A method for constructing a non-small cell lung cancer histological typing multi-modal prediction model

Abstract

The application discloses a non-small cell lung cancer histological typing multi-modal prediction model construction method, and belongs to the field of non-small cell lung cancer diagnosis models. The method comprises the following steps: obtaining multi-modal data of training samples, wherein the multi-modal data comprises clinical features, laboratory features, short-term resting electrocardio signals and histological typing labels, the histological typing labels comprise lung adenocarcinoma labels and lung squamous carcinoma labels; processing the short-term resting electrocardio signals to obtain RR interval sequences; calculating a heart rate variability feature pool based on the RR interval sequences, wherein the heart rate variability feature pool comprises traditional heart rate variability features and nonlinear entropy features, the nonlinear entropy features comprise grid distribution entropy, multi-scale cumulative residual scatter entropy, attention entropy, improved multi-scale Renyi distribution entropy and multi-scale fuzzy scatter entropy; performing feature screening on the heart rate variability feature pool, and constructing a heart rate variability comprehensive score based on the screened core heart rate variability features; fusing the heart rate variability comprehensive score with the clinical features and the laboratory features, training a binary classification prediction model with the histological typing labels as supervision labels, and obtaining the non-small cell lung cancer histological typing multi-modal prediction model.

Description

Technical Field

[0001] This invention relates to the field of diagnostic models for non-small cell lung cancer, specifically to a method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer. Background Technology

[0002] Non-small cell lung cancer (NSCLC) is the most common pathological type of lung cancer, mainly including histological subtypes such as lung adenocarcinoma and lung squamous cell carcinoma. Different histological subtypes differ in molecular characteristics, treatment options, prognosis, and follow-up management; therefore, accurate histological classification of NSCLC is of great significance.

[0003] Currently, the histological classification of non-small cell lung cancer mainly relies on pathological histological examination and immunohistochemical detection. However, in actual clinical practice, some patients may experience insufficient pathological samples, atypical immunohistochemical results, or difficulty in obtaining a clear classification through routine pathology due to factors such as sampling site, biopsy tissue volume, tumor heterogeneity, sample necrosis, or tissue preservation quality. In such cases, relying solely on pathological morphology information may not be sufficient to reliably obtain histological classification results. Therefore, it is necessary to introduce more quantifiable auxiliary information to construct a data processing model that can be used for histological classification prediction.

[0004] Existing lung cancer prediction models are mostly built based on radiomics, serum tumor markers, clinical data, or gene testing data. Among these, clinical and laboratory indicators such as serum carcinoembryonic antigen (CEA), age, sex, and smoking history are readily available, but their information dimensions are limited, making it difficult to fully characterize the impact of tumor burden on the patient's overall physiological regulatory state. On the other hand, heart rate variability (HRV) can be obtained from short-term resting-state electrocardiogram (ECG) signals and is used to reflect the state of autonomic nervous system regulation. Previous studies have suggested a correlation between HRV characteristics and prognosis, inflammatory status, or changes in autonomic nervous function in cancer patients; however, current technologies mainly focus on survival prognosis assessment or risk stratification, and have not fully utilized HRV characteristics to build multimodal prediction models for histological subtyping of non-small cell lung cancer.

[0005] Furthermore, traditional heart rate variability analysis typically focuses on linear indicators in the time and frequency domains, such as SDNN, RMSSD, low-frequency power, and high-frequency power. While these indicators can reflect some characteristics of autonomic nervous activity, they are insufficient for characterizing the nonlinear, non-stationary, and complex changes across time scales present in short-term electrocardiogram signals. Especially under conditions such as tumor-related chronic stress and autonomic nervous system regulatory imbalance, intercardiac interval sequences may exhibit characteristics such as decreased complexity, altered state spatial distribution, and concentration of information in key temporal segments. Relying solely on traditional linear indicators is insufficient to fully extract the aforementioned information.

[0006] Therefore, there is an urgent need for a novel multimodal prediction model for non-small cell lung cancer histological subtyping. This model should incorporate heart rate variability features from short-term resting-state electrocardiogram (ECG) signals, particularly nonlinear entropy features that characterize the complexity of autonomic nervous system regulation, based on clinical data and laboratory indicators. Through feature selection and fusion modeling, a model applicable to the classification and prediction of lung adenocarcinoma and squamous cell carcinoma can be formed. This method would provide a data processing-level auxiliary subtyping tool for scenarios where pathological samples are insufficient or conventional pathological subtyping is difficult. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer.

