature matrix are quantized into a distance value between the two nodes using a weighted Euclidean distance. The weighted Euclidean distance is based on the numerical differences of four feature dimensions: clustering intensity, peak amplitude, duration, and number of involved channels. Each dimension is assigned a weight according to its dispersion across all time points; dimensions with higher dispersion contribute more to the distance calculation, making the distance value more biased towards reflecting the feature differences with the highest discriminative power among the four dimensions. The pairwise distance values ​​between all nodes are filled into a square matrix to form the sample distance metric. The diagonal elements of the square matrix are zero, and the matrix is ​​symmetric about the diagonal, providing a unified interface for accessing node similarity for subsequent coarse clustering, center initialization, and attribution. For example, in a glioma patient, at the end of the fourth course of temozolomide treatment, the white blood cell count and absolute neutrophil count simultaneously dropped to their lowest values, and the neurological function score declined concurrently. This corresponds to nodes with high clustering intensity, involving a large number of channels and large peak amplitudes. These nodes are relatively close to each other in the sample distance metric, but relatively far from the low-toxicity nodes from the first course. This difference in distance structure directly reflects the differentiation of toxicity patterns caused by the cumulative damage from multiple courses of temozolomide treatment. Locally dense regions in the sample distance metric indicate the concentrated distribution of nodes with similar toxicity characteristics. Masked samples within the low-toxicity category are significantly closer to nodes in the high-toxicity category than to nodes within the same category in the sample distance metric. These masked samples typically correspond to transitional nodes where bone marrow suppression indicators are temporarily elevated during a certain course of treatment but have not yet formed a sustained cluster.

[0055] Masking correction factors are generated by detecting high-toxicity feature samples within low-toxicity categories from sample distance metrics. Initial category division uses K-means to perform coarse clustering based on the sample distance metrics. Coarse clustering divides all nodes into several initial categories according to the distance distribution of each node in the sample distance metrics. Low-toxicity categories are determined by those with low clustering strength dimension values ​​in the coarse cluster center feature vectors. Nodes whose clustering strength dimension and peak amplitude dimension values ​​in their feature vectors both exceed one standard deviation of the global mean are identified as high-toxicity feature samples. Nodes within low-toxicity categories that meet this condition are identified as masked samples. The proportion of masked samples in all nodes of a low-toxicity category reflects the severity of masking interference in that category; a higher proportion indicates a greater shift in the cluster center of that category. The masking correction factor uses this proportion as a value to measure the degree of masking in each category. The difference in masking correction factor values ​​between low-toxicity categories reflects the degree of masking in different low-toxicity categories. Categories with higher values ​​receive a greater degree of reverse correction during the dynamic cluster center initialization stage. The masking correction factor for the high toxicity category is set to zero, while each low toxicity category corresponds to a masking correction factor value. In a certain patient, the proportion of masked samples in a certain low toxicity category reaches 25%, corresponding to a high masking correction factor value. The cluster center of this category receives a significant reverse correction during the adaptive initialization phase to offset the pulling effect of the masked samples. After correction, the center position shifts towards the dense area of ​​low clustering intensity nodes, which is closer to the distribution center of the true low toxicity nodes. On the other hand, the correction magnitude for another low toxicity category with a masking correction factor value of only 0.05 is minimal, and the center position remains almost unchanged from the original position of the K-means coarse cluster.

[0056] Dynamic cluster centers are formed by adaptive initialization of centroids using sample distance metrics and masking correction factors. K-means++ selects initial candidate center nodes based on sample distance metrics, calculating the local density of each node from the sample distance metrics. Priority is given to selecting nodes in regions with higher local node density based on the sample distance metrics. This density-first strategy ensures that the initial candidate centers fall as close as possible to the core regions of node clusters, avoiding the initial centers randomly falling into sparse regions, which could lead to slow iteration convergence or getting trapped in local optima. After the candidate nodes are determined, the masking correction factor applies a reverse offset to each low-toxicity category candidate center. The offset direction is the positive direction of the difference between the candidate center vector and the mean of the feature vector of the masked sample for that category. The offset magnitude is equal to the masking correction factor value multiplied by the magnitude of this difference vector. After correction, the candidate centers move away from the direction of masked sample clusters, closer to the centroid of the distribution of true nodes in the low-toxicity category. Candidate centers for the high-toxicity category with a masking correction factor of zero are not corrected and are directly selected using the K-means++ result. The initialization processing method for both categories is adaptively determined by the masking correction factor value, ensuring that high-toxicity category centers remain in the high-cluster density area while low-toxicity category centers are effectively pulled back to the vicinity of the true low-toxicity node distribution centroid. A candidate center for a low-toxicity category, biased towards the high-cluster density area due to the masking sample, is moved to the low-cluster density area after applying a reverse offset using the masking correction factor, correcting the center shift caused by the masking sample. After correction, the inter-class distance between this low-toxicity and high-toxicity category candidate centers is significantly increased, and the feature space boundaries of the two categories are clearer. The set of candidate centers after correction for all categories constitutes dynamic cluster centers. Compared to the standard initialized centers, dynamic cluster centers are closer to the true distribution centroid of each category, and the inter-class distance between dynamic cluster centers is clearer between high- and low-toxicity categories.

