Deep learning-based multi-source element feature and mt hifu efficacy prediction method and system

Abstract

The present application relates to a deep learning-based multi-source element feature and HIFU efficacy prediction method and system for MT, obtains the concentration data of trace elements such as iron, zinc, copper and selenium and the concentration data of MT of different historical patients at multiple time points on a medical platform and pre-processes; extracts multi-dimensional features such as longitudinal dynamic change trend of trace elements, high / low shrinkage rate group difference elements and time-group interaction effect elements, constructs a multi-source element-MT joint feature set; divides it into a training set and a test set, inputs a random forest deep learning model containing recursive feature elimination for training and testing; inputs the model after screening the key features of the patient data to be tested, outputs the efficacy prediction result and evaluates the final efficacy according to the myoma shrinkage rate standard. The system includes multi-source historical data acquisition and preprocessing, multi-source element-MT joint feature construction, deep learning model and efficacy final evaluation module. Through multi-dimensional feature mining and deep learning technology, the present application realizes accurate prediction of HIFU efficacy.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and data processing technology, and in particular relates to a method and system for predicting the efficacy of HIFU based on multi-source element features and MT using deep learning. Background Technology

[0002] Uterine fibroids are the most common benign pelvic tumors among women of reproductive age, with a global prevalence exceeding 70%. This disease exhibits significant clinical heterogeneity, with diverse growth patterns and symptoms, often leading to abnormal uterine bleeding, pelvic compression symptoms, and reproductive dysfunction. Current treatment options include medication, uterine artery embolization, myomectomy, and hysterectomy, but each has its own limitations in indications and risks of complications.

[0003] High-intensity focused ultrasound (HIFU), a completely non-invasive thermal ablation technique, induces coagulative necrosis of fibroid tissue by precisely focusing ultrasound waves onto the target tissue to generate thermal and cavitation effects. Numerous clinical studies have confirmed that HIFU treatment has significant advantages in preserving uterine function, rapidly relieving symptoms, and improving quality of life; fibroid volume can be significantly reduced after treatment, and the symptom improvement rate can reach 79.3%. However, individual differences exist in treatment efficacy, necessitating the establishment of an effective system for predicting and evaluating treatment effectiveness.

[0004] Trace elements play a vital role in maintaining normal human physiological functions, participating in key biological processes such as the regulation of various enzyme activities, redox balance, and immune regulation. In gynecology, the metabolic balance of trace elements such as iron, zinc, copper, and selenium is closely related to the health of the female reproductive system, and their abnormal expression is associated with the occurrence and development of various gynecological diseases. The tissue thermal damage and repair processes induced during HIFU treatment may cause dynamic changes in the metabolism of trace elements in the body; however, there is a certain correlation between these changes and the treatment effect.

[0005] In existing technologies, research on HIFU efficacy prediction mainly focuses on dimensions such as imaging features (e.g., fibroid size, echogenicity, blood flow) and clinical indicators (e.g., patient age, fibroid type). However, these methods have obvious limitations: on the one hand, the acquisition of imaging features depends on specialized equipment and operation, and the predictive sensitivity for early efficacy is insufficient; on the other hand, the singularity of clinical indicators makes it difficult to cover the complex influencing factors of HIFU efficacy, resulting in limited prediction accuracy.

[0006] In recent years, some studies have attempted to explore new directions for predicting therapeutic efficacy from a biochemical perspective. Changes in the concentrations of trace elements (iron, zinc, copper, selenium, etc.) and metallothionein (MT) have been found to be closely related to metabolism, tissue repair, and the biological effects of HIFU treatment. However, current research in this field still suffers from the following key shortcomings: The feature extraction dimension is singular: it only focuses on the absolute value of element concentration at a specific time point, without fully exploring the longitudinal dynamic change trend of elements at multiple time points such as before treatment, 1 day after surgery, 3 months, 6 months, and 1 year, and without deeply analyzing the differences in element concentration between high / low shrinkage rate groups and the interaction effect of time and efficacy group on element concentration, resulting in the biological significance and discriminative power of the features being seriously underestimated.

[0007] Insufficient fusion of multi-source data: The lack of joint feature construction of trace element + MT multi-source biochemical data, the failure to fully utilize the synergistic effect between multiple biomarkers, and the limitation of the information capacity and generalization ability of the prediction model.

[0008] Therefore, this application aims to provide a method and system for predicting the efficacy of HIFU based on multi-source element features and MT using deep learning, so as to achieve accurate and efficient prediction of the efficacy of HIFU. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for predicting the efficacy of HIFU based on multi-source element features and MT using deep learning, so as to achieve accurate and efficient prediction of the efficacy of HIFU.

[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: Firstly, a method for predicting the efficacy of HIFU based on multi-source elemental features and MT (Medium-to-Material) using deep learning is provided, including the following steps: S1: Obtain patient data from the medical platform at at least five time points before treatment, 1 day after HIFU, 3 months, 6 months, and 1 year after treatment, including trace element concentration data and MT concentration data, and perform preprocessing. The trace elements include iron, zinc, copper, and selenium. S2: Based on trace element concentration data, extract trace element features including the longitudinal dynamic change trend of each element, the difference between high / low shrinkage rate groups, and the time-group interaction effect elements; combine with MT concentration data to construct a multi-source element-MT joint feature set; S3: Divide the preprocessed multi-source element-MT joint feature set into a randomized training set and a test set, and input them into the constructed deep learning prediction model for training and testing, respectively. S4: Obtain the multi-source element-MT combined feature data of the patient to be predicted, perform key feature screening on the multi-source element-MT combined feature data, input the screened key features into the trained deep learning prediction model, and output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group or the low shrinkage rate group after HIFU treatment. S5: Based on the prediction results and the fibroid shrinkage rate, a high shrinkage rate is determined according to the criteria of SR≥30% at 3 months post-surgery, SR≥50% at 6 months post-surgery, and SR≥70% at 1 year post-surgery, thus obtaining the final evaluation result of the efficacy of HIFU treatment.

[0011] Preferably, the specific process of extracting trace element features based on trace element concentration data in step S2, including longitudinal dynamic change trend features of each element, inter-group difference features of high / low shrinkage rate elements, and time-group interaction effect elements, is as follows: S21: Extract the longitudinal dynamic trend characteristics of each element: Quantify the concentration change of each element using the same patient + the same element + different time points as the smallest unit: obtain the total change rate, stage change rate, change slope, fluctuation coefficient and recovery coefficient of each element; S22: Extract the characteristic features of differences between high / low shrinkage rate groups: screen for elements with statistically significant differences in concentration among different efficacy groups at the same time point, and quantify the differences; S23: Extraction Time-Group Interaction Effect Element Characteristics: Identify elements whose concentration changes are simultaneously affected by both time and group.

[0012] Preferably, the specific process of step S22 is as follows: S221: Group Identification: Based on the efficacy criteria, all patients were identified as: SR-High group: Meets any of the above stage standards; SR-Low group: Does not meet all of the above stage criteria; S222: Based on the SR-High and SR-Low groups, verify the normality and homogeneity of variance of the two groups of data; S223: Based on the results of normality and homogeneity of variance testing, the difference elements between the two groups are screened by calculating the inter-group difference statistics and the corresponding p-values; S224: Calculate for each historical patient based on the selected differential elements: relative bias between groups, concentration fold ratio, and group discrimination probability.

[0013] Preferably, the specific process of performing the normality test and homogeneity of variance test in step S222 is as follows: S2221: Normality test: ; x ( i ): The concentration value of the i-th element after sorting from smallest to largest; ai Shapiro-Wilk test coefficient; W: The mean concentration of the elements in this group; W: The test statistic; the probability corresponding to the W value. pW If the value is greater than 0.05, the data follows a normal distribution; otherwise, it is not normally distributed. S2222: Homogeneity of variance test: ; in, k =2; N=n1+n2, the total sample size of the two groups; zij The absolute deviation of the i-th sample in the j-th group, i.e. , xij Let i be the concentration of the i-th element in the j-th group. The mean of the j-th group; : The mean of the absolute deviations of the j-th group; The overall mean of all absolute deviations; F: The test statistic, which follows an F-distribution with k−1 and N−k degrees of freedom; the probability corresponding to the F-value. p F If the variance is greater than 0.05, the two groups have homogeneous variances; otherwise, they have heterogeneous variances.

