Medical radiation protection early warning method and device based on personal dose monitoring

Abstract

The invention provides a medical radiation protection early warning method and device based on personal dose monitoring, and is applied to the technical field of data processing. According to the method, the standardized dose monitoring sequence is generated through multi-source data fusion and adaptive filtering, and then the individualized radiation accumulation model is constructed through Monte Carlo simulation and stratified sampling. On the basis of time-dose dependence characteristics, a linear quadratic model is coupled with a dose rate correction term to construct an evaluation equation, and the key organ absorbed dose is calculated in combination with a dynamic dose conversion matrix. Through data comparison, abnormal signals are identified, damage acceleration factors are defined, health archives are combined for grouping, a random forest algorithm is used for constructing a personalized early warning model, finally graded early warning and protection suggestions are output, and a complete early warning chain is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a medical radiation protection early warning method and device based on personal dose monitoring. BACKGROUND

[0002] The existing medical radiation dose monitoring relies on a single device (such as using only a thermoluminescent dosimeter TLD or a real-time radiation detector), and multiple source data collaboration is not achieved. For example, although the thermoluminescent dosimeter can record the cumulative dose, it cannot feedback the dose rate change in real time; the real-time radiation detector can dynamically output the dose rate, but it is difficult to store historical cumulative data for a long time, resulting in defects such as "discontinuous in time dimension" and "incomplete dose information" of the monitoring data, and a comprehensive personal dose data portrait cannot be formed. In the medical scene, there are non-target interference factors such as electromagnetic interference (such as surgical room electrotome equipment, CT machine main control system), space scattered radiation (such as X-ray reflected by wall), etc. The existing technology mostly uses fixed threshold filtering or simple smoothing processing, which cannot adaptively distinguish between "real radiation exposure signal" and "environmental noise", resulting in large dose monitoring data error (up to 15%-20% in some scenes), which affects the accuracy of subsequent dose evaluation.

[0003] In addition, in medical radiation exposure, multiple organs of the human body (such as the thyroid, lens, and hand skin) will be exposed to radiation at the same time, but the existing technology mostly uses a "single organ serial calculation" method, or reduces the calculation accuracy of high dose areas for the sake of efficiency (such as using uniform grid division, without subdividing high-risk areas). For example, the same 10mmx10mm grid is used to calculate the doctor's hand (high dose area) and leg (low dose area) in interventional surgery, resulting in a calculation error of more than 15% for the local hotspot dose of the hand (such as the fingertips), which cannot accurately identify the high-risk area of the key organ.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the disclosed background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] According to one aspect of this application, a medical radiation protection early warning method based on personal dose monitoring is provided, comprising: acquiring real-time radiation exposure data and historical dose records of an individual; using multi-source data fusion technology based on a thermoluminescent dosimeter and a real-time radiation detector, employing an adaptive filtering algorithm to remove environmental interference noise, and generating a standardized dose monitoring sequence; converting the standardized dose sequence into a three-dimensional dose distribution map, processing it with Monte Carlo simulation software, using a stratified sampling algorithm to subdivide high-dose areas, and constructing an individualized radiation accumulation model; setting radiation damage as a time-dose dependent process, using a linear quadratic model coupled with a dose rate correction term to construct a damage assessment equation, extracting weighting factors for different radiation types to construct a dynamic dose conversion matrix, and performing parallel calculations on the dose distribution of multiple organs to obtain... The study collects the absorbed radiation dose of key organs; extracts peak dose, cumulative rate, and organ dose difference values ​​according to radiation exposure scenarios, classifies risk levels, and constructs a multi-dimensional feature matrix containing exposure time, dose rate, and organ sensitivity information as input to the early warning model; compares real-time dose with historical data of the same period to identify abnormal signals such as short-term dose spikes and cumulative dose approaching the threshold, and defines radiation damage acceleration factors using the slope of the dose-effect curve; groups exposed individuals based on their personal health records, uses a random forest algorithm to screen key influencing factors, and integrates dose parameters, physiological indicators, and exposure scenario information to construct a personalized early warning model; based on the risk probability output of the personalized early warning model, and combined with the spatial correlation information between radiation type and exposure site, generates graded early warning signals and early warning information with protective measures recommendations.

[0006] Another aspect of this application discloses a medical radiation protection early warning device based on personal dose monitoring, comprising: an acquisition module for acquiring real-time personal radiation exposure data and historical dose records; using multi-source data fusion technology based on a thermoluminescent dosimeter and a real-time radiation detector; employing an adaptive filtering algorithm to remove environmental interference noise; and generating a standardized dose monitoring sequence; a processing module for converting the standardized dose sequence into a three-dimensional dose distribution map; processing the map using Monte Carlo simulation software; subdividing high-dose areas using a stratified sampling algorithm; and constructing an individualized radiation accumulation model. The module sets radiation damage as a time-dose dependent process, employs a linear quadratic model coupled with a dose rate correction term to construct a damage assessment equation, extracts weighting factors for different radiation types to construct a dynamic dose conversion matrix, and performs multi-organ dose distribution analysis. Parallel computing is used to obtain the radiation absorbed dose of key organs; peak dose, cumulative rate, and organ dose difference values ​​are extracted according to radiation exposure scenarios to classify risk levels, and a multi-dimensional feature matrix containing exposure time, dose rate, and organ sensitivity information is constructed as input for the early warning model; real-time dose is compared with historical data of the same period to identify abnormal signals such as short-term dose spikes and cumulative dose approaching the threshold, and radiation damage acceleration factors are defined using the slope of the dose-effect curve; exposed individuals are grouped based on their personal health records, and key influencing factors are screened using a random forest algorithm, and personalized early warning models are constructed by integrating dose parameters, physiological indicators, and exposure scenario information; based on the risk probability output of the personalized early warning model, combined with the spatial correlation information between radiation type and exposure site, graded early warning signals and early warning information with protective measures recommendations are generated.

[0007] According to another aspect of this application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described medical radiation protection early warning method based on personal dose monitoring by executing the executable instructions.

[0008] This application presents a medical radiation protection early warning method based on personal dose monitoring, aiming to improve the accuracy and personalization of medical radiation protection early warning. First, multi-source data from a thermoluminescent dosimeter and a real-time radiation detector are fused, and adaptive filtering is used to remove interference, generating a standardized dose monitoring sequence. This sequence is then converted into a three-dimensional dose distribution map, and high-dose regions are subdivided through Monte Carlo simulation and stratified sampling to construct an individualized radiation accumulation model. Next, a damage assessment equation is established by coupling a linear quadratic model with a dose rate correction term, and the absorbed dose to key organs is calculated in parallel using a dynamic dose conversion matrix. Abnormal signals are identified by comparing real-time and historical doses, a radiation damage acceleration factor is defined, and a personalized early warning model is constructed by combining health record grouping and using a random forest algorithm to fuse multi-dimensional information. Finally, a graded early warning and protection recommendations are output, forming a complete early warning chain. The device includes an acquisition module and a processing module, which respectively realize data acquisition and processing, and subsequent modeling and early warning functions.

[0009] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0010] Figure 1 The flowchart illustrates a medical radiation protection early warning method based on personal dose monitoring provided in an embodiment of this application; Figure 2 A schematic diagram of a medical radiation protection early warning device based on personal dose monitoring, provided in an embodiment of this application, is shown. Detailed Implementation

[0011] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0012] The following is combined Figure 1 This application describes a medical radiation protection early warning method based on personal dose monitoring according to exemplary embodiments thereof. It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0013] In one embodiment, this application also proposes a medical radiation protection early warning method based on personal dose monitoring. Figure 1 A schematic flowchart of a medical radiation protection early warning method based on personal dose monitoring according to an embodiment of this application is shown.

[0014] S101 acquires real-time personal radiation exposure data and historical dose records. Based on multi-source data fusion technology of thermoluminescent dosimeter and real-time radiation detector, it uses an adaptive filtering algorithm to remove environmental interference noise and generate a standardized dose monitoring sequence.

[0015] In one implementation, real-time personal radiation exposure data and historical dose records are acquired. This step aims to establish a complete foundation of personal radiation dose data, encompassing both real-time dynamic data and historical cumulative data. Specifically, real-time radiation exposure data is collected through personal wearable professional monitoring devices. For example, medical personnel wearing thermoluminescent dosimeters (such as TLD-100) in interventional treatments and radiological examinations can capture radiation energy through material lattice defects, enabling cumulative dose recording of ionizing radiation such as X-rays and gamma rays. Simultaneously, a real-time radiation detector (such as a portable NaI scintillator detector) is used to output the dose rate data of the current environment in real time, ensuring dynamic tracking of instantaneous radiation exposure. Historical dose records are retrieved from hospital radiation safety management systems, personal dose monitoring archives, etc., and include the individual's radiation exposure records for the past 3-5 years, including cumulative dose values ​​for each CT scan, DR examination, and radiotherapy procedure (e.g., a cumulative dose of 5.2 mSv in 2023), single exposure duration (e.g., 45 minutes of exposure during an interventional procedure), and exposure scenario type, providing a historical benchmark for subsequent data comparison and trend analysis.

