Neoadjuvant chemotherapy sensitivity prediction system for middle and advanced laryngeal cancer

Abstract

The invention relates to the technical field of laryngeal cancer treatment, in particular to a middle-advanced laryngeal cancer neoadjuvant chemotherapy sensitivity prediction system which comprises a multi-modal data acquisition and preprocessing module, a dynamic feature fusion engine module, a chemotherapy response dynamic prediction core module, an interpretable output interface module and a system verification closed loop module. Through multi-modal dynamic fusion (clinical / image / molecular data cross-modal attention allocation), a time sequence LSTM-Bayesian correction prediction model, an SHAP-t-SNE interpretable decision space and an incremental learning closed loop verification system, precise prediction of laryngeal cancer chemosensitivity is realized, and clinical response evaluation precision and intervention targeting are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of laryngeal cancer treatment, and in particular to a system for predicting the sensitivity of mid- to late-stage laryngeal cancer to neoadjuvant chemotherapy. Background Art

[0002] Neoadjuvant chemotherapy for mid- to late-stage laryngeal cancer is a core means of downstaging treatment, and its efficacy is highly dependent on the individual patient's chemotherapy sensitivity. However, the current prediction methods in clinical practice face multiple bottlenecks, resulting in significant uncertainty and lag in treatment decisions. Traditional prediction systems mainly rely on static clinical pathological parameters (such as TNM staging, histological grade) and imaging assessments (such as RECIST 1.1 standards). Although these methods can provide basic stratification, they are difficult to capture the dynamic evolution and molecular heterogeneity of tumor biological behavior. For example, although some patients show a reduction in tumor volume on imaging, their actual prognosis has not improved due to residual drug-resistant clones or immunosuppressive microenvironment; conversely, some molecularly sensitive cases may be prematurely judged as ineffective due to delayed imaging assessments. This "phenotype-mechanism disconnect" phenomenon essentially stems from the fact that a single data dimension cannot fully map the temporal and spatial evolution of the tumor.

[0003] Furthermore, the shortcomings of existing technologies in multimodal data integration and dynamic feedback mechanisms are particularly prominent. Clinical indicators, imaging features, and molecular markers characterize tumor characteristics at different scales: clinical parameters reflect the macroscopic course of the disease, imaging omics captures structural changes, and molecular data reveals driving mechanisms. However, current mainstream methods still rely mainly on simple feature splicing or static weight assignment, lacking the ability to model deep correlations between data. For example, ctDNA clearance in the early stages of chemotherapy may indicate treatment response before imaging changes, while tumor volume regression in the late stages requires the combination of molecular residual lesions to assess the risk of recurrence. Due to the rigid fusion strategy, existing models are unable to dynamically adjust their reliance on different modal data during the treatment process, resulting in the submersion or misjudgment of key biological signals. In addition, traditional prediction models are mostly "one-time" classifiers that only output static conclusions based on pre-treatment baseline data and cannot incorporate dynamic indicators generated in real time during the chemotherapy cycle (such as ctDNA fluctuations and changes in metabolic activity). As a result, clinical decisions still rely on periodic retrospective evaluations, making it difficult to achieve precise intervention during the therapeutic window.

[0004] A deeper challenge lies in the operational bottleneck of clinical transformation. On the one hand, the collection, standardization and analysis of multimodal data involve cross-platform technology integration (such as imaging genomics segmentation algorithms, molecular detection sensitivity, and clinical data structuring), and the existing decentralized processing process is difficult to meet real-time requirements; on the other hand, the disconnect between prediction results and clinical action paths results in the inability of model outputs to directly drive treatment strategy adjustments. For example, how to promptly intensify chemotherapy based on early predictions, when to combine targeted drugs or switch to radical surgery, existing technologies lack a closed-loop feedback mechanism and still rely on physician experience and judgment. This "prediction-execution split" not only prolongs the ineffective treatment cycle, but may also aggravate the patient's organ damage and treatment costs.

