Prognostic decision system for unresectable hepatocellular carcinoma

Abstract

The application discloses a prognosis decision system for unresectable hepatocellular carcinoma (uHCC), comprising processing clinical data of uHCC patients to obtain clinical text input label sequences and CECT image block input label sequences; a plurality of prognosis prediction models are trained by using the text input label sequences and the image block input label sequences of a plurality of queues; a decision unit constructs a decision support model based on the plurality of prognosis prediction models, and provides a plurality of treatment decisions. The decision system helps the uHCC patients to construct an optimal treatment decision under a plurality of optional treatment schemes and predict a corresponding prognosis survival score, provides strong support for uHCC personalized precision treatment, and improves the non-surgical treatment effect and survival rate of the patients.

Description

Technical Field

[0001] This application belongs to the field of biomedical technology, specifically relating to a prognostic decision-making system for unresectable hepatocellular carcinoma. Background Technology

[0002] Unresectable hepatocellular carcinoma (uHCC) is a global health challenge, characterized by tumors that cannot be surgically removed due to factors such as tumor size, location, number, vascular invasion, or insufficient liver function reserve. Because the lesions cannot be effectively removed surgically, the prognosis for uHCC is low. Currently, the main treatment strategy for uHCC is intra-arterial interventional therapy (IAIT), including transarterial chemoembolization (TACE) and hepatic arterial chemofusion chemotherapy (HAIC). These strategies rely on the unique heterogeneous imaging biomarkers of each patient. Furthermore, combining TACE or HAIC with molecularly targeted drugs (MTAs) or immune checkpoint inhibitors (ICIs) has shown potential to effectively prolong the survival of uHCC patients.

[0003] Significant heterogeneity exists among uHCC patients, including individual differences in tumor burden, liver function reserve, and general condition. Current non-surgical treatment options lack effective personalized decision-making, often relying on clinical trial treatment plans based on the experience of treating staff, which may fail to meet the needs of diverse patients. Developing treatment plans for individual patients requires comprehensive consideration of factors such as medical history, tumor characteristics, treatment goals, and previous treatments. Current treatment decision-making processes are often complex, requiring multidisciplinary team collaboration and cost-benefit analysis, such as multidisciplinary consultations, which consumes additional medical resources. When medical personnel are insufficient, conventional plans are often adopted, which may not be suitable for the patient's actual situation, making it difficult to effectively improve the success rate of conversion therapy. Therefore, accurate automated decision-making systems are needed to assist treating staff in conducting prognostic assessments under multiple options, thereby improving the efficiency of patient hospital management.

[0004] CN113470782A discloses a clinical prognostic prediction model for unresectable hepatocellular carcinoma (uHCC) before surgical intervention. Based on five patient variables—pre-SLR alpha-fetoprotein (AFP), number of lesions, portal vein tumor thrombus, maximum diameter of the main tumor, and changes in tumor burden after transarterial chemoembolization (TACE)—a Nomogram model was trained to predict the prognosis of patients suitable for salvage hepatectomy after TACE. However, CN113470782A does not consider features in clinical images of uHCC patients for prognostic prediction, such as quantitative analysis of contrast-enhanced CT images (CECT) to measure the volume, location, and morphology of uHCC lesions, thus failing to provide more accurate data support for its predictive model and improve decision-making accuracy. Furthermore, it does not disclose predictions of patients suitable for combined therapy with multiple treatment modalities during the course of the disease and their prognosis, highlighting that receiving TACE monotherapy is not a characteristic of all uHCC patients.

[0005] This application is proposed to address the shortcomings of the above uHCC prognostic methods. Summary of the Invention

[0006] To address the problems existing in the prior art, the purpose of this application is to provide a prognostic decision-making system for unresectable hepatocellular carcinoma (uHCC). This system trains multiple prognostic prediction models using multimodal cohort data of clinical texts and CECT images from uHCC patients, and constructs a modular decision support model using decision units and multiple prognostic prediction models. This system can provide optimal treatment decisions and prognostic survival predictions under various treatment options.

[0007] Specifically, this application relates to the following aspects:

[0008] This application proposes a prognostic decision-making system for unresectable hepatocellular carcinoma, which includes the following functions:

[0009] The sample data preprocessing module acquires clinical text variables of patients with unresectable hepatocellular carcinoma and converts them into text input label sequences through a text embedding unit; it also acquires preoperative CT enhanced images of patients with unresectable hepatocellular carcinoma, segments the preoperative CT enhanced images into arterial phase CT enhanced images and portal venous phase CT enhanced images, and converts the arterial phase CT enhanced images and portal venous phase CT enhanced images into image block input label sequences through an image embedding unit.

[0010] The prognostic prediction model building module inputs text input label sequences and image patch input label sequences into a deep learning model to capture the global relationship between the text input label sequences and image patch input label sequences, thereby obtaining a prognostic prediction model based on a deep learning model to predict patient survival scores.

[0011] The decision support model construction module includes a decision unit, which executes the optimal prognostic decision based on multiple prognostic prediction models. The decision support model construction module obtains multiple prognostic prediction models through training multiple cohorts of unresectable hepatocellular carcinoma patients, and constructs a decision support model based on the decision unit and multiple prognostic prediction models to provide optimal prognostic decision results to patients with unresectable hepatocellular carcinoma.

[0012] According to one embodiment of this application, converting clinical text variables into a text input marker sequence via a text embedding unit includes:

[0013] Clinical text variables are input into text embedding units and linearly mapped; the linearly mapped clinical text variables are then tokenized to obtain clinical text variable token sequences; and position embeddings are added to each clinical text variable token sequence to obtain text input token sequences.

[0014] According to one embodiment of this application, converting arterial phase CT enhanced images and portal venous phase CT enhanced images into image block input marker sequences via an image embedding unit includes:

[0015] Arterial phase CT enhanced images and portal venous phase CT enhanced images are input into the image embedding unit. Feature maps of the arterial phase CT enhanced images and portal venous phase CT enhanced images are obtained through a two-dimensional convolutional layer. Each feature map contains multiple feature map blocks. Position embedding is added to each feature map block to obtain multiple image block input label sequences.

