An artificial intelligence method and system for early risk assessment of ovarian cancer

Abstract

The application relates to the application of artificial intelligence technology in the medical and health field, in particular to an artificial intelligence method and system for early risk assessment of ovarian cancer, which is implemented by a computer device. The modal quality evaluation network directly reflects the reliability of the data at the moment by giving each kind of data at each time point a quantitative quality score. Then, in the multi-modal data fusion model, the quality score is used as an adjustment factor to modulate the attention weight, so that the model can intelligently adjust the weight of different data sources at different time points, automatically enhance the influence of high information quantity data in the fusion process, and suppress the interference of low quality or noise data, significantly enhance the representation ability and signal-to-noise ratio of the fused features, and thus directly improve the accuracy and robustness of the final risk assessment result.

Description

Technical Field

[0001] This application relates to the application of artificial intelligence technology in the field of medical and health care, and in particular to an artificial intelligence method and system for early risk assessment of ovarian cancer. Background Technology

[0002] Ovarian cancer is one of the three major malignant tumors of the female reproductive system. Its early symptoms are not obvious and there is a lack of effective screening methods, which leads to most patients being diagnosed at an advanced stage with a very poor prognosis. With the development of artificial intelligence technology, the use of machine learning or deep learning models to integrate patients' multi-dimensional medical data to achieve early risk identification of ovarian cancer has become a research hotspot in this field. Generally, algorithms automatically learn the complex nonlinear patterns contained in massive amounts of data to build predictive models that surpass traditional medical scoring systems or single biomarker detection.

[0003] However, after in-depth analysis of the existing technology, the applicant found that it failed to fully consider the dynamic nature and heterogeneity of clinical data. Typically, in the data preprocessing stage, features from different data sources are simply concatenated into vectors, or a set of weights fixed after model training is used for weighted summation. Clinically, the quality of data collected at different time points varies significantly. For example, an ultrasound image may contain artifacts due to insufficient patient preparation or incorrect equipment settings, resulting in decreased image quality; a blood test sample may have slight hemolysis, leading to deviations in the measurement values ​​of some indicators; the depth or coverage of gene sequencing may also vary from batch to batch. The existing technology cannot dynamically identify and quantify these differences in data quality, and therefore cannot adaptively adjust the contribution of the data to the model's decision-making process based on the actual reliability of the data at a specific time point. Summary of the Invention

[0004] The objective of this application is to provide an artificial intelligence method for early risk assessment of ovarian cancer. The method is implemented by a computer device and includes the following steps performed by the computer device: acquiring at least one longitudinal medical record of a subject to be assessed, the longitudinal medical record containing multimodal time-series data collected at multiple time points; the multimodal time-series data includes at least two or more modal data selected from medical imaging data, genomic data, and clinical time-series test index data; aligning the multimodal time-series data according to the acquisition timestamps to form a data sequence associated with a time axis; processing each modal data at each time point in the data sequence using a pre-trained modality quality assessment network; the modality quality assessment network receives the modality data itself and its associated metadata as input, the metadata including but not limited to image resolution, sequencing depth, or data integrity markers; through the processing of the modality quality assessment network, generating a modality quality score for each modality data at each time point to quantify its data reliability and information content at that time point; and processing the data sequence... For each modality of data at each time point, a modality-specific encoder matching the data type is used to process and convert it into a high-dimensional modality feature vector. Different types of modality data correspond to modality-specific encoders with different structures. A multimodal data fusion model based on a temporal attention mechanism is constructed, and the modality feature vectors from multiple time points are input into the multimodal data fusion model. Within the multimodal data fusion model, when calculating the contribution of each modality feature vector to the fusion result at different time points through its temporal attention mechanism, the corresponding modality quality score is used as a dynamic adjustment factor to modulate the attention weight. Finally, a unified fusion feature vector is generated for each time point in the data sequence, forming a unified fusion feature vector sequence. The unified fusion feature vector sequence or a single vector aggregated from the unified fusion feature vector sequence is input into an artificial intelligence risk assessment model. Through the processing of the artificial intelligence risk assessment model, a risk assessment value representing the current or specified time point of the assessed object having ovarian cancer is calculated and output.