[0005] In summary, precise management of neoadjuvant chemotherapy for advanced laryngeal cancer urgently requires a breakthrough from the traditional single-dimensional, static, and isolated technical framework. A predictive system must be constructed that deeply integrates multi-scale bioinformation, tracks treatment responses in real time, and dynamically connects clinical decision-making. The core value of this system lies in integrating the "treatment-assessment-optimization" process into a closed loop through data-driven dynamic iteration, ultimately achieving a paradigm shift from empirical chemotherapy to predictive precision intervention. Summary of the Invention

[0006] The purpose of the present invention is to provide a system for predicting the sensitivity of laryngeal cancer to neoadjuvant chemotherapy in middle and late stages. The present invention realizes accurate prediction of the sensitivity of laryngeal cancer to chemotherapy through multimodal dynamic fusion (cross-modal attention allocation of clinical / imaging / molecular data), time series LSTM-Bayesian correction prediction model, SHAP-t-SNE interpretable decision space and incremental learning closed-loop verification system, thereby improving the accuracy of clinical response assessment and intervention targeting.

[0007] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0008] A system for predicting the sensitivity of neoadjuvant chemotherapy for advanced laryngeal cancer includes a multimodal data acquisition and preprocessing module, a dynamic feature fusion engine module, a core module for dynamic prediction of chemotherapy response, an interpretable output interface module, and a system verification closed-loop module.

[0009] The multimodal data acquisition and preprocessing module includes a clinical indicator acquisition unit, an imaging omics processing unit, and a molecular data analysis unit, which are respectively used to acquire structured clinical data, extract imaging omics features, and analyze molecular time series data;

[0010] The dynamic feature fusion engine module includes a multimodal embedding unit, a cross-modal attention allocation unit, and a weight dynamic adjustment unit to achieve adaptive fusion of clinical-imaging-molecular features;

[0011] The chemotherapy response dynamic prediction core module integrates a ctDNA clearance monitoring unit, a tumor volume response unit, and a prediction correction unit to construct a dynamic prediction model with real-time error correction function;

[0012] The explainability output interface module includes a three-dimensional decision space construction unit, a feature attribution analysis unit, and a molecular pathway analysis unit, providing visual decision support;

[0013] The system verification closed-loop module includes an online learning unit, a hardware acceleration unit and a clinical verification unit to achieve continuous optimization and clinical verification of system performance.

[0014] Furthermore, the clinical indicator collection unit includes:

[0015] The collected parameters included TNM stage, HPV / p16 status, ECOG performance score, smoking index, and laryngeal function score before chemotherapy;

[0016] The SMOTE algorithm is used to correct oversampling of data category imbalance;

[0017] The radiomics processing unit includes:

[0018] Based on high-resolution enhanced CT scan data, tumor volume, surface irregularity, CT value heterogeneity, and dynamic enhancement parameters were extracted using the 3D Slicer platform.

[0019] DICOM-RT protocol was used for multi-temporal image spatial registration;

[0020] The molecular data parsing unit includes:

[0021] EGFR amplification levels were detected by ddPCR and ctDNA mutant allele fraction was monitored by NGS panel;

[0022] The ComBat algorithm was used to eliminate sequencing batch effects and generate a log2(x+1) standardized molecular feature matrix.

[0023] Furthermore, in the dynamic feature fusion engine module:

[0024] The multimodal embedding unit includes:

[0025] The clinical indicators are mapped to a 128-dimensional latent space through a fully connected layer;

[0026] Extract 256-dimensional high-level semantic representation of images through 3D ResNet-34 network;

[0027] Capturing the molecular temporal dynamic clearance pattern through BiLSTM network;

[0028] The cross-modal attention allocation unit adopts a multi-head cross-attention mechanism to calculate the dynamic weights between clinical, imaging, and molecular features, satisfying the constraint of α+β+γ=1, where α∈[0.2,0.6], β∈[0.3,0.7], γ∈[0.1,0.5]; and generates a fused feature vector:

[0029] F fusion =αF clinical ⊕βF imaging ⊕γF molecular

[0030] The dynamic weight adjustment unit adjusts the modal weight according to the real-time changes of the ctDNA clearance rate ΔC and the tumor volume reduction rate ΔV, according to the rule that γ increases by 5% for every 10% increase in ΔC.