[0014] Preferably, the specific process of step S223 is as follows: If the distribution is normal and the variance is homogeneous: perform an independent samples t-test. ; Among them, the combined variance The calculation formula is as follows: ; , The mean elemental concentrations for the SR-High and SR-Low groups are respectively. , The variances of the two groups are respectively. To combine variances; t is the test statistic, which follows a t-distribution with degrees of freedom df = n1 + n2 − 2; If the calculated t-value corresponds to a two-tailed probability < 0.05, then this element is a between-group difference element. If the distribution is non-normal and the variance is unequal: First, calculate the sum of the ranks of the two groups: merge and sort the two groups of data, assign ranks according to their positions, and then sum them separately: R1: Rank sum of the SR-High group, n1 is the number of cases in the group; R2: The rank sum of the SR-Low groups, where n2 is the number of cases in the group; Next, calculate the U statistic (take the smaller value to avoid bias): ; ; ; U is the test statistic, reflecting the degree of difference between the rank sums of the two groups. The smaller U is, the greater the difference between the groups. If the probability corresponding to the value of U p U If the value is less than 0.05, then the element is a difference element between groups.

[0015] Preferably, the deep learning prediction model includes: Input layer: Input multi-source element features + MT data; Feature processing layer: Filters core features through RFE; Core layer: consists of 100 decision trees. Each tree is trained based on random feature selection and random sample sampling. The final result is output through a voting mechanism. For example, if 51 trees predict SR-High, then the final prediction is SR-High. Output layer: Binary classification result + predicted probability; Training optimization: Performance is evaluated using out-of-bag error, and hyperparameters including tree depth and node splitting threshold are optimized.

[0016] Preferably, in step S4, the key feature screening of the multi-source element-MT joint feature data to be tested is achieved through recursive feature elimination: S41: Using random forest as the base model for recursive feature elimination, specify the parameters for recursive feature elimination: Initial feature set: that is, all features of the multi-source element-MT joint feature set; Target feature number: determined by hierarchical 5-fold cross-validation. First, multiple candidate values ​​are preset. After performing RFE on each candidate value, the cross-validation AUC is evaluated using the base model. The candidate value with the highest AUC and the fewest feature numbers is selected. Number of features deleted per iteration: 1 least important feature is deleted per iteration; S42: Based on the base model, key features are selected from the multi-source element-MT joint feature data to be tested.

[0017] Preferably, the specific process of step S42 is as follows: S421: Initialize the base model: Set the number of decision trees in the random forest to 100, the tree depth to 8, and the random seed to 42; S422: The importance of the multi-source element-MT combined feature data and efficacy labels to be tested is scored using the base model, and the features with the lowest importance scores are deleted until the number of features is reduced to the target number of features.

[0018] Secondly, a HIFU efficacy prediction system based on deep learning multi-source element features and MT is provided to implement any of the HIFU efficacy prediction methods based on deep learning multi-source element features and MT as described above. This system includes a multi-source historical data acquisition and preprocessing module, a multi-source element-MT joint feature construction module, a deep learning model, and a final efficacy evaluation module, all connected in sequence. These modules work together to automate the entire HIFU efficacy prediction process from data processing to evaluation. The multi-source historical data acquisition and preprocessing module is used to acquire multi-time point data of different historical patients on the medical platform and perform preprocessing. The multi-source element-MT joint feature construction module is used to extract features based on preprocessed trace element concentration data and construct a joint feature set by combining it with preprocessed MT concentration data. The deep learning model is used to output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group or the low shrinkage rate group after HIFU treatment. The final efficacy evaluation module is used to determine the high shrinkage rate based on the prediction results and the fibroid shrinkage rate, and to obtain the final evaluation result of the efficacy of HIFU treatment.

[0019] The beneficial effects of this invention include: 1. Multi-dimensional feature extraction enables in-depth mining and precise characterization of therapeutic effects: By extracting multi-dimensional features such as longitudinal dynamic trends at multiple time points, differences between high / low shrinkage rate groups, and time-group interaction effects, the biological behavior of trace elements and MT during HIFU treatment was fully explored. Compared with the limitations of existing technologies that only focus on the absolute concentration value at a single time point, the feature set of this application has greater biological discriminative power and can accurately capture the intrinsic relationship between element concentration changes and HIFU efficacy, providing high-value input information for subsequent prediction models.

[0020] 2. Multi-source data fusion enhances the information capacity and generalization ability of prediction models: We innovatively constructed a multi-source element-MT joint feature set, combining trace elements and micronutrients (MT), breaking through the limitations of existing technologies that rely on single biomarkers or single-class data. By leveraging the synergistic effects of trace elements and MT in metabolism and tissue repair, we significantly enhanced the information dimensionality and biological interpretability of the feature set. This allows the predictive model to more comprehensively learn the complex correlation between multi-source biomarkers and HIFU efficacy, significantly improving the model's generalization ability and robustness.

[0021] 3. Combining deep learning with key feature selection enables accurate and efficient prediction of treatment efficacy: The combination of a random forest deep learning model and recursive feature elimination (RFE) effectively captures the nonlinear, high-dimensional correlation between multi-source element features and efficacy through a multi-decision tree voting mechanism and out-of-bag error assessment, significantly improving prediction accuracy (e.g., the efficacy prediction accuracy at 3 months, 6 months, and 1 year post-surgery can reach over 90%). On the other hand, RFE filters key features, eliminates redundant information, reduces model complexity, and improves the interpretability of prediction results, making them easier for clinical understanding and application. Attached Figure Description

[0022] Figure 1 Machine Learning-Based Identification of SR-High and SR-Low Groups from HIFU-1D Plasma Specimens. A) ROC curves and AUC values ​​of each model. B) Comparison of ROC performance among models. C) RF had the highest AUC (95.8%). D) Selection of elements used in the RF model. E) Independent validation of the RF model (AUC = 90.5). F) Radar chart summarizing the performance of RF in terms of accuracy, AUC, recall, precision, and F1 score. Larger areas indicate stronger overall performance. Abbreviations: ROC, Receiver Operating Characteristic; AUC, Area Under the ROC Curve; ANN, Artificial Neural Network; DT, Decision Tree; KNN, k-Nearest Neighbors; LR, Logistic Regression; NB, Naïve Bayes; RF, Random Forest; Support Vector Machine.

[0023] Figure 2 Machine Learning-Based Identification of SR-High and SR-Low Groups from HIFU-1D Urine Specimens. A) ROC curves and AUC values ​​of each model. B) Comparison of ROC performance among models. C) ANN had the highest AUC (93.8%). D) Selection of elements used in the ANN model. E) Independent validation of the ANN model (AUC = 90.5). F) Radar chart summarizing the performance of ANN in terms of accuracy, AUC, recall, precision, and F1 score. Larger areas indicate stronger overall performance. Abbreviations: ROC, Receiver Operating Characteristic; AUC, Area Under the ROC Curve; ANN, Artificial Neural Network; DT, Decision Tree; KNN, k-Nearest Neighbors; LR, Logistic Regression; NB, Naïve Bayes; RF, Random Forest; Support Vector Machine.

[0024] Figure 3Machine Learning-Based Identification of SR-High and SR-Low Groups from HIFU-1D Hair Specimens. A) ROC curves and AUC values ​​of each model. B) Comparison of ROC performance among models. C) RF had the highest AUC (81.2%). D) Selection of elements used in the RF model. E) Independent validation of the RF model (AUC = 90.5). F) Radar chart summarizing the performance of RF in terms of accuracy, AUC, recall, precision, and F1 score. Larger areas indicate stronger overall performance. Abbreviations: ROC, Receiver Operating Characteristic; AUC, Area Under the ROC Curve; ANN, Artificial Neural Network; DT, Decision Tree; KNN, k-Nearest Neighbors; LR, Logistic Regression; NB, Naïve Bayes; RF, Random Forest; Support Vector Machine.

[0025] Figure 4 Box plots of hair elemental concentrations stratified by contraction rate (SR_High vs. SR_Low) in patients with uterine fibroids (UF) at 3 months (A, HIFU-3M), 6 months (B, HIFU-6M), and 12 months (C, HIFU-1Y) after HIFU treatment. Only elements with statistically significant differences between groups are shown (*p<0.05, **p<0.01, ***p<0.001).

[0026] Figure 5 Box plots of plasma elemental concentrations in patients with uterine fibroids (UF) stratified by SR_High vs. SR_Low at 3 months (A, HIFU-3M) and 6 months (B, HIFU-6M) after HIFU treatment. Only elements with statistically significant differences between groups are shown (*p<0.05, **p<0.01, ***p<0.001).

[0027] Figure 6 Box plots of urinary elemental concentrations stratified by contraction rate (SR_High vs. SR_Low) in patients with uterine fibroids (UF) at 6 months (HIFU-6M) and 12 months (HIFU-1Y) after HIFU treatment. Only elements with statistically significant differences between groups are shown (*p<0.05, **p<0.01, ***p<0.001).

[0028] Figure 7The longitudinal changes in hair element concentrations in UF patients classified according to high and low shrinkage rates (SR) were measured at four time points (UF, HIFU-3M, HIFU-6M, and HIFU-1Y). Blue triangles represent concentrations in the SR-Low group. Red circles indicate elements with significant time effects, and green circles indicate significant group-time interactions in the SR-High group.

