Tumor chemotherapy effect prediction model construction method and system based on artificial intelligence

Abstract

The invention discloses a tumor chemotherapy effect prediction model construction method and system based on artificial intelligence. The method comprises the following steps: multi-dimensional data acquisition: acquiring tumor biological data, individual clinical heterogeneity data and patient report outcome data of a target cancer type patient; performing feature engineering processing: performing quantitative processing on the individual clinical heterogeneity data to generate quantitative indexes, and performing feature extraction on the tumor biological data to form a tumor microenvironment feature set; basic model training: taking multi-dimensional data of large-sample general cancer species as a training set, adopting a multi-branch feature fusion network to construct a general chemotherapy effect prediction basic model, and taking chemotherapy effect prediction and life quality influence prediction as double optimization targets for training, according to the invention, by collecting tumor biological data, individual clinical heterogeneity data and patient report outcome data, a multi-dimensional data system is constructed, and tumor pathological characteristics and microenvironment states can be comprehensively covered.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, and in particular relates to a method and system for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence. Background Technology

[0002] Currently, predicting the efficacy of chemotherapy in tumors is an important technical direction for decision support in clinical oncology diagnosis and treatment, and existing technologies have formed a certain technical system. At the data level, mainstream methods collect pathological classification data and baseline tumor marker data (such as carcinoembryonic antigen CEA, carbohydrate antigen CA125, etc.) from tumor patients. Some technologies also incorporate pre-treatment tumor imaging data (such as tumor diameter measurements from CT and MRI images) to provide basic data support for prediction. In terms of model construction, traditional machine learning algorithms such as logistic regression and random forests are commonly used. Some technologies also employ basic neural network models, training on historical chemotherapy patient efficacy data (such as objective response rate (ORR) and progression-free survival (PFS) to achieve preliminary predictions of chemotherapy efficacy in new patients. Regarding system application, existing prediction systems mostly have basic data entry modules and result display functions. Some systems can interface with hospital basic information systems to achieve preliminary interaction of a small amount of core data, providing reference for doctors.

[0003] However, existing tumor chemotherapy efficacy prediction technologies still have many limitations, making it difficult to meet the needs of precise and personalized clinical diagnosis and treatment. From a data perspective, existing technologies mostly focus on tumor pathology and baseline biomarkers, failing to cover tumor microenvironment data (such as immune cell infiltration and vascular factor expression), individual clinical heterogeneity data (such as comorbidity course and quantitative information on lifestyle habits), and patient quality of life data, resulting in incomplete data support. In terms of feature processing, individual heterogeneity is mostly judged on a "presence or absence" basis without establishing a quantitative system, and there is a lack of efficient means to extract microenvironment features, resulting in insufficient feature effectiveness. In terms of model functionality, most are single-objective predictions of efficacy, unable to simultaneously predict quality of life, thus disconnecting from the clinical need for "efficacy-risk balance." The small sample size adaptation capability is weak, and for niche cancer types or rare chemotherapy regimens, it is difficult to build high-precision models due to data scarcity. The system's practicality is insufficient, relying heavily on manual data entry, lacking interactive adjustment and dynamic update functions, and difficult to integrate into clinical workflows. These limitations restrict the clinical application value of prediction technologies. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence, which solves the problems of narrow data, coarse features, single target, poor small sample fit, and weak system implementation in the prior art for tumor chemotherapy prediction.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method and system for constructing an artificial intelligence-based model for predicting the efficacy of chemotherapy in tumors, comprising the following steps: S1. Multi-dimensional data acquisition: Collect tumor biology data, individual clinical heterogeneity data, and patient-reported outcome data from patients with the target cancer type; S2. Feature engineering processing: Quantitative processing is performed on the individual clinical heterogeneity data to generate quantitative indicators, and feature extraction is performed on the tumor biological data to form a tumor microenvironment feature set; S3. Basic Model Training: Using a large sample of multi-dimensional data of common cancer types as the training set, a general chemotherapy efficacy prediction basic model is constructed using a multi-branch feature fusion network. The prediction of chemotherapy efficacy and the prediction of the impact on quality of life are trained as dual optimization objectives. S4. Sub-model transfer adaptation: For small sample cancer types or rare chemotherapy regimens, the general chemotherapy effect prediction base model is fine-tuned using a transfer learning algorithm to generate a specific prediction sub-model; S5. Predictive Model Output: Input the multi-dimensional data of the patient to be predicted into the prediction model of the corresponding cancer type or treatment plan, and output the two-dimensional prediction results of efficacy level and quality of life risk level, and mark the key influencing features and weights.