[0016] By integrating the monitoring advantages of different devices, a comprehensive radiation dose dataset is formed. The core advantage of thermoluminescent dosimeters lies in their accurate recording of cumulative dose. For example, if a physician works for 8 hours, a thermoluminescent dosimeter will show a cumulative dose of 0.15 mSv for the day, reflecting the total amount of radiation exposure throughout the day. Real-time radiation detectors, on the other hand, focus on monitoring instantaneous dose rates. For instance, during cardiovascular interventional surgery, they can display the instantaneous dose rate of the surgical area as 2.5 μSv / h in real time, accurately capturing changes in radiation intensity at different stages of the procedure. Multi-source data fusion technology uses algorithms such as timestamp alignment and data correlation analysis to integrate two types of data. For example, the 0.15 mSv cumulative dose from the thermoluminescent dosimeter is decomposed into dose contributions for different time periods (e.g., 0.08 mSv from interventional surgery at 09:00-10:00 and 0.07 mSv from CT scan at 14:00-15:00), and matched with the dose rate curve for the corresponding time period recorded by the real-time radiation detector (e.g., the dose rate reaches 3.2 μSv / h at 09:30). Finally, a dataset containing three-dimensional information of "time point - cumulative dose - instantaneous dose rate" is formed, which reflects both the total amount and dynamic changes.

[0017] An adaptive filtering algorithm is used to remove environmental interference noise to purify the raw data and ensure the accuracy of subsequent analysis. In actual monitoring, real-time radiation detectors are susceptible to environmental interference, which can produce abnormal values. For example, equipment may experience instantaneous data jumps due to collisions or electromagnetic interference (such as electromagnetic radiation from electrosurgical equipment in an operating room). (For example, the dose rate in a normal surgical environment is stable at 0.5-3 μSv / h, but suddenly a value of 10 μSv / h appears.) Such data does not represent true radiation exposure and must be removed using the algorithm. The adaptive filtering algorithm removes noise in the following ways: First, it constructs a normal fluctuation model based on historical data, such as analyzing the dose rate data of the person in the same work environment over the past 3 months to determine the normal fluctuation range (e.g., the dose rate during interventional surgery is usually between 1.2-4.0 μSv / h). When new data is input, the algorithm calculates the deviation between the current value and the model's predicted value in real time. If the deviation exceeds a preset threshold (e.g., exceeding the normal range by 2 times the standard deviation), it is judged as noise. Then, the data is corrected using methods such as linear interpolation and moving average. For example, an outlier of 10 μSv / h is corrected to the average of the data before and after 10 seconds, which is 2.3 μSv / h, to ensure the continuity and authenticity of the data sequence.

[0018] The fused and denoised raw data is converted into a structured sequence in a unified format to provide standardized input for subsequent modeling. Specifically, the data is sliced ​​and integrated at fixed time intervals (e.g., every 5 minutes). For example, in 5-minute intervals, the cumulative dose increment (e.g., cumulative dose increases from 0.02 mSv to 0.03 mSv from 08:30 to 08:35) and the average instantaneous dose rate (e.g., average dose rate of 0.6 μSv / h during this period) are extracted and associated with corresponding timestamps, ultimately outputting standardized sequence data. An example format is "08:30-0.02 mSv-0.5 μSv / h; 08:35-0.03 mSv-0.6 μSv / h; 08:40-0.04 mSv-0.7 μSv / h…", where each time point includes a clear time identifier, the cumulative dose up to that moment, and the average dose rate for the corresponding period. This standardized processing eliminates data format differences, ensuring that subsequent steps such as constructing three-dimensional dose distribution maps and calculating radiation accumulation models can directly call data with a unified structure, improving process coherence and computational efficiency.

[0019] S102, after converting the standardized dose sequence into a three-dimensional dose distribution map, is processed by Monte Carlo simulation software, and a stratified sampling algorithm is used to subdivide the high-dose region to construct an individualized radiation accumulation model.

[0020] In one implementation, standardized dose sequences are converted into three-dimensional dose distribution maps, providing a spatialized dose data foundation for subsequent Monte Carlo simulations. Standardized dose monitoring sequences contain structured data of "timestamp-cumulative dose-instantaneous dose rate" (e.g., "08:30-0.02mSv-0.5μSv / h; 08:35-0.03mSv-0.6μSv / h…"). Initially, this data only reflects dose changes over time and lacks spatial correlation with body parts. To establish spatial mapping, a high-precision three-dimensional model of human anatomy is required. This model can be constructed based on CT scan data of the irradiated individual (e.g., a three-dimensional model of the torso including bones, organs, and tissues), or using standard human phantoms recommended by the International Commission on Radiation Protection (ICRP) (e.g., adult male phantoms, female phantoms), ensuring the anatomical accuracy of the model.

[0021] Spatial allocation of dose data is achieved through spatial interpolation algorithms (such as Kriging interpolation and inverse distance weighted interpolation). First, the main source and propagation path of the dose are determined based on the radiation exposure scenario (e.g., in interventional surgery, the X-ray source is located on the patient's side, and the physician's hand is close to the X-ray source due to the operation). Second, the dose data at each time point in the standardized sequence are allocated to the corresponding areas of the three-dimensional model according to the spatial distance of the exposed site and the radiation attenuation law. For example, in a cardiovascular interventional surgery, a physician's hand is continuously exposed to X-rays from 08:30 to 08:45. Based on the operation record and the location of the X-ray source, the algorithm mainly maps the accumulated dose of 0.05 mSv during this period to the right hand area of ​​the three-dimensional model, while allocating a small dose to the forearm (e.g., 0.01 mSv) according to the attenuation ratio.

[0022] The resulting three-dimensional dose distribution map visually represents the dose distribution of different parts of the body on a three-dimensional human model using color gradients (e.g., blue for low dose and red for high dose), with each spatial point associated with a corresponding timestamp and dose value. For example, the map clearly shows that "at 08:40, the dose in the right palm region reached 0.03 mSv, with an instantaneous dose rate of 0.7 μSv / h," while the doses in other parts of the torso are generally below 0.01 mSv. This map provides a concrete spatial dose benchmark for subsequent Monte Carlo simulations, enabling the simulation process to accurately match the actual exposed anatomical locations and dose characteristics, laying a spatial data foundation for the construction of individualized radiation accumulation models.

[0023] Import the 3D dose distribution map into Monte Carlo simulation software to set up the computational environment for radiation dose accumulation simulation. Select a suitable Monte Carlo simulation software (such as GEANT4 or MCNP). These software programs have high-precision radiation transport simulation capabilities and can accurately calculate the energy deposition process of different radiation particles in human tissues. When importing the previously generated 3D dose distribution map, it is necessary to align the spatial coordinate system of the map (such as 3D coordinates based on Cartesian coordinates) with the software's built-in physical calculation coordinate system through the software's data interface to ensure that the spatial position of the dose data is accurately reproduced in the simulation environment.

[0024] Secondly, configure key simulation parameters to match the actual radiation scenario: The type of radiation particle should be set according to the exposure scenario. For example, interventional surgery and CT scans primarily use X-rays, while nuclear medicine examinations may involve gamma rays. The particle energy range needs to be specified in the software (e.g., X-ray energy 50-150 keV); the number of simulated particles should be set based on a balance between accuracy requirements and computational efficiency, for example, selecting 10. 6 The simulation employs individual particles to ensure statistical significance (keeping dose calculation errors within 5%) while avoiding excessive computational resource consumption. It utilizes radiation-matter interaction models (such as the photoelectric effect and Compton scattering model) within the software to ensure that the attenuation and scattering patterns of simulated particles in human tissues are consistent with actual physical processes. Simultaneously, a human voxel model is imported as the anatomical carrier of radiation effects. The voxel model can use standard phantoms recommended by ICRP (such as the adult male phantom ICRP110), which includes detailed organ and tissue divisions (such as the thyroid gland, lens, and hematopoietic organs). Each voxel (e.g., 1mm × 1mm × 1mm) is labeled with tissue type, density, atomic number, and other physical parameters.