[0016] According to one embodiment of this application, the clinical text variables include baseline variables, laboratory variables, and imaging variables; the baseline variables include age, sex, medical history, ECOG classification, complications, HBV, ascites, and continuous local treatment; the laboratory variables include ALB, TBIL, ALT, AST, PT, INR, PLT, and AFP; the imaging variables include Barcelona clinical stage of liver cancer, diagnosis of portal vein tumor thrombosis, tumor metastasis results, tumor size, number of tumors, vascular invasion, tumor metastasis, and Child-Pugh classification.

[0017] According to one embodiment of this application, the deep learning model consists of m first neural network units and n second neural network units; wherein each text input label sequence and image patch input label sequence is input into the first first neural network unit and processed by the first normalization layer, the first neural network layer, the second normalization layer and the feedforward neural network therein, and then enters the next first neural network unit for processing, until it is processed by the mth first neural network unit; the text input label sequence and image patch input label sequence processed by the m first neural network units are subjected to unified labeling processing to obtain a multimodal unified label sequence that integrates the text input label sequence information and the image patch input label sequence information.

[0018] According to one embodiment of this application, the multimodal unified label sequence is further input into the first second neural network unit and processed through the third normalization layer, the second neural network layer, the fourth normalization layer, and the feedforward neural network therein, and then processed by the next second neural network unit until it is processed by the nth second neural network unit; the multimodal unified label sequence processed by the nth second neural network unit is passed through the classification unit and converted into a prognostic survival score for patients with unresectable hepatocellular carcinoma; the difference between the prognostic survival score of patients with unresectable hepatocellular carcinoma and the true prognostic survival score of patients with unresectable hepatocellular carcinoma is measured by a loss function, and the loss function is minimized to update the first neural network unit and the second neural network unit until the loss value reaches a specified level and a prognostic prediction model is obtained.

[0019] According to one embodiment of this application, multiple prognostic prediction models are obtained by training multiple cohorts of patients with unresectable hepatocellular carcinoma, including:

[0020] A first prognostic prediction model was obtained by training a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC; a second prognostic prediction model was obtained by training a cohort of patients with unresectable hepatocellular carcinoma treated with TACE; a third prognostic prediction model was obtained by training a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC combined with other therapies; and a fourth prognostic prediction model was obtained by training a cohort of patients with unresectable hepatocellular carcinoma treated with TACE combined with other therapies. Further models include:

[0021] The HAIC survival score was obtained through the first prognostic prediction model; the first TACE survival score and the second TACE survival score were obtained through the second prognostic prediction model; the first HC survival score and the second HC survival score were obtained through the third prognostic prediction model; and the first TC survival score and the second TC survival score were obtained through the fourth prognostic prediction model.

[0022] According to one embodiment of this application, the decision-making unit performs the optimal prognostic decision based on multiple prognostic prediction models, including the following steps:

[0023] For patients with unresectable hepatocellular carcinoma, the HAIC survival score was used for prediction. In response to an HAIC survival score lower than MSS1, the first TACE survival score was used for prediction. In response to an HAIC survival score not lower than MSS1, the second TACE survival score was used for prediction. In response to an HAIC survival score not lower than MSS2, the first HC survival score was used for prediction. In response to a second TACE survival score not lower than MSS2, the first TC survival score was used for prediction. In response to a second TACE survival score lower than MSS2, the second HC survival score was used for prediction. For patients with unresectable hepatocellular carcinoma, if the survival score is not lower than MSS3, HC is performed; if the first HC survival score is lower than MSS3, HAIC is performed; if the first TC survival score is not lower than MSS4, TC is performed; if the first TC survival score is lower than MSS4, TACE is performed; if the second HC survival score is not lower than MSS3, HC is performed; if the second HC survival score is lower than MSS3, second TC survival score prediction is performed; and if the second TC survival score is not lower than MSS4, TC is performed.

[0024] According to one embodiment of this application, MSS1 is the median HAIC survival score of the first prognostic prediction model training set; MSS2 is the median TACE survival score of the second prognostic prediction model training set; MSS3 is the median HC survival score of the third prognostic prediction model training set; and MSS4 is the median TC survival score of the fourth prognostic prediction model training set.

[0025] According to one embodiment of this application, HC is HAIC combination therapy, which includes HAIC therapy combined with TKI (tyrosine kinase inhibitor) therapy, ICI (immune checkpoint inhibitor) therapy and / or local therapy; TC is TACE combination therapy, which includes TACE therapy combined with TKI (tyrosine kinase inhibitor) therapy, ICI (immune checkpoint inhibitor) therapy and / or local therapy.

[0026] According to one embodiment of this application, the HAIC survival score, the first TACE survival score, and the second TACE survival score represent the objective response rate (ORR) for patients with unresectable hepatocellular carcinoma; and the first HC survival score, the second HC survival score, the first TC survival score, and the second TC survival score represent the one-year overall survival (OS) for patients with unresectable hepatocellular carcinoma.

[0027] Beneficial effects

[0028] The prognostic decision-making system for unresectable hepatocellular carcinoma described in this application has the following technical advantages:

[0029] 1. This application proposes a deep learning-based uHCC prognostic decision-making system. Addressing the limitations and improvement needs of existing prediction models in decision-making and prognosis, it provides a prognostic prediction model that can globally learn features from medical images and clinical text, learning treatment prognostic feature representations from multimodal medical clinical data. The prognostic prediction model represents the complex relationships between patient baseline characteristics, clinical data, treatment plans, and prognostic outcomes in the form of a neural network, thereby providing an artificial intelligence-like medical atlas in conjunction with the decision-making unit. In this way, the decision support model can comprehensively analyze and utilize this multidimensional information, providing patients with multiple prognostic predictions and, based on these, making optimal treatment decisions.

[0030] 2. The prognostic prediction model proposed in this application learns global relationships in multimodal clinical data to extract comprehensive information from structured clinical text, unstructured text, and CECT image patches. It can analyze rich prognostic feature representations in patient cohorts, thus exhibiting better population generalization, lower methodological and computational complexity, and ensuring high accuracy. In addition to accurately providing personalized combined or single-treatment decisions among four treatment options for uHCC patients, the decision support model can also recommend further conservative treatment for potential patients with unsatisfactory prognoses under multiple treatment options, avoiding treatment bias, alleviating patient suffering, and reducing ineffective economic expenditures. The application of this decision system will help provide personalized treatment decisions for patients, providing strong support for helping patients achieve conversion therapy or further rehabilitation.