[0057] The toxicity category distribution is determined by classifying samples based on dynamic cluster centers. Weighted Euclidean distances are calculated between each node in the toxicity feature matrix and each center vector of the dynamic cluster centers. Nodes are assigned to the category of the center closest to them. The center vectors of the dynamic cluster centers serve as reference anchors for the first round of classification, ensuring the masking correction effect continues into subsequent iterations. After the first round of classification, the average value of the feature vectors of nodes within each category is used to update the cluster centers. The updated centers replace the dynamic cluster centers in the next iteration. This classification and center update process is repeated until the center displacement in two consecutive rounds is less than 0.1% of the initial maximum inter-cluster distance, at which point the iteration converges. After convergence, the average clustering strength of the feature vectors within each category is sorted from low to high. The sorting results are assigned toxicity level labels to each category. The category with the highest toxicity level corresponds to a set of nodes with high clustering strength and peak amplitude dimensions, while the category with the lowest toxicity level corresponds to a set of nodes with low values ​​in both dimensions. Intermediate toxicity levels sequentially cover the transitional toxicity patterns. The toxicity category distribution includes the category attribution and corresponding toxicity level label for each time point. The toxicity category distribution unfolds on the timeline as a sequence of category labels for each node. At the end of the third treatment course, a patient was continuously classified as high-toxicity at multiple points, while at the end of the first treatment course, all points were concentrated in the low-toxicity category. These two sets of nodes exhibit distinctly different label sequences in the toxicity category distribution. The high-toxicity category label appeared continuously at the end of the third treatment course but was almost absent in the first course. The label sequence in the intermediate treatment courses alternated between low-toxicity and moderate-toxicity categories. This label evolution trajectory from low-toxicity to moderate-toxicity and then to high-toxicity accurately records the complete process of the patient's toxicity load continuously accumulating until concentrated exposure as the treatment progressed. The densely populated high-toxicity node areas in the toxicity category distribution serve as the key detection range for responding to retention trends during the scoring retention identification phase.

[0058] In some embodiments, generating a toxicity adaptation identifier by identifying a sudden change in score retention after toxicity category distribution includes: generating a toxicity score intensity distribution based on the toxicity category distribution; identifying response retention trends from the toxicity score intensity distribution to locate adaptation candidate areas; identifying score change periods from the adaptation candidate areas to determine indicator deviation areas; and calibrating toxicity adaptation intensity based on the indicator deviation areas to generate a toxicity adaptation identifier.

[0059] A toxicity score intensity distribution is generated based on the toxicity category distribution. The toxicity score for each node is determined by a weighted sum of the toxicity level label value of its category and the peak amplitude dimension value of that node. The level label value is normalized to the maximum value and mapped to the 0-1 interval before being weighted. The weight of the level label value is 0.6, and the weight of the peak amplitude dimension value is 0.4. The weighted sum is globally normalized and mapped to the 0-1 interval. Higher values ​​indicate a heavier toxicity load. Different nodes within the same category exhibit different score results due to differences in peak amplitude values. This solves the problem that discrete labels cannot reflect the intensity differences of nodes within a category, allowing the toxicity score intensity distribution to preserve the rank relationship between categories while providing fine intensity resolution for each node within a category. The toxicity scores of all nodes in the toxicity category distribution are arranged chronologically to form the toxicity score intensity distribution, continuously showcasing the evolution trajectory of the toxicity load throughout the patient's entire treatment course. High score intervals correspond to treatment stages where high-toxicity category nodes are concentrated. At the end of the fourth course of treatment, a patient's toxicity score intensity distribution consistently remained above 0.8 across multiple nodes, a stark contrast to the low-score periods in the first three courses. The high-score segments in the time series corresponded perfectly to the densely populated high-toxicity category nodes in the toxicity category distribution, mutually reinforcing each other and suggesting a significant increase in toxicity accumulation during this stage. In contrast, the toxicity score intensity distribution in the first course remained relatively stable between 0.2 and 0.35, significantly contrasting with the high-score segments at the end of the fourth course. The difference between these two score levels directly reflects the cumulative effect of multiple treatment courses on the overall toxicity load. The plateau-like high-score segments in the toxicity score intensity distribution over time are the core detection target for response retention trend identification. Segments with consistently high mean scores and low variation between nodes suggest that the body has achieved a certain adaptive homeostasis to the current toxicity level.

[0060] Adaptation candidate regions are identified by analyzing the toxicity score intensity distribution to pinpoint response retention trends. The moving average of each node in the toxicity score intensity distribution is used as a window with a length of 15 nodes. Periods where the mean within the window is 0.5 times the global mean and the standard deviation within the window is 30% lower than the global standard deviation are considered retention candidate windows. These two conditions ensure that the candidate window exhibits both high score levels and low volatility. Segments with only high scores but drastic fluctuations do not meet the low volatility condition, indicating that the segment is still in a period of toxicity volatility rather than a true adaptive steady state. Adjacent retention candidate windows with an interval of no more than 3 nodes are merged into the same retention period. A merged retention period with a time span of at least 5 nodes is confirmed as an adaptation candidate region. Retention periods with a time span of less than 5 nodes are considered transient fluctuations rather than true retention trends and are not included in the adaptation candidate region. The time range of the adaptation candidate region is also extended outward by one sampling interval as an extended window for abrupt change detection, ensuring that the instantaneous transition nodes from the retention platform to abrupt changes are fully included in the slope analysis. In a patient's toxicity score intensity distribution, the mean of 12 consecutive nodes during a certain period of the third course of treatment was consistently higher than the global mean, with extremely low standard deviations between nodes. After merging, this met the time span requirement and was identified as a candidate adaptation region. Within this candidate adaptation region, the score remained consistently around 0.75, fluctuating no more than 0.05, indicating that the body had developed temporary tolerance to continuous toxic stimulation at this stage. Although the toxicity load was high, it no longer caused drastic changes in the indicators in the short term. When multiple candidate adaptation regions exist in the toxicity score intensity distribution, each candidate adaptation region is independently calibrated. The intervals between candidate adaptation regions are usually accompanied by score fluctuations and rebounds, and each candidate adaptation region independently enters the subsequent slope analysis stage.