[0008] The objective of this invention can be achieved through the following technical solutions: A method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer includes the following steps: Acquire multimodal data of training samples, including clinical features, laboratory features, short-term resting-state electrocardiogram signals, and histological classification labels, including lung adenocarcinoma labels and lung squamous cell carcinoma labels; The short-term resting-state electrocardiogram signal is processed to obtain the RR interval sequence; The heart rate variability feature pool is calculated based on the RR interval sequence. The heart rate variability feature pool includes traditional heart rate variability features and nonlinear entropy features. The nonlinear entropy features include grid distribution entropy, multi-scale cumulative residual scattering entropy, attention entropy, improved multi-scale Renyi distribution entropy, and multi-scale fuzzy scattering entropy. Feature filtering is performed on the heart rate variability feature pool, and a comprehensive heart rate variability score is constructed based on the core heart rate variability features obtained from the filtering. The heart rate variability score is fused with the clinical and laboratory characteristics to train a binary classification prediction model with the histological subtyping label as the supervision label, thus obtaining the non-small cell lung cancer histological subtyping multimodal prediction model.

[0009] More specifically: (1) Collect clinical data, laboratory indicators, electrocardiogram data and pathological labels of non-small cell lung cancer patients as training datasets. The pathological labels are lung adenocarcinoma or lung squamous cell carcinoma. The clinical data include age, gender and smoking history. The laboratory indicators include serum carcinoembryonic antigen (CEA) concentration. The electrocardiogram data are short-term resting state electrocardiogram signals, generally lasting 3-5 minutes. (2) The electrocardiogram signal is preprocessed and R-wave is detected. The RR interval sequence is extracted and outlier detection and correction are performed. The corrected RR interval sequence is interpolated and resampled. Multidimensional HRV parameters are calculated from the perspectives of time domain, frequency domain and nonlinear dynamics to construct an HRV feature pool. The HRV feature pool covers traditional HRV indicators and cutting-edge nonlinear entropy indicators. The traditional HRV indicators include, but are not limited to, the standard deviation of normal interval (SDNN), the root mean square of the difference between adjacent normal intervals (RMSSD), cardiac acceleration capacity (AC), cardiac deceleration capacity (DC), low-frequency power (LF), high-frequency power (HF) and Poincaré scatter plot quantization parameters. The cutting-edge nonlinear entropy indicators include grid distribution entropy, multi-scale cumulative residual scatter entropy (MCRDE), attention entropy, improved multi-scale Renyi distribution entropy (MMRDis) and multi-scale fuzzy scatter entropy (MFDE), which are used to comprehensively quantify the dynamic complexity decay of the autonomic nervous system state from multiple dimensions such as probability distribution morphology, cross-time scale complexity, key subsequence significance, generalized information measurement and fuzzy logic robust classification. (3) An embedded feature selection algorithm is used to select the core feature subset most relevant to the pathological label from the generated multidimensional HRV feature pool, and the HRV-score is constructed by weighted summation based on its coefficient in the logistic regression model.

[0010] (4) Construct a histological pathological classification (adenocarcinoma vs. squamous cell carcinoma) prediction model characterized by clinical information (age, gender, smoking history), serum carcinoembryonic antigen (CEA) concentration, and HRV-score, and transform it into a visual nomogram decision tool.

[0011] Optionally, the laboratory indicators mentioned in step (1) need to be preprocessed by logarithmic transformation (log(CEA)) to correct the skewness of the concentration distribution.

[0012] Optionally, the embedded feature selection algorithm in step (3) is LASSO regression, or other wrapper feature selection methods may be used.

[0013] Optionally, the formula for calculating the HRV-score is: HRV-score =-38.111 - 0.015×RMSSD + 0.234×AC + 0.511×HF + 0.456×DET - 0.698×Gridded Distribution Entropy - 8.887×Attention Entropy +5.184MCRDE -0.223MMRDis + 0.752MFDE.