[0031] Furthermore, in the chemotherapy response dynamic prediction core module:

[0032] The ctDNA clearance monitoring unit uses an exponential decay model to calculate the periodic clearance rate ΔC:

[0033] C(t)=C0·e -kt

[0034] Where C0 is the baseline ctDNA concentration, k is the clearance rate constant fitted by nonlinear least squares method;

[0035] The tumor volume response unit is based on the RECIST 1.1 standard, and the volume reduction rate ΔV is calculated by outlining the GTVnx three-dimensional area using MIM software;

[0036] The prediction correction unit generates a compensation factor δ through the LSTM time series network and dynamically adjusts the prediction value according to the Bayesian parameter update formula:

[0037] p(θ|D new )∝p(D new |θ)·p(θ|D old )

[0038] Where D old Basic prediction probability, D new is the corrected probability.

[0039] Furthermore, in the interpretability output interface module:

[0040] The three-dimensional decision space construction unit uses the t-SNE algorithm to project high-dimensional features into the three-dimensional space and define risk thresholds for the sensitive group, critical group, and drug-resistant group;

[0041] The feature attribution analysis unit quantifies the feature contribution through SHAP value and generates an image thermal map to mark high-weight areas with CT value >90HU;

[0042] The molecular pathway analysis unit uses the GSEA algorithm to identify the enrichment status of the NFKB and PI3K / AKT pathways.

[0043] Furthermore, in the system verification closed-loop module:

[0044] The online learning unit integrates new case data through an incremental learning framework, uses an EWMA control chart to monitor model drift, and triggers retraining when MAE is > 0.15 three times in a row;

[0045] The hardware acceleration unit implements parallel computing acceleration based on NVIDIA A100 GPU, with a throughput of >200 samples / second;

[0046] The clinical validation unit records the pathological complete response rate (pCR) and the predictive efficacy AUC indicator, and performs a difference test with the RECIST standard.

[0047] Furthermore, the multi-head cross attention mechanism of the cross-modal attention allocation unit adopts 4 attention heads and the key vector dimension is 64.

[0048] Furthermore, the LSTM timing network is configured with 64 hidden units, and the time window length of the input historical prediction error sequence is 3 chemotherapy cycles.

[0049] Furthermore, the t-SNE algorithm parameters include perplexity 30, learning rate 200, and number of iterations 1000 times.

[0050] Furthermore, the hardware acceleration unit achieves data update delay of <200ms through the Redis database and implements load balancing based on the Kubernetes cluster.

[0051] Beneficial effects of the present invention:

[0052] The present invention constructs a multidimensional feature space to characterize the heterogeneity of tumor biological behavior through the deep integration of clinical indicators, imaging omics and molecular time series data. Clinical indicators use the SMOTE algorithm to perform oversampling correction on the category imbalance problem of clinical parameters, overcome the prediction bias caused by the sample distribution bias of traditional models, and enhance the recognition ability of high-risk subgroups (such as HPV negative, high smoking index). The imaging omics processing unit is based on high-order semantic features such as tumor volume and surface irregularity extracted by the 3D ResNet-34 network, which can quantify the spatial heterogeneity of tumor aggressiveness. For example, the surface irregularity feature reflects the degree of infiltration of the tumor-stroma interface through three-dimensional curvature calculation, while the dynamic enhancement parameter indirectly characterizes the tumor microvessel density and perfusion status through the contrast agent uptake kinetic model. After the molecular time series data is modeled by the BiLSTM network, the dynamic clearance trajectory of the ctDNA mutation allele fraction can be captured. Its implicit gating mechanism can identify nonlinear patterns in molecular responses (such as drug resistance signals in the plateau phase after early rapid clearance). The cross-modal attention allocation unit uses a multi-head cross-attention mechanism to dynamically calculate the contribution weights of clinical, imaging, and molecular modalities, establishing semantic associations between features from different modalities. For example, when ctDNA clearance increases significantly in the early stages of chemotherapy, the query vector of the molecular feature and the key vector of the clinical feature become highly correlated, triggering an increase in the molecular modality weight γ. This allows the model to prioritize early sensitive signals at the molecular level, overcoming the lag inherent in traditional imaging assessment.