[0006] Preferably, in step S1, the tumor biological data includes tumor single-cell sequencing data, tumor spatial transcriptome data, tumor imaging data, and pathological classification data; individual clinical heterogeneity data includes comorbidity course data, lifestyle data, and previous treatment history data; and patient-reported outcome data includes post-treatment pain scores and daily living ability scores.

[0007] Preferably, in step S2, when quantifying individual clinical heterogeneity data, quantitative indicators are generated based on a preset heterogeneity impact index calculation system. In the heterogeneity impact index calculation system, the comorbidity impact index is calculated by multiplying the number of years of comorbidity duration by the corresponding physiological indicator detection value and then dividing by the normal reference value of the physiological indicator. The lifestyle habit impact index is calculated by multiplying the duration of lifestyle habit by the daily frequency and then dividing by the standard frequency threshold of the habit.

[0008] Preferably, in step S2, when extracting features from tumor biological data, the tumor diameter and density features in tumor imaging data are extracted using an image segmentation algorithm, the immune cell infiltration density features in tumor single-cell sequencing data are extracted using a single-cell parsing algorithm, and the spatial clustering degree features of vascular endothelial growth factor expression in tumor spatial transcriptome data are extracted using a spatial clustering algorithm.

[0009] Preferably, in step S3, the multi-branch feature fusion network is a two-branch attention convolutional neural network, where the first branch is used to learn tumor cell features and the second branch is used to learn tumor microenvironment features. The feature weights of the two branches are dynamically allocated through an attention mechanism. The training of the dual optimization objectives adopts a multi-objective optimization algorithm, which is the NSGA-II algorithm. The chemotherapy efficacy prediction includes objective response rate and progression-free survival. The quality of life impact prediction includes the decline in patient-reported outcome score.

[0010] Preferably, in step S4, the transfer learning algorithm adopts a fine-tuning strategy, freezing the bottom common feature extraction layer of the general chemotherapy effect prediction basic model and only updating the upper cancer type or scheme-specific feature layer; the sample size of the large sample of general cancer types meets the convergence requirements of the general chemotherapy effect prediction basic model; the sample size of the small sample of cancer types is less than 5% of the sample size of the large sample of general cancer types; the sample size of the labeled data used for fine-tuning meets the convergence requirements of the dedicated prediction sub-model.

[0011] Preferably, in step S5, the prediction model is a general chemotherapy efficacy prediction basic model or a specific prediction sub-model; the efficacy level includes excellent, moderate, and poor; the quality of life risk level includes low, moderate, and high; key influencing features and weights are obtained through an interpretation algorithm, and features whose weights meet a preset threshold are selected as key influencing features, and the preset threshold is dynamically set according to the accuracy target of the prediction model.

[0012] Preferably, an artificial intelligence-based tumor chemotherapy efficacy prediction system includes: Data Interaction Layer: Adopts standardized medical data interfaces to connect with hospital-related information systems and wearable device data platforms, automatically capturing multi-dimensional patient data; Feature processing module: used for quantifying individual clinical heterogeneity data and extracting features from tumor biological data; Model training module: used to train a general chemotherapy efficacy prediction base model and generate a specific prediction sub-model through transfer learning; Prediction and decision-making module: used to input patient data to be predicted and output two-dimensional prediction results, supporting interactive adjustment of key feature weights; Visualization output module: Used to generate prediction result interpretation reports and display actual treatment data of similar patients.

[0013] Preferably, the standardized medical data interface adopted by the data interaction layer is the HL7FHIR standard; the hospital-related information systems include the hospital information system, laboratory information system, medical image archive and communication system; the data interaction layer supports incremental data reception, can receive new data during the patient's chemotherapy process in real time, and transmit the new data to the model training module to trigger incremental learning and updating of the model without retraining all the data.

[0014] Preferably, the prediction and decision-making module also includes a regimen recommendation unit. When the output efficacy level is poor and the corresponding probability is less than 30%, the regimen recommendation unit automatically calls a multi-center database of similar patient treatment outcomes, selects at least three alternative chemotherapy regimens, outputs the efficacy rate, quality of life risk rate, and recommendation priority of each alternative regimen, and marks the basis for the recommendation. The basis for the recommendation includes the sample size of similar patients and the degree of matching of patient characteristics. The prediction result interpretation report generated by the visualization output module can be exported in PDF format and can be directly viewed in the hospital's relevant information system.

[0015] The technical effects and advantages of the present invention, which is a method and system for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence, are as follows: 1. This invention constructs a multi-dimensional data system by collecting tumor biological data (single-cell sequencing, spatial transcriptomics, imaging, etc.), individual clinical heterogeneity data (comorbidities, lifestyle habits, etc.), and patient-reported outcome data. It can comprehensively cover tumor pathological characteristics, microenvironment status, individual health background, and post-treatment quality of life information, providing more complete data support for accurate prediction and avoiding prediction bias caused by data gaps.

