Transfer learning method for bad outcome prediction of small-sample early-stage cancer patients

Abstract

The invention relates to the technical field of medical treatment, and discloses a transfer learning method for bad outcome prediction of small-sample early-stage cancer patients. According to the method, the neural network risk function is introduced and the transfer learning framework is combined, so that the nonlinear interaction of high-dimensional features can be better fitted, and the prediction accuracy of the small-probability bad outcome of the early-stage tumor is improved. Although the existing RSF can overcome linear hypothesis to a certain extent, the existing RSF is susceptible to noise interference in a small sample queue, and the prediction stability is insufficient. According to the migration survival forest method provided by the invention, by training and migrating structure priori on a large-scale database, a stable tree structure can be kept under a small sample condition, and a split threshold is re-estimated in combination with target data, so that global structure knowledge and local adaptability are taken into consideration, and the reliability of the system is improved. Therefore, the robustness and generalization ability of the model are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a transfer learning method for predicting adverse outcomes in small-sample early-stage cancer patients. Background Technology

[0002] With the aging population and changes in environment and lifestyle, the incidence and mortality rates of malignant tumors in my country are rising year by year, making cancer a major public health problem that seriously threatens human health. In recent years, thanks to the promotion of early screening and the application of new technologies such as circulating tumor DNA (ctDNA), more and more cancer patients are being diagnosed at an early stage (stage I). For example, about 43% of newly diagnosed breast cancer cases are stage I patients, while about 30% of colorectal cancer cases are diagnosed at stage I. For these early-stage patients, after receiving standardized treatment such as radical surgery, their overall survival prognosis is generally good, but some patients still experience adverse outcomes, such as recurrence, metastasis, or even death. For example, in colorectal cancer, about 5%–10% of stage I patients experience recurrence or death within five years. Therefore, how to accurately identify potentially high-risk individuals for adverse outcomes at the early stage of cancer diagnosis is a core issue that urgently needs to be addressed in clinical practice, and it has significant value for the development of individualized treatment strategies and follow-up management.

[0003] While existing technologies have made some progress in predicting tumor survival, they still have several limitations. First, the Cox proportional hazards model is widely used in clinical practice, but it relies on the linear assumption between covariates and the risk function, making it difficult to capture complex nonlinear effects and high-dimensional interactions. Therefore, its accuracy is limited in predicting low-probability adverse outcomes in early-stage cancer patients. Second, the Random Survival Forest (RSF), as a nonparametric method, mainly relies on the log-rank statistic to select split points when nodes split. Although it can overcome the linear limitation to some extent, it is easily affected by noise in early-stage cancer cohorts with small sample sizes, resulting in model instability and poor generalization ability.

[0004] Furthermore, existing technologies typically lack specific optimization for low-probability outcomes in early-stage patients, generally fail to effectively utilize prior knowledge in large-scale public databases (SEER), and lack the application of transfer learning and structural priors. Consequently, they struggle to maintain stable predictive performance in small sample data of early-stage cancer. Therefore, it is crucial to improve the accuracy, robustness, and generalization ability of existing technologies to achieve precise identification of high-risk individuals. Summary of the Invention

[0005] In view of this, the present invention provides a transfer learning method for predicting adverse outcomes in small-sample early-stage cancer patients, in order to solve the above problems.

[0006] To address the above technical problems, this invention provides a transfer learning method for predicting adverse outcomes in small-sample early-stage cancer patients, comprising: Collect cancer patient data from large-scale public source databases as source task data, and collect characteristic data, survival time data, and censoring indication data from a small sample of early-stage cancer patients as target task data; The survival prediction model is pre-trained based on source task data to learn the basic disease risk pattern, and then the pre-trained survival prediction model is fine-tuned or retrained based on target task data. A random survival forest is trained on the source task data and the tree structure prior is extracted. Based on the tree structure prior, the splitting threshold is re-estimated on the target task data and the lower-level nodes are continued to grow to form the target survival forest. Input the target task data into the fine-tuned or retrained survival prediction model and the target survival forest, and output the prediction results of adverse outcomes for early-stage cancer patients. The prediction results were evaluated using the consistency index and the comprehensive Brier score.

[0007] As an alternative approach, the survival prediction model is a parametric survival analysis model, in which the patient risk score output by the model is optimized through a loss function during the pre-training process.