[0053] The core module of dynamic prediction of chemotherapy response of the present invention realizes real-time correction and error compensation of prediction results through LSTM timing network and Bayesian parameter update algorithm. The memory unit of LSTM network stores historical prediction error sequence through cell state, and the synergistic effect of its input gate, forget gate and output gate can adaptively adjust the retention and forgetting ratio of error information. For example, when over-estimation occurs in multiple consecutive chemotherapy cycles, the forget gate will weaken the weight of early errors, while the input gate will strengthen the compensatory effect of recent errors on the current prediction. The Bayesian parameter update formula calculates the posterior probability based on the product of the prior probability distribution and the likelihood function, and dynamically corrects the confidence interval of the prediction probability through evidence accumulation, thereby reducing the risk of misjudgment caused by single measurement noise. The ctDNA clearance monitoring unit adopts an exponential decay model to fit the ctDNA concentration change curve. The physical meaning of its clearance rate constant k reflects the sensitivity of tumor cell clones to chemotherapy drugs. When the k value is lower than the threshold, it indicates the continued existence of residual resistant subclones. The model optimizes the parameter fitting process through nonlinear least squares method to ensure accurate quantification of molecular response dynamics. The tumor volume response unit combines the RECIST 1.1 standard with three-dimensional region of interest delineation and calculates the tumor volume reduction rate by voxel-level grayscale value integration. Its spatial registration algorithm (DICOM-RT protocol) eliminates position deviations of imaging data at different time points to ensure the comparability of volume measurement results.

[0054] The three-dimensional decision space construction unit of the present invention uses the t-SNE algorithm to project high-dimensional features into a low-dimensional visual space. By minimizing the KL divergence, local similarities between high-dimensional data are maintained, ensuring topological continuity between the classification boundaries of sensitive, critical, and resistant groups in three-dimensional space, facilitating clinicians' intuitive understanding of patient risk stratification. The feature attribution analysis unit calculates the contribution of each feature to the prediction result based on SHAP values. Its mathematical principle is to quantify the direction and intensity of the impact of a single feature or feature combination on the model output through the allocation of marginal contributions based on cooperative game theory. For example, a positive SHAP value shift for HPV / p16 status indicates its promotion of chemotherapy sensitivity, while a high SHAP value for the CT heterogeneity feature reflects the significance of necrosis or calcification within the tumor for treatment resistance. The molecular pathway analysis unit uses the GSEA algorithm to identify gene set enrichment trends and assess the correlation between pathway activity and chemotherapy response using permutation tests. Significant enrichment of the NFKB and PI3K / AKT pathways indicates activation of pro-survival signaling pathways, providing a molecular-level theoretical basis for the combined use of pathway inhibitors. Image heat maps locate high-contribution image areas through gradient-weighted class activation mapping (Grad-CAM) technology. Its generation mechanism relies on the gradient return of the last feature map of the convolutional neural network, and generates a visually interpretable thermal distribution through spatial weighted summation to guide the precise delineation of radiotherapy targets.

[0055] The online learning unit of the present invention realizes the dynamic update of model parameters through an incremental learning framework, adopts the elastic weight consolidation algorithm (EWC) to constrain the parameter update direction, retains the key weights of previous knowledge when integrating new case data, and prevents model performance degradation caused by data distribution drift. The EWMA control chart monitors the temporal changes of the model prediction error through the exponentially weighted moving average method. Its control limit (±3σ) is dynamically adjusted based on the standard deviation of historical errors. When it exceeds the limit continuously, the model retraining process is triggered to ensure the system's adaptability to regional case feature differences. The hardware acceleration unit optimizes the computational efficiency of feature fusion and predictive reasoning based on the GPU parallel computing architecture. The thread-level parallel mechanism of its CUDA core significantly improves the execution speed of 3D convolution (image processing) and matrix operations (attention weight calculation), meeting clinical real-time requirements. The clinical validation unit uses the area under the receiver operating characteristic curve (AUC) and pathological complete response rate (pCR) as core evaluation indicators, and verifies the significant difference between the system's prediction results and traditional evaluation methods (RECIST standards) through statistical hypothesis testing. Its internal logic is to use non-parametric tests (such as the DeLong test) to compare the statistical significance of AUC differences, providing evidence-based medicine evidence for the clinical transformation of the system.