[0016] 2. This invention achieves quantitative modeling of comorbidities and lifestyle habits through the "Heterogeneity Impact Index Calculation System," avoiding the coarse processing of individual difference information. At the same time, it extracts microenvironment features such as tumor diameter, immune cell infiltration density, and vascular factor expression clustering degree through professional algorithms, transforming unstructured data into high-value features, significantly improving the correlation between features and chemotherapy effects, providing high-quality input for the model, and optimizing prediction accuracy from the source.

[0017] 3. This invention employs a bi-branch / multi-branch attention network, which can efficiently integrate tumor cell characteristics and microenvironment characteristics and dynamically allocate feature weights. At the same time, it introduces a multi-objective optimization algorithm to collaboratively model "chemotherapy efficacy prediction" and "quality of life impact prediction" and output two-dimensional prediction results. This not only meets the needs of efficacy assessment but also provides early warning of quality of life risks, which is more in line with the actual considerations of doctors' clinical decision-making and fills the gap of the single function of traditional models.

[0018] 3. This invention adopts the architecture of "general base model + sub-model transfer fine-tuning". It uses a large sample of general cancer data to train the base model, and then uses a small amount of labeled data to fine-tune and generate a small sample of cancer / rare treatment-specific sub-model. There is no need to carry out large-scale data collection and training again, which significantly reduces the modeling threshold for small sample scenarios, expands the applicability of the model to niche cancer types and new treatment options, and solves the pain point of "inaccuracy due to insufficient data" in traditional technology.

[0019] 5. The system of this invention seamlessly connects with hospital HIS / LIS / PACS systems through a standardized medical data interface, automatically capturing data and reducing manual operation costs. At the same time, it is designed with an interactive adjustment unit (supporting doctors to manually optimize feature weights), a plan recommendation unit (providing alternative plans for patients with poor treatment results), and a visualization output module (intuitively displaying key influencing factors and similar case references), transforming "predicted results" into "decision support", improving doctors' trust in the predicted results and ease of use, significantly optimizing the clinical decision-making process, and reducing diagnosis and treatment costs.

[0020] 6. The system of this invention supports incremental data reception, can acquire new data in real time during the chemotherapy cycle, and dynamically update the model parameters through incremental learning to correct subsequent prediction results, ensuring that the model is always synchronized with the patient's treatment progress, guaranteeing the accuracy and timeliness of long-term prediction, and avoiding the problem of "one prediction failing" in traditional static models. Attached Figure Description

[0021] Figure 1 This is a simplified flowchart of a method and system for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence, as proposed in this invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0023] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Example

[0024] refer to Figure 1 This embodiment provides a method and system for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model for basic implementation. Specific implementation details include: Purpose of implementation: To predict the efficacy of chemotherapy in non-small cell lung cancer (NSCLC), this method addresses the limitations of traditional models, such as limited data dimensions and poor fit with small samples, and achieves synergistic prediction of efficacy and quality of life.

[0025] System Architecture: Data Interaction Layer: Adopting standardized medical data interfaces, it connects to the hospital information systems, laboratory information systems, medical image archiving and communication systems, and wearable device data platforms of three top-tier hospitals, automatically capturing multi-dimensional patient data.

[0026] Feature processing module: includes a heterogeneity quantification unit (calculates comorbidities and lifestyle impact index) and a tumor microenvironment feature extraction unit (extracts immune cell infiltration, imaging features, etc.).

[0027] Model training module: includes a basic model training unit (training a general chemotherapy efficacy prediction basic model) and a sub-model adaptation unit (generating small-sample cancer-specific sub-models through transfer learning).

[0028] Predictive decision module: includes a dual-objective prediction unit (outputting the risk level of efficacy and quality of life) and an interactive adjustment unit (supporting doctors to manually adjust feature weights).

[0029] Visualization output module: Generates a report explaining the prediction results, showing the distribution of tumor microenvironment characteristics, key influencing factors, and actual efficacy data of similar patients.

[0030] Implementation steps: Step 1: Multi-dimensional data acquisition Data on NSCLC patients from 2019 to 2023 were collected from three top-tier hospitals in China, including a large sample of 120,000 cases of common cancer types (lung cancer and breast cancer) and a small sample of 3,000 cases of cholangiocarcinoma.

[0031] Data type: Tumor biological data: tumor single-cell sequencing data (10xGenomicsChromium platform), tumor spatial transcriptome data (Visium platform), chest CT images (1mm slice thickness), and pathological classification (adenocarcinoma / squamous carcinoma).