[0031] 3. The decision-making system proposed in this application can also help researchers or medical personnel determine the treatment and hospitalization plans for early-stage uHCC patients, thereby rationally organizing the use of medical resources. For example, it can help medical institutions determine the procurement plans for medicines and medical devices based on recent patient cohorts, and coordinate the allocation of inpatient beds in departments based on patient survival scores, thus providing feasible solutions for making full use of limited medical resources. Attached Figure Description

[0032] Figure 1 The illustration shows a schematic diagram of the training process of the prognostic prediction model according to an embodiment of this application, which receives uHCC patient cohort data.

[0033] Figure 2 The figure shows a schematic diagram of the structure of the first neural network unit of the prognostic prediction model according to an embodiment of the present application.

[0034] Figure 3The illustration shows a flowchart of constructing a decision support model using four prognostic prediction models according to an embodiment of this application.

[0035] Figure 4 The illustration shows a flowchart of a decision support model providing treatment decisions and prognostic predictions according to an embodiment of this application.

[0036] Figure 5A The illustration shows a performance comparison of the decision support model according to the embodiments of this application in predicting HAIC and HC prognostic survival scores on the training and validation sets of a patient cohort.

[0037] Figure 5B The illustration shows a performance comparison of the decision support model according to the embodiments of this application in predicting TACE and TC prognostic survival scores on the training and validation sets of a patient cohort. Detailed Implementation

[0038] The present application is further illustrated below with reference to embodiments. It should be understood that the embodiments are only used to further illustrate and explain the present application and are not intended to limit the present application.

[0039] Unless otherwise defined, technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art. While similar or identical methods and materials may be applied in experimental or practical applications, materials and methods are described herein. In case of conflict, the definitions included herein shall prevail. Furthermore, materials, methods, and examples are for illustrative purposes only and are not intended to be limiting. The present application is further described below with reference to specific embodiments, but is not intended to limit the scope of the application.

[0040] definition

[0041] Hepatic artery infusion chemotherapy

[0042] Hepatic artery infusion chemotherapy (HAIC) is a type of intra-arterial interventional therapy (IAIT) that involves directly infusing chemotherapy drugs into the liver via the hepatic artery to achieve therapeutic goals. Specifically, HAIC can be used to treat diseases such as uHCC (ultra-hepatic cancer). Its advantage lies in the higher drug concentration within the hepatic artery, allowing liver cancer tissue to absorb a larger amount of chemotherapy drugs compared to systemic chemotherapy, thereby improving the efficacy of liver cancer treatment. For a period after treatment, uHCC patients may need to follow specific care and rehabilitation guidelines to ensure treatment effectiveness and reduce the occurrence of complications.

[0043] Transhepatic artery chemoembolization

[0044] Transarterial chemoembolization (TACE) is another interventional treatment method for liver cancer within the field of intra-arterial interventional therapy (IAIT). Its principle is to deliver chemotherapy drugs and embolic agents directly to the blood supply artery of the liver tumor via catheter technology, achieving high local concentrations of medication and blocking the tumor's blood supply. However, TACE is not suitable for all liver cancer patients, especially given the significant individual variability reported in Asian clinical reports. Therefore, the decision to undergo TACE treatment requires comprehensive consideration of factors such as the patient's basic characteristics, disease progression, liver function status, and potential complications.

[0045] Objective remission rate

[0046] Objective Response Rate (ORR) is a core indicator used to evaluate the efficacy of cancer treatment. Specifically, ORR refers to the proportion of patients whose tumor volume shrinks to a predetermined value and remains stable for a certain period during cancer treatment. This indicator directly measures the anti-tumor activity of a drug and is an important tool for evaluating efficacy, especially in early-stage clinical trials.

[0047] Overall survival

[0048] Overall survival (OS) is a key indicator for evaluating the efficacy of anti-tumor drugs. Generally, if a drug can significantly prolong a patient's OS, then that drug is considered to have good efficacy. In clinical practice, OS data one year after surgery / treatment is considered to effectively reflect efficacy and is of great significance in guiding treatment decisions. Physicians often determine appropriate treatment plans based on the patient's expected one-year OS.

[0049] This application proposes a prognostic decision-making system for unresectable hepatocellular carcinoma (uHCC), which may include the following functions:

[0050] The sample data preprocessing module acquires clinical text variables from patients with unresectable hepatocellular carcinoma and converts them into text input labeled sequences; and

[0051] Preoperative enhanced CT images of patients with unresectable hepatocellular carcinoma were acquired, segmented into arterial phase enhanced CT images and portal venous phase enhanced CT images, and converted into image blocks for input labeling sequences.

[0052] The prognostic prediction model building module inputs text input label sequences and image patch input label sequences into a deep learning model to capture the global relationship between the text input label sequences and image patch input label sequences, thereby obtaining a prognostic prediction model based on a deep learning model.

[0053] The decision support model construction module includes a decision unit, which executes the optimal prognostic decision based on multiple prognostic prediction models. The decision support model construction module obtains multiple prognostic prediction models through training multiple cohorts of patients with unresectable hepatocellular carcinoma, and constructs a decision support model based on the decision unit and multiple prognostic prediction models to provide optimal prognostic decision results to patients with unresectable hepatocellular carcinoma.

[0054] In this application, preoperative contrast-enhanced computed tomography (CECT) images of uHCC patients, obtained by intravenously injecting contrast agent, increase the contrast between tissues, thereby more clearly displaying the lesions of blood vessels and tissue structures. CECT images of liver vessels show different manifestations in different phases. HCC shows significant enhancement in the arterial phase, exhibiting hypervascularization characterized by arterial blood supply. In the portal venous phase, contrast agent is widely distributed in the hepatic portal venous system, allowing assessment of the enhancement characteristics of HCC lesions, such as tumor size, morphology, growth characteristics, and the presence of distant metastases. Therefore, contrast-enhanced images in the arterial and portal venous phases can provide detailed information on tumor blood supply and surrounding tissues, enabling a more comprehensive assessment of the characteristics and progression of liver cancer. By combining multiple images and providing them to a prognostic prediction model to learn from the rich pathological information, the model can better integrate clinical text to mine highly accurate treatment plans.

[0055] Automatic segmentation methods were used to divide CECT images and obtain arterial and portal venous phase CECT images to help the prognostic prediction model further identify and quantify lesion tissue in both phases, improving the quality of image data input to the prognostic prediction model. Common segmentation methods can be used to segment the uHCC patient images in this application, for example, using an nnUNet model pre-trained on liver tumor images for automatic segmentation. Vascular enhancement exhibits different characteristics in different phases, and images from both phases provide detailed information on tumor blood supply and surrounding tissues, thus contributing to a more comprehensive assessment of the characteristics and progression of liver cancer. This combination of multi-phase hepatic vascular images helps improve the diagnostic and predictive accuracy of the model.