[0029] Figure 8 The longitudinal changes in plasma elemental concentrations in UF patients classified according to high and low systolic rates (SR) were measured at five time points (UF, HIFU-1D, HIFU-3M, HIFU-6M, and HIFU-1Y). Blue triangles represent concentrations in the SR-Low group. Red circles indicate elements with significant time effects, and green circles indicate significant group-time interactions in the SR-High group.

[0030] Figure 9 The longitudinal changes in urinary elemental concentrations in UF patients classified according to high and low systolic rates (SR) were measured at five time points (UF, HIFU-1D, HIFU-3M, HIFU-6M, and HIFU-1Y). Blue triangles represent concentrations in the SR-Low group. Red circles indicate elements with significant time effects, and green circles indicate significant group-time interactions in the SR-High group.

[0031] Figure 10 Pearson correlation analysis was performed on plasma elemental concentrations and clinical symptoms 3 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0032] Figure 11 Pearson correlation analysis was performed on urinary element concentrations and clinical symptoms 3 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0033] Figure 12 Pearson correlation analysis was performed on hair elemental concentrations and clinical symptoms 3 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0034] Figure 13 Pearson correlation analysis was performed on plasma elemental concentrations and clinical symptoms 6 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0035] Figure 14 Pearson correlation analysis was performed on urinary element concentrations and clinical symptoms 6 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0036] Figure 15 Pearson correlation analysis was performed on hair elemental concentrations and clinical symptoms 6 months post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0037] Figure 16 Pearson correlation analysis was performed on plasma elemental concentrations and clinical symptoms one year post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0038] Figure 17 Pearson correlation analysis was performed on urinary element concentrations and clinical symptoms one year post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0039] Figure 18 Pearson correlation analysis was performed on hair elemental concentrations and clinical symptoms one year post-surgery. Blue and red ellipses represent statistically significant negative and positive correlations, respectively (p<0.05). The gray numbers inside the ellipses represent the correlation coefficients.

[0040] Figure 19 Longitudinal changes in oxidative stress markers after HIFU treatment in patients with uterine fibroids. Plasma concentrations of copper-zinc superoxide dismutase (CuZn-SOD, A), malondialdehyde (MDA, B), and metallothionein (MT, C) were measured at the following time points: UF (before treatment), HIFU-1D (1 day post-treatment), HIFU-3M (3 months post-treatment), HIFU-6M (6 months post-treatment), and HIFU-1Y (1 year post-treatment). Different letters indicate significant differences between time points (lowercase: p < 0.05; uppercase: p < 0.01). The levels of CuZn-SOD (D), MDA (E), and MT (F) were compared between the SR_High and SR_Low groups at the HIFU-3M, HIFU-6M, and HIFU-1Y time points. An asterisk indicates a significant difference between groups (p < 0.05). Detailed Implementation

[0041] The following is in conjunction with the appendix Figures 1-19 The present invention will be further described in detail below: See appendix Figure 1As shown, the HIFU efficacy prediction method based on deep learning multi-source element features and MT includes the following steps: S1: Obtain patient data from the medical platform at at least five time points before treatment (UF), 1 day after HIFU (HIFU-1D), 3 months (HIFU-3M), 6 months (HIFU-6M), and 1 year (HIFU-1Y), including trace element concentration data and MT concentration data, and perform preprocessing. Trace elements include iron, zinc, copper, selenium, etc.

[0042] S2: Based on trace element concentration data, extract trace element features including the longitudinal dynamic change trend of each element, the difference between high / low shrinkage rate groups, and the time-group interaction effect elements; combine with MT concentration data to construct a multi-source element-MT joint feature set.

[0043] Among them, the differential elements between the high / low shrinkage rate groups refer to the trace elements that show statistically significant differences in concentration between the two groups after patients are divided into a high shrinkage rate group (SR-High) and a low shrinkage rate group (SR-Low) based on the fibroid shrinkage rate (SR) after HIFU treatment, and detected at the same time points (such as 3 months, 6 months, and 1 year after surgery). The concentration differences of these elements are related to the efficacy of HIFU treatment (the degree of fibroid shrinkage) and are one of the core bases for constructing efficacy prediction models and finding efficacy biomarkers.

[0044] Myoma shrinkage rate (SR) calculation: SR = [1 - (myoma volume at follow-up / myoma volume at baseline before surgery)] × 100%, which is the percentage reduction in myoma volume at follow-up compared to before treatment.

[0045] Grouping criteria: Three months post-surgery: SR ≥ 30% was classified as high shrinkage rate group (SR-High), and SR < 30% was classified as low shrinkage rate group (SR-Low).

[0046] Six months post-surgery: SR ≥ 50% is SR-High, SR < 50% is SR-Low; One year post-surgery: SR ≥ 70% is SR-High, SR < 70% is SR-Low.

[0047] In short, the SR-High group consists of patients who responded well to HIFU treatment (the fibroids shrank significantly), while the SR-Low group consists of patients who responded poorly to treatment (the fibroids shrank less).

[0048] S3: Divide the preprocessed multi-source element-MT joint feature set into a randomized training set and a test set, and input them into the constructed deep learning prediction model for training and testing, respectively.

[0049] S4: Obtain the multi-source element-MT combined feature data of the patient to be predicted, perform key feature screening on the multi-source element-MT combined feature data, input the screened key features into the trained deep learning prediction model, and output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group (SR-High) or the low shrinkage rate group (SR-Low) after HIFU treatment.

[0050] S5: Based on the prediction results, combined with the fibroid shrinkage rate SR=[1-(fibroid volume at follow-up / baseline fibroid volume before surgery)]×100%, a high shrinkage rate is determined according to the criteria of SR≥30% at 3 months after surgery, SR≥50% at 6 months after surgery, and SR≥70% at 1 year after surgery, and the final evaluation result of the efficacy of HIFU treatment is obtained.

[0051] Example 2 Based on Example 1, the specific process of extracting trace element features, including longitudinal dynamic change trend features of each element, inter-group difference element features based on high / low shrinkage rate, and time-group interaction effect elements in step S2 is as follows: S21: Extracting the longitudinal dynamic trend characteristics of each element: Quantifying the concentration changes of each element using the same patient + the same element + different time points as the smallest unit: Obtain the total rate of change, stage rate of change, slope of change, volatility coefficient, and coefficient of restitution for each element: The formula for calculating the total rate of change is as follows: ΔC total =(C 1Y -C UF ) / C UF ×100%, C 1Y =Concentration 1 year post-surgery, C UF =Baseline concentration, which is the initial concentration level of trace elements (iron, zinc, copper, selenium) and MT (metallothionein) in the patient's body before receiving HIFU treatment; Stage Change Rate: The change rate for each stage is obtained based on the technical formula of the total change rate, including ΔC. 1D-UF (Postoperative day 1 - baseline), ΔC 3M-1D (3M-1D), ΔC 6M-3M ΔC 1Y-6M ; where ΔC 1D-UF ΔC represents the rate of change in elemental concentration from before HIFU treatment to one day after the treatment. 3M-1D ΔC represents the rate of change in elemental concentration from one day post-surgery to three months post-surgery. 6M-3M ΔC represents the rate of change in elemental concentration from 3 to 6 months post-surgery. 1Y-6M The percentage change in elemental concentrations from 6 months to 1 year post-surgery; The formula for calculating the volatility coefficient is as follows: CV = (SD / Mean) × 100%, where SD is the standard deviation of concentration at each time point and Mean is the mean; The formula for calculating the coefficient of restitution is as follows: Rec=C 1Y / C 3M ×100%, which is the recovery rate relative to 3M one year after surgery; S22: Extract the characteristic features of differences between high / low shrinkage rate groups: screen for elements with statistically significant differences in concentration among different efficacy groups at the same time point, and quantify the differences; S23: Extraction Time-Group Interaction Effect Element Characteristics: Identifying elements whose concentration changes are simultaneously affected by both time and group: Constructing a linear mixed-effects model: Y ijk =β 0 +β 1 ×Time i +β 2 ×Group j +β 3 × ( Time i ×Group j ) +γ k +ε ijk ; Y ijk For the kth patient at the th i The time point, the first j Concentration of each group; β 0 represents the model intercept (mean baseline concentration for all patients); β 1 represents the time-dependent main effect coefficient, which reflects only the independent effect of "time" on concentration and is independent of the group. β The group-specific main effect coefficient reflects only the independent effect of "group differences" on concentration and is independent of time. β 3 × ( Time i ×Group j The time-group interaction effect term reflects how the effect of time varies with different groups; for example, the element concentration in the SR-High group increases faster over time. β 3 represents the time-group interaction effect coefficient; β3 is a non-preset value and needs to be obtained by fitting the training set data using a linear mixed-effects model: Data input: Standardized element concentrations of all patients in the training set are used as the dependent variable, Time, Group, and Time×Group (interaction term) are used as fixed-effect independent variables, and patient ID is used as a random effect (controlling for individual baseline differences); Model fitting: The model is constructed using the lme4 package in R language: lmer(Concentration~Time+Group+Time:Group+(1|PatientID),data=train_data); Concentration = standardized elemental concentration, Patient ID = unique patient identifier, 1|Patient ID represents the random intercept for each patient; Coefficient extraction: After model fitting, the coefficients corresponding to the "Time:Group" interaction item are output, which are... β 3.