[0032] Individual clinical heterogeneity data: duration of comorbidities (diabetes / COPD), lifestyle habits (smoking / drinking), and previous treatment history (history of targeted therapy).

[0033] Patient-reported outcome data: pain score (NRS scale) and activities of daily living score (ADL scale) 1 month after treatment.

[0034] Step 2: Feature Engineering Processing Quantification of individual clinical heterogeneity: Calculate the impact index of comorbidities (e.g., the impact index of diabetes = number of years of disease × percentage of glycated hemoglobin / 5.6) and the impact index of lifestyle habits (e.g., the impact index of smoking = number of years of smoking × number of cigarettes per day / 20).

[0035] Tumor biological feature extraction: The tumor diameter (longest diameter / shortest diameter) and density (mean HU value) were extracted from CT images using the U-Net algorithm.

[0036] Single-cell sequencing was performed using Seurat software to extract CD8+ T cell infiltration density (mean 25 cells / mm²).

[0037] Spatial transcriptomics was used to extract the spatial clustering degree of VEGF expression using the DBSCAN algorithm (cluster radius 8 μm, minimum number of cluster points 4 cells).

[0038] Step 3: Basic Model Training A dual-branch attention convolutional neural network is used. The first branch inputs tumor pathological classification and gene mutation data, and the second branch inputs CD8+ T cell density and VEGF clustering degree. Weights are allocated through the attention mechanism (the microenvironment branch accounts for 42%).

[0039] With "Objective Remission Rate (ORR) prediction accuracy" and "ADL score decline prediction error" as dual objectives, the NSGA-II algorithm was trained for 200 epochs with a learning rate of 1e-4. The final base model achieved an ORR prediction accuracy of 89.2% and an ADL score prediction error of ±3.5 points.

[0040] Step 4: Sub-model transfer adaptation: For a small sample of 3,000 cases of cholangiocarcinoma, 150 cases of labeled data were collected, and a fine-tuning strategy was adopted: the bottom 80% of the feature layer of the base model was frozen, and only the top 20% of the cholangiocarcinoma-specific layer was updated, with a learning rate of 1.25e-5 (1 / 8 of the base model) and 80 iterations.

[0041] A sub-model specific to cholangiocarcinoma was generated, with an ORR prediction accuracy of 82.6%.

[0042] Step 5: Prediction model output: Input data for a bile duct cancer patient (diabetes impact index 4.8, CD8+ T cell density 18 cells / mm²).

[0043] Output results: The efficacy level is "moderate" (ORR probability 62%), the quality of life risk level is "moderate" (ADL score decreased by 12 points), and the key impact features and weights are (diabetes impact index 31%, CD8+ T cell density 28%). Implementation effect

[0044] The general basic model achieves an ORR prediction accuracy of 89.2% for NSCLC, which is 12.5% ​​higher than the traditional model.

[0045] The ORR prediction accuracy of the bile duct cancer-specific sub-model was 82.6%, which was 14.3% higher than the independent model without transfer (68.3%).

[0046] It achieves synergistic prediction of efficacy and quality of life, with an accuracy rate of 86.8% in predicting the risk level of quality of life. Example

[0047] This embodiment provides a method and system for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model, used for feature engineering optimization. Specific implementation details include: Purpose of implementation: Optimize methods for quantifying individual clinical heterogeneity and extracting tumor microenvironment features to improve the prediction accuracy for HER2-positive breast cancer patients.

[0048] System Architecture: Data interaction layer: Added interfaces for capturing detailed data such as hypertension and smoking intensity.

[0049] Feature processing module: The heterogeneity quantification unit expands the comorbidity type (adds hypertension) and lifestyle habit correction coefficient; the tumor microenvironment feature extraction unit adds immune cell ratio (CD8+ / Treg) and tumor edge features.

[0050] The other modules are the same as in Example 1.

[0051] Implementation steps: Step 1: Multi-dimensional data acquisition: Data from 50,000 HER2-positive breast cancer patients were collected, and data on hypertension duration, smoking intensity (cigarettes per day grade), Treg cell sequencing data, and tumor margin imaging characteristics were added.

[0052] Step 2: Feature Engineering Processing Quantification of individual heterogeneity: The newly added hypertension impact index = hypertension duration × systolic blood pressure / 140 (e.g., 10-year duration + 150 mmHg systolic blood pressure, index = 10 × 150 / 140 ≈ 10.7).

[0053] The smoking impact index is adjusted by a coefficient (1.2 for more than 20 cigarettes per day, and 0.8 for less than 10 cigarettes per day). For example, if you have a 20-year smoking history and smoke 25 cigarettes per day, the index would be 20 × 25 / 20 × 1.2 = 30.