[0056] In particular, CECT images can be preprocessed before segmentation. Preprocessing methods include interpolation, denoising, and voxel intensity quantization to improve image quality and useful information, thereby enhancing the reliability of the segmentation algorithm.

[0057] In this application, the sample data preprocessing module for converting clinical text variables into a text input tag sequence may include:

[0058] Transforming clinical text variables using text embedding methods may include the following steps: linearly mapping the clinical text variables; tokenizing the linearly mapped clinical text variables to obtain a sequence of clinical text variable tokens; and adding positional embeddings to each clinical text variable token sequence to obtain a text input token sequence. By linearly mapping the input clinical text variables, the clinical text variables are converted into word vectors, and the word vectors are segmented to obtain multiple token sequences. It should be noted that the clinical text variables used in this application may include structured and unstructured text variables; therefore, unstructured text variables can be segmented into multiple word units and processed together with structured text variables.

[0059] Specifically, positional embedding is performed on the labeled sequences of clinical text variables. The purpose is to encode the positional information of words and word units in the labeled sequence into vector representations, so that the labeled sequence has complete positional information. In this way, the prognostic prediction model can capture the semantic relationships between words in the text input labeled sequence, such as the relationship between patient baseline characteristics, laboratory indicators, and imaging diagnoses.

[0060] In this application, clinical text variables may include baseline variables, laboratory variables, and imaging variables. Specifically, baseline variables may include: age, sex, medical history, ECOG classification, complications, HBV, ascites, and continuous local treatment. Laboratory variables may include: ALB, TBIL, ALT, AST, PT, INR, PLT, and AFP. Imaging variables may include: Barcelona clinical stage of liver cancer, diagnosis of portal vein tumor thrombosis, tumor metastasis results, tumor size, number of tumors, vascular invasion, tumor metastasis, and Child-Pugh classification.

[0061] In this application, the sample data preprocessing module may convert arterial phase CT enhanced images and portal venous phase CT enhanced images into image block input marker sequences, including:

[0062] Transforming arterial-phase and portal-venous-phase enhanced CT images using image embedding methods may include the following steps: inputting the arterial-phase and portal-venous-phase enhanced CT images into a two-dimensional convolutional layer to obtain feature maps of the arterial-phase and portal-venous-phase enhanced CT images, each feature map containing multiple feature map patches; and adding positional embeddings to each feature map patch to obtain an image patch input label sequence. Specifically, the sliding convolution kernel of the two-dimensional convolutional layer is used to extract local features from the patient's arterial-phase and portal-venous-phase CECT images and form feature maps, each feature map being further divided to contain multiple feature map patches.

[0063] Specifically, since the prognostic prediction model cannot directly understand the two-dimensional spatial structure information between feature patches, it can first utilize labeled embeddings, such as linear layers, to convert feature patches into embedded vector representations. Then, positional embeddings are used to add positional information to the embedding vectors of the feature patches, providing the prognostic prediction model with the two-dimensional spatial positional information of the feature patches. Therefore, the embedding vectors of multiple feature patches, each with added positional embeddings, can be further merged into a sequence, i.e., an image patch input labeled sequence, as input to the prognostic prediction model.

[0064] In this way, the image patch input label sequence contains local features from its corresponding arterial phase CECT image and portal venous phase CECT image. The prognostic prediction model can use the text input label sequence and the image patch input label sequence as a unified label sequence, thereby uncovering the complex relationships between text and image modal data, improving the accuracy of the prognostic prediction model, and avoiding model overfitting.

[0065] Specifically, after adding positional embeddings to each clinical variable label sequence, the model can further include: adding a random deactivation layer to randomly discard some text input label sequence weights during each training iteration of the prognostic prediction model; and after adding positional embeddings to each feature patch, the model can further include adding a random deactivation layer to randomly discard some image patch input label sequence weights during each training iteration of the prognostic prediction model. In this way, when the prognostic prediction model is trained using both clinical variable label sequences and image patch input label sequences as input, randomly discarding some label sequences reduces its dependence on specific features of some label sequences and decreases the number of effective parameters updated during training. Therefore, it can improve the generalization performance of the prognostic prediction model on the external validation set and further avoid overfitting.

[0066] In this application, the prognostic prediction model is trained based on a deep learning model. The deep learning model constructed in this application can consist of m first neural network units and n second neural network units. In one embodiment, the first neural network unit is a transformer-based multimodal input neural network. Each text input label sequence and image patch input label sequence is input into the first first neural network unit, respectively, and passes through the first normalization layer, the first neural network layer, the second normalization layer, and the feedforward neural network therein, before entering the next first neural network unit for processing, until it passes through the last first neural network unit. Figure 1 An example of a deep learning model consisting of two first neural network units and ten second neural network units is given.

[0067] In a first neural network unit, the text input label sequence and the image patch input label sequence are respectively passed through a first normalization layer and then input into the first neural network layer. In the first neural network layer, the deep learning model learns the unimodal representations of the two input label sequences, and also learns the multimodal representations between the two input label sequences. Figure 2 The structure and learning path of a first neural network layer are illustrated, where Q, K, and V represent the first, second, and third vectors of the text input label sequence / image patch input label sequence, respectively, Matrix.Mul represents matrix multiplication, and Scaled Softmax represents the scaled Softmax function.

[0068] In one embodiment, the first neural network layer includes a vector generation layer, a similarity calculation layer, a weight layer, and a weighted layer. The vector generation layer converts each text input tag sequence and each image patch input tag sequence into a first vector, a second vector, and a third vector. The similarity calculation layer acquires and calculates the similarity value between the first vector of the text input tag sequence / image patch input tag sequence and its second vector, and between the second vector of the image patch input tag sequence / text input tag sequence. The similarity value calculation method may include calculating the cosine similarity, Euclidean distance, Pearson correlation coefficient, etc., between the first vector and the second vector.