[0052] Time i This is a time variable used to quantify a continuous variable at different follow-up time points after HIFU treatment. i Representing the i These time points are used to reflect the independent effect of time on element concentration after treatment (i.e., the main effect of time), and are the basic parameters for calculating the longitudinal dynamic change trend of elements (such as the total rate of change). Baseline status of UF (before treatment): Time i The value is 0, which represents the initial reference time point for all patients; The acute reaction period following treatment on the first day after surgery. Time i A value of 1 reflects the acute fluctuation of elements caused by HIFU thermal damage; The early recovery period after 3 months of postoperative treatment. Time i The value is 3, corresponding to the time point for evaluating the efficacy of fibroid shrinkage rate ≥30%; The treatment effect stabilized 6 months after surgery. Time i The value is 6, which corresponds to the time point for determining the efficacy of fibroid shrinkage rate ≥50%, at which point the element concentration enters a steady state. One year post-surgery, long-term follow-up period after treatment. Time i The value was 12, corresponding to the time point for determining the efficacy of fibroid shrinkage rate ≥70%, to verify the long-term efficacy.

[0053] Group jThis is a group variable, a binary variable used to distinguish the treatment groups to which patients belong. j Representing the j The groups are used to reflect the independent effect of different efficacy groups on element concentration (i.e., group main effect), and are the basic parameters for screening elements with differences between high / low shrinkage rate groups; SR-High (high shrinkage rate group) indicates good HIFU treatment effect. Group j The value is assigned to 1, and the classification criteria are: meeting any of the following: SR ≥ 30% at 3 months post-surgery, SR ≥ 50% at 6 months post-surgery, or SR ≥ 70% at 1 year post-surgery; SR-Low (low shrinkage rate group) indicates poor response to HIFU treatment. Group j The value was assigned to 0; the classification criteria were: SR < 30% at 3 months post-surgery, SR < 50% at 6 months post-surgery, and SR < 70% at 1 year post-surgery. γ k For the patient random effect, that is, an independent random intercept assigned to each patient, the training set data is automatically estimated by fitting a linear mixed-effects model; ε ijk Let be the random error, representing the th k The patient in i The time point, the first j For each group, the difference between the actual observed values ​​and the model predicted values ​​of elemental concentration is determined by... β 0、 β 1× Timeᵢ , β 2 ×Groupⱼ , β 3 ×(Timeᵢ×Groupⱼ) , γ k This is jointly determined to address uncontrollable fluctuations that cannot be explained by quantitative models, such as detection errors and short-term physiological fluctuations. Filter elements with significant interaction effects: A Wald test (α=0.05) was performed on the Time×Group interaction term in the model. Elements with p<0.05 after correction were identified as interaction effect elements. For the selected interactive elements, calculate three types of core features: The interaction effect coefficient, which is directly extracted from the model. β The value of 3; Time-group difference series: Calculate the concentration difference between the two groups at each time point: Δ = μ SR-High -μ SR-Low Arranged in chronological order (ΔUF, Δ1D, Δ3M, Δ6M, Δ1Y); Interaction trend bias: Δ pred =C 实际 -( β 0 +β 1 ×Time+β 2 ×Group+γ k ); Characterizing the deviation between actual concentrations and predicted values ​​without interaction effects, the core objective is to isolate the interaction effect term. β 3×( Time×Group The influence of time-group interaction effects was analyzed using the LMM model when screening for elements with time-group interaction effects. The effects of time (different stages after treatment), group (SR-High / SR-Low), and their interaction were examined.

[0054] The specific derivation process is as follows: If we assume no interaction effect, i.e., β3=0, the complete model simplifies to: Y ijk，无交互 =β 0 +β 1 ×Time i +β 2 ×Group j +γ k ; This formula is the theoretical prediction value that only considers the main effect of time, the main effect of group, and the random effect of individual. It can be understood as the conventional trend of element concentration changing with time when the difference in efficacy groups is not considered. The interaction trend bias Δpred is defined as the actual observed concentration (C 实际 = Y ijk ) and the theoretical predictions without interaction effects ( Y ijk，无交互 The difference between Δ and Δ pred =C 实际 - Y ijk，无交互 ; After substituting into the simplified model, we get: Δ pred =C 实际 -( β 0 +β 1 ×Time+β 2 ×Group+γ k ).

[0055] Δ pred The magnitude and sign of the value directly reflect the degree of influence of the time-group interaction on the element concentration, and its physical meaning can be intuitively understood by comparing the high / low shrinkage rate groups.

[0056] Assuming the parameters obtained through LMM model fitting are: β0=12 (baseline concentration), β1=0.8 (concentration increases by an average of 0.8 for every 1 unit of time), β2=1.5 (baseline concentration in the SR-High group is 1.5 higher than that in the SR-Low group), γ k =0.3 (a patient) k Individual random effects); In a certain SR-High group of patients (Group=1), the actual concentration Cactual at 3 months post-surgery (Time=3) was 18. The theoretical predicted value without interaction effect was: β 0 +β 1 ×Time+β 2 ×Group+γk =12+0.8×3+1.5×1+0.3=12+2.4+1.5+0.3=16.2; Interactive trend deviation Δ pred =18-16.2=1.8 (positive value).

[0057] Δ pred A positive value indicates that the actual concentration is higher than the theoretical predicted value without interaction effect. This means that the time × group interaction effect of the patient's group (such as SR-High) caused the concentration to increase additionally. This corresponds to the phenomenon that plasma Se in the SR-High group increases faster over time. That is, the interaction effect promotes the increase in concentration, which may be related to the repair of oxidative stress after HIFU treatment (Se participates in antioxidation).

[0058] Δ pred A negative value indicates that the actual concentration is lower than the theoretically predicted value without interaction effect, which means that the interaction effect causes the concentration to decrease further. For example, in the SR-Low group, plasma Se Δpred = -0.5 3 months after surgery, reflecting its weak oxidative stress repair capacity and the concentration increase did not reach the usual trend.

[0059] Δ pred ≈0: The actual concentration is close to the theoretical prediction, indicating that the element has no significant interaction effect, such as Ca and Cd, which were not screened as elements with interaction effects.

[0060] The specific process of step S22 is as follows: S221: Grouping: Based on efficacy criteria (postoperative 3MSR≥30%, 6M≥50%, 1Y≥70%), all patients were grouped as follows: SR-High group (high shrinkage rate / high efficacy): meets any of the above stage criteria; SR-Low group (low shrinkage rate / low efficacy): did not meet all of the above stage criteria.

[0061] S222: Based on the SR-High and SR-Low groups, verify the normality and homogeneity of variance of the two groups of data: S223: Based on the results of normality and homogeneity of variance tests, the difference elements between the two groups are screened by calculating the inter-group difference statistic and the corresponding p-value: S224: Calculate the following features for each historical patient based on the selected differential elements: Intergroup relative deviation: δ=(C 患者 -μ SR-Low ) / σ SR-Low (μ) SR-Low (where σ is the mean of this element in the low-efficacy group and σ is the standard deviation). Concentration ratio: R=C 患者 / μ SR-High μ SR-High This represents the mean value of the high-efficacy group; Group discrimination probability: Based on this difference element, a logistic regression model is constructed to calculate the probability P(SR-High) that the patient belongs to the SR-High group.

[0062] The specific process of performing the normality test and homogeneity of variance test in step S222 is as follows: S2221: Normality test: ; x ( i ): The concentration value of the i-th element after sorting from smallest to largest (standardized data, i.e., Cstand); ai Shapiro-Wilk test coefficient, when n≤ At 50, by querying statistical tables, the sample size was found to be... n Related, n =30 hours a 1≈0.237、 a 30≈0.041; when the sample size of a single group n When the quantile is >50, the formula based on the normal distribution quantile is used for calculation, as follows: ; z i The first of the standard normal distributions i The expected value of the order statistic, i.e., when the sample is drawn from the standard normal distribution N(0,1), the expected value of the order statistic after sorting. i The theoretical expected value of a sample z k The first of the standard normal distributions k Expected values ​​of a series of ordinal statistics; The mean concentration of this group of elements; W: Test statistic (range 0~1, the closer to 1, the better the normality of the data). If the probability corresponding to the value of W p W If the value is greater than 0.05, the data follows a normal distribution; otherwise, it is not normally distributed. S2222: Homogeneity of variance test: ; in, k =2, only the SR-High group and the SR-Low group; N = n1 + n2, the total sample size of the two groups, n1 = the number of cases in the SR-High group, n2 = the number of cases in the SR-Low group; zij The absolute deviation of the i-th sample in the j-th group, i.e. , xij Let i be the concentration of the i-th element in the j-th group. The mean of the j-th group; : The mean of the absolute deviations of the j-th group; The overall mean of all absolute deviations; F: The test statistic, which follows an F-distribution with k−1 and N−k degrees of freedom; If the probability corresponding to the F value p F If the variance is greater than 0.05, the two groups have homogeneous variances; otherwise, they have heterogeneous variances.