[0054] Tumor microenvironment feature extraction: The density of newly added CD4+ regulatory T cells (Treg) infiltration (mean 12 cells / mm²) was calculated, and the CD8+ / Treg ratio (mean 2.1) was calculated.

[0055] The CT images showed an additional irregularity in the tumor margins (fractal dimension, mean 1.35).

[0056] Steps 3-5: Same as Example 1 (model training and prediction process is consistent) Implementation effect

[0057] The ORR prediction accuracy for HER2-positive breast cancer patients improved from 89.2% to 91.5%, with the "CD8+ / Treg ratio" accounting for 22% of the key features.

[0058] After expanding the feature dimensions, the model's prediction error for patients with complex comorbidities decreased by 18.3%. Example

[0059] This embodiment provides a method and system for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model, used for basic model training and optimization. Specific implementation details include: Purpose of implementation: Optimize the multi-branch feature fusion network structure and multi-objective optimization algorithm to improve the model's ability to fit complex data.

[0060] System Architecture: Model training module: The basic model training unit supports a three-branch network structure (with a newly added patient-reported outcome feature branch) and integrates two multi-objective optimization algorithms, NSGA-II and MOEA / D.

[0061] The other modules are the same as in Example 1.

[0062] Implementation steps: Steps 1-2: Same as Example 1 (data acquisition and feature processing are consistent) Step 3: Basic Model Training Network structure comparison: Experimental group 1 (dual-branch network): Input tumor cell characteristics + microenvironment characteristics, ORR accuracy 89.2%, quality of life prediction error ±3.5 points.

[0063] Experimental group 2 (three-branch network): The patient-reported outcome feature branch was added, with an ORR accuracy of 90.8% and a quality of life prediction error of ±2.8 points.

[0064] Comparison of multi-objective optimization algorithms: Experimental group A (NSGA-II): Inverse Generational Distance (IGD) 0.12.

[0065] Experimental group B (MOEA / D): Overall performance index 0.18.

[0066] Steps 4-5: Same as Example 1 (sub-model adaptation and prediction process are consistent) Implementation effect

[0067] Compared to the two-branch network, the three-branch network improves ORR accuracy by 1.6% and reduces quality of life prediction error by 20%.

[0068] The NSGA-II algorithm outperforms MOEA / D in overall performance and is more suitable for dual-objective collaborative optimization. Example

[0069] This embodiment provides a method and system for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model, used for sub-model transfer and adaptation scenarios. Specific implementation details include: Purpose of implementation: To validate the effectiveness of transfer learning in adapting to small sample cancer types (pancreatic cancer) and rare chemotherapy regimens (lung cancer "pemetrexed + carboplatin").

[0070] System Architecture: Model training module: The sub-model adaptation unit supports different proportions of feature layer freezing (85%-90%) to adapt to small sample and rare scenario scenarios.

[0071] The other modules are the same as in Example 1.

[0072] Implementation steps: Scenario 1: Small sample size of cancer (pancreatic cancer).

[0073] Step 1: Collect general data from a large sample of 120,000 cases and small sample data from 2,000 cases of pancreatic cancer (with 100 cases labeled).

[0074] Step 2: Feature processing is the same as in Example 1 (extracting tumor microenvironment and heterogeneity features).

[0075] Step 3: The basic model is the same as in Example 1.

[0076] Step 4: Transfer strategy: Freeze 90% of the bottom feature layer of the base model, update 10% of the upper layer, iterate for 70 rounds, and use a learning rate of 1e-5.

[0077] Step 5: Output the predicted PFS (progression-free survival) results for pancreatic cancer patients, with an error of ±1.2 months.

[0078] Scenario 2: Rare chemotherapy regimen (lung cancer "pemetrexed + carboplatin").

[0079] Step 1: Collect 800 data points for this scheme (120 data points are labeled).

[0080] Steps 2-3: Same as in scenario 1.

[0081] Step 4: Transfer strategy: Freeze 85% of the bottom feature layer of the base model, update 15% of the upper layer, iterate for 60 rounds, and use a learning rate of 1.5e-5.

[0082] Step 5: Output the ORR prediction results, with an accuracy of 81.7%. Implementation effect

[0083] The pancreatic cancer sub-model had a PFS prediction error of ±1.2 months, which was 52% lower than the independent model (±2.5 months).

[0084] The rare scheme sub-model achieved an ORR accuracy of 81.7%, a 15.8% improvement over the non-transfer model (65.9%). Example

[0085] This embodiment provides a method and system for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model, used for the application and implementation of the prediction system. Specific implementation details include: Purpose of implementation: To validate the application process and practicality of the prediction system in real-world clinical scenarios and improve doctors' decision-making efficiency.