[0069] Furthermore, the weight layer obtains the weight value of the text input label sequence / image patch input label sequence corresponding to each first vector based on the similarity value; the weighting layer, based on the obtained weight value, performs a weighted summation on the third vector of the text input label sequence and image patch input label sequence of each first vector corresponding to the weight value, and outputs text input label sequences and image patch input label sequences with different weight values, so that each text input label sequence and image patch input label sequence has a different contribution to the first neural network unit.

[0070] Therefore, in the first neural network layer, each text input token in the text input token sequence focuses not only on its similarity to other text input tokens but also on its similarity to other image patch input tokens; similarly, each image patch input token also focuses on its relationship with other image patch input tokens and text input tokens. Thus, the first neural network layer helps the deep learning model acquire comprehensive multimodal representations, enabling the trained prognostic prediction model to capture the intricate relationships between patients' clinical variables and medical images, thereby improving its prediction accuracy in complex diagnostic and treatment environments.

[0071] In this application, the first neural network unit is used to help the model better learn the relationships between different modalities, such as the complex intrinsic correlation between image lesions and laboratory parameters, and between baseline features and diagnostic results. This application first uses the first neural network unit to perform interactive learning between input data from different modalities, which the second neural network unit cannot accomplish. The labeled output with complex multimodal features obtained from the first neural network unit is then input into the second neural network unit, which can provide the prognostic prediction model of this application with better model generalization and medical data understanding capabilities.

[0072] Furthermore, the text input label sequence and image patch input label sequence processed by multiple first neural network units are unified to obtain a multimodal unified label that integrates the information from both the text input label sequence and the image patch input label sequence. After passing through the first neural network units, each label in the text input label sequence and image patch input label sequence contains information from all labels with different weights. During the training of the deep learning model, these weights are updated with the first neural network units, reflecting the interactions and dependencies between multimodal medical data. Therefore, the two label sequences can be unified, and their feature representations can be further learned using an attention mechanism. For example, multiple second neural network layers can be used to further obtain the global representation of the multimodal unified label sequence.

[0073] Specifically, in this application, each multimodal unified label sequence can be further input into a first second neural network unit and processed through a third normalization layer, a second neural network layer, a fourth normalization layer, and a feedforward neural network therein, before entering the next second neural network unit for processing, until it is processed by the last second neural network unit. Finally, the multimodal unified label sequence processed by multiple second neural network units can be input into a classification unit to convert the label sequence into a prognostic survival score as output. Specifically, the classification unit may include multiple feedforward neural networks, such as a multilayer perceptron, and includes a classifier, such as a sigmoid activation function of logistic regression, to convert the label sequence into a prognostic survival score for patients with unresectable hepatocellular carcinoma.

[0074] In this application, a loss function can be used to measure the distance between the predicted prognostic survival score and the actual prognostic survival score of patients with unresectable hepatocellular carcinoma. The loss function is minimized by minimizing this distance to update the first and second neural network units until the loss value reaches a predetermined level, thus obtaining the prognostic prediction model. For example, the label weights in the first and second neural network units can be updated using a cross-entropy loss function to obtain the prognostic prediction model.

[0075] In this application, the decision support model construction module can construct a decision support model based on a decision unit and multiple prognostic prediction models. These multiple prognostic prediction models may include a first prognostic prediction model trained on a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC; a second prognostic prediction model trained on a cohort of patients with unresectable hepatocellular carcinoma treated with TACE; a third prognostic prediction model trained on a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC combined with therapy (hereinafter referred to as HC); and a fourth prognostic prediction model trained on a cohort of patients with unresectable hepatocellular carcinoma treated with TACE combined with therapy (hereinafter referred to as TC).

[0076] It is important to note that the four prognostic prediction models can be configured to output HAIC survival scores, TACE survival scores, HC survival scores, and TC survival scores, respectively. The decision support model can predict the TACE survival score of uHCC patients multiple times, for example, predicting the first and second TACE survival scores; similarly, the decision support model can also predict the HC and TC survival scores multiple times, for example, predicting the first and second HC survival scores and the first and second TC survival scores. The survival scores predicted by the decision support model for uHCC patients can include HAIC ORR, TACE ORR, HC one-year OS, and TC one-year OS, such as... Figure 3 As shown, these serve as the basis for predicting treatment options for uHCC patients.

[0077] In this application, the decision-making unit of the decision support model can perform actions based on the four prognostic prediction models described above, such as... Figure 4 The following example illustrates the steps for optimal prognostic decision-making to obtain the optimal prognostic decision outcome for uHCC patients:

[0078] Predicting survival using the HAIC score for patients with unresectable hepatocellular carcinoma;

[0079] In response to an HAIC survival score lower than MSS1, a first TACE survival score is used to predict the survival of patients with unresectable hepatocellular carcinoma; in response to an HAIC survival score not lower than MSS1, a second TACE survival score is used to predict the survival of patients with unresectable hepatocellular carcinoma.

[0080] In response to patients with a first TACE survival score below MSS2, HAIC was performed on unresectable hepatocellular carcinoma patients, and first HC survival score prediction was performed simultaneously.

[0081] In response to a second TACE survival score not lower than MSS2, TACE was performed on patients with unresectable hepatocellular carcinoma, and the first TC survival score was predicted simultaneously; in response to a second TACE survival score lower than MSS2, the second HC survival score was predicted for patients with unresectable hepatocellular carcinoma.

[0082] In response to a first HC survival score of not less than MSS3, HC was switched to be performed for patients with unresectable hepatocellular carcinoma; in response to a first HC survival score of less than MSS3, HAIC was performed for patients with unresectable hepatocellular carcinoma.

[0083] In response to a first TC survival score of not less than MSS4, TC is performed on patients with unresectable hepatocellular carcinoma; in response to a first TC survival score of less than MSS4, TACE is performed on patients with unresectable hepatocellular carcinoma.

[0084] In response to a second HC survival score not lower than MSS3, HC is performed on patients with unresectable hepatocellular carcinoma; in response to a second HC survival score lower than MSS3, second TC survival score is used to predict the survival of patients with unresectable hepatocellular carcinoma; and

[0085] TC is performed in patients with unresectable hepatocellular carcinoma in response to a second TC survival score of not less than MSS4.

[0086] In this application, MSS1 is the median HAIC survival score of the first prognostic prediction model training set; MSS2 is the median TACE survival score of the second prognostic prediction model training set; MSS3 is the median HC survival score of the third prognostic prediction model training set; and MSS4 is the median TC survival score of the fourth prognostic prediction model training set.

