Method for predicting the efficacy of umbilical cord blood natural killer cell treatment of neuroblastoma

Abstract

The present application provides a method for predicting the efficacy of umbilical cord blood natural killer cell treatment of neuroblastoma, belonging to the technical field of medical prediction and evaluation. The present application solves the clinical need for early and accurate prediction of the efficacy of immunotherapy for neuroblastoma. The method comprises: obtaining peripheral blood test data of a neuroblastoma patient before multiple reinfusion treatment nodes of umbilical cord blood natural killer cell treatment; based on the processed node-level peripheral blood test data, at least one dynamic linkage feature reflecting the dynamic linkage relationship of the indicators in the patient's body is calculated; at least one static feature index is obtained; the dynamic linkage feature and the static feature index are input into a pre-trained efficacy prediction model to obtain the response probability of the patient to umbilical cord blood natural killer cell immunotherapy. The present application accurately predicts the efficacy of umbilical cord blood natural killer cell treatment of neuroblastoma by analyzing the peripheral blood test data of the patient at multiple time points and constructing dynamic linkage features.

Description

Technical Field

[0001] This invention relates to the field of medical prediction and assessment technology, specifically to a method for predicting the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma. Background Technology

[0002] Neuroblastoma is a highly malignant solid tumor originating from the neural crest cells of the embryonic period, and is one of the most common extracranial solid tumors in infants and young children. Its clinical manifestations and biological behavior are heterogeneous, accounting for approximately 8%-10% of childhood malignancies and 15% of childhood cancer-related deaths, with a high recurrence rate and poor prognosis. According to the international neuroblastoma risk stratification system, even with multimodal treatments such as surgical resection, multi-drug chemotherapy, radiotherapy, and autologous hematopoietic stem cell transplantation, the long-term overall survival rate of high-risk patients remains below 50%. Furthermore, relapsed / refractory neuroblastoma has a poor prognosis; approximately 20% of high-risk neuroblastoma patients are refractory after standard induction therapy, and 50% relapse after achieving remission. Therefore, improving patient prognosis remains a significant challenge in clinical research.

[0003] In recent years, cell immunotherapy has emerged as a novel treatment strategy for improving the prognosis of high-risk, relapsed / refractory neuroblastoma patients due to its unique advantages in targeting and eliminating tumor cells and overcoming drug resistance. However, significant individual differences exist in its clinical efficacy, making the identification of patients who will benefit from treatment a key issue in treatment decision-making. Currently, there is a lack of cost-effective and convenient efficacy prediction models in clinical practice, making it difficult to meet the needs of individualized efficacy assessment in cell therapy and traditional treatments. Although existing methods such as MYCN status assessment and MIBG imaging are valuable for risk stratification and monitoring, their reliance on technological platforms and resources leads to varying accessibility across different centers, limiting their widespread application and timely assessment and individualized adjustment of efficacy responses during treatment. Therefore, identifying beneficiaries early in treatment and developing more cost-effective and convenient efficacy prediction methods are clinical needs for improving the prognosis of high-risk neuroblastoma patients.

[0004] While peripheral blood inflammatory markers have shown potential in predicting the prognosis of various cancers, their association with efficacy in umbilical cord blood natural killer cell therapy for neuroblastoma remains unclear. Therefore, this method integrates peripheral blood test data from eight pre-infusion umbilical cord blood natural killer cell therapies to construct a predictive model for the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma. Compared to existing prediction methods that rely on advanced technologies, this method, based on peripheral blood test data, offers advantages such as economy and simplicity, facilitating clinical translation and providing evidence-based support for early identification of patients who will benefit from umbilical cord blood natural killer cell therapy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by proposing a solution. To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: A method for predicting the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma, the method comprising: Peripheral blood test data of neuroblastoma patients before multiple reinfusion treatment nodes of umbilical cord blood natural killer cell therapy were obtained, and the data were subjected to node alignment and standardization to form node-level peripheral blood test data. Based on the processed node-level peripheral blood detection data, at least one dynamic linkage feature reflecting the dynamic linkage relationship of indicators in the patient's body is calculated. The dynamic linkage feature is constructed based on the synergistic changes of at least two detection indicators at multiple time points in the same patient. Obtain at least one static feature indicator; The dynamic linkage features and the static feature indicators are input into a pre-trained efficacy prediction model to obtain the patient's response probability to umbilical cord blood natural killer cell immunotherapy.

[0006] Preferably, the peripheral blood test data includes complete blood count (CBC) test data and biochemical test data; and / or The efficacy prediction model is a logistic regression model with L2 regularization, a support vector machine, or other statistical learning or machine learning model that can output probability results.