[0086] System Architecture: Data interaction layer: adopts HL7FHIR standard interface, latency <10 seconds, supports 200 concurrent users.

[0087] Predictive decision-making module: A new treatment plan recommendation unit has been added to automatically filter alternative chemotherapy plans.

[0088] Visual output module: Supports PDF export and direct viewing within the hospital system.

[0089] The other modules are the same as in Example 1.

[0090] Implementation steps: Step 1: System Deployment Server configuration: Intel Xeon Gold 6348 CPU, NVIDIA A100 GPU, interface with hospital HIS / LIS / PACS system.

[0091] Step 2: Clinical Application Process: Doctors submit a "chemotherapy efficacy prediction" application in the HIS system, and the system automatically retrieves patient data (no manual entry required).

[0092] The feature processing module completes the calculation of heterogeneity index (such as COPD impact index 5.3) and microenvironment feature extraction within 10 seconds.

[0093] The model training module calls the lung cancer sub-model, and the prediction and decision module outputs the results ("excellent" efficacy, "low" risk to quality of life).

[0094] Doctors can adjust the weight of "EGFR mutation" from 25% to 30% through an interactive adjustment unit, and the system updates the ORR probability in real time (78%→82%).

[0095] The visualization output module generates a report, including a TME heatmap and a feature weight bar chart, matching 15 similar patients in our hospital (actual ORR was 86.7% consistent with the prediction).

[0096] Step 3: Recommended Solution: When a patient's predicted treatment outcome is "poor" (ORR 28%), the protocol recommendation unit calls upon a multi-center database and outputs three alternative protocols: Option 1: "Paclitaxel + Cisplatin", with an efficacy rate of 58% (210 similar patients, 89% match rate); Option 2: "Osimertinib", with an efficacy rate of 72% (180 similar patients, 85% match rate); Option 3: "Pemetrexed maintenance therapy", with an efficacy rate of 49% (150 similar patients, 82% match). Implementation effect

[0097] The system's average response time is less than 30 seconds, a 97.5% improvement over traditional manual analysis (2 hours).

[0098] Doctors rated the confidence level in the prediction results (on a Likert scale of 5) at 4.2, and the adoption rate of the recommended treatment plan was 78%.

[0099] This comparison provides traditional forecasting methods, including: Traditional method: Data dimensions: Only tumor pathological classification and baseline tumor markers (such as CEA) are used.

[0100] Model structure: Single logistic regression model, predicting only ORR (single objective).

[0101] Small sample adaptation: No transfer learning, independent modeling for small sample cancer types.

[0102] Performance comparison (taking cholangiocarcinoma prediction as an example): Regarding ORR prediction accuracy, the traditional method achieves only 68.3%, while the present invention (Example 1) reaches 82.6%, an improvement of 14.3%. Regarding PFS prediction error, the traditional method has an error of ±3.8 months, while the present invention has an error of only ±1.5 months, a reduction of 60.5%. Regarding quality of life risk prediction coverage, the traditional method lacks this predictive function entirely, with a coverage rate of 0%, while the present invention achieves 100% coverage. Regarding small sample modeling time, the traditional method requires 72 hours to complete small sample cancer modeling, while the present invention only requires 8 hours, improving modeling efficiency by 88.9%.

[0103] This invention, through multi-dimensional data fusion, multi-branch model training, and transfer learning adaptation, significantly outperforms traditional methods in terms of prediction accuracy, functional completeness, and efficiency, and is more in line with actual clinical needs.

[0104] Compared with Comparative Example 1, Examples 1-5 demonstrate a comprehensive breakthrough in five core dimensions—data dimension, feature processing, model performance, small sample adaptation, and clinical application—from core technical solutions to clinical implementation effects. The specific differences are as follows: In terms of data dimensions, Comparative Example 1 relies only on two basic data types: tumor pathological classification and baseline tumor markers (such as CEA). The data dimension is singular and cannot reflect individual patient differences and the state of the tumor microenvironment. In contrast, Examples 1-5 construct a multi-dimensional data system that not only covers tumor biological data (single-cell sequencing, spatial transcriptomics, CT images, etc.) and individual clinical heterogeneity data (comorbidities, disease course, lifestyle habits, etc.), but also includes patient-reported outcome data (pain score, daily living ability score). Example 2 further expands the data to include more detailed data such as hypertension course and smoking intensity. The data dimension is increased by 4-5 times compared to traditional methods, laying a data foundation for accurate prediction.

