Active learning-driven clinical prospective sample collection optimization method and system, and storage medium

Abstract

The invention provides an active learning-driven clinical prospective sample collection optimization method and system and a storage medium, which are suitable for multi-center, high-cost and low-positive-rate clinical research scenes, such as laboratory self-research detection (LDT), real-world research (RWS) and group-entering optimization of clinical tests. According to the method, on the basis of omics data of candidate individuals, such as a nuclear magnetic resonance spectrogram and current prediction model output, three-element scores of uncertainty U, sample diversity D and clinical utility C of the candidate individuals are calculated, the model performance, calibration capability and external generalization can be improved while the cost is predicted to be controlled, and particularly, the coverage capability of rare subtypes is enhanced; the method is suitable for prospective group-entering scheduling of LDT, clinical test and real world research.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence auxiliary clinical research and medical data acquisition and scheduling, and particularly relates to a proactive learning driven clinical prospective sample collection optimization method and system and a storage medium. BACKGROUND

[0002] In precision medicine research, prospective sample collection faces challenges such as high cost, long cycle, and low positive rate. Traditional convenience sampling or quota enrollment is prone to cause problems such as sample redundancy, insufficient coverage of rare subtypes, and poor model generalization ability.

[0003] Existing active learning methods are mostly used for image or text annotation, lack the modeling ability for clinical practical constraints (budget, quota, time window, and ethical priority), and usually only optimize a single objective (such as uncertainty), making it difficult to be applied in real medical scenarios.

[0004] Therefore, there is an urgent need for a prospective sampling framework that takes into account model learning efficiency, data quality, clinical value, and operational feasibility to support LDT high-threshold application scenarios. SUMMARY

[0005] Technical solution: In order to solve the above technical problems, the present application specifically proposes a proactive learning driven clinical prospective sample collection optimization method, and the specific steps include: Embedding the omics data of the candidate individual and the initial prediction model to perform prediction distribution; Constructing a three-element comprehensive scoring function score(x) = αU(x) + βD ( x )+ γC ( x ), and combining AUC , ECE and the rare subtype recall target to form a multi-objective optimization problem; set the constraint conditions including total budget, multi-center quota, enrollment time window, minimum coverage ratio of population / subtype, and ethical priority, solve and obtain the optimal enrollment list and scheduling; After collection, the sample is added to the experience buffer pool, and the multi-armed bandit or Bayesian optimization is used to update the selection strategy to form a closed loop iteration, and the sample information gain is updated under the online update and closed loop mechanism; Among them U ( x i ) is an uncertainty scoring function, which is based on at least one of prediction entropy, confidence interval width, or enhanced consistency difference, D( x i) is a diversity score function, which is a minimum distance or kernel density estimation function based on the embedding space of the candidate individual and the sampled set, C( x i ) is a clinical utility score function, which is a function based on cost-benefit function, risk threshold sensitivity or high-risk population priority.

[0006] As an improvement, the multi-objective optimization includes solving by integer linear programming, greedy approximation or Lagrangian relaxation.

[0007] As an improvement, the omics data is completed by serum nuclear magnetic resonance detection NMR spectrum, including 600 MHz acquired on a superconducting nuclear magnetic resonance instrument, using NOESY PR1D pulse sequence, Fourier transform, phase correction, baseline correction and TSP 0.84 ppm normalized, 0.5-12.0 ppm region is intercepted, 12800 data points are reserved, and is recorded as ; wherein the initial prediction model adopts 1 D - CNN and Attention structure.