[0008] As an alternative approach, the tree structure prior is extracted by analyzing the frequency of occurrence of the first k layers of splitting feature combinations of each survival tree trained in the statistical source task, and determining the high-frequency splitting feature combinations as the tree structure prior; where k is an adjustable hyperparameter.

[0009] As an alternative approach, the process of re-estimating the split threshold involves calculating the log-rank statistic based on the target task data to obtain the optimal split threshold.

[0010] As an alternative approach, the cumulative risk function of the leaf nodes of the target survival forest is estimated using Nelson-Aalen, and it is also used to calculate the cumulative risk based on event data and risk set data of samples within the leaf nodes.

[0011] As an alternative approach, when fine-tuning or retraining the pre-trained survival prediction model, the model parameters obtained from the source task training are used as initial parameters, and the parameters are adjusted in combination with the target task data to adapt to the characteristics of small sample early cancer patient data.

[0012] As an optional approach, the consistency index is used to assess the model's ability to distinguish patient risk rankings; the combined Brier score is used to assess the model's predictive calibration performance throughout the follow-up period.

[0013] The beneficial effects of this invention are as follows: This invention, by introducing a neural network risk function and combining it with a transfer learning framework, can better fit the nonlinear interactions of high-dimensional features, thereby improving the prediction accuracy of low-probability adverse outcomes in early-stage tumors. Secondly, while RSF can overcome the linear assumption to some extent, it is susceptible to noise interference in small sample cohorts, resulting in insufficient prediction stability. The transfer survival forest method proposed in this embodiment, by training on a large-scale database and transferring structural priors, can maintain a stable tree structure under small sample conditions. Furthermore, by re-estimating the splitting threshold using target data, it achieves a balance between global structural knowledge and local adaptability, thus significantly improving the model's robustness and generalization ability. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the transfer learning method for predicting adverse outcomes in small-sample early-stage cancer patients according to the present invention. Figure 2 This is a schematic diagram of the migration survival forest model structure of this invention when K=3; Figure 3 This is a line graph showing that TSF outperforms RSF in terms of consistency index under different training set sizes according to the present invention; Figure 4 This is a line graph showing that the TSF is lower than the RSF on IBS under different training set sizes according to the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of the present invention to enable the reader to better understand the present invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0016] Please see Figure 1 This embodiment provides a transfer learning method for predicting adverse outcomes in a small sample of early-stage cancer patients, including: Collect cancer patient data from large-scale public source databases as source task data, and collect characteristic data, survival time data, and censoring indication data from a small sample of early-stage cancer patients as target task data; The survival prediction model is pre-trained based on source task data to learn the basic disease risk pattern, and then the pre-trained survival prediction model is fine-tuned or retrained based on target task data. A random survival forest is trained on the source task data and the tree structure prior is extracted. Based on the tree structure prior, the splitting threshold is re-estimated on the target task data and the lower-level nodes are continued to grow to form the target survival forest. Input the target task data into the fine-tuned or retrained survival prediction model and the target survival forest, and output the prediction results of adverse outcomes for early-stage cancer patients. The prediction results were evaluated using the consistency index and the comprehensive Brier score.

[0017] In implementing the above method, this embodiment first collects data. On the one hand, it collects cancer patient data from a large-scale public source database as source task data. The large-scale public source database is preferably the SEER database, which contains rich clinical data of cancer patients and can provide sufficient prior knowledge support for the model. On the other hand, it collects feature data, survival time data, and censoring indication data of a small sample of early-stage cancer patients as target task data, which is used to adapt the model to the training environment of the small-sample early-stage cancer prediction scenario.

[0018] Next, the survival prediction model undergoes pre-training and fine-tuning or retraining. This survival prediction model is a parametric survival analysis model. During the pre-training phase, it is trained based on collected source task data to learn the basic disease risk patterns. During training, a loss function is used to optimize the patient risk score output by the model. This loss function measures the deviation between the model's predicted risk score and the actual patient survival, ensuring that the pre-trained model accurately captures the general patterns of survival risk for cancer patients. Here, a negative log-partial likelihood function is used to calculate the loss and optimize the model. In this embodiment, the input is assumed to be a patient feature vector. Survival time is The deletion indicator is The neural network outputs a risk score. Its corresponding negative log-partial likelihood function is defined as:

[0019] in Indicates in The sample set remains at the risk concentration at all times. Through transfer learning, this invention utilizes parameters learned from the source task. As the initial value, then based on the target task data Fine-tuning is performed to maintain stable prediction performance even in small sample scenarios.