[0014] Optionally, the specific definitions of the five frontier nonlinear entropy indices mentioned in step (2) are as follows: (1) Grid distribution entropy: By mapping the vector after the RR interval sequence is reconstructed in phase space to the preset gridded distribution space, the distribution probability of data points in each grid is statistically analyzed and Shannon entropy is calculated to quantify the uniformity of the probability distribution of autonomic neural regulation in the state space. The lower the value, the more the system gathers to a few abnormal states, reflecting the loss of regulatory flexibility. (2) Multiscale cumulative residual spread entropy (MCRDE): Based on multiscale coarse-grained calculation, spread entropy is calculated and the entropy values ​​of each scale are accumulated and summed to simultaneously capture the cross-scale dynamic complexity from second-level respiratory sinus arrhythmia to minute-level body temperature and humoral regulation rhythm. As a comprehensive quantitative indicator of the global regulatory adaptability of the autonomic nervous system, the chronic sympathetic hyperactivity and vagal withdrawal caused by tumor burden cause the entropy values ​​of multiple scales to decrease synchronously. MCRDE is extremely sensitive to this cross-scale rigidity. (3) Attention entropy: Drawing on the attention mechanism in deep learning, the entropy value is calculated by analyzing the distribution concentration and significance of key subsequences or feature segments in the RR interval sequence. Instead of treating all sequence segments equally, it automatically identifies and highlights key time periods with high information density, revealing deep temporal correlation patterns that traditional entropy indicators cannot capture. The higher the value, the more dispersed the information distribution and the weaker the regularity of the regulation pattern. (4) Improved multi-scale Renyi distribution entropy (MMRDis): Based on the moving average multi-scale decomposition, Renyi entropy is introduced to replace the traditional Shannon entropy. The sensitivity to rare events is controlled by parameter q, providing a generalized multi-scale complexity measure, and completely avoiding the "undefined" output problem that may occur in the traditional entropy in short-time RR interval sequences from the root of the algorithm. (5) Multiscale fuzzy spread entropy (MFDE): Based on multiscale spread entropy, a fuzzy membership function is introduced to replace the hard category division with a smooth membership degree, making the transition boundary between adjacent states more continuous and natural, significantly improving the robustness of entropy value to physiological signal noise, maintaining a stable complexity estimate within a wide parameter range, and enhancing the ability to distinguish between different autonomic nervous function states.

[0015] Optionally, the prediction model in step (4) is a logistic regression model, or other traditional statistical models or machine learning models may be used.

[0016] The aforementioned visual nomogram decision-making tool is characterized by: (1) Variable scoring scale: The top of the nomogram is the total score scale (0-100 points). Below it are rows of predictor variables, and each row of variable values ​​has a clear mapping relationship with the top score scale, including age: continuous variable (e.g., 25-85 years old), each age corresponds to a specific score; gender: categorical variable (male / female), each category corresponds to a specific score; smoking history: categorical variable (yes / no), each category corresponds to a specific score; log(CEA): continuous variable (e.g., 0-7.5 ng / mL after logarithmic transformation), each value corresponds to a specific score; HRV-score.

[0017] (2) Total score calculation and probability mapping: Based on the specific values ​​of the five variables of the patient, find the corresponding individual scores on the nomogram; add up all the individual scores to get the patient's total score (corresponding to the total score scale on the nomogram, for example, the range of 0-260 points); project the total score vertically to the lung squamous cell carcinoma prediction probability scale at the bottom (for example, the range of 0.1%-90%), and you can directly read the individualized prediction probability of the patient being diagnosed with lung squamous cell carcinoma.

[0018] (3) Clinical threshold determination: The nomogram provides exemplary decision thresholds. For example, when a patient's total score is ≥ 200, the probability of diagnosis of squamous cell carcinoma of the lung is > 60%; when the total score is > 220, the probability of diagnosis is > 90%. Clinicians can combine this probability with the actual situation to formulate subsequent diagnosis or treatment strategies. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a nomogram of lung squamous cell carcinoma probabilities predicted based on a joint model.

[0021] Figure 2 This is a calibration curve for the combined prediction model of squamous cell carcinoma of the lung.

[0022] Figure 3 This is a decision curve analysis diagram for a lung squamous cell carcinoma prediction model.

[0023] Figure 4 The plot shows the receiver operating characteristic curves for Model 1 through Model 4 in the test set. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] In some embodiments of the present invention, a method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer is disclosed, specifically including the following steps: Step 1: Training Data Collection Clinical data, laboratory indicators, heart rate RR interval data, and histopathological classification of patients with non-small cell lung cancer were obtained as training set data. To ensure that the baseline status of biomarkers was not affected by intervention, patients had no history of other malignant tumors and had not received any form of anti-tumor treatment (including but not limited to surgery, radiotherapy, chemotherapy, targeted therapy, or immunotherapy). To exclude severe interference with heart rate variability (HRV) signals, patients with clinically diagnosed coronary artery disease, post-coronary artery stent implantation status, or severe arrhythmias (such as atrial fibrillation or frequent premature ventricular contractions) were excluded. In addition, patients with ectopic heartbeats exceeding 10%, diagnosed with poorly differentiated lung cancer, or lacking key clinical data were also excluded.