[0056] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. DETAILED DESCRIPTION

[0057] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0058] Example 1

[0059] The system for predicting sensitivity of advanced laryngeal cancer to neoadjuvant chemotherapy described in this embodiment includes:

[0060] Multimodal data acquisition and preprocessing module:

[0061] This module achieves standardized processing of multi-source heterogeneous data by integrating clinical indicators, imaging genomics and molecular data. The clinical indicator acquisition unit strictly follows the NCCN guidelines to collect structured data including TNM staging (UICC 8th edition), HPV / p16 status, and ECOG physical fitness score, and uses the SMOTE algorithm to oversample subgroups with insufficient sample size to solve the problem of class imbalance. The imaging genomics processing unit is based on high-resolution CT arterial phase enhanced scan data (layer thickness ≤ 1mm), using the 3D Slicer platform to outline the tumor volume (VOI), extract 42 quantitative features such as surface irregularity (Sphericity), CT value heterogeneity (GLCM contrast) and dynamic enhancement parameters (Wash-inrate slope), and complete the spatial alignment of multimodal imaging data through the DICOM-RT protocol. The molecular data analysis unit uses ddPCR to quantitatively detect EGFR amplification levels and NGSpanel to monitor ctDNA mutation allele fraction (MAF), eliminates batch effects through the ComBat algorithm, and performs log2(x+1) transformation to generate a standardized molecular feature matrix to ensure cross-platform data comparability.

[0062] Dynamic feature fusion engine module:

[0063] This module uses a cross-modal attention mechanism to achieve dynamic allocation and fusion of feature weights. The multimodal embedding unit maps clinical indicators to the latent space through a fully connected layer (128 nodes, ReLU activation). The imaging features are extracted through a 3D ResNet-34 network with 256-dimensional high-order semantic representations. The molecular time series data is captured through a BiLSTM network (64 hidden layer units) to capture the dynamic clearance pattern of ctDNA. The cross-modal attention allocation unit constructs a multi-head cross-attention mechanism (4 heads, key dimension 64), calculates the association weights between clinical, imaging, and molecular features through the QKV transformation matrix, and applies the constraint condition of α+β+γ=1 to generate the fused feature vector F.fusion =αF clinical ⊕βF imaging ⊕γF molecular The dynamic weight adjustment unit dynamically optimizes weight allocation according to the real-time changes in ctDNA clearance rate (ΔC) and tumor volume reduction rate (ΔV) during the chemotherapy cycle according to preset rules (for example, for every 10% increase in ΔC, the molecular weight γ increases by 5%), thereby achieving adaptive fusion of multimodal features.

[0064] Core module for dynamic prediction of chemotherapy response:

[0065] This module builds a real-time prediction and correction system based on dual-channel biomarker feedback. The ctDNA clearance monitoring unit establishes an exponential decay model C(t) = C0·e -kt The clearance rate constant k was fitted using the nonlinear least squares method (goodness of fit R 2 >0.85), calculate the cycle clearance rate ΔC=(C pre -C post ) / C pre The tumor volume response unit was based on the RECIST 1.1 standard. The GTVnx three-dimensional area was delineated by MIM software and the volume reduction rate ΔV was calculated. pre -V post ) / V pre The prediction correction unit introduces the LSTM time series network (64 hidden units), inputs the historical prediction error sequence to generate the compensation factor δ, and combines the Bayesian parameter update formula p(θ|D new )∝p(D new |θ)·p(θ|D old ) Dynamically adjust the basic prediction value and finally output the corrected chemotherapy sensitivity probability y final =y base ·(1+δ·ΔC·ΔV), which reduced the prediction error MAE to 0.11 (95% CI 0.09-0.13) after the second chemotherapy cycle.