[0025] Spatial matching between 3D dose maps and voxel models is achieved through a software interface: A coordinate transformation algorithm is used to precisely align high-dose regions marked in the map (such as the hand and chest) with the corresponding organ / tissue coordinate ranges in the phantom. For example, the spatial coordinates (X: 120-150mm, Y: 300-330mm, Z: 50-80mm) of the "right palm region" in the map are bound to the voxel coordinate range of the right palm in the phantom. After matching, the software can directly call the dose data in the map as the initial radiation source term, and the voxel model serves as the target for radiation, forming a complete computational chain of "dose input - particle transport - energy deposition".

[0026] The establishment of this computing environment ensures that Monte Carlo simulation can reproduce the spatial distribution characteristics of radiation dose based on real anatomical structures. For example, the dose accumulation process of X-rays in the surgeon's hand during interventional surgery can be calculated by the energy deposition of the corresponding region in the voxel model, providing an accurate simulation platform for subsequent high-dose region subdivision and individualized radiation accumulation model construction.

[0027] Based on the dose difference characteristics of radiation exposure areas, a stratified sampling algorithm was selected as the core method for regional subdivision to determine the identification and division rules for high-dose areas. A three-dimensional dose distribution map clearly presents the radiation dose distribution characteristics of different parts of the human body through color gradients and numerical annotations. Taking a physician interventional surgery scenario as an example, significant differences in dose distribution can be visually observed in the map: the chest, due to its proximity to the radiation source and relatively weak protective shielding, has a cumulative dose concentrated in the range of 0.1-0.3 mSv; the hands, due to direct exposure to the radiation field from operating instruments, also have a dose in the 0.1-0.3 mSv range; while the legs, due to their greater distance from the radiation source and shielding by lead aprons and other protective equipment, have a dose of only 0.01-0.05 mSv. This difference reflects the spatial specificity of radiation exposure, providing data support for the accurate identification of high-risk areas.

[0028] Based on the aforementioned dose distribution characteristics, the rule for identifying high-dose areas is set as "continuous areas with dose values ​​exceeding 0.1 mSv". This threshold is set with reference to the dose constraint principles of occupational radiation protection (such as one percent of the annual effective dose limit), while also incorporating actual clinical exposure data—dose areas above 0.1 mSv typically overlap significantly with radiation-sensitive organs (such as the thyroid gland and lens in the chest) or high-frequency operating areas (such as the hands), representing high-risk areas for radiation damage. This rule allows for the automatic delineation of key areas such as the chest and hands from the 3D atlas, eliminating redundant calculations for low-dose areas such as the legs.

[0029] For the identified high-dose areas, a stratified sampling algorithm is used for refinement: First, the area is divided into different levels according to the dose value, such as 0.1-0.2 mSv as the first level and 0.2-0.3 mSv as the second level. The level division is gradually refined as the dose increases to ensure that the stratification density of the high-dose area is higher than that of the low-dose area. Second, each level is divided into grids with a spatial resolution of 5 mm × 5 mm × 5 mm. This resolution is much higher than that of low-dose areas (such as the leg area using a 20 mm × 20 mm × 20 mm grid), which can capture the subtle differences between local hot spot doses (such as 0.28 mSv at the fingertips of the hand) and the surrounding tissues.

[0030] The core advantage of the stratified sampling algorithm lies in "allocating computational resources on demand": by dividing high-dose regions into multi-level, high-density grids, subsequent Monte Carlo simulations can allocate more computational samples to these regions (e.g., increasing the number of simulated particles per grid by 50%), thereby keeping the dose calculation error within 3%. Meanwhile, the sparse grid setting in low-dose regions reduces unnecessary computations, balancing simulation accuracy and efficiency. This approach ensures the accuracy of simulations in high-risk regions while avoiding the surge in computational costs caused by using high-precision grids across the entire region, providing a scientific and efficient foundation of detailed data for individualized radiation accumulation models.

[0031] Stratified sampling was used to further subdivide the high-dose region, improving the accuracy and spatial resolution of dose simulation in this area. For the identified high-dose region in the chest (0.2-0.3 mSv), a stratified sampling algorithm was employed to further subdivide the grid to 2mm × 2mm × 2mm, while the low-dose region in the legs maintained a 10mm × 10mm × 10mm grid. In the simulation, more simulation particles were assigned to the subdivided high-dose grid (e.g., 5000 particles per grid), and 1000 particles per grid were assigned to the low-dose region. This reduced the dose calculation error in the chest region from ±5% to ±2%, significantly improving the simulation accuracy in high-risk areas.

[0032] The software generates individualized radiation accumulation models containing dose characteristics of subdivided regions, forming a quantitative analysis platform that meets the needs of radiation damage assessment. The subdivided 3D mesh is differentiated according to dose levels: high-dose areas (such as the chest and hands) use a fine 5mm×5mm×5mm mesh, while low-dose areas use a coarser mesh. This structure provides a foundation for accurately simulating the trajectory and energy transfer of radiation particles in human tissue. During the simulation, the software tracks the complete path of each particle from the radiation source to its absorption or scattering by human tissue, calculating the energy deposition in the mesh along its path, based on preset radiation particle types (such as X-rays) and physical action models (such as the photoelectric effect and Compton scattering). For example, in the subdivided mesh of the thyroid region in the chest, the software simulates 10... 6 The transport process of each X-ray particle was statistically analyzed to obtain the cumulative dose of each grid. The core thyroid region grid received a cumulative dose of 0.28 mSv due to direct irradiation, while the dose of adjacent local hotspot grids (such as those near the edge of the radiation source) increased to 0.32 mSv due to particle scattering and superposition. The dose of the surrounding muscle tissue grids decreased to 0.15 mSv due to radiation attenuation. These data accurately reflect the differences in radiation distribution in different tissues.

[0033] After the simulation, the software integrates parameters such as cumulative dose, energy deposition, and particle flux from each grid according to organ and tissue categories to form an individualized radiation accumulation model. This model is stored in structured data format, containing dose distribution characteristics for various human organs (such as the thyroid gland, lens, and hematopoietic organs) and tissues (such as muscles and bones). It includes not only the overall cumulative dose but also detailed parameters such as local hotspot dose and dose volume distribution (e.g., the percentage of the thyroid gland with a dose exceeding 0.2 mSv). For example, the model can explicitly label "the average cumulative dose of the thyroid gland is 0.25 mSv, 90% of the volume dose is below 0.3 mSv, and 10% of the hotspot area reaches 0.32 mSv," while simultaneously recording corresponding dose data for tissues such as the skin of the hand and the muscles of the chest.

[0034] This individualized cumulative radiation model directly serves the subsequent calculation of absorbed dose to critical organs: because the dose parameters of each organ in the model are precisely correlated with anatomical structures, dose values ​​of radiation-sensitive organs such as the thyroid and lens can be directly extracted without additional spatial matching steps. This quantitative data provides a core basis for radiation damage assessment—for example, by comparing the cumulative dose of the thyroid with the organ dose limits recommended by the ICRP (such as the annual dose limit of 20 mSv for the thyroid), a preliminary assessment of radiation damage risk can be made. Simultaneously, it lays the data foundation for subsequent calculations of damage severity using a linear quadratic model and the construction of early warning models, ensuring the consistency and accuracy of the entire radiation protection early warning process.

[0035] S103 sets radiation damage as a time-dose dependent process, uses a linear quadratic model coupled with a dose rate correction term to construct a damage assessment equation, extracts weighting factors for different radiation types to construct a dynamic dose conversion matrix, performs parallel calculations on dose distribution of multiple organs, and obtains the radiation absorption dose of key organs.

[0036] In one implementation, radiation damage is defined as a time-dose dependent process. A linear quadratic model coupled with a dose rate correction term is used to construct a damage assessment equation. The dose rate correction term reflects the impact of different irradiation rates on the degree of damage, making the equation more closely reflect the time-cumulative characteristics of radiation damage. The occurrence and development of radiation damage are the result of the combined effects of time and dose. The damage effects on human tissues differ significantly between short-duration high-dose irradiation and long-duration low-dose irradiation; this characteristic is the core basis for constructing the damage assessment equation. The linear quadratic model is a classic model for radiation damage assessment. Its basic formula is "damage probability = αD + βD²" (where D is the total radiation dose, α and β are tissue-specific coefficients, α reflects the linear damage component, and β reflects the secondary damage component), which describes the relationship between total dose and damage probability. However, this model does not consider the influence of dose rate (the dose received per unit time) on damage. In practice, under the same total dose, high-dose-rate irradiation may result in more significant damage due to insufficient tissue repair time, while low-dose-rate irradiation allows tissue more time to repair sublethal damage, resulting in a relatively weaker damage effect. A dose rate correction term k needs to be introduced to improve the model. The formula for the correction term is set as "k=1 / (1+γ・R)" (R is the dose rate, γ is the rate sensitivity coefficient, which is obtained by fitting experimental data. The γ value is different for different tissues, such as γ=0.05h / μSv for hematopoietic tissue).