[0008] As an improvement, the step of calculating the uncertainty score U( x i ) is: first, the prediction entropy and the enhanced consistency difference are fused to calculate (1) is the probability that the sample belongs to class c and is defined as follows: (2) wherein represents the class label, represents the input sample, represents the class; U entropy is the prediction entropy, which represents the uncertainty of the prediction distribution, and is commonly used to measure the dispersion degree of the output distribution of the model; in active learning, the sample with larger prediction entropy indicates that the model is uncertain about its classification result, then, noise is added to SNR and the phase is fine-tuned to generate n enhanced samples , and the prediction distribution variance (3) wherein, is the prediction result of the model for the original sample ; is the prediction result of the augmented sample ( ); denotes the computation of variance, which measures the volatility of the prediction result of the augmented sample.

[0009] Finally, the final uncertainty score is composed of two parts: the prediction entropy and the augmented sample variance, as follows (4) where is the prediction entropy, is the augmented sample variance, and are self-defined weight coefficients that control the contribution of the two parts to the final uncertainty score.

[0010] As an improvement, when calculating the diversity score D ( x i ), an embedding model SimCLR pre-trained using the ϕ framework is used to output , and the diversity score is calculated as the minimum distance normalized value: (5) is a smoothing term; C is the index set of the collected samples; A is the candidate sample set used for normalization; is the Euclidean distance between the sample and the closest sample in the collected set; is the maximum value of the minimum distance of all samples in the candidate set, which is used as a normalization factor to map the score to the [0,1] interval.

[0011] As an improvement, when calculating the clinical utility score C( x i ), set (6) where is the basic clinical utility value, which is used to ensure that all samples have basic clinical value; is the upper bound of clinical utility; is the clinical cost utility term, which is used to quantify the cost-benefit ratio of the sample ; The clinical cost utility term includes sampling cost, risk threshold sensitivity, and population value, wherein the sampling cost is at least one of the cost of nuclear magnetic resonance, the cost of reagents, and the cost of manpower.

[0012] As an improvement, the method for solving the set constraints includes selecting from a candidate library. k Patients were enrolled, and binary variables were defined. , indicating whether to select the first i For each sample, a target function is defined, namely, maximizing the total score: (7) Then, the specific constraint conditions for establishing the constraint mechanism are as follows: (Total budget less than or equal to 250,000) (8) k =1,2,3 (Central quota: 15 / 20 / 15) (9) (Rare subtype ≥ 8 people) (10) =0, if (Time window ≤ 14 days) (11) (High-risk group ≥ 5 people) (12) Based on the constraints described above, solve for the objective function using... Gurobi 9.5 Solve this integer linear programming problem and output the optimal grouping list and scheduling suggestions.

[0013] As a specific embodiment of the present invention, the present invention also provides an active learning-driven clinical prospective sample collection optimization model, the model including a data access module, a preprocessing module, a scoring calculation module, an optimization solution module, and a scheduling synchronization module; wherein the data access module supports multiple formats. NMR Data is uploaded and transmitted to the preprocessing module; the preprocessing module automatically performs phase correction, baseline correction, and other functions on the input data. TSP Normalize the data, and then transmit the processed data to the scoring calculation module; The scoring calculation module performs a ternary comprehensive scoring function calculation and transmits the results to the optimization solution module; the optimization solution module then calls a mathematical optimization solver. Gurobi Alternatively, an open-source solver can be used to solve the problem, and the optimal solution is transmitted to the scheduling synchronization module; the scheduling synchronization module is used to interface with the scheduling system and output the optimal schedule; The ternary comprehensive scoring function is the ternary comprehensive scoring function in any of the above-mentioned active learning-driven clinical prospective sample collection optimization methods. score(x).

[0014] As an improvement, it also includes an audit log module and a visualization interface; the audit log module is used to record each parameter modification, list generation, and manual intervention, and supports export; the visualization interface is used to display the ternary score distribution, constraint satisfaction status, and historical trends.

[0015] As another specific embodiment of the present invention, the present invention also provides a storage medium having a computer program stored thereon, which, when executed, implements the steps of any of the above-described optimization methods.