[0063] The specific process of step S223 is as follows: If the distribution is normal and the variance is homogeneous: perform an independent samples t-test. ; Among them, the combined variance The calculation formula is as follows: ; , The mean elemental concentrations for the SR-High and SR-Low groups are respectively. , The variances of the two groups are respectively. To combine variances; t is the test statistic, which follows a t-distribution with degrees of freedom df = n1 + n2 − 2; If the calculated t-value corresponds to a two-tailed probability < 0.05, then this element is a between-group difference element; If the distribution is non-normal and the variance is unequal: First, calculate the sum of the ranks of the two groups: merge and sort the two groups of data, assign ranks according to their positions, and then sum them separately: R1: Rank sum of the SR-High group, n1 is the number of cases in the group; R2: The rank sum of the SR-Low groups, where n2 is the number of cases in the group; Next, calculate the U statistic (take the smaller value to avoid bias): ; ; ; U is the test statistic, reflecting the degree of difference between the rank sums of the two groups. The smaller U is, the greater the difference between the groups. If the probability corresponding to the value of U p U If the value is less than 0.05, then the element is a difference element between groups; For example: SR-High group (n1=30): mean Se concentration The value was 18 μg / L, and the variance was... It is 2.25; SR-Low group (n2=35): mean Se concentration 12 μg / L, variance It is 2.56; Premise test: Shapiro-Wilk test p W =0.12>0.05 (normal), Levene test p F =0.35>0.05 (homogeneous variance); Calculate the combined variance: ; Calculate the t-value: ; Look up the t-distribution table: degrees of freedom df =63, t =15.79 corresponds to the original p value p raw <0.001; The conclusion is that Se was the differentiating factor between the SR-High and SR-Low groups at 3 months post-surgery.

[0064] The deep learning prediction model includes: Input Layer: Input multi-source elemental features + MT data; where multi-source elemental features are extracted from trace element concentration data in step S2, including longitudinal dynamic change trend features of each element, inter-group difference elemental features based on high / low shrinkage rate, and time-group interaction effect elements; MT data is the concentration of metallothionein (MT) in plasma. The core function of the input layer is to convert multi-source elemental features and MT data into a structured format that the model can compute, ensuring seamless integration of data with subsequent feature selection and integrated decision-making processes.

[0065] Feature processing layer: Filters core features through RFE; Using random forest as the base model, a recursive process of multiple training iterations, feature importance ranking, and removal of the least important features is employed to retain the feature subset that optimizes model performance, as detailed below: Step 1: Input the initial feature set into the random forest base model, and output the importance score of each feature after training. The score is based on the reduction of the Gini coefficient of the decision tree node split. The higher the score, the greater the contribution to the efficacy prediction. Step 2: Delete the 1-2 features with the lowest importance scores. The number of features to be deleted each time is determined by cross-validation to avoid losing effective features due to excessive deletion in a single step. Step 3: Retrain the base model with the remaining features, repeat the scoring-deletion process until the number of features drops to the optimal value with the highest cross-validation AUC, such as filtering from the initial 30-dimensional features to 8-dimensional core features.

[0066] Adaptability to multi-source features: During the screening process, features that are strongly correlated with the efficacy of HIFU are retained due to their high importance scores, while redundant features are deleted, resulting in the final selection of a subset of core features.

[0067] Core layer: consists of 100 decision trees. Each tree is trained based on random feature selection and random sample sampling. The final result is output through a voting mechanism. For example, if 51 trees predict SR-High, then the final prediction is SR-High. Each decision tree is trained based on random feature selection and random sample sampling to ensure diversity among trees (reducing the variance of the ensemble model): Random sampling of samples: The Bootstrap sampling method is used to randomly select n1 samples from the training set (repeated sampling is allowed) as training data for a single tree. The samples that are not selected (about 37%) are called "out-of-bag samples" and are used for subsequent performance evaluation. Feature random selection: When each tree splits at each node, n1 features are randomly selected from the core feature subset. 1 / 2 For example, three features are randomly selected from the eight core features. The optimal splitting threshold is found based solely on these features to avoid the problem of strong features dominating all trees.

[0068] A single decision tree is constructed using a recursive binary search method: starting from the root node, the optimal splitting feature and threshold are selected based on the principle of minimizing the Gini coefficient, until the leaf nodes contain samples of the same group (SR-High or SR-Low) or the preset tree depth is reached; during prediction, the feature vector of the patient to be predicted is traversed from the root node to the leaf node, and the group of the leaf node is the prediction result of the single tree, SR-High=1 or SR-Low=0.

[0069] After training 100 decision trees, prediction results are output for the same patient to be predicted, and the final prediction group is determined by a majority voting mechanism. If ≥51 trees are predicted as SR-High, then the ensemble result is SR-High; If less than 51 trees are predicted as SR-High, then the ensemble result is SR-Low; This mechanism offsets the prediction bias of a single tree through collaborative decision-making among multiple trees. For example, if a tree misclassifies a patient group due to sampling bias, the correct predictions of other trees can cover the bias and improve the overall prediction accuracy.

[0070] Output layer: Binary classification result (SR-High / SR-Low) + predicted probability; Binary classification results output The system can directly output the group label "SR-High" or "SR-Low", which corresponds to the criteria in the technical disclosure document that "SR ≥30% at 3 months post-surgery, ≥50% at 6 months, and ≥70% at 1 year are considered high efficacy". For example, outputting "SR-High" indicates that the model predicts that the fibroid shrinkage rate after HIFU treatment for this patient can reach the high efficacy standard.

[0071] Predicted probability output: The probability of belonging to the SR-High group is calculated based on the voting results of the core layer. The formula is: P(SR-High) = number of trees predicted as SR-High / total number of trees (100). For example, if 55 trees are predicted as SR-High, then P(SR-High) = 55%. The higher the probability value, the higher the confidence of the prediction result. Clinicians can combine this probability to develop individualized follow-up plans. For example, patients with P(SR-High) = 85% can have their follow-up intervals appropriately extended.

[0072] Example 3 Based on Example 1 or Example 2, the key feature screening of the multi-source element-MT joint feature data to be tested in step S4 is achieved through recursive feature elimination: S41: Using random forest as the base model for recursive feature elimination, specify the parameters for recursive feature elimination: Initial feature set: that is, all features of the multi-source element-MT joint feature set; Target feature number: determined by hierarchical 5-fold cross-validation. First, multiple candidate values ​​are preset. After performing RFE on each candidate value, the cross-validation AUC is evaluated using the base model. The candidate value with the highest AUC and the fewest feature numbers is selected. Number of features deleted per iteration: 1 least important feature is deleted per iteration; S42: Based on the base model, key features are selected from the multi-source element-MT joint feature data to be tested.

[0073] The specific process of step S42 is as follows: S421: Initialize the base model: Set the number of decision trees in the random forest to 100, the tree depth to 8, and the random seed to 42; S422: The importance of the target multi-source element-MT combined feature data and efficacy label (SR-High=1 or SR-Low=0) is scored using the base model, and the feature with the lowest importance score is deleted until the number of features is reduced to the target number of features.

[0074] A deep learning-based multi-source element feature and MT-based HIFU efficacy prediction system is used to implement any of the deep learning-based multi-source element feature and MT-based HIFU efficacy prediction methods described above. It includes a multi-source historical data acquisition and preprocessing module, a multi-source element-MT joint feature construction module, a deep learning model, and a final efficacy evaluation module, all connected in sequence. These modules work together to automate the entire HIFU efficacy prediction process from data processing to evaluation. The multi-source historical data acquisition and preprocessing module is used to acquire multi-time point data of different historical patients on the medical platform and perform preprocessing. The multi-source element-MT joint feature construction module is used to extract features based on preprocessed trace element concentration data and construct a joint feature set by combining it with preprocessed MT concentration data. The deep learning model is used to output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group or the low shrinkage rate group after HIFU treatment. The final efficacy evaluation module is used to determine the high shrinkage rate based on the prediction results and the fibroid shrinkage rate, and to obtain the final evaluation result of the efficacy of HIFU treatment.