[0037] The complete injury assessment equation is "Injury probability = (αD + βD²) ・ k", where the correction term k adjusts the injury probability corresponding to the total dose, reflecting the influence of the dose rate. Taking a physician receiving a dose of 0.5 mSv as an example: when irradiated at a low dose rate (R1) of 5 μSv / h, substituting into the correction term formula, if γ = 0.05 h / μSv, then k1 = 1 / (1 + 0.05 × 5) = 0.8, and the injury probability = (α × 0.5 + β × 0.5²) × 0.8; when irradiated at a high dose rate (R2) of 20 μSv / h, k2 = 1 / (1 + 0.05 × 20) = 0.5, and the injury probability = (α × 0.5 + β × 0.5²) × 0.5. Calculations show that the probability of damage from high-dose-rate irradiation is 1.6 times that of low-dose-rate irradiation (the reciprocal of 0.5 / 0.8 = 0.625 is 1.6), directly reflecting the amplifying effect of irradiation rate on damage. This equation design with coupled correction terms retains the description of the total dose effect by the linear quadratic model while quantifying the influence of the time dimension through a dynamic adjustment factor k. This makes the damage assessment results more closely resemble the dynamic process of "dose-time-repair" in radiation biology, providing a scientific model basis for the accurate calculation of subsequent critical organ damage risks.

[0038] Weighting factors for different radiation types are extracted to construct a dynamic dose conversion matrix. This matrix is ​​used to uniformly convert various radiation doses into equivalent biological doses, ensuring the comparability and summation of multi-source radiation doses. Different radiation types cause significantly different damage effects on organisms due to differences in energy transfer mechanisms and ionization densities. This difference can be quantified by relative biological efficacy (RBE)—the higher the RBE value, the more severe the biological damage under the same physical dose. For example, alpha rays have a high ionization density, with an RBE value of 20; X-rays and gamma rays are low LET (linear energy transfer density) radiation, both with an RBE value of 1; neutrons, because they can directly induce nuclear reactions, have an RBE value of 10.

[0039] To achieve a unified assessment of multi-source radiation dose, a dynamic dose conversion matrix needs to be constructed. Using X-rays as the baseline (with a weighting factor of 1), the RBE values ​​of other radiation types are incorporated into the matrix as conversion coefficients, forming a correspondence between "radiation type and conversion coefficient" (e.g., alpha rays correspond to 20, neutrons to 10, and gamma rays to 1). The dynamic nature of this matrix is ​​reflected in its ability to update the RBE values ​​according to the latest radiation protection standards (such as ICRP publications), ensuring the scientific validity of the conversion coefficients.

[0040] In practical applications, if a physician is simultaneously exposed to multiple types of radiation during work—such as 0.1 mSv of alpha rays (originating from radiopharmaceuticals) during a nuclear medicine examination and 0.3 mSv of gamma rays (scattered from equipment) during a CT scan—the two types of doses can be uniformly converted into equivalent biological doses using a dynamic dose conversion matrix: the alpha ray contribution dose is 0.1 mSv × 20 = 2 mSv, the gamma ray contribution dose is 0.3 mSv × 1 = 0.3 mSv, and the total equivalent biological dose is 2 + 0.3 = 2.3 mSv.

[0041] This conversion method solves the problem that doses of different types of radiation cannot be directly added together, enabling radiation exposure in multiple scenarios (such as simultaneous interventional surgery X-rays and radionuclide therapy gamma rays) to be compared and evaluated through a unified biological dose index. This provides a standardized quantitative basis for subsequent calculation of absorbed doses to critical organs and risk warning, ensuring the accuracy and consistency of multi-source radiation exposure assessment.

[0042] The individualized radiation accumulation model is input into the damage assessment equation, and multi-threaded parallel computing technology is used to simultaneously calculate the dose distribution of multiple organs, obtaining the radiation absorbed dose of key organs, including the average absorbed dose of the organ, the local hot spot dose, and the dose-volume histogram parameters. The individualized radiation accumulation model already includes refined dose distribution data (such as cumulative dose and energy deposition of each subdivided grid) for key organs / tissues such as the thyroid, lens, and hand skin. These data are precisely correlated with human anatomical structures, laying the foundation for targeted calculation of damage risk for each organ. The input damage assessment equation is a linear quadratic model coupled with a dose rate correction term (damage probability = (αD + βD²)·k), where α and β are organ-specific coefficients (e.g., thyroid α = 0.1 / Sv, β = 0.05 / Sv², lens α = 0.05 / Sv², β = 0.02 / Sv²), which can be dynamically adjusted according to the differences in radiation sensitivity of different organs.

[0043] To improve computational efficiency, an 8-thread parallel computing technique is employed. Specifically, each thread independently handles the dosage calculation for 1-2 organs. Threads synchronize key parameters (such as total dose D and dose rate R) through shared memory to avoid data conflicts. Thread 1 processes thyroid data, thread 2 processes lens data, thread 3 processes hand skin data, and the remaining threads are reserved for or handle dosage calculations for minor tissues (such as muscles and bones). This parallel architecture reduces the multi-organ computation time from 15 minutes with a single thread to 2 minutes, significantly improving process efficiency.

[0044] During the calculation, the subdivided grid data of each organ was substituted into the equation one by one, and finally integrated into the quantitative parameters at the organ level. Specifically, for the thyroid gland: the average absorbed dose was 0.25 mSv obtained by calculating the dose data of its 200 subdivided grids. Among them, the five grids closest to the radiation source formed local hot spots due to particle scattering and superposition, with the highest dose reaching 0.32 mSv, reflecting the overall exposure level of the thyroid gland and local high-risk areas; for the lens: the calculation results showed an average dose of 0.18 mSv. Through dose-volume histogram analysis (statistical analysis of the volume percentage of different dose intervals), it was found that 90% of the lens volume had a dose below 0.2 mSv, and only 10% of the edge areas had a slightly higher dose, reflecting the uniformity of dose distribution within the organ; for the skin of the hand: as a high-frequency operation site, its average dose was 0.22 mSv. The fingertip grid showed a hot spot dose of 0.28 mSv due to direct exposure, which is consistent with the actual scenario of the hand being close to the radiation source in clinical operations.

[0045] These parameters encompass overall organ exposure (average dose), local high-risk points (hotspot dose), and dose distribution uniformity (dose-volume histogram parameters), comprehensively characterizing the radiation exposure features of key organs. For example, although the 0.32 mSv hotspot dose of the thyroid gland does not exceed the ICRP-recommended annual dose limit (20 mSv), its radiation sensitivity, combined with this data, provides core information for subsequent risk grading. The dose-volume parameters of the lens help assess the potential risk of radiation-related diseases such as cataracts. Ultimately, these quantitative results are integrated into a structured dataset, directly serving as input parameters for the early warning model, ensuring the accuracy and relevance of subsequent risk assessments.

[0046] S104 extracts peak dose, cumulative rate, and organ dose difference values ​​according to radiation exposure scenarios, classifies risk levels, and constructs a multi-dimensional feature matrix containing exposure time, dose rate, and organ sensitivity information as input for the early warning model.

[0047] In one implementation, characteristic parameters are extracted for different radiation exposure scenarios (such as interventional surgery, CT scans, and nuclear medicine procedures). Specifically, in the interventional surgery scenario: due to close-range manipulation, the peak dose to the physician's hands reaches 0.3 mSv, with a cumulative rate of 0.05 mSv / hour; the organ dose difference between the thyroid and hands is 0.2 mSv (0.3 mSv for hands vs. 0.1 mSv for thyroid). In the CT scan scenario: the technician's chest is affected by scattered radiation, with a peak dose of 0.15 mSv and a cumulative rate of 0.02 mSv / hour; the organ dose difference between the lens and the chest is 0.05 mSv (0.15 mSv for chest vs. 0.1 mSv for lens). These parameters reflect the intensity, rate, and organ-specific differences in radiation exposure under different scenarios, providing a basis for risk grading.