[0026] Clinical data included age, sex, body mass index (BMI), Eastern Cooperative Oncology Group Performance Status Scale (ECOG PS), pain level assessed using a 4-point verbal rating scale (VRS), smoking history (defined as yes / no), history of hypertension, history of diabetes, presence of pneumonia, number of organ metastases, and status of visceral metastases versus non-visceral metastases.

[0027] All patients were diagnosed with primary lung adenocarcinoma (LUAD) or lung squamous cell carcinoma (LUSC) by histopathology and immunohistochemistry (IHC).

[0028] Serum tumor marker detection and processing: Carcinoembryonic antigen (CEA) was used as the core laboratory variable. Approximately 3 mL of EDTA-K2 anticoagulated peripheral venous blood was collected from patients before treatment. Complete blood count (CBC) was performed using a Sysmex XE-2100 fully automated hematology analyzer. CEA concentration was quantitatively detected using an Abbott I2000 fully automated chemiluminescence immunoassay analyzer via chemiluminescent microparticle immunoassay. All reagents, calibrators, and quality control materials were sourced directly from the original manufacturer. The testing process was performed by professionally trained laboratory personnel and kept blinded from clinical information. To correct for skewed distribution, the raw CEA concentration values ​​were transformed using the natural logarithm to generate the log(CEA) variable for subsequent analysis.

[0029] Heartbeat RR interval signal acquisition: A 5-minute resting-state electrocardiogram (ECG) signal was acquired for each patient using a single-lead ECG recorder. ECG data generation conditions included: sampling rate 400 Hz, single lead, bandpass frequency 0.5–44 Hz, and R-wave detection using BioSPPy. Before acquisition, it was confirmed that the patient had not smoked, consumed alcohol, ingested caffeine, or engaged in strenuous exercise within 30 minutes. During acquisition, the patient was placed in a supine position, maintaining quietness and stable breathing. The raw ECG signal was preprocessed by bandpass filtering and baseline correction before R-wave detection to obtain the heartbeat RR interval time series. A time-varying threshold automatic correction algorithm was used to clean the signal to remove potential noise and ectopic beats.

[0030] Step 2: Heart Rate Variability Analysis A thorough analysis was conducted on the corrected RR interval sequence obtained in step 1 to quantify the state of autonomic nervous function from multiple perspectives: (1) Signal preprocessing: First, the non-uniformly sampled RR interval sequence is resampled using cubic spline interpolation to generate a uniform time sequence of 4 Hz to meet the requirements of frequency domain analysis.

[0031] (2) Calculation of time-domain features: The calculation includes traditional indicators such as mean heart rate (Mean HR), standard deviation of all sinus intervals (SDNN), and root mean square of the difference between adjacent sinus RR intervals (RMSSD). In particular, to assess the heart's ability to accelerate and decelerate in response to instantaneous changes, cardiac deceleration capacity (DC) and cardiac acceleration capacity (AC) were calculated.

[0032] (3) Frequency domain feature calculation: Fast Fourier Transform (FFT) was applied to the homogenized RR interval sequence for spectral estimation. The absolute power values ​​of very low frequency power (VLF: 0–0.04 Hz), low frequency power (LF: 0.04–0.15 Hz), and high frequency power (HF: 0.15–0.4 Hz), as well as the LF / HF ratio, were calculated. Among them, LF and HF power were used as indirect quantitative indicators of sympathetic and vagal nerve activity, respectively.

[0033] (4) Nonlinear feature calculation: In addition to obtaining graphical quantitative parameters such as SD1, SD2, SI, and PI from the Poincaré scatter plot, this step focuses on integrating the following five cutting-edge nonlinear entropy indices to characterize the complexity decay of autonomic neural regulation from multiple dimensions: (a) Grid distribution entropy: First, the phase space of the RR interval sequence is reconstructed (embedding dimension m=3, delay time τ=1). The reconstructed vectors are mapped to a predefined gridded distribution space, and the distribution probability of data points within each grid is statistically analyzed. Then, the Shannon entropy is calculated. Grid distribution entropy quantifies the uniformity of the distribution of the autonomic nervous system in the state space. When the autonomic nervous system has good flexibility, the system state is relatively uniformly distributed in the phase space, and the entropy value is high. When tumor burden induces sympathetic hyperactivity and vagal withdrawal leads to a rigidity in the regulatory pattern, the system state clusters in a few regions, and the entropy value decreases accordingly.