[0105] At the feature processing level, Comparative Example 1 lacks a systematic feature quantification and extraction process, only using data as raw input; Examples 1-2 innovatively design a "heterogeneity impact index calculation system" to achieve quantitative modeling of comorbidities (such as diabetes and hypertension) and lifestyle habits (such as smoking). It also extracts microenvironment features such as tumor diameter, immune cell infiltration density, and VEGF expression clustering degree through algorithms such as U-Net, Seurat, and DBSCAN. Example 2 adds features such as CD8+ / Treg ratio and tumor edge irregularity. The feature effectiveness is significantly improved compared with traditional methods, reducing the prediction error of the model for complex individuals by 18.3%.

[0106] In terms of model performance, Comparative Example 1 uses a single logistic regression model, which can only predict ORR with a single objective (accuracy of 68.3%) and has no quality of life prediction function. Examples 1-3 construct a two-branch / three-branch attention convolutional neural network and combine it with the NSGA-II multi-objective optimization algorithm to achieve dual-objective prediction of "treatment efficacy + quality of life". Example 1 achieves an ORR prediction accuracy of 89.2% for NSCLC. Example 3 further improves the accuracy to 90.8% through a three-branch network and reduces the quality of life prediction error to ±2.8 points. Compared with traditional methods, it achieves a double breakthrough in prediction accuracy and functional integrity.

[0107] The ability to adapt to small samples is the key difference between the two: Comparative Example 1 uses independent modeling for small sample cancer types (such as cholangiocarcinoma and pancreatic cancer), which takes 72 hours and has an ORR accuracy of only 68.3%; Examples 1 and 4 use transfer learning strategies to freeze 80%-90% of the feature layers of the general base model and only update the specific layers, which improves the ORR accuracy of the cholangiocarcinoma sub-model to 82.6%, reduces the PFS prediction error of pancreatic cancer to ±1.2 months, shortens the small sample modeling time to 8 hours, and improves efficiency by 88.9%, solving the pain points of "difficult modeling and low accuracy" in small sample scenarios of traditional methods.

[0108] In clinical application, Comparative Example 1 required manual data entry and analysis, taking an average of 2 hours, and lacked decision support functions; Example 5, on the other hand, uses the HL7FHIR interface to seamlessly connect with the hospital's HIS / LIS / PACS system, automatically captures data, has an average response time of less than 30 seconds, and also supports interactive adjustment of feature weights and recommendation of alternative solutions (such as recommending 3 types of solutions for patients with poor treatment outcomes in Example 5). The doctor trust score reached 4.2 points (Likert 5 scale), and the solution adoption rate was 78%, which significantly improved the efficiency of clinical decision-making compared to traditional methods.

[0109] In summary, the traditional method in Comparative Example 1, due to its narrow data, coarse features, simple model, poor adaptability, and difficulty in implementation, can no longer meet the needs of precision chemotherapy. However, the present invention embodied in Examples 1-5, through multi-dimensional data fusion, refined feature processing, high-performance model construction, efficient small-sample adaptation, and clinical system integration, achieves a leap from "single prediction" to "precise decision support," fully meeting the actual needs of clinical practice.

[0110] The above embodiments can be implemented in whole or in part by software, hardware, firmware or other arbitrary combinations. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product.

[0111] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0112] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0113] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0114] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing an artificial intelligence-based model for predicting the efficacy of tumor chemotherapy, characterized in that, Includes the following steps: S1. Multi-dimensional data acquisition: Collect tumor biology data, individual clinical heterogeneity data, and patient-reported outcome data from patients with the target cancer type; S2. Feature engineering processing: Quantitative processing is performed on the individual clinical heterogeneity data to generate quantitative indicators, and feature extraction is performed on the tumor biological data to form a tumor microenvironment feature set; S3. Basic Model Training: Using a large sample of multi-dimensional data of common cancer types as the training set, a general chemotherapy efficacy prediction basic model is constructed using a multi-branch feature fusion network. The prediction of chemotherapy efficacy and the prediction of the impact on quality of life are trained as dual optimization objectives. S4. Sub-model transfer adaptation: For small sample cancer types or rare chemotherapy regimens, the general chemotherapy effect prediction base model is fine-tuned using a transfer learning algorithm to generate a specific prediction sub-model; S5. Predictive Model Output: Input the multi-dimensional data of the patient to be predicted into the prediction model of the corresponding cancer type or treatment plan, and output the two-dimensional prediction results of efficacy level and quality of life risk level, and mark the key influencing features and weights.

2. The method for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence as described in claim 1, characterized in that, In step S1, tumor biological data include tumor single-cell sequencing data, tumor spatial transcriptome data, tumor imaging data, and pathological classification data; individual clinical heterogeneity data include comorbidity course data, lifestyle data, and previous treatment history data; and patient-reported outcome data include post-treatment pain scores and daily living ability scores.