[0048] Based on the extracted parameters, a three-tiered risk standard is established: High risk: peak dose > 0.25 mSv or cumulative rate > 0.04 mSv / hour (e.g., hand exposure during interventional surgery); Medium risk: 0.1 mSv < peak dose ≤ 0.25 mSv and 0.02 mSv / hour < cumulative rate ≤ 0.04 mSv / hour (e.g., chest exposure for technicians during CT scans); Low risk: peak dose ≤ 0.1 mSv and cumulative rate ≤ 0.02 mSv / hour (e.g., routine DR examinations). This risk level classification allows for the quantitative categorization of radiation hazards in different scenarios.

[0049] A multidimensional feature matrix containing exposure time, dose rate, and organ sensitivity information is constructed as input to the early warning model. Specifically, each row of the matrix corresponds to a radiation exposure event, and the columns include the following dimensions: exposure time (e.g., "Interventional surgery: 2024-05-20 09:00-10:30"); dose rate (e.g., average dose rate of interventional surgery 2.5 μSv / h); organ sensitivity (thyroid sensitivity coefficient 1.2, lens 1.0, hand skin 0.8, set based on ICRP organ weighting factor); risk level label (e.g., "high risk").

[0050] Examples of matrices are shown in Table 1.

[0051] Table 1: Matrix Examples This matrix integrates multi-dimensional information to provide structured input for personalized early warning models, ensuring that the models can comprehensively assess the impact of scenario, time, dosage, and organ differences on risk.

[0052] S105 compares real-time dose with historical data from the same period to identify abnormal signals such as short-term dose spikes and cumulative dose approaching the threshold, and uses the slope of the dose-effect curve to define the radiation damage acceleration factor.

[0053] In one implementation, real-time dose data is compared with historical dose data for the same period, and the deviation and trend of the two types of data are extracted to generate a dose dynamic comparison chart. Real-time dose data of a physician in May 2024 (e.g., daily cumulative dose: May 1st 0.12 mSv, May 2nd 0.15 mSv, etc.) is selected and compared with the same period in 2023 (May 1st 0.08 mSv, May 2nd 0.09 mSv, etc.). The deviation value is calculated (May 1st deviation +0.04 mSv, May 2nd deviation +0.06 mSv), and the trend is extracted (the average daily dose for the first 10 days of May 2024 was 0.13 mSv, an increase of 44% compared to 0.09 mSv in the same period of 2023). This data is then used to generate a dose dynamic comparison chart in the form of a line graph, with the horizontal axis representing the date and the vertical axis representing the cumulative dose. The deviation and trend differences are visually displayed through a double line graph (real-time vs. historical).

[0054] Based on the dose dynamic comparison chart, a short-term dose surge threshold and a cumulative dose warning line are set to identify abnormal signals exceeding the thresholds. Abnormal signals include short-term dose surge signals and cumulative dose approaching the threshold signal. Combining the historical data fluctuation range in the chart (e.g., the maximum daily increase of 0.03 mSv in the same period of 2023), the short-term dose surge threshold is set as "a daily dose increase ≥ 0.05 mSv compared to the previous day"; referencing 80% of the occupational annual dose limit (20 mSv), the cumulative dose warning line is set at 16 mSv. If the physician's dose reaches 0.2 mSv on May 15th, an increase of 0.08 mSv compared to May 14th (0.12 mSv), a short-term dose surge signal is triggered; by November, the cumulative dose reaches 15.8 mSv, approaching the 16 mSv warning line, triggering the cumulative dose approaching the threshold signal. Both types of signals are marked as abnormal.

[0055] A dose-response curve is introduced, and the slope of the curve is calculated in different dose ranges. The slope value is defined as the radiation damage acceleration factor, which is used to quantify the rate at which changes in radiation dose affect the risk of damage. The dose-response curve is a core tool in the field of radiation protection for describing the relationship between radiation dose and biological damage risk. Its horizontal axis represents the cumulative radiation dose (in mSv), and the vertical axis represents the probability of occurrence of a specific injury (such as thyroid cancer or cataracts) (in %). The shape of the curve varies with the type of radiation, organ sensitivity, and dose range, and generally exhibits the characteristic of "flat in the low-dose range and steep in the high-dose range"—this is because when the dose exceeds a certain threshold, the tissue repair mechanism saturates, and the rate of damage risk increases significantly with increasing dose. Taking the thyroid cancer risk curve as an example, this curve is constructed based on a large amount of epidemiological data (such as health monitoring data of radiation workers) and can reflect the probability change of thyroid cancer under different cumulative doses.

[0056] To accurately quantify the rate at which dose changes affect risk, the slope of the curve needs to be calculated for different dose ranges. Specifically, in the low-dose range (0-50 mSv): the radiation dose has not yet exceeded the threshold of tissue repair capacity, and the risk of damage increases slowly with increasing dose. Calculations using the curve derivative show that the slope in this range is 0.02% / mSv, meaning that for every 1 mSv increase in cumulative dose, the risk of thyroid cancer increases by 0.02%. For example, when the dose increases from 10 mSv to 11 mSv, the risk increases from 0.2% to 0.22%, showing a stable linear relationship with the dose increment. In the medium-to-high-dose range (50-100 mSv): as the dose accumulates, sub-lethal tissue damage gradually accumulates, and repair mechanisms are unable to fully compensate, leading to a faster rate of risk increase. The curve slope increases to 0.05% / mSv, meaning that for every 1 mSv increase in dose, the risk increases by 0.05%. For example, when the dose is increased from 60 mSv to 61 mSv, the risk increases from 3.0% to 3.05%, which is 2.5 times the increase in the low-dose range.

[0057] These slope values ​​are defined as radiation damage acceleration factors, which physically represent the "rate of change in damage risk caused by a unit dose change." A higher acceleration factor indicates a more significant impact of dose fluctuations on risk within that dose range. For example, an acceleration factor of 0.02 in the 0-50 mSv range means that the "driving effect" of dose changes on risk is relatively weak in this stage; while an acceleration factor of 0.05 in the 50-100 mSv range indicates that dose increments need to be more strictly controlled in this stage to avoid a rapid increase in risk. This setting aligns with the fundamental principle of "dose-repair-threshold" in radiation biology and provides a quantitative benchmark for subsequent risk assessment of abnormal signals. When anomalies such as short-term dose spikes or cumulative doses approaching the threshold are identified, the degree of risk impact can be quickly calculated using the acceleration factor of the dose range in which it occurs (e.g., a 0.5 mSv spike in the 50-100 mSv range results in a risk increment of 0.5 × 0.05% = 0.025%), ensuring the scientific rigor and dynamic adaptability of risk assessment.

[0058] A correlation analysis was conducted between abnormal signals and radiation damage acceleration factors to generate an anomaly-acceleration factor correlation matrix, providing dynamic parameter support for subsequent early warning models. The risk impact of abnormal signals needs to be calculated in conjunction with the radiation damage acceleration factor within their dose range. The core logic is "dose change × acceleration factor = risk impact increment". This correlation considers both the dose amplitude of the abnormal signal and the sensitivity of risk at that dose level, making risk assessment more targeted. The correlation of short-term dose surge signals is as follows: On May 15, a physician's daily dose was detected to have increased by 0.08 mSv compared to the previous day. Combined with their cumulative dose (current total dose 1.2 mSv), the range was determined to be 0-50 mSv, corresponding to an acceleration factor of 0.02% / mSv. The risk impact increment was calculated as: 0.08 mSv × 0.02% / mSv = 0.0016%, meaning that this surge increased the risk of radiation-related diseases such as thyroid cancer by 0.0016%. This result reflects the immediate impact of short-term high-dose fluctuations on risk; the larger the sudden increase and the higher the acceleration factor in the interval, the more significant the risk increase.

[0059] The correlation of the cumulative dose approaching the threshold signal is as follows: In November, the physician's cumulative dose reached 15.8 mSv, only 0.2 mSv away from the preset warning line (16 mSv), still within the 0-50 mSv range, with an acceleration factor of 0.02% / mSv. The calculated risk increment is: 0.2 mSv × 0.02% / mSv = 0.0004%, indicating that the risk growth potential corresponding to the remaining dose space is 0.0004%. This correlation reflects the cumulative risk effect when the cumulative dose approaches the threshold; even a small increment requires early warning to control subsequent exposure.

[0060] The above correlation results were organized into a structured matrix, which clearly presents the key parameters of various abnormal signals. The results are shown in Table 2.

[0061] Table 2: Key parameters of various abnormal signals The core value of this matrix lies in transforming different types of abnormal signals (short-term surges, cumulative approach) into a unified "risk impact increment" indicator, which facilitates horizontal comparison of subsequent early warning models. The acceleration factor and risk increment in the matrix are directly used as input parameters for the early warning model, enabling the model to adjust the early warning level according to the actual risk level of the signal (e.g., a higher risk increment for a short-term surge can trigger a more urgent early warning). If abnormal signals in a higher dose range (e.g., 50-100 mSv) are subsequently detected, they can be directly included in the matrix and associated with the corresponding acceleration factor (0.05) to ensure coverage of anomalies across the entire dose range.