[0066] Explainable output interface module:

[0067] This module provides three-dimensional visualization and attribution analysis to enhance the interpretability of clinical decisions. The three-dimensional decision space construction unit uses the t-SNE algorithm (perplexity 30, learning rate 200) to project high-dimensional features into three-dimensional space, dividing risk areas into sensitive (>0.7), critical (0.4-0.7), and resistant (<0.4) groups. The feature attribution analysis unit quantifies the contribution of each feature using SHAP values ​​(e.g., PIK3CA mutation SHAP value = 0.23±0.07) and generates image heat maps to annotate high-weighted regions with CT values ​​>90HU (pixel-level accuracy error <0.1mm). The molecular pathway analysis unit uses the GSEA algorithm (1000 permutation tests) to identify the enrichment status of the key NFKB and PI3K / AKT pathways (FDR <0.05), providing a molecular basis for the analysis of resistance mechanisms.

[0068] System verification closed-loop module:

[0069] This module ensures continuous system optimization and real-time responsiveness through online learning and hardware acceleration. The online learning unit deploys an incremental learning framework, integrating new case data reviewed by three associate chief physicians in a double-blind manner. An EWMA control chart (λ = 0.2, control limits ±3σ) is used to monitor model drift. Model retraining is triggered when the MAE exceeds a threshold of 0.15 three consecutive times. The hardware acceleration unit utilizes NVIDIA A100 GPUs for parallel computing (throughput > 200 samples / second), a Redis database to ensure data update latency < 200ms during chemotherapy cycles, and a Kubernetes cluster for load balancing (QPS > 500). The clinical validation unit recorded the pathological complete response rate (pCR = 57.1%) and predictive efficacy (AUC = 0.89, 95% CI 0.85-0.93) guided by the system, which significantly improved compared to the traditional RECIST criteria (ΔAUC = +0.21, P < 0.01), validating the system's clinical translational value.

[0070] Example 2

[0071] The system for predicting the sensitivity of mid- to late-stage laryngeal cancer to neoadjuvant chemotherapy described in this embodiment includes a multimodal data acquisition and preprocessing module, a dynamic feature fusion engine module, a core module for dynamic prediction of chemotherapy response, an interpretable output interface module, and a system verification closed-loop module, wherein:

[0072] The multimodal data acquisition and preprocessing module includes:

[0073] The clinical indicator collection unit collects TNM stage, HPV / p16 status, ECOG physical fitness score, smoking index, and pre-chemotherapy laryngeal function score, and uses the SMOTE algorithm to correct data category imbalance;

[0074] The radiomics processing unit extracts tumor volume, surface irregularity, CT value heterogeneity, and dynamic enhancement parameters based on high-resolution CT enhanced scan data using the 3DSlicer platform, and performs spatial registration using the DICOM-RT protocol.

[0075] The molecular data analysis unit uses ddPCR to detect EGFR amplification levels and NGS panels to monitor ctDNA mutation allele fractions, and uses the ComBat algorithm to eliminate sequencing batch effects and generate a log2(x+1) standardized molecular feature matrix;

[0076] The dynamic feature fusion engine module includes:

[0077] The multimodal embedding unit maps clinical indicators to a 128-dimensional latent space through a fully connected layer, extracts 256-dimensional high-order semantic representations of image features through a 3DResNet-34 network, and captures dynamic clearance patterns of molecular time series data through a BiLSTM network;

[0078] The cross-modal attention allocation unit calculates the dynamic weights between clinical, imaging, and molecular features through a multi-head cross-attention mechanism (4 heads, key dimension 64), satisfies the constraint of weight coefficient α+β+γ=1, and generates a fused feature vector:

[0079] F fusion =αF clinical ⊕βF imaging ⊕γF molecular

[0080] (α, β, γ) are the contribution weights of clinical, imaging, and molecular modalities, α∈[0.2, 0.6], β∈[0.3, 0.7], γ∈[0.1, 0.5];

[0081] The dynamic weight adjustment unit adjusts the weights of each modality according to the real-time changes in the ctDNA clearance rate ΔC and the tumor volume reduction rate ΔV during the chemotherapy cycle according to preset rules (for every 10% increase in ΔC, the molecular weight γ increases by 5%).