[0016] Beneficial effects: This invention proposes a proactive learning-driven prospective sample acquisition optimization method, which constructs an "uncertainty ( U ) - Diversity ( D - Clinical efficacy ( C The "three-element scoring system" combines multi-objective optimization and explicit constraint modeling to generate an executable enrollment list and schedule, and supports online strategy updates, forming a closed-loop iteration.

[0017] Furthermore, this invention is expected to improve model performance, calibration capability, and external generalization while controlling costs, especially enhancing coverage of rare subtypes, and is suitable for... LDT Prospective enrollment scheduling for clinical trials and real-world studies. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall optimization method of the present invention.

[0019] Figure 2 This is the ternary scoring calculation process of the present invention.

[0020] Figure 3 This is a schematic diagram of the multi-objective and constraint optimization of the present invention.

[0021] Figure 4 This invention provides an online update mechanism for the closed-loop structure.

[0022] Figure 5 This is a schematic diagram of the performance trend of the present invention, where the green line at the top represents the performance trend of the active learning of the present invention, and the red line at the bottom represents the baseline of random / convenience grouping. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] This invention discloses an active learning-driven method, system, and storage medium for optimizing prospective clinical sample collection, suitable for multi-center, high-cost, and low-positive-rate clinical research scenarios, such as laboratory-developed tests (…). LDT Real-world studies RWS This method optimizes enrollment in clinical trials. It calculates the uncertainty based on candidate individuals' omics data, such as NMR spectra, and the output of current predictive models. U Sample diversity D Clinical efficacy C Tripartite rating.

[0025] Construct a comprehensive scoring function score(x) = αU (x)+ βD (x)+ γC (x), where α + β + γ =1, and combine the model performance objective with the rare subtype recall rate to form a multi-objective optimization problem; under constraints such as budget, central quota, time window, population / subtype coverage, and ethical priority, solve the optimal enrollment list and scheduling; after collection, the labels are fed back to the experience buffer pool, and the selection strategy is dynamically updated using a multi-armed gambling machine or Bayesian optimization to form a closed-loop iteration.

[0026] In this invention, candidate individuals x Based on its omics data embedding and model prediction distribution, a ternary score is calculated as a specific implementation method. U(x) Uncertainty scoring functions, such as prediction entropy and enhanced consistency differences; D(x) Diversity scoring functions, such as minimum distance in the embedding space and kernel density estimation; C(x) Clinical utility scoring functions, such as prioritizing high-risk populations and cost-benefit ratios.

[0027] The final result is a comprehensive score and multi-objective optimization: score(x) = aU(x) + bD(x) + gC(x) The primary goal is to maximize the total score, while also taking into account... AUC , ECE Recall of rare subtypes.

[0028] See Figure 2 The diagram shows a flowchart for calculating the ternary score. Specific embodiments are described below. This implementation method is applicable to non-alcoholic fatty liver disease (NAFLD). NAFLD This is a multicenter prospective study of early subtyping. The candidate cohort includes participants from multiple collaborating medical institutions, all of whom have completed serum nuclear magnetic resonance (¹H-NMR) testing.

[0029] NMRThe spectrum at 600 MHz Data acquired on a superconducting nuclear magnetic resonance spectrometer, using NOESY PR1D Pulse sequence, after Fourier transform, phase correction, baseline correction and TSP (0.84) ppm After normalization, truncate to 0.5–12.0. ppm The region retains 12,800 data points, denoted as... .

[0030] Initial prediction model Using 1 D - CNN + Attention The structure directly processes the original spectrum, specifically including: (1) Input layer: 12800 × 1; (2) Convolutional block 1: 32 convolutional kernels, size 64, stride 2, ReLU Activation, followed by max pooling, 2x downsampling; (3) Convolutional block 2: 64 convolutional kernels, size 32, stride 2, ReLU Activation, pooling 2x; (4) Convolutional block 3: 128 convolutional kernels, size 16, stride 2, ReLU Activation, pooling 2x; (5) Flattening layer: output dimension 128 × 16 = 2048; (6) Attention layer: 4-head scaled dot product attention, hidden dimension 128, output 128-dimensional vector; (7) Fully connected layer: 128 → 64 → 2, Softmax Output the class probability.