[0062] Through this process, abnormal signals are upgraded from qualitative descriptions to quantitative risk parameters, providing precise dynamic input for personalized early warning models. This enables early warning results to reflect both the objective characteristics of abnormal events and their actual impact on human health, ultimately enhancing the scientific rigor and practicality of radiation protection early warning.

[0063] S106, combining personal health records to group exposed individuals, uses a random forest algorithm to screen key influencing factors, and integrates dose parameters, physiological indicators and exposure scenario information to construct a personalized early warning model.

[0064] In one implementation, key health indicators affecting radiation sensitivity are extracted based on individual health records. Individuals exposed are then categorized into high-sensitivity, moderate-sensitivity, and low-sensitivity groups according to their radiation tolerance levels, with each group corresponding to a unique sensitivity label. A multi-indicator clustering model (hierarchical clustering algorithm) is employed. Core health indicators include hematological indicators such as white blood cell count (normal range 3.5-9.5 × 10⁻⁶). 9 / L), lymphocyte percentage (20%-50%); genetic factors include DNA repair gene polymorphism (such as XRCC1 gene typing); underlying diseases include hyperthyroidism, diabetes, etc. (weight 1.2).

[0065] Indicator data of 300 medical staff were extracted from their health records and divided into three groups using hierarchical clustering (Euclidean distance, Ward's link method): High-sensitivity group (label S1): white blood cell count <4.0×10⁻⁶. 9 / L and carrying the XRCC1 mutant gene (e.g., 20 people); Intermediately sensitive group (tag S2): indicators in the middle range (e.g., 180 people); Lowly sensitive group (tag S3): white blood cell count >6.0×10 9 / L and no radiation-sensitive diseases (e.g., 100 people).

[0066] The dose parameters, physiological indicators, and exposure scenarios of different sensitivity groups were compared and analyzed to generate characteristic difference data between the groups. A characteristic difference analysis model (ANOVA, significance level p<0.05) was used, with the following comparison dimensions: dose parameter: thyroid absorbed dose (mean 0.22 mSv in the high-sensitivity group vs. 0.25 mSv in the low-sensitivity group); physiological indicator: the lymphocyte ratio in the high-sensitivity group (mean 28%) was significantly lower than that in the low-sensitivity group (35%); exposure scenario: the interventional surgery participation rate in the high-sensitivity group (60%) was higher than that in the low-sensitivity group (30%). A difference data table was generated, annotating the significant differences between group S1 and group S3 (such as lymphocyte ratio and interventional surgery frequency) to provide distinguishing features for model training.

[0067] The dataset integrates and processes feature difference data, dose parameters, physiological indicators, and exposure scenario information to generate a multi-dimensional model training dataset. The dose parameters include the absorbed dose to critical organs and the radiation damage accelerator factor. A feature fusion model (structured data concatenation) is employed. The dataset fields include dose parameters: absorbed dose to critical organs (e.g., thyroid 0.25 mSv), radiation damage accelerator factor (0.02); physiological indicators: white blood cell count, genotyping (binary encoding: mutation=1, wild=0); exposure scenarios: surgical type (unique heat encoding: interventional=1, CT=0), exposure duration (hours); and labels: sensitivity grouping (S1 / S2 / S3). The dataset contains 1000 samples, each a 12-dimensional feature vector (e.g., [0.25,0.02,3.8,1,1,2.5,...,S1]), and is divided into training and validation sets in a 7:3 ratio.

[0068] Based on a multi-dimensional model training dataset, the Random Forest algorithm was used to select key factors that significantly influence early warning results, and a preliminary personalized early warning model incorporating decision tree ensembles was constructed. The model employs the Random Forest algorithm, a decision tree-based ensemble learning model. Its core structure is a parallel ensemble of multiple independent decision trees. The basic unit consists of 100 CART (Classification and Regression Tree) decision trees. Each tree obtains independent samples from the training set through bootstrap sampling (sampling with replacement) to ensure diversity among trees. For classification tasks (such as risk level classification), a majority voting method is used; for regression tasks (such as risk probability prediction), a mean method is used, ultimately outputting a comprehensive result. During splitting, each tree randomly selects a subset of features (the subset size is the square root of the total number of features, such as randomly selecting 3-4 dimensions from 12 features) to avoid overfitting.

[0069] The number of decision trees (100) was determined by monitoring out-of-bag (OOB) error: as the number of trees increases, the OOB error gradually decreases and tends to stabilize (when the number of trees reaches 100, the error stabilizes at 8.5%). Further increasing the number of trees offers limited improvement in accuracy but increases computational cost; therefore, 100 trees were set as the optimal value. Feature selection mechanism (Gini coefficient screening, threshold 0.01): The Gini coefficient measures the impurity of a feature in classifying a sample; the calculation formula is... (in (This represents the probability that a sample belongs to class k). The model retains only features where the Gini coefficient decreases by ≥0.01. The key factors and their importance were finally selected as follows: Radiation damage acceleration factor (importance 0.25): directly reflects the rate at which dose changes affect risk; White blood cell count (0.20): as a physiological marker of radiation sensitivity, individuals with low white blood cell counts are more sensitive to radiation; Interventional procedure duration (0.18): the longer the duration, the higher the cumulative dose, and the risk increases linearly; Other minor factors: thyroid absorbed dose (0.15), XRCC1 genotyping (0.12), etc. Output format: The model output is "probability of radiation damage risk to key organs" (e.g., thyroid damage risk %), ranging from 0-100%, accurate to one decimal place, to facilitate subsequent risk level classification (e.g., >30% is high risk).

[0070] Taking a multi-dimensional dataset of 300 medical staff (including a high-sensitivity group S1, a medium-sensitivity group S2, and a low-sensitivity group S3) as an example: In the 30% validation set (90 people), the overall accuracy of the model reached 89%, meaning that 80 people were correctly predicted to have the risk level. The confusion matrix showed that the high-sensitivity group (S1) had the highest recognition rate (92%), with only 2 people misclassified as medium-sensitivity group (S2), a misclassification rate of 8%, demonstrating accurate identification of high-risk groups. The recognition rate of the medium-sensitivity group (S2) was 88%, and the recognition rate of the low-sensitivity group (S3) was 86%, showing good overall balance. Through partial dependency graph analysis, when the acceleration factor increased from 0.02 to 0.05, the risk probability of group S1 increased by an average of 15%, verifying the significant impact of key factors on the model output. This model, through the stability of ensemble learning and the targeted nature of feature selection, achieved differentiated risk assessment for different radiation-sensitive populations, providing core algorithmic support for subsequent personalized early warning.

[0071] The initial personalized early warning model was cross-validated and its parameters were tuned to generate the final personalized early warning model. Five-fold cross-validation was performed by randomly dividing the dataset into five mutually exclusive subsets (each subset containing 20% ​​of the samples). One subset was selected sequentially as the validation set, and the remaining four were used as the training set. This process was repeated five times, and the average of the five results was taken as the model performance metric. This method effectively avoids the bias caused by a single partition and ensures the reliability of parameter tuning.

[0072] To address potential overfitting issues in the initial model (e.g., overfitting the training set but showing large prediction deviations on new data), two core parameters were optimized: the maximum tree depth was adjusted from the default "unlimited" to 15. Excessively deep decision trees can lead to overfitting to detailed data (e.g., capturing noise rather than patterns). Limiting the maximum depth to 15 preserves the hierarchical relationships of key features (e.g., the transmission chain of acceleration factor → dose accumulation → risk probability) while avoiding redundancy, allowing the model to focus more on generalizable patterns. The minimum number of samples per leaf node was set to 5. Leaf nodes are the final output units of the decision tree. This parameter ensures that each leaf node contains at least 5 samples, avoiding extreme predictions due to insufficient sample size (e.g., accidental results with only 1 sample), and improving the model's adaptability (generalization ability) to new data.

[0073] After parameter optimization, the model's performance indicators were significantly improved. The accuracy increased from 89% in the initial model to 91%, meaning that the proportion of correctly predicting radiation damage risk levels in randomly selected test samples increased by 2 percentage points. The AUC value (area under the ROC curve) increased from 0.91 to 0.93, which is closer to 1.0 (the ideal value), indicating that the model's ability to distinguish between high-risk and low-risk groups has been enhanced, especially in borderline cases (such as samples with risk probabilities close to the threshold).