[0005] By adopting the above technical solution, the modal quality assessment network assigns a quantified quality score to each type of data at each time point. This quantified quality score directly reflects the reliability of the data at that moment. Subsequently, in the multimodal data fusion model, this quality score is used as a modulating factor to modulate the attention weights, enabling the model to dynamically adjust the weights of different data sources at different time points. During the fusion process, it automatically enhances the influence of high-information data while suppressing the interference of low-quality or noisy data, thereby enhancing the representation ability and signal-to-noise ratio of the fused features and directly improving the accuracy and robustness of the final risk assessment results.

[0006] Optionally, the method further includes the following steps: inputting the unified fused feature vector sequence into a hierarchical time series prediction model in chronological order; the hierarchical time series prediction model includes a bottom-level time series network for capturing short-term feature patterns and a top-level time series network for learning long-term risk evolution trends; and generating a prediction result representing the risk evolution trend of the object to be evaluated at one or more future time points through the processing of the hierarchical time series prediction model.

[0007] By adopting the above technical solution, the hierarchical time series prediction model captures fluctuations in short-term characteristic patterns through the bottom layer network and learns the macro trend of long-term risk evolution through the top layer network. This enables the prediction of the patient's risk change trajectory over a future period, providing clinicians with highly valuable prospective information that allows them to anticipate the direction of risk evolution. This gives them the opportunity to take more proactive intervention measures, such as adjusting the follow-up frequency or initiating preventive treatment, rather than simply reacting to the current risk status.

[0008] Optionally, the modality-specific encoder specifically includes: for medical imaging data, a convolutional neural network as its modality-specific encoder; for genomics data, a graph neural network or an autoencoder as its modality-specific encoder; and for clinical time-series test index data, a recurrent neural network or a one-dimensional convolutional network as its modality-specific encoder.

[0009] By adopting the above technical solutions and assigning dedicated encoders with different structures to different types of modal data, the unique semantic features within each modality can be preserved and extracted to the maximum extent. For example, using convolutional neural networks to process medical image data can effectively capture its spatial structure and texture features; using graph neural networks or autoencoders to process genomic data can uncover the interactions or expression patterns between genes; and using recurrent neural networks to process clinical time-series test indicators can capture the dynamic patterns of indicator changes over time. Compared with using a single general encoder or simple preprocessing methods, this provides richer information and more accurate representations for subsequent fusion and evaluation steps.

[0010] Optionally, the step of aligning the multimodal time-series data according to the acquisition timestamp to form a data sequence associated with the time axis specifically includes: when performing time-series alignment, filling in missing data points in the data sequence using a preset interpolation algorithm and generating a missing mask representing the missing data position; the missing mask, as an additional feature, is input together with the corresponding modal data into the modality-specific encoder for processing.

[0011] By adopting the above technical solution, and by imputing missing data and generating a missing mask during the data acquisition stage, the problem of missing values ​​that are common in real-world data can be handled more elegantly. The imputation algorithm fills in the data sequence, ensuring the integrity of the model input, while the missing mask, as an additional feature input, clearly tells the model which data points are original and which are inferred. This allows the modality-specific encoder to learn the patterns that the missing data itself may contain, avoiding the noise introduced by improper imputation from misleading the model and enhancing the model's ability to process incomplete data.

[0012] Optionally, after generating the risk assessment value or risk evolution trend prediction result, the method further includes a step of performing multi-level interpretability analysis, specifically including: applying a preset attribution algorithm to calculate the contribution value of the multimodal time series data to the finally generated risk assessment value or risk evolution trend prediction result; aggregating the contribution values ​​of all original modal data originating from the same modality to obtain the overall contribution of each modality data to the risk assessment value or risk evolution trend prediction result; and cross-validating the contribution value of each modality data by analyzing the attention weight distribution of the time series attention mechanism; and analyzing the hierarchical time series prediction model to identify and highlight one or more historical time points that have a decisive impact on the risk evolution trend prediction result.