[0007] Preferably, the at least one dynamic linkage feature is constructed based on an index pair consisting of platelet-weighted hematocrit and basophil percentage.

[0008] Preferably, the dynamic linkage feature includes: The dynamic response slope of the percentage of basophils relative to changes in platelet-weighted hematoma is denoted as dyn_slope_BA_on_PCT, and its calculation methods include Theil-Sen estimator, linear regression slope, or robust regression slope. And / or, The proportion of plateletcrit and basophil percentage changes in the same direction in adjacent time periods is denoted as dyn_same_dir_rate_BA_PCT.

[0009] Preferably, the dyn_slope_BA_on_PCT is calculated using the Theil-Sen estimator, including: For multiple time points of the same patient, calculate the local slope of the platelet hematocrit and the percentage of basophils between any two time points; The median of all valid local slopes is taken as dyn_slope_BA_on_PCT.

[0010] Preferably, the static characteristic indicators include at least one of the following: mean hemoglobin; standard deviation of red blood cell distribution width; serum amyloid protein; sodium ion concentration.

[0011] More preferably, the acquisition of peripheral blood test data of neuroblastoma patients before multiple reinfusion treatment nodes includes: acquiring clinical peripheral blood test data of patients before eight reinfusion treatment nodes.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements a method for predicting the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the efficacy of umbilical cord blood natural killer cells in the treatment of neuroblastoma.

[0014] To achieve the technical objective of this invention, this invention also provides a device for predicting the efficacy of umbilical cord blood natural killer cell immunotherapy for neuroblastoma, the device comprising: The data processing module is used to acquire peripheral blood test data of neuroblastoma patients before multiple reinfusion treatment nodes of umbilical cord blood natural killer cell therapy, and to perform node alignment and standardization processing on the data. The feature construction module is used to calculate at least one dynamic linkage feature reflecting the dynamic linkage relationship of indicators in the patient's body based on the processed data, and to obtain at least one static feature indicator. The prediction module is used to input the dynamic linkage features and the static feature indicators into a pre-trained efficacy prediction model to output the patient's response probability to treatment.

[0015] Compared with the prior art, the present invention achieves the following technical effects: ① Data modeling based on peripheral blood test indicators. The data used in this method mainly comes from routine blood tests and biochemical tests performed during clinical diagnosis and treatment. There is no need to add special test items, special test equipment or high-cost molecular detection methods, which reduces the data acquisition cost required for auxiliary efficacy assessment. It can complete the construction of the input features required for efficacy prediction model, and has good feasibility and strong clinical application value.

[0016] ② This method can identify dynamic structures related to treatment efficacy in noisy clinical settings. In cell therapy, patients' hematological indicators are easily affected by factors such as chemotherapy-induced bone marrow suppression, infection-related inflammatory responses, and individual differences in baseline status. This means that increases or decreases in a single indicator at a single time point often lack specificity and fail to consistently reflect the true therapeutic effect. Unlike methods based solely on a single static indicator, this method focuses on the dynamic linkage structure of time series within the patient, identifying a stable coupling relationship between PCT and BA% in the effective group and a weaker one in the ineffective group, thereby extracting signals that are more relevant to treatment efficacy. This dynamic linkage feature reflects the systemic impact of treatment on the body's physiological state more effectively than a single static indicator, and its combination with static feature indicators further improves the specificity of efficacy identification.

[0017] ③ The final model exhibited a high Area Under the Receiver Operating Characteristic Curve (AUC) of 0.935 and an Average Precision (AP) of 0.964 in a small sample size of 39. The Brier score, used to measure the calibration and accuracy of probability predictions, was 0.107, indicating that the model has good discriminative ability under repeated cross-validation. In contrast, the tree model showed more significant deviations in some probability intervals, indicating insufficient stability of probability estimation in small sample settings. Meanwhile, the permutation label experiment also showed that the model's performance stemmed from genuine structural associations rather than random fitting.

[0018] ④ It is interpretable and easy to use in clinical practice. The overall method adopts a dynamic feature-driven and relatively simple logistic regression model with L2 regularization. Its coefficients can directly quantify the direction and degree of influence of each input feature on the efficacy prediction results, making it easy to explain which features are the key factors affecting efficacy judgment. Compared with tree models, it is less prone to overfitting in small sample clinical data scenarios and is more suitable for actual clinical research and deployment. In addition, it can be further combined with the influence direction and prediction basis of SHAP (SHapley Additive exPlanation) output features, which is beneficial to providing understandable and traceable auxiliary references for clinical diagnosis and treatment. Attached Figure Description