[0074] The optimized model implements differentiated early warning threshold adjustments for different sensitivity groups. Optimization for the highly sensitive group (S1 group) is particularly crucial. The early warning threshold for S1 group is reduced by 10% compared to the initial model (e.g., from "risk probability > 30% triggers early warning" to "> 27% triggers early warning"), ensuring that this group is identified in the early stages of risk accumulation, buying time for protective measures. The final personalized early warning model still contains 100 decision trees, but the splitting logic of each tree is simpler (maximum depth 15), and the leaf nodes are more robust (≥ 5 samples). This retains the stability of ensemble learning while achieving "accurate early warning without lag and no risk omission" through parameter constraints. This process, through scientific validation methods and targeted parameter adjustments, enables the model to maintain overall performance while strengthening the protection of highly sensitive groups, providing reliable algorithmic support for personalized early warning in medical radiation protection.

[0075] S107 generates graded early warning signals and protective measures recommendations based on the risk probability output of the personalized early warning model and the spatial correlation information between radiation type and exposure site.

[0076] In one implementation, the risk probability output by the personalized early warning model is associated with the spatial correlation information of radiation type and exposure site. The risk probability corresponds one-to-one with the spatial location of the corresponding radiation type and exposure site in the 3D human body model. This model is a spatial data association model, employing a three-layer architecture: a coordinate reference layer, a rule mapping layer, and an association output layer. The coordinate reference layer uses the voxel coordinates of the 3D human body model as a reference (e.g., a 1mm×1mm×1mm voxel grid from an ICRP standard phantom), employing a Cartesian coordinate system (X-axis for left-right, Y-axis for front-back, and Z-axis for up-down), with coordinate accuracy controlled within ±1mm to ensure each voxel unit has a unique and accurate spatial identifier. The rule mapping layer has a built-in association rule engine that presets the mapping logic of "risk probability - radiation type - spatial coordinates." For example, when the risk probability points to the thyroid gland, it automatically matches its standard coordinate range in the model and associates it with common radiation types in that area (e.g., X-rays in interventional surgery). The associated output layer outputs the matching results in the form of structured data, including three core fields: risk probability value, radiation type label, and spatial coordinate range, ensuring that the downstream visualization module can directly call them.

[0077] A Cartesian coordinate system is used, with the X, Y, and Z axes set to an accuracy of ±1mm. This parameter references the standard resolution of human voxel models (1mm voxel can clearly distinguish fine parts such as the thyroid gland and the hand), ensuring that the coordinate range can accurately define specific organs or tissues. For example, the standard coordinate range of the thyroid gland in an adult male phantom is predefined as X: 50-70mm, Y: 120-140mm, and Z: 80-100mm. This range is determined based on the anatomical data of the ICRP110 phantom and closely matches the anatomical location of the actual human thyroid gland.

[0078] The rule design follows the clinical correlation between "location-radiation type". For example, the high risk probability of the right palm is usually associated with "interventional surgery X-ray" (because the doctor's hand is close to the radiation source during the operation); the risk probability of the thyroid region of the chest is mostly associated with "CT scan scattered X-ray" or "interventional surgery direct X-ray". By pre-setting these clinical scenario correlations, the model can quickly bind the risk probability to the corresponding radiation type without manual intervention.

[0079] During an interventional procedure, a physician's personalized early warning model output "Right palm risk probability 28%". The spatial correlation mapping model's processing is as follows: Coordinate matching: Based on the exposed area "right palm", the model automatically retrieves the pre-stored standard coordinate range of the right palm: X:120-150mm, Y:300-330mm, Z:50-80mm. This range covers the voxel area of ​​the right palm from the fingertips to the base of the palm. Radiation type correlation: Based on the rule engine, "right palm" is strongly correlated with the "interventional procedure" scenario, therefore the radiation type is matched as "interventional procedure X-ray". The final correlation result is "risk probability 28% → interventional procedure X-ray → X:120-150mm, Y:300-330mm, Z:50-80mm", achieving precise binding of the three. This process ensures that the abstract risk probability data is given specific spatial location and radiation source information, providing a precise spatial benchmark for subsequent 3D visualization marking and protection suggestion generation, upgrading the early warning information from "numerical" to "spatial", which is more in line with the actual needs of clinical protection.

[0080] Based on associated information, high-risk exposure areas are color-coded in a 3D human body model to visually represent the spatial distribution of areas with different risk levels. A 3D visualization marking model (a color mapping system based on voxel rendering) is used, with parameters set to classify risk levels as follows: low risk (<20%, blue), medium risk (20%-50%, yellow), and high risk (>50%, red); rendering accuracy: 2mm×2mm×2mm voxel resolution for high-risk areas and 5mm×5mm×5mm for low-risk areas.

[0081] The physician's thyroid region has a risk probability of 35% (medium risk), which is marked in yellow in the 3D model. The right palm has a risk probability of 28% (medium risk), which is also marked in yellow, while the leg has a risk probability of 5% (low risk), which is marked in blue, visually presenting the spatial distribution of high-risk areas.

[0082] The risk levels of each exposed site were quantified and organized to generate a list of data including the name of the exposed site, radiation type, risk level, and warning prompts. A structured list generation model (field mapping and rule engine) was used. The list fields and rules are as follows: the exposed site is named according to anatomical standards (e.g., "thyroid gland" or "right palm skin"); the warning prompts are automatically generated based on the risk level and radiation type (e.g., for high-risk X-ray exposure, the prompt is "Immediately reduce operation time and enhance shielding"). An example list is shown in Table 3.

[0083] Table 3: Quantification of Risk Levels for Each Exposed Site By integrating information such as risk probability and exposure site distribution, an individualized risk assessment table and dose cumulative trend chart are generated to quantitatively display the evolution of radiation risk. A quantitative report generation model (statistical analysis and time series visualization engine) is used, with parameters including: assessment table dimensions: cumulative dose this week, dose to key organs, and deviation compared with the same period in history; trend chart parameters: the horizontal axis is time (last 30 days), the vertical axis is cumulative dose (mSv), and the curve marks the risk level threshold line (e.g., the dose value corresponding to 20% risk).

[0084] The assessment form shows "This week's cumulative dose is 0.8 mSv, and the thyroid dose is 0.35 mSv (an increase of 12% compared to last week)"; the blue curve in the trend chart represents the actual dose, and the red dashed line represents the medium-risk threshold (1.2 mSv), visually demonstrating the trend of the dose approaching the threshold.

[0085] This system integrates a 3D human body visualization model, a risk level list, an individualized assessment table, and trend charts to form a complete multi-dimensional early warning information output. The key represents the exposed site identifier, and the value includes the risk probability, radiation type, protective recommendations, and related visualization data for the corresponding site. A multi-source data fusion model (key-value pair structured storage) is used. The output structure is: Key: Unique identifier of the exposed site (e.g., "THY-001" represents the thyroid gland); Value: Includes risk probability, radiation type, protective recommendations, 3D coordinates, and visualization rendering parameters. Specifically, {"THY-001":{"Risk Probability":"35%","Radiation Type":"X-ray","Protective Recommendation":"Lead neck brace + shorten operation time to 15 minutes / time","3D Coordinates":"X:50-70,Y:120-140,Z:80-100","Rendering Color":"#FFFF00" (yellow)}, "HAND-002":{"Risk Probability":"28%","Radiation Type":"X-ray","Protection Recommendation":"0.5mm lead gloves + operating distance ≥30cm","3D Coordinates":"X:120-150,Y:300-330,Z:50-80","Rendering Color":"#FFFF00" (Yellow)}}. This output integrates spatial visualization, quantitative data, and action recommendations, providing healthcare workers with intuitive and actionable radiation protection guidance.

[0086] In one implementation, such as Figure 2 As shown, this application also provides a medical radiation protection early warning device based on personal dose monitoring, comprising: The acquisition module 201 is used to acquire real-time personal radiation exposure data and historical dose records. Based on the multi-source data fusion technology of thermoluminescent dosimeter and real-time radiation detector, an adaptive filtering algorithm is used to remove environmental interference noise and generate a standardized dose monitoring sequence. Processing module 202 is used to convert standardized dose sequences into three-dimensional dose distribution maps, process them using Monte Carlo simulation software, subdivide high-dose areas using a stratified sampling algorithm, and construct individualized radiation accumulation models. Radiation damage is defined as a time-dose dependent process; a damage assessment equation is constructed using a linear quadratic model coupled with a dose rate correction term; weighting factors for different radiation types are extracted to construct a dynamic dose conversion matrix; parallel calculations are performed on multi-organ dose distributions to obtain the radiation absorbed dose of key organs; peak dose, accumulation rate, and organ dose difference values ​​are extracted according to radiation exposure scenarios to classify risk levels; a multi-dimensional feature matrix containing exposure time, dose rate, and organ sensitivity information is constructed as input to the early warning model; real-time doses are compared with historical data to identify abnormal signals such as short-term dose spikes and cumulative doses approaching thresholds; the slope of the dose-effect curve is used to define radiation damage acceleration factors; exposed individuals are grouped based on their personal health records; a random forest algorithm is used to screen key influencing factors; and dose parameters, physiological indicators, and exposure scenario information are integrated to construct personalized early warning models; based on the risk probability output of the personalized early warning model, combined with the spatial correlation information between radiation type and exposure site, graded early warning signals and protective measure recommendations are generated.