[0034] (b) Multiscale Cumulative Residual Spread Entropy (MCRDE): First, the RR interval sequence is coarsely processed across multiple scales (scale factor τ = 1~10). Spread entropy is calculated at each scale, and finally, the entropy values ​​at each scale are accumulated and summed. Spread entropy maps signal values ​​to a finite number of categories through a normal cumulative distribution function, and then calculates the diversity of the arrangement patterns of the category sequences, which has the advantages of fast computation speed and robustness to noise. On this basis, MCRDE adds multiscale analysis capabilities, which can simultaneously capture the cross-scale dynamic complexity from second-level respiratory sinus arrhythmia to ten-second to minute-level body temperature and humoral regulation rhythms. Tumor burden-induced chronic sympathetic hyperactivity and vagal retreat cause the signal entropy values ​​at multiple time scales to decrease synchronously. MCRDE is extremely sensitive to this "cross-scale rigidity" and can serve as a comprehensive quantitative indicator of the decline in the global regulatory adaptive capacity of the autonomic nervous system.

[0035] (c) Attention Entropy: Drawing inspiration from the attention mechanism in deep learning architectures such as Transformer, entropy is defined by calculating the distribution of attention weights among different subsequences or feature segments in the RR interval sequence. Specifically, the RR interval sequence is divided into several fixed-length time windows, and the self-attention weight matrix between each time window is calculated. The entropy value is then calculated based on the probability distribution of this weight matrix. Unlike traditional entropy, attention entropy can automatically identify and highlight key periods with high information density, revealing deep temporal correlation patterns that traditional entropy indicators cannot capture. A higher value indicates a more dispersed distribution of key information along the time axis and a weaker regularity in the autonomic neural regulation pattern.

[0036] (d) Improved Multi-Scale Renyi Distribution Entropy (MMRDis): Based on the moving average multi-scale decomposition (scale factor τ = 1~10), Renyi entropy is introduced to replace the traditional Shannon entropy in calculating the distribution entropy. Renyi entropy controls the sensitivity to rare events in the probability distribution through the parameter q—highlighting rare patterns when q < 1 and frequent patterns when q > 1, thus providing a more generalized perspective on complexity measurement than Shannon entropy. More importantly, MMRDis completely avoids the "undefined" output problem that may occur in traditional entropy in short RR interval sequences (3-5 minutes) from the algorithmic root, ensuring that effective and stable complexity estimates can be output for each patient's HRV data, which is particularly suitable for the short ECG data scenario collected in this invention.

[0037] (e) Multiscale Fuzzy Spread Entropy (MFDE): Building upon traditional multiscale spread entropy, fuzzy membership functions replace hard category divisions. Specifically, when mapping signal values ​​to categories, instead of using a step function for an "either / or" division, a Gaussian fuzzy membership function is employed, allowing each signal value to be assigned to multiple adjacent categories with different membership degrees, thus achieving a smooth transition between state boundaries. This improvement significantly enhances the robustness of entropy values ​​to physiological signal noise (such as baseline drift residue and electromyographic interference), maintains stable complexity estimation across a wide range of embedding dimensions and number of categories, and strengthens the distinction between different autonomic nervous system states.

[0038] (5) Feature pool construction: All HRV parameters calculated above—including time domain indicators (SDNN, RMSSD, AC, DC, etc.), frequency domain indicators (LF, HF, LF / HF, etc.), traditional nonlinear indicators (SD1, SD2, SI, PI) and five frontier entropy indicators (grid distribution entropy, MCRDE, attention entropy, MMRDis, MFDE)—are pooled into a multidimensional HRV feature pool for subsequent feature selection and model construction.

[0039] Step 3: Feature standardization, selection, and HRV-score construction This step aims to screen the core biomarkers most relevant to the histological classification of NSCLC from the high-dimensional HRV feature pool and construct an integrated HRV comprehensive score.

[0040] (1) Feature standardization: Z-score standardization (subtracting the mean and dividing by the standard deviation) is performed on all continuous variables in the multidimensional HRV feature pool to eliminate the incomparability caused by the difference in units between different indicators and to ensure the stability of subsequent feature selection and coefficient estimation. In this embodiment, the Z-score is calculated as follows: Z=(X-μ) / σ, where X is the original data, and μ and σ are the mean and standard deviation of the corresponding features in the training set, respectively.

[0041] (2) Feature Selection: LASSO (Minimum Absolute Shrinkage and Selection Operator) regression is used for feature selection on the training set data. LASSO automatically selects features by applying an L1 regularization term to the loss function, which can precisely compress the regression coefficients of a large number of irrelevant or redundant features to zero during model training. The optimal regularization parameter λ is selected through 10-fold cross-validation, and the features with non-zero coefficients are finally determined to constitute the core feature subset. In this embodiment, after LASSO selection, several of the aforementioned five frontier entropy indicators, together with traditional indicators (such as cardiac acceleration capacity AC, high-frequency power HF, etc.), are retained as the core autonomic neural features with the most predictive value.