[0075] Example 4 In this embodiment, another approach is implemented, including 63 patients with uterine fibroids who underwent HIFU treatment and 41 healthy controls. ICP-MS / MS technology was used to quantitatively analyze elemental concentrations at five time points: before treatment, 1 day post-treatment, 3 months, 6 months post-treatment, and 1 year post-treatment. Logistic regression, mixed-effects models, correlation analysis, and machine learning algorithms were combined to screen key elements associated with disease risk and treatment efficacy.

[0076] Plasma Ti, Tl, Sb, and hair Mo, La were selected as disease diagnostic markers. Acute fluctuations in plasma Se, S, and Zn were observed after HIFU, followed by long-term changes related to oxidative stress recovery. The high shrinkage rate group exhibited significant time-group interactions in plasma Se and Sn, urinary Co, P, S, and hair Zn. The dynamic changes in metallothionein (MT) elevation and malondialdehyde (MDA) decline were synchronized with the fibroid regression process, revealing a synergistic regulatory mechanism of the body's antioxidant adaptation.

[0077] Subject inclusion: This invention employs a prospective cohort design, with subject enrollment completed at Chongqing HIFU Hospital from April 2022 to March 2025. The study population was established after rigorous screening and included two independent groups: a group of patients with uterine fibroids who met the indications for HIFU treatment, and a healthy control group with matched baseline characteristics. Before the start of the study, all subjects were fully informed about the study protocol and signed standardized informed consent forms. This invention strictly sets the subject selection criteria: (1) the experimental group consisted of patients with uterine fibroids who were diagnosed by MRI or ultrasound and were receiving HIFU treatment for the first time; (2) the control group consisted of healthy volunteers whose uterine morphology and function were confirmed to be normal by a systematic physical examination. To minimize the influence of confounding factors, uniform exclusion criteria were set: age ≥50 years, postmenopausal status, pregnancy or lactation, active infectious diseases, history of alcohol or drug dependence, confirmed hemolytic anemia, and any history of malignant tumors. After rigorous screening, the final study sample included 65 patients with uterine fibroids and 40 healthy controls.

[0078] Therapeutic efficacy evaluation system: (1) Imaging data acquisition: A 1.5T superconducting magnetic resonance imaging system was used. The scanning protocol covered the following sequences: T2 weighted imaging used a repetition time of 4692 ms, an echo time of 75 ms, and a slice thickness of 5 mm; T1 weighted imaging (including plain scan and enhancement) was set with a repetition time of 196 ms (plain scan) / 4 ms (enhanced scan), an echo time of 8.18 ms (plain scan) / 2 ms (enhanced scan), and a slice thickness of 6 mm (plain scan) / 5 mm (enhanced scan).

[0079] (2) Volume quantification: The baseline fibroid volume and the follow-up volume were automatically calculated using the MicroSea-HIFU three-dimensional imaging system. The fibroid shrinkage rate SR = [1 - (fibroid volume at follow-up / baseline fibroid volume before surgery)] × 100% (3) Efficacy grading: Patients were divided into low shrinkage rate group and high shrinkage rate group at each time point. The grouping criteria were 30%, 50%, and 70% at three months, six months, and one year after treatment, respectively.

[0080] Sample collection, storage, and pretreatment procedures: In the uterine fibroid group, serial sampling was performed at six specific time points (UF, HIFU-1D, HIFU-3M, HIFU-6M, HIFU-1Y), with no hair samples collected at the HIFU-1D time point. Healthy control volunteers received a single baseline sample at the same time.

[0081] (1) The sampling procedure is as follows: Collect 10 mL of peripheral blood using an anticoagulation vacuum tube containing EDTA-heparin sodium; simultaneously collect 10 mL of midstream urine sample into a standard centrifuge tube; aseptically collect 40-50 hairs from 0.5 cm from the hair root, wrap them in aluminum foil and mark the location of the hair root. All samples are transported to the central laboratory via dry ice within 6 hours after collection.

[0082] (2) Sample processing followed a uniform standard: blood samples were centrifuged at 3150 rpm for 10 minutes (room temperature) to separate plasma, and 500 µL were aliquoted into dedicated 1.5 mL cryovials for storage; urine samples were centrifuged at 10000 rpm for 10 minutes at 4°C, and 1 mL of supernatant was aliquoted. All aliquoted samples were stored at -80°C, while hair samples were stored at -20°C.

[0083] (3) The pretreatment stage follows a standardized procedure: After thawing the cryopreserved plasma and urine samples in an ice bath, accurately measure 50 μL of the sample and 300 μL of high-purity nitric acid (65% concentration) and transfer them together into a 15 mL polytetrafluoroethylene digestion container. Process the samples for 30 minutes using a professional microwave digestion device at a working frequency of 2450 MHz and rated power. After the digestion process is completed, wait for the sample to cool to room temperature, add 5.65 mL of ultrapure water to make up to a total volume of 6 mL, and obtain a test solution with a final nitric acid concentration of approximately 5%. Hair samples are treated with a special pretreatment scheme: a 3 cm section of hair is cut off from 0.5 cm from the hair root, immersed in acetone solution (the liquid surface completely covers the sample, soaking for 10 minutes), rinsed thoroughly with deionized water, and finally purified again with acetone and dried overnight at room temperature. Accurately weigh 10.0 ± 0.5 mg of the pretreated hair sample, add 500 μL of concentrated nitric acid, and complete the processing under the same microwave digestion parameters. Finally, 9.5 mL of ultrapure water was used to bring the volume to 10 mL to obtain a test sample with consistent acidity.

[0084] ICP-MS / MS analysis: (1) Preparation of standards: Agilent multi-element standard solution (covering 52 trace elements), ANPEL ICP-MS grade nitric acid digestion system and Millipore 18.2 MΩ•cm ultrapure water system were used. The preparation of standard curves included: a mixed standard working solution with 7 calibration curves (0.1-2500 μg / L) established by 5% nitric acid gradient dilution; special calibration solutions for macroelements Na / K (0.003-30 μg / mL) and P / Fe / Ca (0.001-10 μg / mL); and Bi / In / Rh mixed internal standard solution (5 μg / mL).

[0085] (2) Elemental determination: The ICP-MS / MS technology platform was used, and its multi-mode detection system integrates a quaternary collision reaction cell of helium / high-purity helium / ammonia-helium mixture / oxygen. The platform is equipped with an SPS4 automatic sample introduction system and combined with a T-type mixer to achieve online internal standard synchronous calibration.

[0086] Machine learning development and validation for uterine fibroids: To address the differences in characteristics of different machine learning models when handling classification tasks, this invention employs seven algorithms to construct a binary classification model suitable for multi-source samples: Artificial Neural Network (ANN), Decision Tree (DT), K Nearest Neighbor (KNN), Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM).

[0087] The raw trace element data, after log2 transformation and Z-score standardization preprocessing, was randomly divided into training and test sets. Key features were selected using recursive feature elimination (RFE), and a supervised learning model was built using R language toolkits such as Caret and NeuralNet. Hierarchical 5-fold cross-validation was used to optimize hyperparameters, and internal validation was performed using the test set. Variable contribution was evaluated by ranking features by importance. Model performance was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC); a higher AUC value indicates stronger model discrimination ability.

[0088] To validate the optimal model, external validation was performed using an independent cohort of samples. This external validation was achieved by applying key features identified in the initial model to this independent dataset. Performance metrics included true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), defined as follows: Accuracy = (TP + TN) / (TP + FP + TN + FN); Sensitivity = TP / (TP + FN); Specificity = TN / (TN + FP); Recall = TP / (TP + FN); F1 score = 2 × (Precision × Recall) / (Precision + Recall); Positive predictive value (precision) = TP / (TP + FP); Negative predictive value = TN / (TN + FN).

[0089] ELISA quantification of MT and CuZn-SOD levels: The concentration of malondialdehyde (MDA) in plasma samples was determined using a competitive ELISA method. Microplates were coated with anti-MDA antibody, and biotinylated MDA was used as a competitor. 50 µL of plasma sample and standards were added to the wells, followed by 50 µL of horseradish peroxidase-labeled streptavidin. After incubation and washing, 50 µL of substrate solution was added, and the reaction was terminated by adding 50 µL of acidic solution. The optical density at 450 nm was measured, and the MDA concentration was calculated using a standard curve.

[0090] The concentrations of MT and CuZn-SOD in plasma samples were quantified using a sandwich ELISA method. Microplates were pre-coated with anti-human MT or anti-human CuZn-SOD antibodies. 50 µL of plasma sample and the corresponding standards were added to the wells, followed by 100 µL of HRP-labeled anti-human MT or anti-human CuZn-SOD antibody. After incubation and washing, 100 µL of substrate solution was added, and the reaction was terminated by adding 50 µL of acidic solution. The optical density at 450 nm was measured, and the concentrations of MT and CuZn-SOD were calculated using a standard curve.