[0061] For example, the step of identifying the index deviation region from the score abrupt change period in the adaptation candidate region includes: performing a score time-series slope calculation on the adaptation candidate region to generate a slope change sequence; locating the time node where the absolute value of the slope exceeds the threshold based on the slope change sequence to generate a mutation candidate point set; performing dual verification of the duration and amplitude of the mutation candidate point set to filter effective mutation periods; and labeling the index deviation region according to the distribution range of the effective mutation periods.

[0062] The slope change sequence is generated by calculating the temporal slope of the adaptation candidate region. The instantaneous slope of each node is obtained by dividing the difference in toxicity score intensity distribution between adjacent nodes by the time interval. Positive values ​​indicate an increase in score at that node, while negative values ​​indicate a decrease. The larger the absolute value, the more drastic the change. The absolute value of the slope is generally low during the plateau phase of the adaptation candidate region, while the absolute value increases sharply near the sudden change node, forming a prominent peak. The difference in the absolute value of the slope between the two types of segments is the core signal for identifying the sudden change node. The instantaneous slopes of each node are arranged in chronological order to form a slope change sequence. The sequence length is one element less than the number of nodes in the adaptation candidate region. Segments with generally low absolute values ​​in the sequence correspond to the score stagnation phase, while positions with sudden increases in absolute values ​​correspond to the nodes where the score changes drastically. In a patient's adaptation candidate region slope change sequence, the absolute values ​​of the first 10 elements were all below 0.05, while the absolute value of the 11th element suddenly increased to 0.35. This sudden increase formed a significant peak in the slope change sequence, with a seven-fold difference in absolute value compared to the first 10 elements, indicating a significant abrupt change in toxicity score at this point. The alternation of positive and negative slopes reflects the temporal distribution of directional changes in the score within the adaptation candidate region. The pattern of a sudden jump after a sustained low absolute value in the slope change sequence is a typical characteristic distinguishing a true abrupt change from random fluctuations. Short-term consecutive positive slope peaks indicate a rapid increase in score, while short-term consecutive negative slope peaks indicate a rapid decrease in score. These two patterns are marked and retained in the slope change sequence with different symbols. The global mean and standard deviation of the absolute values ​​of each element in the slope change sequence jointly determine the detection threshold of the mutation candidate point set. The lower the average level of the absolute value of the slope in the plateau phase, the lower the detection threshold of the mutation candidate points, and the sensitivity automatically increases with the sequence's own distribution.

[0063] Based on the slope change sequence, time nodes where the absolute value of the slope exceeds a threshold are used to generate a mutation candidate point set. The detection threshold is set as the mean of the absolute values ​​of all elements in the slope change sequence plus twice the standard deviation. This setting places approximately 95% of normal fluctuation elements below the threshold, retaining only high-slope nodes that are statistically significantly deviated from the overall distribution. When there are large differences in slope fluctuation amplitude among different patients, the detection standard is automatically adjusted according to the sequence's own distribution, avoiding the problem of a large number of false positives for patients with high slope baselines or false negatives for true abrupt changes in patients with low slope baselines caused by a fixed threshold. Time nodes in the slope change sequence whose absolute value exceeds the detection threshold are extracted one by one, and the slope value, slope direction, and position on the adaptation candidate region time axis of each node are recorded. These three pieces of information together constitute the mutation candidate point description for that node. The set of descriptions for all nodes exceeding the threshold constitutes the mutation candidate point set. In a patient's adaptation candidate region slope change sequence, there are 3 positive and 1 negative threshold nodes. The mutation candidate point set contains 4 records. The 3 positive records appear consecutively in a segment where the score rises rapidly, corresponding to a rapid jump in toxicity score intensity from 0.72 to 0.89 in this segment. The 1 negative record appears at an adjacent node after the rise ends, corresponding to a slight drop in score of 0.04. The number of positive records in the mutation candidate point set is significantly greater than that of negative records, suggesting that the adaptation candidate region is mainly characterized by abrupt changes. The appearance of the negative record identifies a brief adjustment node after the abrupt rise. The two types of records together completely describe the temporal distribution characteristics of the main body of the abrupt change and the subsequent adjustment tail segment.

[0064] The set of candidate mutation points is subjected to dual verification of duration and amplitude to screen for effective mutation periods. The duration verification requires that the candidate mutation points appear at least twice on the time axis, either continuously or with an interval of no more than one node. Isolated single-node mutations are judged as occasional disturbances, and isolated candidate points that do not meet the duration requirement are removed from the set of candidate mutation points. The amplitude verification requires that the absolute value of the difference between the first and last nodes of the toxicity score intensity distribution within the corresponding time period exceeds 30% of the average score within the adaptive candidate region. This absolute value of the difference between the first and last nodes is used as the unified definition value of the amplitude change ΔS of the indicator deviation area. If the amplitude change is insufficient, although the slope is abnormal, the overall score change is limited, and it is also removed from the set of candidate mutation points. The dual verification ensures that the final effective mutation period has sufficient time span and amplitude support. Mutation candidate points that pass both validations are confirmed as valid mutation nodes. Adjacent valid mutation nodes on the timeline are merged into a single valid mutation period. The start and end nodes of a valid mutation period are taken as the earliest and latest valid mutation node positions within the merged range, respectively. The direction of abrupt change is determined by the slope direction of the majority of valid mutation nodes within the period. When the number of nodes in the positive and negative directions is equal, the one with the larger absolute value of the slope is used. In a patient's mutation candidate point set, a group of 3 positive and 1 negative overthreshold candidate points underwent double validation. The positive candidate points all met the duration condition, and their corresponding amplitude changes reached 42% of the average score of the adaptation candidate region. They were merged into a single effective mutation period of 3 consecutive nodes. The negative candidate point's amplitude change was only 6% of the average score of the adaptation candidate region, and it did not meet the amplitude validation condition, so it was removed. The time range of the valid mutation period determines the core boundary of the abrupt change interval. The average absolute value of the slope of each node within the valid mutation period is denoted as the average rate of change R of the deviation area of ​​the indicator. R reflects the rate intensity of the abrupt change process; the larger the R value, the more rapid the abrupt change.