[0013] By adopting the above technical solutions and introducing multi-level interpretability analysis, the transparency and clinical credibility of the model decision-making process are improved. Feature-level attribution analysis can clearly identify which specific biomarkers or imaging features lead to high-risk predictions; modality-level attribution analysis can quantify the overall contribution of different data sources, helping doctors understand the model's dependence on different examinations; and time-series attribution analysis can pinpoint the key time points that have a decisive impact on the risk evolution trend, enabling doctors to understand, verify, and even question the model's output, thereby establishing trust in the model's results and promoting its practical application in clinical decision-making.

[0014] Optionally, the method further includes the following steps: integrating the results of feature-level attribution analysis, modality-level attribution analysis, and time-series attribution analysis to generate an integrated interpretability report; the interpretability report includes at least a waterfall chart to show the contribution of key risk factors, a pie chart or bar chart to show the overall contribution of each modality, and a visualization chart that highlights key impact time points on the risk evolution curve.

[0015] By adopting the above technical solutions, complex attribution data is transformed into visual charts and summaries that are easy for clinicians to understand. For example, waterfall charts intuitively show the positive and negative contributions of each risk factor, pie charts clearly compare the overall impact of each modality, and highlighted risk evolution curves mark key turning points in the disease process. The analysis results of the model are presented in a highly integrated and user-friendly manner, reducing the cognitive load on doctors interpreting the model output, enabling them to quickly grasp the key points, and providing a direct and actionable basis for formulating subsequent diagnosis and treatment plans.

[0016] Optionally, the specific method of the time-series attribution analysis is as follows: by calculating the gradient of the risk evolution trend prediction result with respect to the unified fusion feature vector sequence, or by analyzing the time attention weights inside the hierarchical time-series prediction model, to identify the key time points that have a significant impact on the risk evolution trend prediction result.

[0017] By adopting the above technical solutions, and by calculating the gradient of the prediction result with respect to the input sequence, or by directly analyzing the time attention weights within the model, it is possible to quantitatively identify which past time points or events have the greatest impact on future risk prediction.

[0018] Optionally, the risk evolution trend prediction result specifically involves: outputting a risk probability distribution for at least one preset time point in the future, and calculating and outputting a confidence interval corresponding to the risk probability distribution based on a preset statistical method.

[0019] By adopting the above technical solutions, a quantitative measure of uncertainty is provided for the prediction results. The confidence interval can tell the doctor how confident the model is in predicting the future. A wider confidence interval may mean that the current data is insufficient to make a very certain prediction, prompting the doctor to collect more information or adopt a more conservative observation strategy. This honest expression of uncertainty enables clinical decisions to be based on a more comprehensive risk assessment.

[0020] Optionally, the multimodal data fusion model based on the temporal attention mechanism is specifically implemented as a multi-head attention module.

[0021] By adopting the above technical solution, the current advanced sequence processing architecture is used to perform the fusion task. The self-attention mechanism can capture the long-distance dependency between any two positions in the sequence, so that the model can not only focus on the correlation between different modalities at the same time point, but also capture the complex cross-influence between different time points or different modalities at the same time point.

[0022] The second objective of this application is to provide an artificial intelligence system for early risk assessment of ovarian cancer, comprising multiple functional modules configured to perform the above-described methods. These functional modules include at least: a data acquisition and processing module for acquiring and aligning multimodal time-series data from multiple time points for at least one subject to be assessed; a modality quality assessment module for generating a modality quality score that quantifies the reliability and information content of each modality data point; a feature representation module for processing each modality data into a modality feature vector; an adaptive fusion module for adaptively fusing the modality feature vector sequence into a unified fused feature vector sequence based on a temporal attention mechanism and in conjunction with the modality quality score; and a risk assessment module for processing and outputting an early risk assessment value characterizing the subject to be assessed as having ovarian cancer based on the unified fused feature vector sequence. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the artificial intelligence method for early risk assessment of ovarian cancer in this application.

[0024] Figure 2 This is a block diagram of the artificial intelligence system for early risk assessment of ovarian cancer in this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] like Figure 1 As shown, this application provides an artificial intelligence method and system for early risk assessment of ovarian cancer, which is implemented by computer equipment, such as one or more servers configured with a central processing unit, a graphics processing unit, memory and storage, or a cloud computing platform. The method includes the following steps performed by the computer equipment.