[0042] (3) HRV-score construction: The selected core HRV features are weighted and summed according to their standardized regression coefficients obtained in the logistic regression model with histological classification (LUAD vs. LUSC) as the dependent variable.

[0043] Step 4: Multimodal adaptive fusion modeling The HRV-score is fused with the key clinical variables (age, sex, smoking history) and laboratory variables (log(CEA)) from step 3 after preprocessing to construct the final classification prediction model.

[0044] (1) Feature vector construction: The feature vector of each sample consists of six variables: age, gender (coded as 0 / 1), smoking history (coded as 0 / 1), log(CEA), HRV-score (or its core components AC and HF, depending on the model design), and HRV-score (a continuous variable, calculated in step 3).

[0045] (2) Model training: The multivariate logistic regression model is preferred for training because its coefficients have clear clinical interpretability and are easy to convert into nomograms. The model uses LUAD / LUSC as the binary dependent variable.

[0046] (3) Explanation of model scalability: The multimodal fusion framework described in this invention is not limited to logistic regression and can also be adapted to other machine learning algorithms. For example, when it is necessary to capture high-order nonlinear interactions between features, ensemble learning algorithms such as random forest or XGBoost can be used for training; when the data scale is large and deeper feature interaction learning is required, deep learning methods such as Transformer or Mamba can be introduced, and HRV-score can be used as a guiding signal to achieve adaptive fusion of multimodal data and higher-dimensional feature interaction learning.

[0047] Step 5: Building and Visualizing Nomogram Tools Based on the optimal logistic regression model obtained in step 4, a visual clinical decision-making tool—the nomogram—is constructed.

[0048] (1) Graphical structure generation: In the R language environment, use the nomogram function in the rms package to generate a nodal plot (corresponding to the appendix). Figure 1 The lon plot, from top to bottom, includes: Total Points: the top bar, ranging from 0 to 100 points; Variable rows: sequentially listing each predictor variable and its corresponding score mapping, including: age (continuous variable, example range 25-85 years), gender (categorical variable: male / female), smoking history (categorical variable: yes / no), log(CEA) (continuous variable, example range 0-7.5), and HRV-score; Total Points: located below the variable rows, with an example range of 0-260 points; Risk of LUSC: the bottom bar, displaying the corresponding predicted probability of squamous cell lung cancer (example range 0.1% - 90%). (2) Usage and Interpretation Methods: Individual Score: Based on the specific values ​​of each patient's variables, locate the corresponding variable row and project it vertically upwards to the top "Points" scale to read the individual score; Total Score Calculation: Add the individual scores of all variables to obtain the patient's total score; Probability Mapping: Project the total score vertically to the bottom "Risk of LUSC" probability scale to directly read the individualized predicted probability of the patient being diagnosed with squamous cell carcinoma of the lung; Decision Threshold Example: The nomogram provides reference clinical thresholds. For example, when the patient's total score is ≥ 200, the probability of a LUSC diagnosis is >60%; when the total score is >220, the probability of a diagnosis is >90%. Clinicians can use this probability to assist in decision-making.

[0049] Step 6: Model Calibration and Clinical Applicability Assessment A comprehensive performance evaluation was conducted on the final model and nomogram to ensure that they have good calibration accuracy and clinical applicability.

[0050] (1) Discrimination assessment: The ability of the model to distinguish between lung adenocarcinoma and lung squamous cell carcinoma was assessed by the area under the receiver operating characteristic curve (AUC-ROC). At the same time, the C-statistic (C-index) was calculated as a comprehensive measure of discrimination.

[0051] (2) Calibration evaluation: Internal verification was performed using the Bootstrap resampling method (1000 resamplings), and the calibration curve was plotted using the calibrate function (see attached diagram). Figure 2The calibration curve is plotted with the model-predicted probability on the horizontal axis and the actual observed incidence of squamous cell carcinoma of the lung on the vertical axis. Ideally, the curve should coincide with the 45° diagonal. A Hosmer-Lemeshow goodness-of-fit test was also performed, with P > 0.05 indicating no significant deviation between the model predictions and actual values, and thus good calibration.

[0052] (3) Clinical Decision Curve Analysis (DCA): Plotting decision curves (corresponding to the attached diagram). Figure 3 The model's clinical net benefit was evaluated at different threshold probabilities. The horizontal axis of the curve represents the threshold probability, and the vertical axis represents the net benefit. The decision curve of this model was compared with two extreme strategies—"treating all patients as lung adenocarcinoma" and "treating all patients as lung squamous cell carcinoma." If the net benefit curve of this model is higher than that of the two extreme strategies over a wide range of threshold probabilities, it demonstrates that the model has good clinical applicability and added value, and can bring substantial information gain to clinical decision-making.