[0065] The index divergence region is generated based on the distribution range of the effective mutation periods. The start and end nodes of the effective mutation periods determine the core range of the abrupt change interval. The index divergence region is extended to both sides of the effective mutation period by one sampling interval. The transition features near the start and end of the abrupt change are thus fully preserved without being truncated due to the narrow boundary. The scores at the beginning and end nodes of the extended region are usually still in a stable state before and after the abrupt change, providing a stable boundary reference point for calculating the difference between the beginning and end of the abrupt change amplitude. When there are multiple effective mutation periods within a single adaptation candidate region, the corresponding index divergence region is independently calibrated for each effective mutation period. When the interval between two adjacent index divergence regions is less than two nodes, they are merged into a continuous region. The start and end boundaries after merging are taken from the earliest start node and the latest end node of the two original regions, respectively, to prevent fragmented records from occurring between adjacent abrupt change regions due to the influence of the sampling interval. A patient had two effective mutation periods within a candidate adaptation region, separated by only one node. Before merging, the score at the starting node of the first segment was 0.72, rising to 0.80 at the ending node, and slightly decreasing to 0.77 at the interval node. The score at the starting node of the second segment continued to rise to 0.85. These four consecutive nodes constituted a complete step-like surge. After merging, a single indicator deviation region covering the complete abrupt change was formed. The magnitude of the change within this region was calculated as the absolute value of the difference between the scores at the beginning and end nodes (0.85 minus 0.72, equaling 0.13), fully recording the actual magnitude of the abrupt change. Without merging, the magnitude of the change calculated separately for each segment would be too small, potentially falling below the significance threshold of the toxicity adaptation strength formula and thus underestimated. After processing all candidate adaptation regions, the deviation regions for each indicator were identified, with the starting and ending node time coordinates, the direction of the abrupt change, the magnitude of the change ΔS, and the average rate of change R as the main attributes. ΔS was the absolute value of the difference in toxicity scores between the beginning and end nodes of the region, and R was the average of the absolute values ​​of the slopes of each node within the region.

[0066] Toxicity adaptation intensity is calibrated and a toxicity adaptation label is generated based on the deviation region of the indicator. The magnitude change ΔS and the average rate of change R in the deviation region of the indicator are jointly used to determine the adaptation intensity value, ΔS_norm=(ΔS−ΔS_min) / (ΔS_max−ΔS_min), R_norm=(R−R_min) / (R_max−R_min), where ΔS_min and ΔS_max are the minimum and maximum values ​​of the magnitude change of all indicators in the deviation region, and R_min and R_max are the minimum and maximum values ​​of the average rate of change of all indicators in the deviation region. After normalization, ΔS_norm and R_norm are both mapped to the interval of 0 to 1. The adaptation intensity A is determined by the formula A=0.5×ΔS_norm+0.5×R_norm, and the value of A ranges from 0 to 1. The higher the value of A, the greater the magnitude and the more rapid the rate of change. The adaptation intensity A of each indicator deviation area is divided into three levels according to the value: A above 0.7 is judged as high intensity sudden change, between 0.4 and 0.7 is judged as medium intensity sudden change, and below 0.4 is judged as low intensity sudden change. High intensity sudden change indicates that the body's adaptation to toxic stimulation has been severely disrupted, while medium and low intensity sudden change indicates that the adaptation state has been locally disturbed but the overall homeostasis has not completely collapsed. The toxicity adaptation markers record the time coordinates, direction of change, intensity of adaptation, and start and end points of the corresponding candidate adaptation areas for each indicator's deviation area. The direction of change is divided into two categories: sudden increase and sudden decrease. If a patient's deviation area for a certain indicator has a large amplitude and high rate of change, and the corresponding adaptation intensity A exceeds 0.7, it is judged as a high-intensity sudden increase type sudden change. The complete time coordinates from the start point to the end point of this area are recorded, and the range of the candidate adaptation area that triggered this sudden change is marked, thus generating a complete high-intensity sudden increase type toxicity adaptation marker entry. Another patient has a medium-intensity sudden decrease type sudden change with an adaptation intensity A of 0.52 during the same course of treatment, indicating that after a brief period of toxicity relief, the body's high toxicity adaptation state has loosened locally, but the fundamental toxicity level has not yet fallen back.

[0067] Adverse reaction atlases are constructed by extracting toxicity gradient features from toxicity adaptation markers. Toxicity gradient features describe the amplitude evolution of each abrupt change event in the toxicity adaptation markers at both local and global time scales. Local gradients are directly assigned values ​​based on the amplitude change in the intensity distribution of toxicity scores within the deviation regions of each indicator, reflecting the amplitude intensity of a single abrupt change event. Global gradients are obtained by linearly regressing the amplitude change sequence of all indicator deviation regions onto the treatment progress; the regression slope coefficient serves as the global gradient value. A positive slope indicates that the abrupt change amplitude continues to increase with the progress of treatment, while a negative slope indicates that the abrupt change amplitude tends to narrow. The absolute value of the slope reflects the significance of this increasing or decreasing trend. The adverse reaction atlas uses the time axis of the toxicity adaptation markers as the main axis, filling in the corresponding local gradient values ​​at the deviation regions of each indicator. A global gradient trend line is overlaid on the entire map. The directional attributes of each abrupt change event are preserved as sudden increases or decreases. The start and end nodes of the adaptation candidate areas in the toxicity adaptation markers are preserved on the atlas as context anchors for the corresponding abrupt change events, enabling the adverse reaction atlas to simultaneously carry amplitude information, directional information, and context information. One patient experienced a sudden and severe increase in adverse reactions at the end of the fourth treatment cycle. The corresponding local gradient peak value was the highest in the entire treatment cycle on the adverse reaction map. In the first treatment cycle, the local gradient values ​​of the divergence areas of each indicator were all below 0.2. In the third treatment cycle, a moderate local gradient value of about 0.45 appeared. At the end of the fourth treatment cycle, it suddenly increased to 0.82. The global gradient trend line showed a significant positive slope. The linear regression coefficient between the treatment progress and the magnitude of the sudden change reached 0.31, indicating that the patient's adverse reaction magnitude showed a significant cumulative aggravation as the treatment progressed. The end of the fourth treatment cycle was the core period of concentrated outbreak of adverse reactions in the entire radiotherapy and chemotherapy cycle. In contrast, the slope of the global gradient trend line of another patient was close to zero, indicating that the magnitude of the sudden changes in each treatment cycle remained stable and the overall adverse reactions were under control.