[0020] After pre-training is completed, the survival prediction model is fine-tuned or retrained based on the target task data. During fine-tuning or retraining, the model parameters obtained from the source task training are used as initial parameters, and the parameters are adjusted in combination with the target task data to adapt the model to the data characteristics of a small sample of early cancer patients, alleviate the problems of model overfitting and training instability under small sample conditions, and improve the model's adaptability in the target scenario.

[0021] Subsequently, a migration survival forest is constructed. First, a random survival forest is trained on the source task data. After training, the tree structure prior is extracted. Specifically, the extraction method involves statistically analyzing the frequency of the first k layers of splitting feature combinations in each survival tree trained on the source task, identifying high-frequency splitting feature combinations as the tree structure prior, where k is an adjustable hyperparameter. Then, based on the tree structure prior, the splitting threshold is re-estimated on the target task data. The re-estimation process calculates the log-rank statistic based on the target task data. This statistic measures the event differences between sample subsets under different splitting methods, thereby determining the optimal splitting threshold.

[0022] Traditional Random Survival Forest (RSF) selects the optimal splitting feature and threshold based on the log-rank test when splitting nodes. Specifically, for a given candidate splitting feature... and split points Divide the dataset into a left subset and right subset Its log-rank statistic is:

[0023] in Indicates the left subset within the time interval The number of events within, The expected number of events, This is a variance estimate. It is achieved by maximizing... To determine the optimal split point.

[0024] Based on this, this embodiment proposes a migration survival forest (TSF). First, a random survival forest is trained on the source task, denoted as . In front of each tree Layer splitting feature combination is denoted as Its empirical distribution is as follows:

[0025] And it serves as a structural prior for migration. This is an indicator function. Parameters The depth of the migration, i.e., how many layers of the split structure are fixed from the source task, is essentially an adjustable hyperparameter. When the target data is small, a smaller value is preferable. It can provide good generalization; when there is sufficient target data, a larger [size / scale] can be selected. Alternatively, no restrictions may be set to preserve more structural information. After determining the splitting threshold, lower-level nodes continue to grow, eventually forming the target survival forest.

[0026] In constructing the target survival forest, the cumulative risk function of its leaf nodes is estimated using the Nelson-Aalen method. This method calculates the cumulative risk based on event data and risk set data of samples within the leaf nodes, providing a foundation for subsequent survival probability prediction.

[0027] in Indicates the number of events. This indicates the number of people in the risk group.

[0028] Please see Figure 2 , Figure 2 This demonstrates the overall workflow of TSF: the left side shows the forest trained on the source data (source forest), and the high-frequency tree structures are statistically analyzed; the right side shows the forest trained on the target task based on... The sampled prototype tree, front The layer feature combination is fixed, and only the splitting threshold is re-determined by the target data, and subsequent nodes grow independently.

[0029] After model construction and optimization are completed, the prediction stage begins. The target task data is input into the fine-tuned or retrained survival prediction model and the constructed target survival forest. Through the synergistic effect of the two models, the prediction results of adverse outcomes for early-stage cancer patients are output. These adverse outcomes include recurrence, metastasis, and death of early-stage cancer patients, thereby identifying high-risk individuals.

[0030] Finally, the prediction results were evaluated using two metrics: the concordance index (c-index) and the integrated Brier score (IBS). The concordance index assesses the model's ability to distinguish patient risk rankings; the integrated Brier score assesses the model's predictive calibration performance throughout the follow-up period, quantifying performance by measuring the deviation between the model's predicted survival probability and the patient's actual survival. The IBS is defined as follows:

[0031] in, Individuals represented by model predictions At any moment The probability of survival, For indicator functions, This represents the maximum follow-up time. A smaller IBS value indicates better overall predictive performance of the model.

[0032] This evaluation process ensures the reliability and practicality of the prediction results and verifies the effectiveness of the model in predicting adverse outcomes in a small sample of early-stage cancer patients.

[0033] This evaluation process ensures the reliability and practicality of the prediction results and verifies the effectiveness of the model in predicting adverse outcomes in a small sample of early-stage cancer patients.