[0091] Experimental results: Using machine learning to differentiate patients with high / low tumor shrinkage rates one day after HIFU treatment: To predict the impact of trace elements in samples collected one day post-surgery on the shrinkage rate of uterine fibroids, we selected 19 trace elements from plasma samples, 27 from urine samples, and 5 from hair samples, and evaluated and ranked the performance contribution of each machine learning algorithm for these specific trace elements. Figure 1 & Figure 2 & Figure 3 It should be noted that hair samples cannot reflect elemental changes over several days, therefore we only collected a hair sample once before the HIFU treatment.

[0092] For plasma samples, the random forest model showed a high AUC value (0.96). Figure 1 A&B). For urine samples, the artificial neural network model had the highest AUC value (0.94). Figure 2 A&B). Furthermore, for hair samples, the random forest model again yielded the best AUC value (0.81). Figure 3 (A&B). To further validate the performance of the best AUC machine learning model with the highest-ranking elements, we ran the model again using plasma, urine, and hair samples independently collected from 10 patients one day post-surgery. In independent validation with plasma samples, the random forest model (AUC=0.91) achieved the following results: precision=1.000, recall=0.778, F1 score=0.789, accuracy=0.800. Figure 1 E&F). When using urine samples, the artificial neural network model (AUC=0.91) achieved the following results: precision = 0.714, recall = 0.833, F1 score = 0.761, accuracy = 0.700. Figure 2 E&F). In contrast, for hair samples, the validation results of the random forest model (AUC=0.91) were: precision = 0.429, recall = 1.000, F1 score = 0.750, accuracy = 0.600 (E&F). Figure 3 (E&F). Therefore, for samples taken one day after HIFU treatment, the artificial neural network model applied to urine samples and the random forest model applied to hair and plasma samples performed best in distinguishing between subjects with high / low tumor shrinkage rates. The main advantage of artificial neural networks lies in their ability to automatically extract complex features, thereby enhancing the flexibility and accuracy of data analysis.

[0093] Longitudinal changes in hair, plasma, and urine elemental composition between high / low shrinkage rate uterine fibroid groups over one year The study examined longitudinal changes in elements in hair, plasma, and urine samples from patients with uterine fibroids of high and low shrinkage rates over one year. Significant differences in the concentrations of 6, 9, and 15 elements were observed over time in the hair, plasma, and urine samples, respectively.

[0094] In hair samples ( Figure 7 Al, Gd, and Pd showed a consistent pattern of first increasing and then decreasing, with Gd in the low shrinkage rate group being significantly higher than that in the high shrinkage rate group at HIFU-1Y. Figure 4 C). Sn exhibits a pattern of first rising, then falling, and finally rising again. Notably, Zn in the high shrinkage rate group shows a unique pattern of first falling, then rising, and then falling again. In particular, Zn shows a significant time- and group-related interaction effect.

[0095] In plasma samples ( Figure 8 The concentrations of K, P, S, and Zn remained relatively high in the high shrinkage rate group, exhibiting similar temporal trends to the low shrinkage rate group. Among these elements, the P concentration in the high shrinkage rate group was significantly higher than that in the low shrinkage rate group at HIFU-6M. Figure 5 B). Similarly, the concentrations of S and Zn in the high shrinkage group were also significantly higher than those in the low shrinkage group at HIFU-3M and HIFU-6M. Figure 5 A&B). From preoperative assessment to HIFU-3M, the trends for Li and Se were consistent between the two groups ( Figure 8 At HIFU-6M, the concentrations of both Li and Se were significantly higher in the low shrinkage group. Figure 5B). During the recovery period (3-12 months after HIFU), Mg, Sn, and V showed similar trends, with plasma concentrations in the high shrinkage rate group ultimately being higher at 12 months. Figure 8 It is noteworthy that Se and Sn in uterine fibroid patients of different shrinkage rates showed significant time- and group-related interaction effects.

[0096] In urine samples ( Figure 9 The concentrations of B and K remained consistently high in the high shrinkage rate group, while Y and Sm remained relatively low, with a consistent trend across the shrinkage rate groups. Ce, La, and Se decreased from their preoperative peak levels to a secondary increase postoperatively, subsequently converging with the trends of the low shrinkage rate group during the recovery period. The initial concentrations of Co, K, Mo, Na, P, and S were relatively low in the high shrinkage rate group but showed a fluctuating upward trend, eventually exceeding those of the low shrinkage rate group. Among these elements, the concentrations of K, Na, P, and S in the high shrinkage rate group were significantly higher than those in the low shrinkage rate group at HIFU-6M. Figure 6 A). Similarly, the Co concentration in the high shrinkage group was also significantly higher with HIFU-1Y ( Figure 6 (B) Of particular note is the profound time- and group-related interaction effect observed in the patterns of Co, P, and S. In contrast, elements including Ca, Cd, and Ti exhibit irregular concentration variations without a consistent trend.

[0097] Correlation between elemental concentrations and clinical symptoms at HIFU-3M, HIFU-6M, and HIFU-1Y The association between elemental concentrations in plasma, urine, and hair and uterine fibroid-related symptoms and tumor shrinkage rate was assessed at 3, 6, and 12 months post-HIFU.

[0098] HIFU-3M ( Figure 10 Plasma Mn and V were positively correlated with fatigue and lower abdominal discomfort, while Co, Mn, and V were correlated with daytime urinary frequency. Fatigue and nocturia were negatively correlated with Sr. In urine ( Figure 11 Li and Na were positively correlated with prolonged menstruation and lower abdominal discomfort, while Nd was negatively correlated with menstrual blood clots. In hair ( Figure 12 Menorrhagia was positively correlated with P, urinary frequency was positively correlated with Zn, while clots were negatively correlated with Mn and V.

[0099] HIFU-6M ( Figure 13 Plasma levels of vitamin V were associated with lower abdominal discomfort and menorrhagia. Tumor shrinkage rate was positively correlated with Ca, Co, Cu, Mg, Na, P, and S. Menstrual blood clots were associated with Mo and Si, while prolonged menstruation was negatively correlated with Fe, and irregular menstrual cycles were negatively correlated with Sn. In urine ( Figure 14Menorrhagia is positively correlated with Cd, while clot formation is correlated with Ce, Cu, Se, Sr, Ti, U, V, Y, and Zn. Fatigue is correlated with B and K. In hair ( Figure 15 Cd and P were associated with menorrhagia, S with both clots and menorrhagia, and Li with abdominal discomfort. Tumor shrinkage rate was negatively correlated with Cr.

[0100] During HIFU-1Y ( Figure 16 Plasma Ba levels were positively correlated with menorrhagia and clot formation. Irregular menstruation was associated with Co, Fe, Li, V, and Y, while tumor shrinkage rate was negatively correlated with Ag. Urinary frequency was negatively correlated with Sr and Zn. In urine ( Figure 17 Mg was positively correlated with urinary frequency, while Ca, K, and Mg were positively correlated with fatigue; Mo was negatively correlated with clots. Tumor shrinkage rate was positively correlated with Co and negatively correlated with Y. In hair ( Figure 18 Clots were associated with Mo, nocturia with La, Li, and Tl, and urinary frequency and fatigue were negatively correlated with Zn. Abdominal discomfort was negatively correlated with Al, Cd, Sn, and Pb, and tumor shrinkage rate was negatively correlated with La.

[0101] Longitudinal analysis of plasma oxidative stress markers and comparison between high / low reduction rate groups: Oxidative stress markers, including MDA, MT, and CuZn-SOD, were analyzed in plasma samples collected at six time points (UF, HIFU-1D, HIFU-3M, HIFU-6M, and HIFU-1Y).

[0102] Compared with the control group, CuZn-SOD levels gradually increased from UF to HIFU-6M, and then decreased slightly at HIFU-1Y, but were still significantly higher than baseline levels. Figure 19 A). MDA levels showed a significant decreasing trend from UF to HIFU-6M, reaching their lowest point at HIFU-6M, with a slight rebound at HIFU-1Y, but still significantly lower than the control group levels. Figure 19 B). Conversely, MT levels rise steadily from UF, peaking at HIFU-6M, and then slightly decline at HIFU-1Y. Figure 19 C).

[0103] Furthermore, oxidative stress markers were compared between the high and low shrinkage rate groups at these three time points. At HIFU-3M, there was a significant difference in MT levels between the high and low shrinkage rate groups ( Figure 19 F), while no significant differences were found in MDA or CuZn-SOD levels at any time point (F). Figure 19These findings suggest that MT levels may serve as a potential biomarker for distinguishing between high and low shrinkage rates after HIFU treatment, while MDA and CuZn-SOD levels did not show significant differences between the two groups.