[0068] Step S15: Based on the toxicity response parameters and the toxicity feature matrix, identify the weight inversion feature to determine the toxicity sensitive weight. Construct a graded early warning rule based on the toxicity sensitive weight and the compensation exhaustion early warning indicator. Perform risk mapping on the graded early warning rule and the adverse reaction spectrum to output the adverse reaction early warning result.

[0069] In some embodiments, determining toxicity-sensitive weights based on weight inversion features identified by the toxicity response parameters and the toxicity feature matrix includes: performing multi-scale decomposition on the toxicity response parameters and the toxicity feature matrix to form hierarchical loss values; identifying decoupling segments in the toxicity feature matrix where loss decreases but toxicity is not alleviated to generate decoupling weights; performing adaptive weighting based on the decoupling weights and the hierarchical loss values ​​to form a decoupling loss; and performing gradient updates based on the decoupling loss to determine toxicity-sensitive weights.

[0070] Multi-scale decomposition of toxicity response parameters and toxicity feature matrices is used to generate hierarchical loss values. The prediction residuals of the toxicity response parameters are extracted for each of the short-term, medium-term, and long-term scales. These prediction residuals are normalized and mapped to the 0-1 interval. The toxicity feature matrix is ​​expanded into sub-matrices at the corresponding scale, and the values ​​of each dimension of the sub-matrices are also normalized to the 0-1 interval. The deviation between the two is calculated under a unified dimension as the loss metric for that scale. Each of the three scales independently forms a loss metric sequence, preserving the perceived error information at each time granularity. The loss metrics at each scale differ significantly in magnitude. The short-term scale has a lower overall magnitude due to its narrow time window, while the long-term scale has a higher overall magnitude due to its coverage of the entire treatment course. Direct superposition would cause the long-term scale loss to dominate the optimization direction and mask the short-term acute toxicity signal. Therefore, the loss metrics at each scale must be normalized within the scale to eliminate magnitude differences before they can be equally weighted in the hierarchical dimension. After normalization, the loss measures at each scale are arranged hierarchically to form hierarchical loss values. The number of hierarchical loss values ​​equals the number of decomposition scales. Each level corresponds to a normalized loss vector at a specific time granularity. Higher short-term hierarchical loss values ​​indicate a significant bias in the current weighting configuration's perception of acute toxicity, while higher long-term values ​​indicate a significant bias in the perception of cumulative trends. In one patient, the residuals of the toxicity response parameters and toxicity feature matrix after short-term decomposition were significantly larger, corresponding to short-term hierarchical loss values ​​that were significantly higher than those in the intermediate and long-term hierarchical levels. This suggests that the current weighting system for this patient has a significant bias in capturing acute toxicity fluctuations after a single dose, while its perception of cumulative trends over multiple treatment courses is relatively accurate.

[0071] Decoupling weights are generated by identifying decoupling segments in the toxicity feature matrix where loss decreases but toxicity remains unresolved. The node-level loss metric for each time point in the toxicity feature matrix is ​​determined by the weighted Euclidean distance between the node's feature vector and the mean of the feature vectors of its preceding nodes within a sliding window. This distance is used to identify the node distribution pattern of the decoupling segment and, together with the hierarchical loss value formed across scales, describes the loss distribution. Both are linked through node position alignment during the adaptive weighting phase. The toxicity level is determined by whether the node's aggregation intensity and peak amplitude values ​​simultaneously exceed the global mean of their respective dimensions; a node is considered highly toxic when both dimensions exceed the global mean. When the sliding mean of the loss metric decreases within five consecutive nodes and the determination of highly toxic nodes has not been alleviated during the same period, this five-node window is considered a decoupling window. Adjacent decoupling windows are merged into decoupling segments. The decoupling strength of each decoupling segment is determined by the magnitude of loss decrease and the ratio of nodes continuously occupied by highly toxic nodes. For example, in a glioma patient, the toxicity feature matrix showed a continuous decrease in loss metrics across eight consecutive nodes to 60% of its initial value during a certain period in the third treatment cycle. Simultaneously, the aggregation intensity and peak amplitude dimensions consistently exceeded their respective global averages, indicating a persistently high toxicity determination and high decoupling strength. The corresponding channel weight inversion phenomenon was continuously reinforced during the original optimization process without correction. The decoupling weights are indexed using the channels and time nodes of the toxicity feature matrix. Each position is filled with the decoupling strength value of the corresponding decoupling segment. Non-zero decoupling weight positions identify the spatiotemporal regions where weight inversion features are concentrated. The distribution differences of decoupling weight channels reflect which indicator dimensions are most prominently affected by weight inversion. Channels with higher decoupling weight values ​​receive a larger loss amplification coefficient during the adaptive weighting stage.