[0053] like Figure 4 Tables 1 and 2 show that, to evaluate the predictive ability of different feature combinations for histological subtyping of non-small cell lung cancer, four candidate models—Model 1, Model 2, Model 3, and Model 4—were constructed based on the aforementioned training samples. Model 1 was the clinical baseline model, incorporating only clinical features, including age, sex, and smoking history. Model 2 was a combined model incorporating clinical features and carcinoembryonic antigen (CEA). Model 3 was a combined model incorporating clinical features and heart rate variability (based on the aforementioned HRV-score). Model 4 was a multimodal combined model incorporating clinical features, CEA, and the aforementioned HRV-score. All models used the standard binary logistic regression algorithm. Lung squamous cell carcinoma was used as the positive category, and lung adenocarcinoma as the negative category. The area under the receiver operating characteristic (AUC), accuracy, sensitivity, and specificity of each model were calculated on both the training and test sets. The DeLong test was used to compare the AUC differences between Model 2-Model 4 and Model 1. The ROC curves for the test set are shown below. Figure 4 As shown, the specific evaluation results are shown in Tables 1 and 2.

[0054] In the training set, Model 1 had an AUC of 0.745, with a 95% confidence interval of 0.638–0.852, and accuracy, sensitivity, and specificity of 0.650, 0.950, and 0.500, respectively. Model 2 had an AUC of 0.953, with a 95% confidence interval of 0.900–1.000, and accuracy, sensitivity, and specificity of 0.917, 0.900, and 0.925, respectively; its AUC relative to Model 1 had a DeLong test p-value of 0.002. Model 3 had an AUC of 0.897, with a 95% confidence interval of 0.821–0.974, and accuracy, sensitivity, and specificity of 0.800, 0.950, and 0.725, respectively; its AUC relative to Model 1 had a DeLong test p-value less than 0.001. Model 4 has an AUC of 0.963 with a 95% confidence interval of 0.922–1.000, and accuracy, sensitivity, and specificity of 0.900, 0.950, and 0.875, respectively. Its AUC relative to Model 1 has a DeLong test p-value of less than 0.001. These results indicate that in the training data of this embodiment, Models 2–4 all have higher discrimination than Model 1, with Model 4 achieving the highest training set AUC.

[0055] In the test set, according to the results listed in Table 2, Model 1 had an AUC of 0.700, a 95% confidence interval of 0.418–0.841, and accuracy, sensitivity, and specificity of 0.667, 0.800, and 0.600, respectively. Model 2 had an AUC of 0.840, a 95% confidence interval of 0.632–1.000, and accuracy, sensitivity, and specificity of 0.800; its AUC relative to Model 1 had a DeLong test p-value of 0.247. Model 3 had an AUC of 0.800, a 95% confidence interval of 0.564–1.000, and accuracy, sensitivity, and specificity of 0.733, 0.600, and 0.800, respectively; its AUC relative to Model 1 had a DeLong test p-value of 0.219. Model 4 had an AUC of 0.880, a 95% confidence interval of 0.697–1.000, and accuracy, sensitivity, and specificity of 0.800. Its AUC relative to Model 1 showed a DeLong test p-value of 0.135. Model 4 achieved the highest AUC in the test data of this embodiment while maintaining a relative balance between sensitivity and specificity, indicating that the model has a certain ability to distinguish between lung adenocarcinoma and lung squamous cell carcinoma.

[0056] Table 1. Comparison of classification performance of Model 1–Model 4 on the training set. Table 2 Comparison of classification performance of Models 1-4 on the test set In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer, characterized in that, Includes the following steps: Acquire multimodal data of training samples, including clinical features, laboratory features, short-term resting-state electrocardiogram signals, and histological classification labels, including lung adenocarcinoma labels and lung squamous cell carcinoma labels; The short-term resting-state electrocardiogram signal is processed to obtain the RR interval sequence; The heart rate variability feature pool is calculated based on the RR interval sequence. The heart rate variability feature pool includes traditional heart rate variability features and nonlinear entropy features. The nonlinear entropy features include grid distribution entropy, multi-scale cumulative residual scattering entropy, attention entropy, improved multi-scale Renyi distribution entropy, and multi-scale fuzzy scattering entropy. Feature filtering is performed on the heart rate variability feature pool, and a comprehensive heart rate variability score is constructed based on the core heart rate variability features obtained from the filtering. The heart rate variability score is fused with the clinical and laboratory characteristics to train a binary classification prediction model with the histological subtyping label as the supervision label, thus obtaining the non-small cell lung cancer histological subtyping multimodal prediction model.

2. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, The clinical characteristics include age, sex, and smoking history, and the laboratory characteristics include serum carcinoembryonic antigen (CEA) concentration, and the serum CEA concentration is logarithmically transformed to obtain the log(CEA) variable.

3. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, The short-term resting-state electrocardiogram signal is processed to obtain the RR interval sequence, including: Bandpass filtering and baseline correction were performed on the short-time resting-state electrocardiogram signal; R-wave detection was performed on short-term resting-state ECG signals after bandpass filtering and baseline correction to obtain the initial RR interval sequence; Outlier detection and correction are performed on the initial RR interval sequence, and the corrected RR interval sequence is then resampled by interpolation.

4. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, The traditional heart rate variability features include time-domain features, frequency-domain features, and Poincaré scatter plot quantization features. The time-domain features include SDNN, RMSSD, cardiac acceleration capacity, and cardiac deceleration capacity. The frequency-domain features include low-frequency power, high-frequency power, and the ratio of low-frequency power to high-frequency power. The Poincaré scatter plot quantization features include SD1, SD2, stress index, and Poincaré index.

5. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, Calculating the nonlinear entropy value features includes: The phase space of the RR interval sequence is reconstructed, the reconstructed vector is mapped to a gridded distribution space, and the grid distribution entropy is calculated based on the distribution probability of data points in each grid. The RR interval sequence is subjected to multi-scale coarse-grained processing. The spread entropy is calculated at each scale, and the spread entropy at each scale is accumulated and summed to obtain the multi-scale cumulative residual spread entropy. The RR interval sequence is divided into multiple time windows, the self-attention weight matrix between the multiple time windows is calculated, and the attention entropy is calculated based on the probability distribution of the self-attention weight matrix. The improved multi-scale Renyi distribution entropy is obtained by calculating the distribution entropy using Renyi entropy based on the moving average multi-scale decomposition. In the calculation of multi-scale scatter entropy, a fuzzy membership function is used to classify the signal values ​​to obtain the multi-scale fuzzy scatter entropy.

6. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, The heart rate variability feature pool is subjected to feature filtering, and a comprehensive heart rate variability score is constructed based on the core heart rate variability features obtained from the filtering, including: Z-score standardization is performed on the continuous variables in the heart rate variability feature pool. An embedded feature selection algorithm was used to screen core heart rate variability features related to the histological typing label from the Z-score-normalized heart rate variability feature pool; Based on the regression coefficients of the core heart rate variability features in the logistic regression model, the core heart rate variability features are weighted and summed to obtain the comprehensive heart rate variability score.

7. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 6, characterized in that, The embedded feature selection algorithm is LASSO regression. The LASSO regression determines the regularization parameter through cross-validation and identifies the heart rate variability features with non-zero regression coefficients as the core heart rate variability features.

8. The method for constructing a multimodal prediction model for histological subtyping of non-small cell lung cancer according to claim 1, characterized in that, It also includes: generating a nomogram based on a trained binary classification prediction model, wherein the nomogram includes a variable score scale, a total score scale, and a category probability scale corresponding to the clinical features, the laboratory features, and the comprehensive heart rate variability score.

9. A device for constructing a multimodal predictive model for histological subtyping of non-small cell lung cancer, characterized in that, include: The data acquisition module is used to acquire multimodal data of training samples. The multimodal data includes clinical features, laboratory features, short-term resting-state electrocardiogram signals, and histological classification labels, including lung adenocarcinoma labels and lung squamous cell carcinoma labels. An electrocardiogram (ECG) processing module is used to process the short-term resting-state ECG signal to obtain an RR interval sequence; The feature extraction module is used to calculate the heart rate variability feature pool based on the RR interval sequence. The heart rate variability feature pool includes traditional heart rate variability features and nonlinear entropy features. The nonlinear entropy features include grid distribution entropy, multi-scale cumulative residual scattering entropy, attention entropy, improved multi-scale Renyi distribution entropy, and multi-scale fuzzy scattering entropy. The scoring construction module is used to filter features from the heart rate variability feature pool and construct a comprehensive heart rate variability score based on the core heart rate variability features obtained from the filtering. The model training module is used to fuse the heart rate variability comprehensive score with the clinical features and the laboratory features, train a binary classification prediction model with the histological subtyping label as the supervision label, and obtain the non-small cell lung cancer histological subtyping multimodal prediction model.

10. 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 steps of the method according to any one of claims 1 to 8.

Images