[0104] This invention successfully established an early prediction model for the efficacy of HIFU based on trace elements, demonstrating that machine learning algorithms can effectively identify characteristic elemental spectra in different biological samples to predict the fibroid shrinkage rate in different patients after HIFU surgery. Model validation showed that plasma (RF model), urine (ANN model), and hair (RF model) can all serve as effective media for predicting tumor shrinkage rates.

[0105] Further analysis using elemental dynamics revealed that HIFU treatment induced acute and long-term elemental reprogramming: an immediate increase in plasma selenium (Se) was accompanied by a decrease in zinc (Zn) and sulfur (S), reflecting oxidative stress and immune response; while sustained changes in selenium, zinc, sulfur, and cobalt (Co) paralleled tissue repair and fibroid regression processes lasting up to twelve months. Characteristic elemental analysis showed that plasma selenium, hair chromium (Cr) and lanthanum (La), and urinary cobalt were associated with tumor shrinkage and symptom improvement, identifying potential non-invasive biomarkers for efficacy assessment. These findings support a unified theoretical framework: HIFU-induced thermal ablation triggers oxidative stress and immune activation, driving elemental redistribution to restore redox balance, promote metabolic adaptation, and ultimately achieve fibroid resorption. Monitoring these multi-matrix elemental characteristics provides a new approach for personalized prognostic assessment and optimization of HIFU treatment.

Claims

1. A method for predicting the efficacy of HIFU based on multi-source element features and MT using deep learning, characterized in that, Includes the following steps: S1: Obtain patient data from different historical patients at at least five time points before and after HIFU treatment on the medical platform, including trace element concentration data and MT concentration data, and perform preprocessing. S2: Based on trace element concentration data, extract trace element features including the longitudinal dynamic change trend of each element, the difference between high / low shrinkage rate groups, and the time-group interaction effect elements; combine with MT concentration data to construct a multi-source element-MT joint feature set; S3: Divide the preprocessed multi-source element-MT joint feature set into a randomized training set and a test set, and input them into the constructed deep learning prediction model for training and testing, respectively. S4: Obtain the multi-source element-MT combined feature data of the patient to be predicted, perform key feature screening on the multi-source element-MT combined feature data, input the screened key features into the trained deep learning prediction model, and output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group or the low shrinkage rate group after HIFU treatment. S5: Based on the prediction results and the fibroid shrinkage rate, the final evaluation result of the HIFU treatment efficacy is obtained by judging the high shrinkage rate.

2. The HIFU efficacy prediction method based on deep learning multi-source element features and MT according to claim 1, characterized in that, The specific process of extracting trace element features based on trace element concentration data in step S2, including longitudinal dynamic change trend features of each element, inter-group difference features of elements with high / low shrinkage rates, and time-group interaction effect elements, is as follows: S21: Extract the longitudinal dynamic trend characteristics of each element: Quantify the concentration change of each element using the same patient + the same element + different time points as the smallest unit: obtain the total change rate, stage change rate, change slope, fluctuation coefficient and recovery coefficient of each element; S22: Extract the characteristic features of differences between high / low shrinkage rate groups: screen for elements with statistically significant differences in concentration among different efficacy groups at the same time point, and quantify the differences; S23: Extraction Time-Group Interaction Effect Element Characteristics: Identify elements whose concentration changes are simultaneously affected by both time and group.

3. The HIFU efficacy prediction method based on deep learning multi-source element features and MT according to claim 2, characterized in that, The specific process of step S22 is as follows: S221: Group Identification: Based on the efficacy criteria, all patients were identified as: SR-High group: Meets the standards for any stage; SR-Low group: Does not meet all stage criteria; S222: Based on the SR-High and SR-Low groups, verify the normality and homogeneity of variance of the two groups of data; S223: Based on the results of normality and homogeneity of variance testing, the difference elements between the two groups are screened by calculating the inter-group difference statistics and the corresponding p-values; S224: Calculate for each historical patient based on the selected differential elements: relative bias between groups, concentration fold ratio, and group discrimination probability.

4. The HIFU efficacy prediction method based on deep learning multi-source element features and MT according to claim 3, characterized in that, The specific process of performing the normality test and homogeneity of variance test in step S222 is as follows: S2221: Normality test: ； x ( i ): The number after sorting from smallest to largest i Concentration values ​​of each element; ai Shapiro-Wilk test coefficient; The mean concentration of this group of elements; W: Test statistic; the probability corresponding to the W value. p W If the value is greater than 0.05, the data follows a normal distribution; otherwise, it is not normally distributed. S2222: Homogeneity of variance test: ； in, k =2; N=n1+n2, the total sample size of the two groups; z ij The absolute deviation of the i-th sample in the j-th group, i.e. , xij Let i be the concentration of the i-th element in the j-th group. The mean of the j-th group; : The mean of the absolute deviations of the j-th group; The overall mean of all absolute deviations; F: The test statistic, which follows an F-distribution with k−1 and N−k degrees of freedom; the probability corresponding to the F-value. p F If the variance is greater than 0.05, the two groups have homogeneous variances; otherwise, they have heterogeneous variances.

5. The HIFU efficacy prediction method based on deep learning multi-source element features and MT according to claim 3, characterized in that, The specific process of step S223 is as follows: If the distribution is normal and the variance is homogeneous: perform an independent samples t-test. ； Among them, the combined variance The calculation formula is as follows: ; , The mean elemental concentrations for the SR-High and SR-Low groups are respectively. , The variances of the two groups are respectively. To combine variances; t is the test statistic, which follows a t-distribution with degrees of freedom df = n1 + n2 − 2; If the calculated t-value corresponds to a two-tailed probability < 0.05, then this element is a between-group difference element. If the distribution is non-normal and the variance is unequal: First, calculate the sum of the ranks of the two groups: merge and sort the two groups of data, assign ranks according to their positions, and then sum them separately: R1: Rank sum of the SR-High group, n1 is the number of cases in the group; R2: The rank sum of the SR-Low groups, where n2 is the number of cases in the group; Calculate the U statistic again: ； ； ； U is the test statistic; If the probability corresponding to the value of U p U If the value is less than 0.05, then the element is a difference element between groups.

6. The method for predicting HIFU efficacy based on deep learning-based multi-source element features and MT according to claim 1, characterized in that, The deep learning prediction model includes: Input layer: Input multi-source element features + MT data; Feature processing layer: Filters core features through RFE; Core layer: Consists of 100 decision trees, each trained based on random feature selection and random sample sampling, and finally outputs the result through a voting mechanism; Output layer: Binary classification result + predicted probability; Training optimization: Performance is evaluated using out-of-bag error, and hyperparameters including tree depth and node splitting threshold are optimized.

7. The HIFU efficacy prediction method based on deep learning multi-source element features and MT according to claim 1, characterized in that, In step S4, key feature filtering of the multi-source element-MT joint feature data to be tested is achieved through recursive feature elimination: S41: Using random forest as the base model for recursive feature elimination, specify the parameters for recursive feature elimination: Initial feature set: that is, all features of the multi-source element-MT joint feature set; Target feature number: determined by hierarchical 5-fold cross-validation. First, multiple candidate values ​​are preset. After performing RFE on each candidate value, the cross-validation AUC is evaluated using the base model. The candidate value with the highest AUC and the fewest feature numbers is selected. Number of features deleted per iteration: 1 least important feature is deleted per iteration; S42: Based on the base model, key features are selected from the multi-source element-MT joint feature data to be tested.

8. The method for predicting the efficacy of HIFU based on multi-source element features and MT according to claim 7, characterized in that, The specific process of step S42 is as follows: S421: Initialize the base model: Set the number of decision trees in the random forest to 100, the tree depth to 8, and the random seed to 42; S422: The importance of the multi-source element-MT combined feature data and efficacy labels to be tested is scored using the base model, and the features with the lowest importance scores are deleted until the number of features is reduced to the target number of features.

9. A deep learning-based multi-source element feature and MT-based HIFU efficacy prediction system, used to implement the deep learning-based multi-source element feature and MT-based HIFU efficacy prediction method according to any one of claims 1-8, characterized in that, It includes a multi-source historical data acquisition and preprocessing module with sequentially connected signals, a multi-source element-MT joint feature construction module, a deep learning model and final efficacy evaluation module. The modules work together to automate the entire process of HIFU efficacy from data processing to evaluation. The multi-source historical data acquisition and preprocessing module is used to acquire multi-time point data of different historical patients on the medical platform and perform preprocessing. The multi-source element-MT joint feature construction module is used to extract features based on preprocessed trace element concentration data and construct a joint feature set by combining it with preprocessed MT concentration data. The deep learning model is used to output the prediction result of whether the patient to be predicted belongs to the high shrinkage rate group or the low shrinkage rate group after HIFU treatment. The final efficacy evaluation module is used to determine the high shrinkage rate based on the prediction results and the fibroid shrinkage rate, and to obtain the final evaluation result of the efficacy of HIFU treatment.

Images