[0031] The model is used in publicly available metabolomics datasets (such as...) Phen-X Pre-trained on a subset, then fine-tuned for 5 rounds after transferring to this task, initially... AUC It is approximately 0.71, indicating potential for further optimization.

[0032] In this embodiment, for each candidate Perform the following ternary score calculation: Uncertainty score Diversity score Clinical efficacy score The final comprehensive score is obtained. Specifically, it is calculated by fusing prediction entropy and enhanced consistency difference: (1) At the same time, Add slight noise, for example SNR =40 dB And fine-tune the phase, for example, to ±0.1. rad In this embodiment, five enhanced samples are selected for generation. Calculate the variance of its predicted distribution: (2) (3) Furthermore, using SimCLR Frame pre-trained embedding model Output Let the set of collected samples be... C Its embedding is ,in To smooth out the terms and prevent division by zero, the diversity score is the minimum distance normalized value: (4) Furthermore, taking into account the history of diabetes: +0.3; BMI ≥ 30: +0.2; ALT >60 U / L +0.2; Cost deduction: ;final: (5) Finally, initial weights are set as α=0.4, β=0.3, and γ=0.3. The uncertainty score is the information entropy of the probability distribution output by the model, calculated using the following formula: Its maximum value is ( C (To ensure consistency of measurement, a normalization factor is introduced to normalize the uncertainty score to the interval [0,1], as follows:) (6)

[0033] Therefore, the overall score is: ;in α + β + γ =1, and α , β , γ ≥0. The initial weight is set to... α =0.4, b = 0.3, g = 0.3, It can be dynamically adjusted based on subsequent feedback.

[0034] See Figure 3 The diagram shown is a flowchart of the multi-objective and constraint optimization process in this invention. First, select the next batch of 50 patients (k=50) from the candidate pool and define binary variables. , indicating whether to select the first i We have 100 samples. Then, we define the objective function: maximizing the total score. (7) Then, a constraint mechanism is established, and the constraint conditions are set as follows in this embodiment: (Total budget less than or equal to 250,000) (8) k =1,2,3 (Central quota: 15 / 20 / 15) (9) (Rare subtype ≥ 8 people) (10) =0, if (Time window ≤ 14 days) (11) (High-risk group ≥ 5 people) (12) In this invention, the objective function is solved according to the above constraints using... Gurobi 9.5 Solve this integer linear programming problem with an average solution time of less than 10 seconds, and output the optimal group list and scheduling suggestions.

[0035] See Figure 4 The diagram shown is a logic diagram of the online update and closed-loop structure of this invention. After each round of data collection, the following updates are performed: (1) Model update: new samples are added to the training set, and the model is updated accordingly. Fine-tuning 3 rounds ( Adam, lr =5e-5); (2) Embedded model update: for ϕ Comparative learning and fine-tuning ( batch =32); (3) Strategy update: First define the reward function. (13) Then use Bayesian optimization (Gaussian process + EI The acquisition function searches for the optimal weights α, β, γ ∈ [0.1, 0.8], iterates 5 times, completes the update, and finally sorts by center. BMI , ALT The group entry rounds are stored hierarchically, supporting coverage statistics and updating the experience buffer pool.