[0082] The chemotherapy response dynamic prediction core module includes:

[0083] The ctDNA clearance monitoring unit calculates the periodic clearance rate ΔC based on the exponential decay model:

[0084] C(t)=C0·e -kt

[0085] C0 is the baseline ctDNA concentration, k is the clearance rate constant, and the nonlinear least squares method is used to fit the ctDNA. When k < 0.1 day -1 When marked as low clearance group;

[0086] Tumor volume response unit: The tumor volume reduction rate ΔV was calculated using the RECIST 1.1 standard, and the GTVnx three-dimensional region of interest was delineated based on the MIM software;

[0087] The prediction correction unit inputs the historical prediction error sequence through the LSTM time series network, generates the compensation factor δ, and dynamically adjusts the prediction value according to the Bayesian parameter update formula:

[0088] p(θ|D new )∝p(D new |θ)·p(θ|D old )

[0089] Final output:

[0090] y final =y base (1+δ·ΔC·ΔV)

[0091] y base is the basic prediction probability, which is obtained by fusion feature F fusion After the full connection layer output, y final is the corrected probability;

[0092] The interpretability output interface module includes:

[0093] The three-dimensional decision space construction unit was constructed, and the t-SNE algorithm (perplexity 30, learning rate 200) was used to project the high-dimensional features into the three-dimensional space, and the risk thresholds of the sensitive group (>0.7), critical group (0.4-0.7), and resistant group (<0.4) were defined;

[0094] The feature attribution analysis unit quantifies the contribution of each feature using the SHAP value (PIK3CA mutation SHAP value = 0.23 ± 0.07) and generates an image thermal map to mark high-weight areas with CT values ​​> 90 HU;

[0095] In the molecular pathway analysis unit, the GSEA algorithm (1000 permutation tests) was used to identify the enrichment status of the NFKB and PI3K / AKT pathways (FDR < 0.05);

[0096] The system verification closed-loop module includes:

[0097] The online learning unit integrates new case data through an incremental learning framework and uses an EWMA control chart (λ = 0.2, control limits ± 3σ) to monitor model drift. Model retraining is triggered when MAE is > 0.15 three times in a row;

[0098] Hardware acceleration unit, based on NVIDIA A100 GPU to achieve parallel computing acceleration (throughput > 200 samples / second), and through the Redis database to ensure data update latency < 200ms;

[0099] The clinical validation unit recorded the pathological complete response rate (pCR) and predictive efficacy AUC under the guidance of the system, and performed a difference test with the traditional RECIST standard (P<0.01).

[0100] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A system for predicting sensitivity of neoadjuvant chemotherapy for advanced laryngeal cancer, characterized by: It includes a multimodal data acquisition and preprocessing module, a dynamic feature fusion engine module, a chemotherapy response dynamic prediction core module, an interpretable output interface module, and a system verification closed-loop module, which are connected in sequence. The multimodal data acquisition and preprocessing module includes a clinical indicator acquisition unit, an imaging omics processing unit, and a molecular data analysis unit, which are respectively used to acquire structured clinical data, extract imaging omics features, and analyze molecular time series data; The dynamic feature fusion engine module includes a multimodal embedding unit, a cross-modal attention allocation unit, and a weight dynamic adjustment unit to achieve adaptive fusion of clinical-imaging-molecular features; The chemotherapy response dynamic prediction core module integrates a ctDNA clearance monitoring unit, a tumor volume response unit, and a prediction correction unit to construct a dynamic prediction model with real-time error correction function; The explainability output interface module includes a three-dimensional decision space construction unit, a feature attribution analysis unit, and a molecular pathway analysis unit, providing visual decision support; The system verification closed-loop module includes an online learning unit, a hardware acceleration unit and a clinical verification unit to achieve continuous optimization and clinical verification of system performance.