[0072] Decoupling loss is formed by adaptive weighting based on decoupling weights and hierarchical loss values. Adaptive weighting aligns decoupling weights and hierarchical loss values ​​across time nodes and channels. The weight of nodes corresponding to high decoupling weight positions is increased in the hierarchical loss value, amplifying the loss contribution of the decoupling segment in the optimization objective, while the loss weight of nodes outside the decoupling segment remains unchanged. The weighting uses D(c,i)×L_i as the decoupling loss component for each node and channel position, where D(c,i) is the decoupling weight value at node i in channel c, and L_i is the weighted average of the hierarchical loss value across all layers at that node. The weighted average is determined with short-term, medium-term, and long-term weights of 0.5, 0.3, and 0.2, respectively, to highlight the contribution of acute toxicity changes to the loss. The product of D(c,i) and L_i is constrained by both the magnitude of the decoupling weight and the order of the hierarchical loss value, preventing a single factor from dominating the overall distribution of the decoupling loss. The decoupling loss is formed by normalizing the decoupling loss components of all nodes. The main difference between the decoupling loss and the standard loss function is that the loss contribution of nodes in the decoupling segment is significantly amplified, while the contribution of nodes outside the decoupling segment changes little. In a certain patient, the decoupling weight is relatively high in the third treatment segment of two channels. The hierarchical loss value of the corresponding node, after being amplified by multiplication, accounts for a significantly higher proportion in the decoupling loss than other nodes. The gradients of these two channels receive greater correction in subsequent updates. However, the decoupling weight of another channel of the same patient is close to zero. Even if the hierarchical loss value of this channel is large, its contribution to the decoupling loss after multiplication is minimal, reflecting the differentiated processing of channels with different degrees of disturbance by adaptive weighting.

[0073] The decoupling loss is used to determine toxicity-sensitive weights through gradient updates. The partial derivatives of the decoupling loss with respect to the weights of each indicator dimension form a gradient vector. The sign and magnitude of each component of the gradient vector jointly indicate the correction direction and strength of each indicator weight. After the contribution of nodes in the decoupling section to the decoupling loss is amplified, the magnitude of the gradient components of the corresponding channels increases accordingly, giving indicators with severe weight reversal a stronger reverse correction push. The learning rate is dynamically adjusted based on the global variance of the decoupling loss. When the variance is large, the learning rate is appropriately reduced to prevent oscillations caused by the gradients of high-weight nodes in the decoupling section. When the variance is small, the learning rate is appropriately increased to accelerate the correction speed of weight reversal indicators. This dynamic adjustment ensures that gradient updates can still converge stably even when the distribution of decoupling loss varies significantly. Weight updates are performed simultaneously across all indicator dimensions. Convergence is determined when the change in weights across all dimensions is less than 0.001 for two consecutive rounds. After convergence, the weight values ​​of each indicator dimension constitute the toxicity-sensitive weights. For example, during a glioma patient receiving concurrent chemoradiotherapy with temozolomide, the absolute neutrophil count showed a continuous decrease in decoupling during the third and fourth cycles, but the toxicity level did not decrease. During the gradient update process, the gradient component of this dimension was continuously amplified and corrected. After convergence, the toxicity-sensitive weight of this dimension was corrected from the initial 0.12 to 0.31, reflecting that its early warning value for bone marrow reserve depletion was significantly enhanced. On the other hand, the hemoglobin concentration dimension only showed scattered decoupling windows throughout the process, with a low degree of decoupling and the gradient direction continuously indicating that the weight should be contracted. Finally, it was reduced from 0.18 to 0.09. The overall distribution of the toxicity-sensitive weights underwent a significant dimensional rearrangement compared to the initial weights, with high-weight dimensions concentrated on the most prominent indicators in the decoupling segment.

[0074] A tiered early warning rule is constructed based on toxicity sensitivity weights and compensatory depletion early warning indicators. Toxicity sensitivity weights provide the true contribution weight of each indicator to the toxicity early warning, while compensatory depletion early warning indicators provide the tiered status of bone marrow reserve depletion. Both jointly constrain the early warning triggering conditions from two dimensions: indicator perception accuracy and bone marrow reserve status. The tiered early warning rule is divided into three levels: Level 1 (high toxicity risk), Level 2 (medium toxicity risk), and Level 3 (low toxicity risk). The core trigger quantity for each level is the comprehensive toxicity score T_w, which is obtained by weighted summing of the current normalized values ​​of each indicator and the corresponding dimension values ​​of the toxicity sensitivity weights. The value ranges from 0 to 1, with higher values ​​indicating a heavier current toxicity load. The warning level is determined jointly by the severity of the compensated exhaustion warning indicator and T_w (the body's vital signs). A Level 1 warning is triggered when the severity of compensated exhaustion is high and T_w exceeds 0.7; a Level 2 warning is triggered when the severity of compensated exhaustion is high and T_w is between 0.5 and 0.7, or when the severity is moderate and T_w exceeds 0.65, or when the severity is low but T_w exceeds 0.8; in other cases, a Level 3 warning is triggered when T_w exceeds 0.45, and no warning is triggered when T_w does not exceed 0.45. One patient has a high severity warning indicator and a current T_w of 0.78, triggering a Level 1 warning; another patient has a low severity warning indicator but a high T_w of 0.82. Due to adequate bone marrow reserve, the combined conditions only meet the Level 2 warning threshold. The two patients, with the same high T_w level, receive different warning levels due to the difference in the severity of the compensated exhaustion warning indicator, demonstrating the corrective effect of the two-dimensional joint constraint on overly sensitive warning triggering.