[0036] As another specific embodiment of the present invention, the present invention provides an active learning-driven clinical prospective sample collection optimization model, the model including a data access module, a preprocessing module, a scoring calculation module, an optimization solution module, and a scheduling synchronization module; wherein the data access module supports multiple formats. NMR Data is uploaded and transmitted to the preprocessing module; the preprocessing module automatically performs phase correction, baseline correction, and other functions on the input data. TSP Normalize the data, and then transmit the processed data to the scoring calculation module; The scoring calculation module performs a ternary comprehensive scoring function calculation and transmits the results to the optimization solution module; the optimization solution module then calls a mathematical optimization solver. Gurobi The solution is obtained by either an open-source solver or a third-party solution, and the optimal solution is transmitted to the scheduling synchronization module. The scheduling synchronization module interfaces with the scheduling system and outputs the optimal schedule. The mathematical optimization solver... Gurobi By supporting integer linear programming ( Integer Linear Programming, ILP ), mixed integer programming Mixed Integer Programming , MIP This method provides efficient solutions to optimization problems such as resource scheduling, combinatorial optimization, and constraint satisfaction. The ternary comprehensive scoring function is the same as the ternary comprehensive scoring function used in the aforementioned active learning-driven clinical prospective sample collection optimization method. score(x) .

[0037] Furthermore, it also includes an audit log module and a visualization interface; the audit log module is used to record each parameter modification, list generation, and manual intervention, and supports export; the visualization interface is used to display the ternary score distribution, constraint satisfaction status, and historical trends.

[0038] In this invention, the data access module can support formats including fid 、. txt 、. csv Furthermore, the system meets the ethical review requirements for transparency, operability, and data security. Example 1

[0039] This embodiment tests the performance in a simulation environment using the optimization method and comparative experimental design of the present invention. Specifically, it includes the following multi-round simulation comparative experiments. Experimental settings: 50 patients are selected in each round, and a total of 6 rounds are run; the evaluation indicators include: (1) model AUC and ECE (2) The cumulative number of rare subtype recalls; (3) The total cost and the central quota satisfaction rate. The comparative experimental method adopts random selection, uncertainty priority, and U+D weighting, while the present invention adopts U+D+C+constraint optimization; the specific test results are shown in Table 1 below.

[0040] Table 1 Test Results

[0041] Note: The above values ​​are reasonable predictions and preliminary simulation results based on the evolution of existing active learning systems. Actual performance may vary depending on data quality, model structure, and parameter configuration.

[0042] In this invention, as can be seen from Table 1: [the following parameters are considered:] random sampling, uncertainty priority, and... U +D Compared to weighted methods, the method proposed in this invention... U + D + C The ternary scoring combined with the constraint optimization strategy significantly improves the model's discriminative ability while maintaining low cost and satisfying the multi-center quota. AUC Improved to 0.83–0.84), calibration performance ( ECE The cost was reduced to 0.10–0.12, and more efficient coverage of rare subtypes was achieved (cumulative recall of 20–23 cases, more than 50% higher than other methods), validating its comprehensive advantages in prospective clinical sample collection. It is expected that this invention will significantly improve model performance and rare subtype coverage while controlling costs.

[0043] See Figure 5 As shown, Figure 5 As shown, with increasing cost or sample size, the active learning method proposed in this invention (green curve) significantly outperforms the baseline method (red curve) in model performance and converges faster. This indicates that this invention can achieve superior model performance under the same budget, possessing higher sample efficiency and clinical application value.

[0044] Example 2: Specific Calculation Example To more intuitively demonstrate the specific implementation process of the active learning-driven clinical prospective sample collection optimization method of the present invention, the following will use a specific example to illustrate how to perform uncertainty scoring. Diversity score and clinical utility score The calculation is performed, and a comprehensive score is finally obtained. .

[0045] Suppose we are conducting a study on non-alcoholic fatty liver disease (NAFLD) NAFLD A multicenter prospective study of early subtyping. The candidate cohort included participants from 5 collaborating medical institutions, each of whom had completed serum MRI (¹H- ). NMR ) detection. NMR The spectrum at 600 MHz Data acquired on a superconducting nuclear magnetic resonance spectrometer, using NOESY PR1D The pulse sequence, after Fourier transform, phase correction, baseline correction, and... TSP After normalization, 12,800 data points were retained as the input feature vector. The following steps involve constructing the initial prediction model and calculating the ternary score.