[0087] Specifically, the prognostic prediction model can also output the risk group of patients under different treatment modalities. Specifically, the decision-making system of this application can use the prognostic prediction model to classify patients with HAIC survival scores no lower than MSS1, TACE survival scores no lower than MSS2, first / second HC survival scores no lower than MSS3, and first / second TC survival scores no lower than MSS4 into a low-risk patient group, and patients with the opposite survival scores into a high-risk patient group. The decision unit can further use the classification results to determine the optimal treatment prognosis for the patient. For patients whose predicted outcomes for all treatment modalities are in the high-risk patient group, the decision support model can provide the degree of risk induced by the treatment, avoiding serious problems.

[0088] In this application, HC refers to HAIC combination therapy, which may include one or more of HAIC combined with TKI therapy, ICI therapy, and local therapy; TC refers to TACE combination therapy, which may include one or more of TACE combined with TKI therapy, ICI therapy, and local therapy. TKIs and ICIs are often used as part of combination therapy, in conjunction with other major treatment modalities, such as radiotherapy and chemotherapy, to improve treatment efficacy and prolong patient survival. Combining the two main treatment modalities HAIC and TACE with the aforementioned anticancer drugs and local therapy can effectively improve prognosis; however, drug tolerance varies highly among patients, and the efficacy of combination therapy also varies from person to person.

[0089] In the overall uHCC population, HAIC and TACE combination therapy have been shown to have statistical value in improving patient survival. Therefore, when the prognosis of single therapy is poor, different forms of combination therapy are considered. However, some patient groups still do not experience survival benefits under combination therapy, and may even experience side effects. The decision support model in this application can capture the optimal path for patients among multiple treatment modalities. This means that it can directly determine the treatment plan without trying each one individually, avoiding ineffective treatment and reducing unnecessary drug side effects and the patient's economic burden. For example, if a patient is not suitable for any of the single therapies, and the decision support model finds that combination therapy also cannot produce significant efficacy, the decision unit can not recommend any treatment plan, reminding the patient not to expect subsequent targeted immunotherapy or molecularly targeted drugs to avoid potential economic losses and complications. On this basis, the decision support model can further suggest to the patient that a more aggressive treatment or prevention plan should be implemented, emphasizing that the conservative treatment plan provided by the model may lead to an increased risk of death, requiring further strengthening of monitoring and support maintenance systems.

[0090] Example

[0091] Example 1: Patient cohort design and data preprocessing

[0092] From January 2014 to November 2021, a total of 4,773 patients with untreated hepatic cerebrovascular disease (uHCC) who had undergone initial transarterial chemoembolization (TACE) or liver-associated intrahepatic chemoembolization (HAIC) at twelve tertiary hospitals underwent follow-up examinations. 1,448 registered follow-up patients from three of these hospitals were selected and assigned to a predetermined training and internal validation set for training and optimizing the prognostic prediction model; 277 patients from the other nine hospitals were assigned to an external validation set. All patients were diagnosed according to guidelines 1 and 2 of the European Association for the Study of the Liver (EASL) and the American Association for the Study of the Liver (AASLD), defined as having multiple intrahepatic metastases, distant metastases, or tumors larger than 10 cm. In addition, some suspected cases were confirmed by image-guided needle biopsy.

[0093] Specifically, all 1725 uHCC patients (mean age (52.8±11.5) years; 1529 males) underwent preoperative CECT scans, were identified as meeting the IAIT criteria, and were included in the cohort. First, the cohort was divided into two groups based on the interventional treatment method: the TACE group (816 patients) and the HAIC group (909 patients). Within these two groups, further subdivision was performed based on the combined systemic therapy received, resulting in the TC group (314 patients) and the HC group (443 patients) for those receiving systemic therapy. Among these patients, 70% / 30% were randomly assigned to the training group and the internal validation group. The remaining patients were assigned to the external validation set. The prognostic prediction models for different cohorts were reinforced using the training set to fine-tune the model parameters. This helped enhance the ability of each prognostic prediction model to identify subtle variations in complex patient heterogeneity, thereby improving the model's accuracy.

[0094] To ensure the consistency and reliability of the image segmentation algorithm and prognostic prediction model, several preprocessing steps were performed on the patient's CECT images. Specifically, one-dimensional spline interpolation was performed on all CECT images, and the slice thickness was standardized to 1 mm. In addition, the voxel intensity was limited to the range of -50 to 350 HU to eliminate non-liver and non-tumor tissues, thereby improving the sensitivity and accuracy of subsequent segmentation for further analysis.

[0095] In this embodiment, the CECT images are finally segmented using an nnU-Net model pre-trained on the Liver and Tumor Segmentation (LITS) challenge dataset. nnU-Net processes the arterial and portal venous phase CECT images of each patient in the cohort to provide liver and tumor contours. Based on the location of the largest tumor area in the image, an information slice (containing the slice with the largest tumor area) is selected from the arterial and portal venous phase CECT images of each patient; using a bounding box covering the entire liver region of the selected slice, the image is cropped to a size of 224mm*224mm and contains multiple patches to serve as the image input for the prognostic prediction model.

[0096] Example 2: Construction and Performance Verification of Prognostic Prediction Models for Four Treatment Methods

[0097] All four initial prognostic prediction models used preoperative CECT images of each patient from the four cohorts in Example 1, along with corresponding baseline and clinical parameters, as input. It is worth noting that in this example, the image segmentation results were further manually modified and corrected by two senior reviewers with over ten years of expertise in liver imaging to improve accuracy. A total of 3450 CECT slices from the four cohorts were used to train and validate the four independent prognostic prediction models.

[0098] Each prognostic prediction model was trained using a predetermined training set queue. Specifically, for each patient, demographic baselines and quantitative clinical parameters were processed to obtain the input labels, and preoperative CECT images of each patient underwent two preprocessing steps to obtain image patch input label sequences. Global features of the input data were extracted sequentially through two first neural network units and ten second neural network units, and dimensionality reduction was performed together with clinical variables to obtain a unified label sequence. The predicted patient prognostic results were then output using a logistic regression model. The model was continuously trained using the dimensionality-reduced features, and the first and second neural network units were updated using the cross-entropy loss function.

[0099] Compared to traditional CNN methods, prognostic prediction models provide prognostic predictions by uniformly processing multimodal inputs, eliminating the need for extensive fine-tuning at each data processing stage. Furthermore, the multimodal attention mechanism of prognostic prediction models provides an additional consideration of the interrelationships between multimodal inputs at a global level. In contrast, non-uniform methods, such as traditional CNN methods, only consider the relationships between global representations (i.e., the features reduced during dimensionality reduction).