[0019] For ease of explanation, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0020] Figure 1 This is a flowchart of the method for predicting the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma according to the present invention. Detailed Implementation

[0021] The following are specific embodiments of the present invention, described in conjunction with the accompanying drawings, to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments. Specific details, such as particular configurations, are provided in the following description merely to aid in a comprehensive understanding of the embodiments of the present invention. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0023] Example 1 Step 1: Data Governance This method is based on predicting the efficacy of natural killer cell immunotherapy for neuroblastoma derived from umbilical cord blood. First, clinical peripheral blood test data corresponding to eight infusion treatment nodes were acquired and standardized to form node-level peripheral blood test data for subsequent feature analysis. Peripheral blood test data were obtained from the hospital's electronic medical record system, laboratory database, and treatment follow-up records, covering multiple time-point test results associated with each infusion treatment node throughout the patient's entire treatment cycle, including complete blood count and biochemical test data. Efficacy labels were derived from clinical evaluation results, including complete response (CR), sustained complete response (SCR), partial response (PR), progressive disease (PD), and stable disease (SD). During model training, supervised learning labels were constructed based on previous clinical efficacy grouping information, where CR, SCR, and PR corresponded to the effective group, and PD and SD corresponded to the ineffective group.

[0024] Step 2: Data Analysis Given that single static indicators are easily affected by background noise such as chemotherapy-induced bone marrow suppression and infection-related inflammatory responses, making accurate prediction of efficacy difficult, this method employs a patient-level analysis strategy to identify dynamic linkage features related to efficacy. Within each patient, the strength of synergistic changes between indicator pairs is measured based on multi-timepoint detection results before infusion therapy, and the results are aggregated and compared among patients to minimize the confounding effect of inter-patient differences on dynamic association estimation. This allows for the identification of dynamic linkage features that characterize the changes in the patient's physiological state during treatment. Specifically, the linkage structure of different detection indicators during the same patient's treatment process is the primary analysis object, rather than analyzing only the rise and fall of a single detection indicator. Simultaneously, the following screening conditions for dynamic linkage features are set: the sample size of both the effective and ineffective groups is not less than 10; the absolute value of the linkage strength within the effective group is not less than 0.6; and the absolute value of the difference in linkage strength between the effective and ineffective groups is not less than 0.3.

[0025] Through the above analysis, platelet-tumor hematocrit (PCT) and basophil percentage (BA%) were identified as a key indicator pair significantly correlated with therapeutic efficacy. This pair does not belong to the same biological family and lacks a proportional closure or other mathematical relationship. Specifically, this pair exhibits a strong dynamic coupling relationship within the effective group, while the coupling relationship is weaker in the ineffective group. Dynamic linkage features were constructed around this key indicator pair: the slope of BA% response to changes in PCT (denoted as dyn_slope_BA_on_PCT) and the proportion of BA% and PCT changing in the same direction (denoted as dyn_same_dir_rate_BA_PCT). Among them, dyn_slope_BA_on_PCT is calculated using the Theil-Sen estimator, which means that the efficacy prediction device calculates the slope between each pair of time points for the same patient at multiple time points, and uses the median of all slopes as the final slope value. This quantifies the directionality and intensity of the two indicators changing with treatment in the patient's body, which can filter out the interference of occasional abnormal points on trend judgment, solve the problem that a single static indicator is difficult to predict efficacy stably, and improve the trend representation stability of multi-time point hematological test data when occasional abnormal points exist (see Formula 1).

[0026]

[0027] in, For the first i , No. j PCT at each time point is a known quantity. For the first i , No. j BA% at each time point is a known quantity. This represents the local slope, and the intermediate computational cost. Ultimately, all valid values ​​will be calculated. The median is taken as dyn_slope_BA_on_PCT.

[0028] Meanwhile, dyn_same_dir_rate_BA_PCT is defined as the ratio of the number of time periods in which both PCT and BA% increase or decrease simultaneously within each valid adjacent time period to the total number of valid time periods. This feature reflects the consistency of the changes in the two indicators, and can stably distinguish between the effective and ineffective groups in small sample, multi-time-point clinical data, reducing the impact of incidental noise and extreme values.

[0029] The dynamic linkage features constructed by the efficacy prediction method can not only improve the accuracy and stability of efficacy prediction, but also have strong interpretability: they reflect the strength and consistency of the changing trends of PCT and BA%, and the two work together to characterize the dynamic response of the patient's physiological system during treatment, thus providing reliable input for the model.