[0075] The adverse reaction (AR) spectrum is risk-mapped according to the tiered early warning rules, and the AAR results are output. The local gradient values ​​and sudden change direction labels at each time point in the AAR spectrum are matched node-by-node according to the tiered early warning rules. Nodes with high local gradients and abrupt increases in the direction of change have correspondingly higher comprehensive toxicity scores (T_w), which, combined with the severity of the compensation depletion early warning indicator, determine whether a Level 1 or Level 2 early warning is triggered. Nodes with low local gradients or abrupt decreases in the direction of change have lower T_w, triggering a Level 3 early warning or not. After each node completes the level matching according to the tiered early warning rules, the three pieces of information—the early warning level determined by the tiered early warning rules, the comprehensive toxicity score, and the status of the compensation depletion early warning indicator—together constitute the AAR result entry for that node. All node entries are arranged in chronological order to form a complete AAR result. At the end of the fourth course of treatment, a patient's adverse reaction profile showed a persistently high local gradient with a sudden increase. Simultaneously, the compensation exhaustion warning indicator was at a high severity level. Risk mapping revealed consecutive Level 1 warnings triggered during this period, suggesting the need for clinical intervention before the next course of treatment. Conversely, during the first course of treatment, the same patient's local gradients were generally low across all nodes, and the compensation exhaustion warning indicator had not yet been triggered. The adverse reaction warning results only showed scattered Level 3 warning entries. The comparison of these two warning distributions clearly illustrates the evolution of the patient's adverse reaction risk as treatment progressed. The presence of three or more consecutive Level 1 warnings in the adverse reaction warning results indicates that the adverse reaction had reached a persistently high-risk level during that period. The overall warning level sequence throughout the entire treatment course reflects the dynamic distribution characteristics of the patient's adverse reaction risk during the radiotherapy and chemotherapy cycle.

[0076] To implement the above-described method embodiments, a method for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma is provided to achieve the corresponding functions and technical effects. See also... Figure 2 , Figure 2 This application provides a structural block diagram of a glioma radiotherapy and chemotherapy adverse reaction monitoring and early warning system, comprising:

[0077] Data acquisition module 201 is used to acquire blood routine data and neurological function score data, and to perform time-series alignment of the blood routine data and the neurological function score data to form a set of physiological dynamic parameters;

[0078] Feature extraction module 202 is used to perform channel-by-channel trend baseline correction and fusion on the physiological dynamic parameter set to generate a toxicity response map, implement cross-index abnormal enhancement based on the toxicity response map to generate toxicity response parameters, and extract bone marrow index sudden drop and rebound features from the blood routine data to generate a compensatory exhaustion early warning indicator.

[0079] The segment positioning module 203 is used to extract the time period when the liver and kidneys deviate from the toxicity index using the toxicity response map to locate the toxicity occurrence segment, identify the toxicity retention and aggregation time period based on the toxicity occurrence segment to generate a toxicity retention and aggregation identifier, and perform cross-dimensional coupling based on the toxicity retention and aggregation identifier and the toxicity response parameters to form a toxicity feature matrix.

[0080] The map construction module 204 is used to perform abnormal pattern clustering on the toxicity feature matrix to determine the toxicity category distribution, generate toxicity adaptation identifiers by identifying abrupt changes in scoring retention segments in the toxicity category distribution, and extract toxicity gradient features based on the toxicity adaptation identifiers to construct an adverse reaction map.

[0081] The early warning output module 205 is used to determine the toxicity sensitivity weight based on the weight inversion feature of the toxicity response parameter and the toxicity feature matrix, construct a graded early warning rule based on the toxicity sensitivity weight and the compensation exhaustion early warning identifier, and perform risk mapping on the graded early warning rule and the adverse reaction map to output the adverse reaction early warning result.

[0082] The aforementioned glioma radiotherapy and chemotherapy adverse reaction monitoring and early warning system can implement one of the glioma radiotherapy and chemotherapy adverse reaction monitoring and early warning methods described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining contents of this application embodiment can be referred to the contents of the above method embodiments, and will not be repeated in this embodiment.

[0083] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of adverse reactions to radiotherapy and chemotherapy for glioma, characterized in that, include: Collect routine blood data and neurological function score data, and perform time-series alignment of the routine blood data and the neurological function score data to form a set of dynamic physiological parameters; The physiological dynamic parameter set is fused by channel-wise trend baseline correction to generate a toxicity response map. Based on the toxicity response map, cross-index abnormality enhancement is performed to generate toxicity response parameters. The bone marrow index sudden drop and rebound characteristics are extracted from the blood routine data to generate a compensatory exhaustion early warning indicator. The toxicity response map is used to extract the time period when the liver and kidneys deviate from the toxicity index to locate the toxicity occurrence segment. Based on the toxicity occurrence segment, the toxicity retention and accumulation period period is identified to generate a toxicity retention and accumulation identifier. Based on the toxicity retention and accumulation identifier and the toxicity response parameters, a toxicity feature matrix is ​​formed by cross-dimensional coupling. Anomaly pattern clustering is performed on the toxicity feature matrix to determine the toxicity category distribution. A toxicity adaptation identifier is generated by identifying abrupt changes in scoring after stagnation in the toxicity category distribution. Toxicity gradient features are extracted based on the toxicity adaptation identifier to construct an adverse reaction map. Based on the toxicity response parameters and the weight inversion feature identified by the toxicity feature matrix, a toxicity-sensitive weight is determined. A graded early warning rule is constructed based on the toxicity-sensitive weight and the compensation exhaustion early warning indicator. The graded early warning rule and the adverse reaction map are then used for risk mapping to output the adverse reaction early warning result.

2. The method according to claim 1, characterized in that, The step of performing channel-by-channel trend baseline correction and fusion of the physiological dynamic parameter set to generate a toxicity response map includes: The physiological dynamic parameter set is decomposed channel by channel to generate channel toxicity response sequences; A baseline shift factor is generated by identifying the baseline shift of adjacent periodic channels from the channel toxicity response sequence. The baseline shift factor is used to recalibrate the channel toxicity response sequence to form calibrated toxicity characteristics. Toxicity response maps are generated by multi-scale cascade fusion based on the calibrated toxicity characteristics.