[0034] Based on the above approach, this embodiment proposes a survival prediction method combining transfer learning and structural prior constraints. It provides a practical technical solution to the problems of difficulty in predicting low-probability adverse outcomes in early-stage cancer patients, limited sample size, and insufficient generalization performance of traditional methods. Compared with existing Cox regression models that rely on linear assumptions and the shortcomings of Random Survival Forest (RSF) models, which are susceptible to noise interference and lack robustness in small sample conditions, this embodiment significantly alleviates overfitting and model instability by pre-training on a large-scale database to learn disease-related basic risk patterns and then fine-tuning or retraining on a small sample target cohort. Furthermore, this embodiment proposes the Transfer Survival Forest (TSF) method, which extracts high-frequency splitting structures as prior information in the source task and re-estimates the splitting threshold in the target task to continue growth. This allows the model to inherit global structural knowledge while flexibly adapting to local data features, maintaining good predictive performance even in small sample scenarios.

[0035] In terms of experimental results, this embodiment not only performs better on the consistency index (Ctd), but also demonstrates better robustness on the overall Brier score. For example... Figure 3 As shown, under different training set sizes, the TSF method (blue line) consistently outperforms the traditional RSF method (red line) in terms of Ctd (calculus potential), especially under small sample conditions, where the difference is even more pronounced. Furthermore, the TSF method's performance approaches or even surpasses that of the source task model (dashed line), indicating that transfer priors can effectively improve risk ranking capabilities. Figure 4 Furthermore, it is shown that the TSF method consistently outperforms the RSF method in terms of IBS, demonstrating higher reliability and lower error in prediction results throughout the overall follow-up period. Combining these two indicators, it can be seen that this embodiment, under small sample conditions, improves Ctd by an average of approximately 3–5 percentage points (a relative improvement of approximately 7.7%) compared to traditional methods, while significantly reducing IBS, fully demonstrating the dual advantages of this method in terms of discriminative power and calibration performance.

[0036] In summary, the key technical point protected in this embodiment lies in proposing a migration survival forest method suitable for predicting early-stage tumors in small samples by introducing a transfer learning framework and structural prior constraints. Its significant advantages in accuracy, robustness, and clinical applicability are verified through a dual-indicator system and experimental data. Furthermore, this embodiment can also predict low-probability adverse outcomes in early-stage tumor patients by introducing data from other modalities (such as medical imaging, genomics, or transcriptomics data).

[0037] The embodiments of the present invention have been described in detail above. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A transfer learning method for predicting adverse outcomes in small-sample early-stage cancer patients, characterized in that, include: Collect cancer patient data from large-scale public source databases as source task data, and collect characteristic data, survival time data, and censoring indication data from a small sample of early-stage cancer patients as target task data; The survival prediction model is pre-trained based on the source task data to learn the basic disease risk pattern, and then the pre-trained survival prediction model is fine-tuned or retrained based on the target task data. A random survival forest is trained on the source task data and the tree structure prior is extracted. Based on the tree structure prior, the splitting threshold is re-estimated on the target task data and the lower-level nodes are continued to grow to form the target survival forest. The target task data is input into the fine-tuned or retrained survival prediction model and the target survival forest, and the adverse outcome prediction results for early-stage cancer patients are output. The prediction results were evaluated using the consistency index and the comprehensive Brier score.

2. The method according to claim 1, characterized in that, The survival prediction model is a parametric survival analysis model, and the patient risk score output by the model is optimized through a loss function during the pre-training process.

3. The method according to claim 1, characterized in that, The extracted tree structure priors include the frequency of occurrence of the first k layers of splitting feature combinations of each survival tree trained in the statistical source task, and the high-frequency splitting feature combinations are determined as tree structure priors; where k is an adjustable hyperparameter.

4. The method according to claim 1, characterized in that, The process of re-estimating the split threshold includes calculating the log-rank statistic based on the target task data to obtain the optimal split threshold.

5. The method according to claim 1, characterized in that, The cumulative risk function of the leaf nodes of the target survival forest is estimated using Nelson-Aalen, and it is also used to calculate the cumulative risk based on event data and risk set data of samples within the leaf nodes.

6. The method according to claim 1, characterized in that, When fine-tuning or retraining the pre-trained survival prediction model, the model parameters obtained from the source task training are used as initial parameters, and the parameters are adjusted in combination with the target task data to adapt to the characteristics of small sample early cancer patient data.

7. The method according to claim 1, characterized in that, The consistency index is used to assess the model's ability to distinguish patient risk rankings; the composite Brier score is used to assess the model's predictive calibration performance throughout the follow-up period.

Images