[0030] Furthermore, changes in hemoglobin (HGB) across tumor types are often associated with tumor prognosis. Since patient samples are measured at multiple time points, the mean HGB is used as a comprehensive indicator of hemoglobin levels during treatment to obtain a more robust overall representation, reducing the impact of fluctuations in single measurements and differences in sampling time points on modeling. Red cell distribution width standard deviation (RDW-SD), as a comprehensive indicator reflecting red cell volume heterogeneity, characterizes pre-treatment red cell volume heterogeneity, capturing inherent individual hematological differences and reducing the reverse causal effects of treatment interventions. Compared to the red cell distribution width coefficient of variation (RDW-CV), RDW-SD describes dispersion on an absolute scale, avoiding the structural coupling introduced by mean corpuscular volume (MCV) normalization. This helps reduce derived redundancy and improve model interpretability; therefore, it is included in the model as a prognostic feature. Serum amyloid A (SAA) is an acute-phase inflammatory response protein, primarily synthesized in the liver in response to stimulation by inflammatory cytokines. In neuroblastoma, SAA may not originate from the malignant cells themselves, but rather be a secondary consequence of the invasive phase of the disease. This secondary inflammatory response can drive tumor progression by influencing the tumor immune microenvironment, and the introduction of SAA as an inflammation-related feature into the model helps to supplement the representation of changes in the inflammatory state by dynamic linkage features. Sodium ions (Na+) are essential indicators for maintaining extracellular fluid balance and electrolyte homeostasis. In the tumor microenvironment, disturbances in their concentration and distribution activate a series of molecular pathways, collectively driving the invasion, metastasis, and deterioration of solid tumors. In malignant tumors such as neuroblastoma, abnormal Na+ not only affects the tumor microenvironment but may also further influence treatment efficacy and prognosis by regulating the growth and metastatic capabilities of tumor cells. Therefore, SAA and Na+ are also included as input features in the model.

[0031] Step 3: Build a machine learning model Based on a combination of two dynamic linkage indicators and four static indicators, specifically, the six input indicators are: the slope of BA% relative to PCT (dyn_slope_BA_on_PCT), the proportion of BA% and PCT changing in the same direction (dyn_same_dir_rate_BA_PCT), hemoglobin (HGB), red blood cell distribution width standard deviation (RDW-SD), sodium ion (Na+), and serum amyloid-acid (SAA). HGB, Na+, and SAA are represented by the mean of eight pre-treatment values; RDW-SD is represented by the pre-treatment value. Based on this, a logistic regression model with L2 regularization is established to output the probability of effective treatment for the patient. (See Formulas 2 and 3). Wherein, The input features are known quantities; These are the model coefficients. The intercept is... These are known quantities determined through model training; This is the logarithmic probability value; The predicted probability of a patient responding effectively is given, and the output result is given. Finally, based on the optimal threshold determined by the Youden index, the treatment response probability is converted, and the response level information and binary classification auxiliary assessment results are output.

[0032]

[0033] ∴

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] Example 2 This embodiment provides another optional implementation method for constructing dynamic linkage features.

[0040] In Example 1, the dynamic linkage feature is constructed based on the specific indicator pair of PCT and BA%. However, the core idea of ​​this invention lies in utilizing the dynamic linkage relationship between indicators at multiple time points, rather than being limited to a specific indicator pair.

[0041] In alternative approaches, efficacy prediction devices can use a similar screening strategy to identify other target indicator pairs with significant efficacy differentiation capabilities from a broader database of complete blood count and biochemical indicators. For example, possible indicator pairs include: the ratio of absolute neutrophil count to absolute lymphocyte count, platelet count, and C-reactive protein.

[0042] For any selected target indicator pair (X, Y), the efficacy prediction device can construct dynamic linkage features in a variety of ways other than the Theil-Sen slope, including but not limited to: correlation coefficient, consistency coefficient, synergistic volatility, trend score, etc. These features can also capture the synergistic change patterns between indicators and can be used as input to the model.

[0043] Example 3 This embodiment provides another alternative implementation of the prediction model.

[0044] In Example 1, a logistic regression model with L2 regularization was ultimately adopted. The logistic regression model has the advantages of strong coefficient interpretability and relatively low overfitting on small samples.

[0045] In alternative solutions, efficacy prediction devices can utilize other machine learning models with classification capabilities, as long as they can output probability predictions or classification confidence scores. For example, a support vector machine (SVM) can be used in conjunction with a probability calibration method to output the probability of belonging to each category.