[0100] In this embodiment, the prediction performance of the trained prognostic prediction model for OOR and one-year OS under different uHCC treatment regimens was examined. Its prediction performance was compared with that of a standard visual transducer (ViT) using only images as input, a pre-trained convolutional neural network ResNet-50 extracting features from CECT images, and two models built based on clinical variables considered in existing clinical decision-making protocols (Clinical Model-1: Cox survival model; Clinical Model-2: logistic regression model). The prediction evaluation parameters of the multiple models are shown in Table 1 below:

[0101] Table 1. Comparison of prediction performance between four prognostic prediction models and existing prediction models.

[0102]

[0103]

[0104] Among deep learning methods, traditional CNN methods exhibited the lowest regularization capability, with the lowest AUC on the external validation set in the TC, HAIC, and HC groups (TACE: 0.695 [95% CI, 0.621, 0.762], p = 0.44; HAIC: 0.718 [95% CI, 0.611, 0.809], p = 0.286; TC: 0.637 [95% CI, 0.498, 0.761], p = 0.218; HC: 0.685 [95% CI, 0.528, 0.817], p = 0.534). In model comparisons across all four treatment modalities, the prognostic prediction model clearly demonstrated the highest performance compared to other deep learning methods, especially on the external validation set. In the internal and external validation sets, the AUCs for the ORR of the TACE prognostic prediction model were 0.788 (95% CI, 0.707, 0.856) and 0.761 (95% CI, 0.693, 0.820), respectively. In the internal and external validation sets, the AUCs for the ORR of the HAIC prognostic prediction model were 0.854 (95% CI, 0.791, 0.904) and 0.805 (95% CI, 0.707, 0.881), respectively. In the internal and external validation sets, the AUCs for the OS of the TC prognostic prediction model were 0.811 (95% CI, 0.677, 0.907) and 0.736 (95% CI, 0.608, 0.841), respectively. In the internal and external validation sets, the AUCs for overall survival (OS) of the HC prognostic prediction models were 0.837 (95% CI, 0.737, 0.910) and 0.748 (95% CI, 0.599, 0.865), respectively. Finally, the results, validated by the DeLong test, showed that each prognostic prediction model significantly outperformed the commonly used clinical model-1 in all validation cohorts (P<0.05).

[0105] Furthermore, in the HAIC, TACE, and TC cohorts, there were no significant differences in ORR and OS calculated from the training set, internal validation set, and external validation set of the prognostic prediction models (p>0.05), indicating that the models have good generalization ability. In the HC cohort, there was a significant difference between the training set and internal validation set (p=0.019), but no significant difference between the training set and external validation set, further validating the usability of the prognostic prediction models in real-world medical scenarios. In the training datasets of the four prognostic prediction models, the median survival scores (MSS) for the HAIC, TAIC, HC, and TC cohorts were 0.3368 (MSS1), 0.4035 (MSS2), 0.6458 (MSS3), and 0.6130 (MSS4), respectively. These parameters will serve as auxiliary values ​​for the decision-making unit to determine the optimal treatment plan.

[0106] Example 3: Construction and Performance Verification of Decision Support Model

[0107] In Example 2, a decision support model was constructed using four independent prognostic prediction models and decision units obtained from training four cohorts. Based on the survival score calculated by each prognostic prediction model, a comprehensive auxiliary treatment decision was provided for uHCC patients.

[0108] Specifically, the survival score of patients given by the prognostic prediction model is compared with the median survival score of the training set cohort of the prognostic prediction model. Based on whether the patient's survival score reaches the median survival score, patients are divided into a high-risk group (patient survival score not reaching the median survival score) and a low-risk group (patient survival score reaching or exceeding the median survival score). In this embodiment, the decision support model further provides the decision results, namely, the patient's OOR / one-year OS. When the recommendation is consistent with the doctor's actual treatment choice, the patient's overall treatment response and one-year survival rate are significantly improved compared to when the decision support model's recommendation is inconsistent with the doctor's actual treatment choice. Simultaneously, in the group where the decision support model does not recommend IAIT (i.e., does not recommend any treatment option), the response rate and survival rate are also significantly lower than those in the group where the decision is consistent. Table 2. Figure 5A and Figure 5B We statistically analyzed the responses and survival benefits of patient groups in an external validation set when the predicted decision and the actual decision were consistent.

[0109] Table 2 shows the actual objective response rate / one-year overall survival rate for different cohorts where the decision support model and clinicians' recommendations are consistent, and the actual overall response rate / overall one-year survival rate for cohorts where the decision support model's recommendations are inconsistent or not recommended.

[0110]

[0111] For both the TACE and HAIC groups, the ORR was significantly higher when the actual treatment strategy aligned with the predictions of the decision support model (p<0.001). A significant difference was observed when the decision support model recommended against IAIT compared to when the actual treatment strategy aligned with the decision support model's recommendation. The TC group also showed a similar trend, with a significantly higher OS rate when the decision support model's recommendation aligned with the actual treatment decision. For the HC group, a larger patient sample size is needed to demonstrate statistical significance in proving its predictive accuracy.

[0112] Furthermore, compared to patients receiving only IAIT, those receiving combination therapy had longer survival, indicating a statistically significant improvement in overall patient survival. This suggests that for high-risk uHCC patients, TACE or HAIC combination therapy should be recommended as a first-line treatment, prolonging overall survival (OS). Finally, for patients predicted to experience out-of-orthognathic disease (OOR) after initial IAIT alone, the decision support model can remind them that additional targeted immunotherapy or molecularly targeted drugs are unnecessary, avoiding potential financial losses and complications. For high-risk patients identified by the prognostic prediction model, and for patients recommended against any IAIT, the decision support model can advise timely adjustments to other treatment options, postoperative prophylaxis, or monitoring strategies, emphasizing the need for intensive monitoring and other types of adjuvant systemic therapy due to their higher risk of death.

[0113] In summary, the prognostic prediction model proposed in this application will serve as an immediately implementable clinical tool to aid in prognostic diagnosis, providing risk stratification and accurate prognosis for physicians and patients. Furthermore, the application of decision support models, which manipulate the survival scores derived from individual prognostic prediction models through decision paradigm manipulation, represents a promising potential method for optimizing personalized treatment plans in the prognostic management of uHCC patients.