3. The method for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence as described in claim 1, characterized in that, In step S2, when quantifying individual clinical heterogeneity data, quantitative indicators are generated based on a preset heterogeneity impact index calculation system. In the heterogeneity impact index calculation system, the comorbidity impact index is calculated by multiplying the number of years of comorbidity duration by the corresponding physiological indicator detection value and then dividing by the normal reference value of the physiological indicator. The lifestyle habit impact index is calculated by multiplying the duration of lifestyle habit by the daily frequency and then dividing by the standard frequency threshold of the habit.

4. The method for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence as described in claim 1, characterized in that, In step S2, when extracting features from tumor biological data, the tumor diameter and density features in the tumor imaging data are extracted by image segmentation algorithm, the immune cell infiltration density features in the tumor single-cell sequencing data are extracted by single-cell parsing algorithm, and the spatial clustering degree features of vascular endothelial growth factor expression in the tumor spatial transcriptome data are extracted by spatial clustering algorithm.

5. The method for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence as described in claim 1, characterized in that, In step S3, the multi-branch feature fusion network is a two-branch attention convolutional neural network. Its first branch is used to learn tumor cell features, and its second branch is used to learn tumor microenvironment features. The feature weights of the two branches are dynamically allocated through the attention mechanism. The training of the dual optimization objectives adopts a multi-objective optimization algorithm, which is the NSGA-II algorithm. The chemotherapy efficacy prediction includes objective response rate and progression-free survival. The quality of life impact prediction includes the decline in patient-reported outcome score.

6. The method for constructing a tumor chemotherapy efficacy prediction model based on artificial intelligence as described in claim 1, characterized in that, In step S4, the transfer learning algorithm adopts a fine-tuning strategy, freezing the bottom common feature extraction layer of the general chemotherapy effect prediction basic model and only updating the upper cancer type or treatment-specific feature layer; the sample size of the large sample of general cancer types meets the convergence requirements of the general chemotherapy effect prediction basic model; the sample size of the small sample of cancer types is less than 5% of the sample size of the large sample of general cancer types; the sample size of the labeled data used for fine-tuning meets the convergence requirements of the dedicated prediction sub-model.

7. The method for constructing an artificial intelligence-based tumor chemotherapy efficacy prediction model as described in claim 1, characterized in that, In step S5, the prediction model is either a general chemotherapy efficacy prediction basic model or a specific prediction sub-model; the efficacy level includes excellent, moderate, and poor; the quality of life risk level includes low, moderate, and high; key influencing features and weights are obtained through an interpretation algorithm, and features whose weights meet a preset threshold are selected as key influencing features. The preset threshold is dynamically set according to the accuracy target of the prediction model.

8. An artificial intelligence-based tumor chemotherapy efficacy prediction system, characterized in that, The system includes: Data Interaction Layer: Adopts standardized medical data interfaces to connect with hospital-related information systems and wearable device data platforms, automatically capturing multi-dimensional patient data; Feature processing module: used for quantifying individual clinical heterogeneity data and extracting features from tumor biological data; Model training module: used to train a general chemotherapy efficacy prediction base model and generate a specific prediction sub-model through transfer learning; Prediction and decision-making module: used to input patient data to be predicted and output two-dimensional prediction results, supporting interactive adjustment of key feature weights; Visualization output module: Used to generate prediction result interpretation reports and display actual treatment data of similar patients.

9. The tumor chemotherapy efficacy prediction system based on artificial intelligence as described in claim 8, characterized in that, The standardized medical data interface adopted by the data interaction layer is the HL7FHIR standard; the hospital-related information systems include the hospital information system, laboratory information system, medical image archive and communication system; the data interaction layer supports incremental data reception, can receive new data during the patient's chemotherapy process in real time, and transmit the new data to the model training module to trigger incremental learning and updating of the model without retraining all the data.

10. The tumor chemotherapy efficacy prediction system based on artificial intelligence as described in claim 8, characterized in that, The prediction and decision-making module also includes a regimen recommendation unit. When the output efficacy level is poor and the corresponding probability is less than 30%, the regimen recommendation unit automatically calls the multicenter similar patient treatment outcome database, selects at least 3 alternative chemotherapy regimens, outputs the efficacy rate, quality of life risk rate and recommendation priority of each alternative regimen, and marks the basis for the recommendation. The recommendations are based on factors including the size of similar patient samples and the degree of matching of patient characteristics. The prediction result interpretation report generated by the visualization output module can be exported as a PDF and can be directly viewed in the hospital's relevant information system.

Images