[0046] Dynamic linkage features refer to features constructed based on the direction, magnitude, consistency, or correlation of changes in at least two detection indicators at multiple time points for the same patient. Target indicator pairs refer to a combination of at least two detection indicators that satisfy a preset efficacy discrimination capability. Regarding feature selection and construction, the efficacy prediction model constructs two dynamic features, dyn_slope_BA_on_PCT and dyn_same_dir_rate_BA_PCT, around the dynamic linkage relationship between platelet-weighted hematocrit (PCT) and basophil percentage (BA%). In alternative implementations, the indicator combination used to characterize the efficacy-related dynamic structure is not limited to PCT and BA%, but can also be selected from routine blood tests, biochemical tests, or combinations thereof, to identify other indicator pairs with efficacy discrimination capabilities. Correspondingly, the construction method of dynamic features is not limited to dynamic change slope and the proportion of changes in the same direction; correlation coefficients, change magnitudes, change rates, volatility, consistency coefficients, trend scores, or other feature forms capable of characterizing the linkage relationship between indicators at multiple time points can also be used. The calculation method for the dynamic slope is not limited to using the Theil-Sen estimator; it can also use linear regression slope, robust regression slope, or other calculation methods that can reflect the direction and intensity of changes between indicators.

[0047] In terms of model implementation, a logistic regression model with L2 regularization is used as the final efficacy prediction model to output the patient's effective efficacy probability and classification results. In alternative implementations, the prediction model is not limited to logistic regression; other statistical learning models or machine learning models with classification capabilities, such as support vector machines, can also be used. Preferably, a model that can output probability results and has good interpretability is adopted.

[0048] Those skilled in the art to which this application pertains may modify or supplement the specific embodiments described or use similar methods to replace them, but without departing from the inventive concept of this application or exceeding the scope defined by the appended claims.

Claims

1. A method for predicting the efficacy of umbilical cord blood natural killer cell therapy for neuroblastoma, characterized in that, The method includes: Peripheral blood test data of neuroblastoma patients before multiple reinfusion treatment nodes of umbilical cord blood natural killer cell therapy were obtained, and the data were subjected to node alignment and standardization to form node-level peripheral blood test data. Based on the processed node-level peripheral blood detection data, at least one dynamic linkage feature reflecting the dynamic linkage relationship of indicators in the patient's body is calculated. The dynamic linkage feature is constructed based on the synergistic changes of at least two detection indicators at multiple time points in the same patient. Obtain at least one static feature indicator; The dynamic linkage features and the static feature indicators are input into a pre-trained efficacy prediction model to obtain the patient's response probability to umbilical cord blood natural killer cell immunotherapy.

2. The method according to claim 1, characterized in that, The peripheral blood test data includes complete blood count data and biochemical test data; and / or The efficacy prediction model is a logistic regression model with L2 regularization, a support vector machine, or other statistical learning or machine learning model that can output probability results.

3. The method according to claim 1 or 2, characterized in that, The at least one dynamic linkage feature is constructed based on an index pair consisting of platelet-weighted hematocrit and basophil percentage.

4. The method according to claim 3, characterized in that, The dynamic linkage feature includes: The dynamic response slope of the percentage of basophils relative to changes in platelet-weighted hematoma is denoted as dyn_slope_BA_on_PCT, and its calculation methods include Theil-Sen estimator, linear regression slope, or robust regression slope. And / or, The proportion of plateletcrit and basophil percentage changes in the same direction in adjacent time periods is denoted as dyn_same_dir_rate_BA_PCT.

5. The method according to claim 4, characterized in that, The dyn_slope_BA_on_PCT is calculated using the Theil-Sen estimator, including: For multiple time points of the same patient, calculate the local slope of the platelet hematocrit and the percentage of basophils between any two time points; The median of all valid local slopes is taken as dyn_slope_BA_on_PCT.

6. The method according to claim 1, characterized in that, The static characteristic indicators include at least one of the following: Mean hemoglobin level; Red blood cell distribution width standard deviation; Serum amyloid protein; Sodium ion concentration.

7. The method according to claim 1, characterized in that, The acquisition of peripheral blood test data for neuroblastoma patients prior to multiple reinfusion therapy points includes: Obtain clinical peripheral blood test data for patients before eight reinfusion treatment nodes.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

10. A device for predicting the efficacy of umbilical cord blood natural killer cell immunotherapy for neuroblastoma, characterized in that, The device includes: The data processing module is used to acquire peripheral blood test data of neuroblastoma patients before multiple reinfusion treatment nodes of umbilical cord blood natural killer cell therapy, and to perform node alignment and standardization processing on the data. The feature construction module is used to calculate at least one dynamic linkage feature reflecting the dynamic linkage relationship of indicators in the patient's body based on the processed data, and to obtain at least one static feature indicator. The prediction module is used to input the dynamic linkage features and the static feature indicators into a pre-trained efficacy prediction model to output the patient's response probability to treatment.

Images