2. The system according to claim 1, wherein The clinical indicator collection unit includes: The collected parameters included TNM stage, HPV / p16 status, ECOG performance score, smoking index, and laryngeal function score before chemotherapy; The SMOTE algorithm is used to correct oversampling of data category imbalance; The imaging omics processing unit includes: Based on high-resolution enhanced CT scan data, tumor volume, surface irregularity, CT value heterogeneity, and dynamic enhancement parameters were extracted using the 3D Slicer platform. DICOM-RT protocol was used for multi-temporal image spatial registration; The molecular data parsing unit includes: EGFR amplification levels were detected by ddPCR and ctDNA mutant allele fraction was monitored by NGS panel; The ComBat algorithm was used to eliminate sequencing batch effects and generate a log2(x+1) standardized molecular feature matrix.

3. The system according to claim 1, wherein: In the dynamic feature fusion engine module: The multimodal embedding unit includes: The clinical indicators are mapped to a 128-dimensional latent space through a fully connected layer; Extract 256-dimensional high-level semantic representation of images through 3D ResNet-34 network; Capturing the molecular temporal dynamic clearance pattern through BiLSTM network; The cross-modal attention allocation unit adopts a multi-head cross-attention mechanism to calculate the dynamic weights between clinical, imaging, and molecular features, satisfying the constraint of α+β+γ=1, where α∈[0.2,0.6], β∈[0.3,0.7], γ∈[0.1,0.5]; and generates a fused feature vector: The dynamic weight adjustment unit adjusts the modal weight according to the real-time changes of the ctDNA clearance rate ΔC and the tumor volume reduction rate ΔV, according to the rule that γ increases by 5% for every 10% increase in ΔC.

4. The system according to claim 1, wherein In the chemotherapy response dynamic prediction core module: The ctDNA clearance monitoring unit uses an exponential decay model to calculate the periodic clearance rate ΔC: C(t)=C0·e -kt Where C0 is the baseline ctDNA concentration, k is the clearance rate constant fitted by nonlinear least squares method; The tumor volume response unit is based on the RECIST 1.1 standard, and the volume reduction rate ΔV is calculated by outlining the GTVnx three-dimensional area using MIM software; The prediction correction unit generates a compensation factor δ through the LSTM time series network and dynamically adjusts the prediction value according to the Bayesian parameter update formula: p(θ|D new )∝p(D new |θ)·p(θ|D old ) Where D old Basic prediction probability, D new is the corrected probability.

5. The system according to claim 1, wherein: In the explainability output interface module: The three-dimensional decision space construction unit uses the t-SNE algorithm to project high-dimensional features into the three-dimensional space and define risk thresholds for the sensitive group, critical group, and drug-resistant group; The feature attribution analysis unit quantifies the feature contribution through SHAP value and generates an image thermal map to mark high-weight areas with CT value >90HU; The molecular pathway analysis unit uses the GSEA algorithm to identify the enrichment status of the NFKB and PI3K / AKT pathways.

6. The system according to claim 1, wherein: In the system verification closed-loop module: The online learning unit integrates new case data through an incremental learning framework, uses an EWMA control chart to monitor model drift, and triggers retraining when MAE is > 0.15 three times in a row; The hardware acceleration unit implements parallel computing acceleration based on NVIDIA A100 GPU, with a throughput of >200 samples / second; The clinical validation unit records the pathological complete response rate (pCR) and the predictive efficacy AUC indicator, and performs a difference test with the RECIST standard.

7. The system according to claim 3, wherein: The multi-head cross attention mechanism of the cross-modal attention allocation unit adopts 4 attention heads and the key vector dimension is 64.

8. The system according to claim 4, wherein: The LSTM timing network is configured with 64 hidden units, and the time window length of the input historical prediction error sequence is 3 chemotherapy cycles.

9. The system according to claim 5, wherein: The t-SNE algorithm parameters include perplexity 30, learning rate 200, and number of iterations 1000 times.

10. The system according to claim 6, wherein: The hardware acceleration unit uses the Redis database to achieve data update delay of <200ms and implements load balancing based on the Kubernetes cluster.

Images