3. The method according to claim 1, characterized in that, The method of generating toxicity response parameters by performing cross-index anomaly enhancement based on the toxicity response map includes: The toxicity response map is subjected to dual-path separation of channel response and time-series response to generate an initial anomaly map; Based on the initial anomaly map, a baseline non-recovery weight is generated by identifying channels that have not recovered after treatment. A corrected anomaly map is formed by weakly enhancing the initial anomaly map using the baseline unrecovered weights. High-response regions are screened based on the corrected anomaly map to generate toxicity response parameters.

4. The method according to claim 1, characterized in that, The step of using the toxicity response map to extract the time periods when the liver and kidneys deviate from the toxicity indicators and to locate the toxicity occurrence segment includes: The toxicity response map is mapped to the clinical indicator dimension space to generate an indicator distribution prediction map; Based on the predicted distribution map of the indicators, the normal regions of liver and kidney indicators are identified to obtain candidate time periods for deviation. The rate correction period is formed by constraining the rate of change of the indicator through the aforementioned candidate deviation period. Local peak screening is performed during the rate correction period to locate the toxicity-causing segment.

5. The method according to claim 1, characterized in that, The process of generating a toxicity adaptation identifier by identifying abrupt changes in the score retention segment within the toxicity category distribution includes: A toxicity score intensity distribution is generated based on the aforementioned toxicity category distribution; Identify response retention trends and locate adaptation candidate regions from the toxicity score intensity distribution; Identify regions where indicators deviate from their performance by analyzing periods of sudden score changes within the candidate adaptation region. Toxicity adaptation intensity is calibrated and a toxicity adaptation label is generated based on the deviation area of ​​the aforementioned indicator.

6. The method according to claim 1, characterized in that, The step of determining the toxicity category distribution by performing abnormal pattern clustering on the toxicity feature matrix includes: Generate a sample distance metric based on the toxicity feature matrix; A masking correction factor is generated from high-toxicity feature samples within the low-toxicity category detected by the sample distance metric. The sample distance metric and the masking correction factor are used to perform adaptive initialization of the center points to form dynamic cluster centers; The distribution of toxicity categories is determined by classifying samples based on the dynamic cluster centers.

7. The method according to claim 1, characterized in that, The step of determining toxicity-sensitive weights based on the weight inversion features identified by the toxicity response parameters and the toxicity feature matrix includes: The toxicity response parameters and the toxicity feature matrix are decomposed into hierarchical loss values ​​using a multi-scale method. Decoupling weights are generated from decoupling segments where loss decreases but toxicity is not alleviated, identified from the toxicity feature matrix. An adaptive weighting is performed based on the decoupling weights and the hierarchical loss values ​​to form the decoupling loss; The toxicity-sensitive weights are determined by gradient update based on the decoupling loss.

8. The method according to claim 3, characterized in that, The step of weakly enhancing the initial anomaly map using the baseline unrecovered weights to form a corrected anomaly map includes: Regional dissimilarity analysis is performed on the baseline unrecovered weights and the initial anomaly map to generate an enhanced priority map; A migration enhancement factor is generated by locating adjacent periodic priority offset regions from the enhancement priority map; An enhanced anomaly map is formed by adaptive differential stretching of the initial anomaly map using the migration enhancement factor. Response equalization processing is performed based on the enhanced anomaly graph to form a corrected anomaly graph.

9. The method according to claim 5, characterized in that, The step of identifying the index deviation region from the period of sudden score change in the adaptation candidate region includes: The scoring time-series slope is calculated for the adaptation candidate region to generate a slope change sequence; Based on the slope change sequence, locate the time node where the absolute value of the slope exceeds the threshold and generate a set of mutation candidate points; The set of mutation candidate points is subjected to dual verification of duration and amplitude to screen for effective mutation periods; The deviation region of the generated index is determined based on the distribution range of the effective mutation period.

10. A monitoring and early warning system for adverse reactions to radiotherapy and chemotherapy for glioma, characterized in that, include: The data acquisition module is used to collect blood routine data and neurological function score data, and to perform time-series alignment of the blood routine data and the neurological function score data to form a set of physiological dynamic parameters; The feature extraction module is used to perform channel-by-channel trend baseline correction and fusion on the physiological dynamic parameter set to generate a toxicity response map, implement cross-index abnormal enhancement based on the toxicity response map to generate toxicity response parameters, and extract bone marrow index sudden drop and rebound features from the blood routine data to generate a compensatory exhaustion early warning indicator. The segment positioning module is used to extract the time period when the liver and kidneys deviate from the toxicity index using the toxicity response map to locate the toxicity occurrence segment, identify the toxicity retention and aggregation time period based on the toxicity occurrence segment to generate a toxicity retention and aggregation identifier, and perform cross-dimensional coupling based on the toxicity retention and aggregation identifier and the toxicity response parameters to form a toxicity feature matrix. The map construction module is used to perform abnormal pattern clustering on the toxicity feature matrix to determine the toxicity category distribution, generate toxicity adaptation identifiers by identifying abrupt change segments after score retention in the toxicity category distribution, and extract toxicity gradient features based on the toxicity adaptation identifiers to construct an adverse reaction map. The early warning output module is used to determine the toxicity sensitivity weight based on the weight inversion feature of the toxicity response parameter and the toxicity feature matrix, construct a graded early warning rule based on the toxicity sensitivity weight and the compensation exhaustion early warning identifier, and perform risk mapping on the graded early warning rule and the adverse reaction map to output the adverse reaction early warning result.

Images