[0046] Uncertainty score: Assume that for a specific candidate individual, the probability distribution of the initial prediction model output is shown in Table 2.

[0047] Table 2 Probability distribution of the initial prediction model output

[0048] (2) Calculate the individual's uncertainty score. as follows:

[0049] Substituting the above probabilities into the formula, we get: The normalization factor ,therefore, .

[0050] (3) Diversity scoring: using SimCLR Frame pre-trained embedding model ϕ For all sampled samples and current candidate individuals Encode them to obtain their embedding vectors. Assume the collected sample set is... C Its embedding representation is ,in For candidate individuals Its embedding vector is .

[0051] calculate With the sampled set C The diversity score is obtained by taking the minimum Euclidean distance of each sample and normalizing it. : ;

[0052] The calculation result is Furthermore, The value range is (0,1).

[0053] Clinical utility score Taking into account factors such as history of diabetes, BMI, and ALT, the following weights are set as shown in Table 3.

[0054] Table 3 Weight List

[0055] After deducting the corresponding cost factors, assuming the calculated clinical utility score is... .

[0056] (5) Overall score: based on the initial weight settings , When the value is 0.75, the overall score is calculated as follows: =0.4×0.923+0.3×0.75+0.3×0.6=0.7892 This score indicates that, considering uncertainty, diversity, and clinical utility, the sample... This is a high-priority inclusion subject. This scoring mechanism effectively selects the most valuable samples for prospective studies, thereby improving model performance and coverage of rare subtypes.

[0057] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An active learning-driven method for optimizing prospective clinical sample collection, characterized in that, The specific steps include: Predicting distributions based on omics data embeddings of candidate individuals and initial prediction models; Constructing a ternary comprehensive scoring function score ( x i )=α U ( x i )+β D ( x i )+γ C ( x i ), and combined AUC , ECE The problem involves a multi-objective optimization problem, which is formed by recalling rare subtypes. Constraints are set, including total budget, multi-center quota, enrollment time window, minimum coverage ratio of population / subtype, and ethical priority. The optimal enrollment list and schedule are obtained by solving the problem. After collection, the samples are added to the experience buffer pool, and the selection strategy is optimized and updated using a multi-armed gambling machine or Bayesian optimization to form a closed-loop iteration, and online updates and sample information gain updates are performed under the closed-loop mechanism. in U ( x i () is an uncertainty scoring function, which is based on at least one of prediction entropy, confidence interval width, or enhanced consistency difference. D ( x i ) is the diversity scoring function, and is a function that estimates the minimum distance or kernel density between a candidate individual and the sampled set in the embedding space. C ( x i ) is a clinical utility scoring function, which is a function based on cost-benefit function, risk threshold sensitivity, or priority of high-risk groups.

2. The active learning-driven clinical prospective sample collection optimization method according to claim 1, characterized in that, The multi-objective optimization includes solutions obtained through integer linear programming, greedy approximation, or Lagrange relaxation.

3. The active learning-driven clinical prospective sample collection optimization method according to claim 1, characterized in that, Omics data were used to complete serum MRI detection. NMR Spectra, including those acquired on a 600 MHz superconducting nuclear magnetic resonance spectrometer, were obtained using... NOESYPR1D Pulse sequence, after Fourier transform, phase correction, baseline correction and TSP 0.84 ppm After normalization, the range is truncated to 0.5–12.

0. ppm The region retains 12,800 data points, denoted as... The initial prediction model uses 1D- CNN and Attention structure.