Claims

1. A prognostic decision-making system for unresectable hepatocellular carcinoma, the decision-making system comprising: The sample data preprocessing module acquires clinical text variables from patients with unresectable hepatocellular carcinoma and converts them into text input marker sequences. Preoperative enhanced CT images of patients with unresectable hepatocellular carcinoma were acquired, and the preoperative enhanced CT images were segmented into arterial phase enhanced CT images and portal venous phase enhanced CT images and converted into image block input label sequences. The prognostic prediction model construction module inputs the text input label sequence and the image patch input label sequence into a deep learning model to capture the global relationship between the text input label sequence and the image patch input label sequence, thereby obtaining a prognostic prediction model based on the deep learning model to predict the patient's survival score. The deep learning model is constructed by... m The first neural network unit and n The system consists of several second neural network units. Each text input tag sequence and each image patch input tag sequence is input into the first first neural network unit and processed by the first normalization layer, the first neural network layer, the second normalization layer, and the feedforward neural network therein, respectively, before entering the next first neural network unit for processing, until the second first neural network unit is reached. m The first neural network unit processes the data, and then... m The text input label sequence and the image block input label sequence processed by the first neural network unit are subjected to unified labeling processing to obtain a multimodal unified label sequence that integrates text input label sequence information and image block input label sequence information; A decision support model construction module includes a decision unit, which executes an optimal prognostic decision based on multiple prognostic prediction models; the decision support model construction module obtains multiple prognostic prediction models through training multiple cohorts of unresectable hepatocellular carcinoma patients, and constructs a decision support model based on the decision unit and the multiple prognostic prediction models to provide optimal prognostic decision results to patients with unresectable hepatocellular carcinoma; The decision support model construction module trains multiple prognostic prediction models through multiple cohorts of patients with unresectable hepatocellular carcinoma, including: a first prognostic prediction model trained through a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC, used to obtain the HAIC survival score; a second prognostic prediction model trained through a cohort of patients with unresectable hepatocellular carcinoma treated with TACE, used to obtain the first TACE survival score and the second TACE survival score; a third prognostic prediction model trained through a cohort of patients with unresectable hepatocellular carcinoma treated with HAIC combined therapy, used to obtain the first HC survival score and the second HC survival score; and a fourth prognostic prediction model trained through a cohort of patients with unresectable hepatocellular carcinoma treated with TACE combined therapy, used to obtain the first TC survival score and the second TC survival score. The HAIC survival score, the first TACE survival score, and the second TACE survival score represent the objective response rate for patients with unresectable hepatocellular carcinoma; the first HC survival score, the second HC survival score, the first TC survival score, and the second TC survival score represent the one-year overall survival for patients with unresectable hepatocellular carcinoma. The decision-making unit performs optimal prognostic decision-making based on multiple prognostic prediction models, including: predicting the HAIC survival score for patients with unresectable hepatocellular carcinoma; predicting the first TACE survival score for patients with unresectable hepatocellular carcinoma in response to the HAIC survival score being lower than MSS1; predicting the second TACE survival score for patients with unresectable hepatocellular carcinoma in response to the HAIC survival score being not lower than MSS1; predicting the first HC survival score for patients with unresectable hepatocellular carcinoma in response to the first TACE survival score being lower than MSS2; predicting the first TC survival score for patients with unresectable hepatocellular carcinoma in response to the second TACE survival score being not lower than MSS2; and predicting the first TC survival score for patients with unresectable hepatocellular carcinoma in response to the second TACE survival score being lower than MSS2. The second HC survival score is used for prediction; in response to the first HC survival score being not lower than MSS3, HC is performed on patients with unresectable hepatocellular carcinoma; in response to the first HC survival score being lower than MSS3, HAIC is performed on patients with unresectable hepatocellular carcinoma; in response to the first TC survival score being not lower than MSS4, TC is performed on patients with unresectable hepatocellular carcinoma; in response to the first TC survival score being lower than MSS4, TACE is performed on patients with unresectable hepatocellular carcinoma; in response to the second HC survival score being not lower than MSS3, HC is performed on patients with unresectable hepatocellular carcinoma; in response to the second HC survival score being lower than MSS3, a second TC survival score is used to predict the survival of patients with unresectable hepatocellular carcinoma; in response to the second TC survival score being not lower than MSS4, TC is performed on patients with unresectable hepatocellular carcinoma. MSS1 is the median HAIC survival score in the training set of the first prognostic prediction model; MSS2 is the median TACE survival score in the training set of the second prognostic prediction model; MSS3 is the median HC survival score in the training set of the third prognostic prediction model; MSS4 is the median TC survival score in the training set of the fourth prognostic prediction model; HC is HAIC combination therapy, which includes HAIC treatment combined with tyrosine kinase inhibitor therapy, immune checkpoint inhibitor therapy, and / or local therapy; TC is TACE combination therapy, which includes TACE treatment combined with tyrosine kinase inhibitor therapy, immune checkpoint inhibitor therapy, and / or local therapy.

2. The decision-making system according to claim 1, wherein, The clinical text variables include baseline variables, laboratory variables, and imaging variables; The baseline variables include age, sex, medical history, ECOG classification, complications, HBV, ascites, and continuous local treatment; The laboratory variables include ALB, TBIL, ALT, AST, PT, INR, PLT, and AFP; The imaging variables include Barcelona clinical stage of liver cancer, diagnosis of portal vein tumor thrombosis, tumor metastasis, tumor size, number of tumors, vascular invasion, tumor metastasis, and Child-Pugh classification.

3. The decision-making system according to claim 1, wherein: The multimodal unified label sequence is further input into the first second neural network unit and processed through the third normalization layer, the second neural network layer, the fourth normalization layer, and the feedforward neural network therein, before entering the next second neural network unit for processing, until the [number missing]th ... n Each of the second neural network units processes; Will pass n The multimodal unified label sequence processed by the second neural network unit is passed through a classification unit and converted into a prognostic survival score for patients with unresectable hepatocellular carcinoma; The distance between the prognostic survival score of the patient with unresectable hepatocellular carcinoma and the true prognostic survival score of the patient with unresectable hepatocellular carcinoma is measured by a loss function. The loss function is minimized to update the first neural network unit and the second neural network unit until the loss value reaches a specified level and the prognostic prediction model is obtained.

Images