[0087] The computer-readable storage medium provided in the above embodiments of this application and the medical radiation protection early warning method based on personal dose monitoring provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the application stored therein.

[0088] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the medical radiation protection early warning method, electronic device, electronic device, and readable storage medium based on personal dose monitoring are basically similar to the embodiments of the medical radiation protection early warning method based on personal dose monitoring described above, and are therefore described simply. Relevant parts can be referred to in the descriptions of the embodiments of the medical radiation protection early warning method based on personal dose monitoring described above.

Claims

1. A medical radiation protection early warning method based on personal dose monitoring, characterized in that, include: Acquire real-time personal radiation exposure data and historical dose records. Based on multi-source data fusion technology of thermoluminescent dosimeter and real-time radiation detector, use adaptive filtering algorithm to remove environmental interference noise and generate standardized dose monitoring sequence. After converting the standardized dose sequence into a three-dimensional dose distribution map, it was processed by Monte Carlo simulation software, and a stratified sampling algorithm was used to subdivide the high-dose region to construct an individualized radiation accumulation model. Radiation damage is defined as a time-dose dependent process. A linear quadratic model coupled with a dose rate correction term is used to construct a damage assessment equation. Weighting factors for different radiation types are extracted to construct a dynamic dose conversion matrix. The dose distribution of multiple organs is calculated in parallel to obtain the radiation absorbed dose of key organs. Based on the radiation exposure scenario, peak dose, cumulative rate, and organ dose difference values ​​are extracted, risk levels are classified, and a multi-dimensional feature matrix containing exposure time, dose rate, and organ sensitivity information is constructed as input for the early warning model. By comparing real-time dose with historical data from the same period, abnormal signals such as short-term dose spikes and cumulative dose approaching the threshold are identified, and the radiation damage acceleration factor is defined using the slope of the dose-effect curve. Individuals exposed to the disease are grouped based on their personal health records. A random forest algorithm is used to screen key influencing factors, and a personalized early warning model is constructed by integrating dose parameters, physiological indicators and exposure scenario information. Based on the risk probability output of the personalized early warning model, and combined with the spatial correlation information between radiation type and exposure site, early warning information with graded early warning signals and protective measures recommendations is generated.

2. The method as described in claim 1, characterized in that, After converting the standardized dose sequence into a three-dimensional dose distribution map, it was processed using Monte Carlo simulation software. A stratified sampling algorithm was then used to subdivide the high-dose region, constructing an individualized radiation accumulation model, including: The standardized dose sequence is converted into a three-dimensional dose distribution map, providing a spatial dose data basis for subsequent Monte Carlo simulations; Import the three-dimensional dose distribution map into the Monte Carlo simulation software to build a computational environment for cumulative radiation dose simulation; Based on the dose difference characteristics of radiation exposure areas, a stratified sampling algorithm was selected as the core method for regional subdivision to determine the identification and division rules of high-dose areas. Stratified sampling and subdivision processing are performed for high-dose areas to improve the accuracy and spatial resolution of dose simulation in these areas; The software generates an individualized radiation accumulation model that includes dose characteristics of subdivided regions, forming a quantitative analysis carrier that meets the needs of radiation damage assessment.

3. The method as described in claim 1, characterized in that, Radiation injury is defined as a time-dose dependent process. A linear quadratic model coupled with a dose rate correction term is used to construct a damage assessment equation. Weighting factors for different radiation types are extracted to construct a dynamic dose conversion matrix. Parallel calculations are performed on the dose distribution of multiple organs to obtain the radiation absorbed dose of key organs, including: Radiation damage is defined as a time-dose dependent process. A linear quadratic model coupled with a dose rate correction term is used to construct a damage assessment equation. The dose rate correction term is used to reflect the effect of different irradiation rates on the degree of damage, making the equation more consistent with the time-cumulative characteristics of radiation damage. Weighting factors for different radiation types are extracted to construct a dynamic dose conversion matrix. This matrix is ​​used to uniformly convert various radiation doses into equivalent biological doses, ensuring the comparability and accumulation of multi-source radiation doses. The individualized radiation accumulation model is input into the damage assessment equation, and multi-threaded parallel computing technology is used to perform synchronous calculations on the dose distribution of multiple organs to obtain the radiation absorbed dose of key organs, including the average absorbed dose of the organ, the local hot spot dose, and the dose-volume histogram parameters.

4. The method as described in claim 1, characterized in that, By comparing real-time dose with historical data from the same period, abnormal signals such as short-term dose spikes and cumulative dose approaching the threshold are identified. Radiation damage acceleration factors are defined using the slope of the dose-effect curve, including: By comparing real-time dose data with historical dose data from the same period, the deviation values ​​and trends of the two types of data are extracted, and a dynamic dose comparison chart is generated. Based on the dose dynamic comparison spectrum, a short-term dose surge threshold and a cumulative dose warning line are set to identify abnormal signals that exceed the threshold. Abnormal signals include short-term dose surge signals and cumulative dose approaching the threshold signals. A dose-response curve is introduced, and the slope of the curve is calculated in different dose ranges. The slope value is defined as the radiation damage acceleration factor, which is used to quantify the rate of effect of radiation dose change on damage risk. Correlation analysis was performed on anomalous signals and radiation damage acceleration factors to generate anomaly-acceleration factor correlation matrix, providing dynamic parameter support for subsequent early warning models.

5. The method as described in claim 1, characterized in that, Individuals exposed to the virus were grouped based on their personal health records. A random forest algorithm was used to screen key influencing factors, and personalized early warning models were constructed by integrating dose parameters, physiological indicators, and exposure scenario information. These models included: Based on individual health records, key health indicators affecting radiation sensitivity were extracted, and the exposed individuals were divided into high-sensitivity, medium-sensitivity, and low-sensitivity groups according to their radiation tolerance levels. Each group was assigned a unique sensitivity label. Comparative analysis was conducted on dose parameters, physiological indicators, and exposure scenario information of different sensitive groups to generate characteristic difference data between groups; The feature difference data, dose parameters, physiological indicators and exposure scenario information are integrated and processed to generate a multi-dimensional model training dataset, in which the dose parameters include the absorbed dose of key organs and radiation damage acceleration factor. Based on the multi-dimensional model training dataset, the random forest algorithm was used to screen out the key factors that have a significant impact on the early warning results, and a preliminary personalized early warning model including decision tree ensemble was constructed. Cross-validation and parameter tuning of the preliminary personalized early warning model are performed to generate the final personalized early warning model.

6. The method as described in claim 5, characterized in that, Based on the risk probability output of the personalized early warning model, and combined with the spatial correlation information between radiation type and exposure site, a graded early warning signal and protective measure recommendations are generated, including: The risk probability output by the personalized early warning model is associated with the spatial correlation information of radiation type and exposure site. The risk probability corresponds one-to-one with the spatial location of the corresponding radiation type and exposure site in the three-dimensional human body model. Based on the associated information, high-risk exposure areas are marked with color grading in the 3D human body model to intuitively present the spatial distribution of areas with different risk levels. The risk levels of each exposed site are quantified and organized to generate a list of data including the name of the exposed site, radiation type, risk level, and early warning prompts; By integrating information such as risk probability and exposure site distribution, a personalized risk assessment table and dose accumulation trend chart are generated to quantitatively display the evolution of radiation risk. By integrating a 3D human body visualization model, a risk level list, an individualized assessment table, and trend charts, a complete multi-dimensional early warning information output is formed. The key is the exposure site identifier, and the value is the risk probability, radiation type, protection recommendations, and visualized related data for the corresponding site.

7. A medical radiation protection early warning device based on personal dose monitoring, characterized in that, This method is used to implement the medical radiation protection early warning method based on personal dose monitoring as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the medical radiation protection early warning method based on personal dose monitoring as described in any one of claims 1 to 6 by executing the executable instructions.

Images