4. The active learning-driven clinical prospective sample collection optimization method according to claim 1, characterized in that, Calculate the uncertainty score U( x i The steps are as follows: First, the predicted entropy and the enhanced consistency difference are fused to calculate... （1） for The probability that a sample belongs to class c is defined as follows: （2） in Indicates category label, Indicates the input sample. Indicates category; U entropy Prediction entropy represents the uncertainty of the prediction distribution and is often used to measure the dispersion of the model's output distribution; then, for Add noise SNR And fine-tune the phase to generate n One enhanced sample Calculate the variance of the predicted distribution. （3） in It is the model's application to the original samples The prediction results; It is an augmented sample. ( The prediction results are represented by the variance, which is used to measure the volatility of the prediction results of the enhanced sample. Finally, the final uncertainty assessment It consists of two parts: prediction entropy and enhanced sample variance, as follows: （4） in It is predicting entropy. It is to enhance sample variance. and These are custom weighting coefficients that control the contribution of these two parts to the final uncertainty score.

5. The active learning-driven clinical prospective sample collection optimization method according to claim 1, characterized in that, Calculate diversity score D ( x i When ), it is done by using SimCLR Frame pre-trained embedding model ϕ Output The diversity score is calculated as the minimum distance normalized value: （5） For smoothing terms; C This is the set of indices for the collected samples; A The set of candidate samples used for normalization; For the sample The Euclidean distance to the closest sample in the collected set; The maximum value of the minimum distance among all samples in the candidate set is used as a normalization factor to map the score to the interval [0,1].

6. The active learning-driven clinical prospective sample collection optimization method according to claim 1, characterized in that, Calculate the clinical utility score C( x i When setting （6） in, The basic clinical utility value is used to ensure that all samples have basic clinical value. This is the upper limit of clinical efficacy; This is a clinical cost-utility item used to quantify samples. Cost-benefit ratio; The clinical cost-utility items include sampling cost, risk threshold sensitivity, and population value, wherein the sampling cost is at least one of MRI cost, reagent cost, and labor cost.

7. The active learning-driven clinical prospective sample collection optimization method according to claim 1 or 2, characterized in that, The methods for solving the given constraints include selecting from a candidate library. k Patients were enrolled, and binary variables were defined. , indicating whether to select the first i For each sample, a target function is defined, namely, maximizing the total score: （7） Then, the specific constraint conditions for establishing the constraint mechanism are as follows: (Total budget less than or equal to 250,000) (8) k =1,2,3 (Central quota: 15 / 20 / 15) (9) (Rare subtype ≥ 8 people) (10) =0, (Time window ≤ 14 days) (11) (High-risk group ≥ 5 people) (12) Based on the constraints described above, solve for the objective function using... Gurobi 9.5 Solve this integer linear programming problem and output the optimal group list and scheduling suggestions.

8. An active learning-driven clinical prospective sample collection optimization model, characterized in that, The model includes a data access module, a preprocessing module, a scoring calculation module, an optimization solution module, and a scheduling synchronization module; wherein the data access module supports multiple formats. NMR Data upload, transmitting data to the preprocessing module; The preprocessing module automatically performs phase correction, baseline correction, and TSP normalization on the input data, and then transmits the processed data to the scoring calculation module. The scoring calculation module performs a ternary comprehensive scoring function calculation and transmits the results to the optimization solution module; the optimization solution module then calls a mathematical optimization solver. Gurobi Alternatively, an open-source solver can be used to solve the problem, and the optimal solution is transmitted to the scheduling synchronization module; the scheduling synchronization module is used to interface with the scheduling system and output the optimal schedule; The ternary comprehensive scoring function is the ternary comprehensive scoring function in the active learning-driven clinical prospective sample collection optimization method described in any of claims 1-7. score(x) .

9. The clinical prospective sample collection optimization model according to claim 8, characterized in that, It also includes an audit log module and a visualization interface; the audit log module is used to record each parameter modification, list generation, and manual intervention, and supports export; the visualization interface is used to display the ternary score distribution, constraint satisfaction status, and historical trends.

10. A storage medium having a computer program stored thereon, which, when executed, implements the steps of the optimization method according to any one of claims 1-7.

Images