[0027] S01: Obtain at least one longitudinal medical record of the subject to be evaluated. The longitudinal medical record contains multimodal time-series data collected at multiple time points. The multimodal time-series data includes at least two or more modal data selected from medical imaging data, genomic data, and clinical time-series test index data. The multimodal time-series data is then time-aligned according to the collection timestamp to form a data sequence associated with the time axis.

[0028] Understandably, computer equipment can connect to hospital information systems, laboratory information systems, image archiving and communication systems, and electronic medical record systems via interfaces to automatically or semi-automatically extract historical medical data of designated individuals to be evaluated.

[0029] Specifically, medical imaging data can include, but is not limited to, digital images obtained through various imaging techniques; for example, transvaginal or transabdominal ultrasound images, which can be two-dimensional B-mode grayscale images, Doppler ultrasound images containing blood flow information, or elastography maps assessing tissue stiffness; in addition, it can also include pelvic computed tomography (CT) images, such as plain CT images, contrast-enhanced CT images, and radiomics feature sets extracted from CT images; another important type of image is magnetic resonance imaging (MRI) images, such as T1-weighted imaging, T2-weighted imaging, diffusion-weighted imaging, and sequences of dynamic contrast-enhanced MRI images; each type of imaging data is acquired in its specific format, such as DICOM format, and contains rich metadata, such as scan parameters, slice thickness, equipment model, etc.

[0030] Genomic data provides information at the molecular level. It can be one-time whole-exome sequencing or whole-genome sequencing data. Its original form may be a FASTQ file, which is then compared and variant detected to generate a VCF file, recording the gene mutation information of the object to be evaluated, especially germline or somatic mutations of genes related to ovarian cancer, such as BRCA1, BRCA2, TP53, etc. It can also be targeted sequencing data, focusing on known cancer-related gene sets. In addition, genomic data can also include gene expression profile data, such as the expression levels of thousands of genes in specific tissues (e.g., tumor tissue or blood samples) obtained through RNA sequencing or gene chips. These data are usually in the form of a matrix, with rows representing genes, columns representing samples, and values ​​representing expression levels.

[0031] Clinical time-series test data reflect the changes in the physiological state of the subject being evaluated over time. Clinical time-series test data includes, but is not limited to, the detection values ​​of a series of serum tumor markers, such as cancer antigen 125, human epididymal protein 4, carcinoembryonic antigen (CEA), alpha-fetoprotein, and lactate dehydrogenase. The detection values ​​of these indicators are recorded with the patient's regular follow-up examinations, forming a time series. In addition to tumor markers, it may also include various indicators from routine examinations such as complete blood count and comprehensive biochemistry, such as white blood cell count, hemoglobin concentration, liver function indicators, and kidney function indicators. These clinical time-series test data are usually stored in a database in the form of key-value pairs of numerical values ​​and collection time.

[0032] After acquiring the aforementioned heterogeneous data, time-series alignment is required. The purpose of time-series alignment is to unify these data collected at different times and at different frequencies onto a common timeline. For example, a fixed time window can be set, such as every three months. Then, all the data of each object to be evaluated can be assigned to the corresponding time window according to its original collection timestamp. If there are multiple similar detections within a time window, a specific strategy can be adopted for merging, such as taking the average value, the maximum value, or the last detection value.

[0033] It is understandable that when performing temporal alignment, for missing data points in the data sequence, an interpolation algorithm can be used to fill them and generate a missing mask representing the location of the missing data. The missing mask can be used as an additional feature and input into the modal encoder along with the corresponding modal data for processing.

[0034] Understandably, data gaps frequently occur in real-world clinical scenarios. For example, a patient might miss a follow-up appointment, resulting in a missing cancer antigen 125 value at a certain time point. To construct a complete data sequence for subsequent model processing, these missing data points need to be imputed. Imputation algorithms can be simple, including forward imputation (filling missing values ​​with the previous valid observation) or backward imputation (filling with the next valid observation); or more complex algorithms, including linear imputation (estimating missing values ​​by linear interpolation between two valid observations) or spline imputation (fitting a smoother curve); for multivariate time series, model-based imputation methods can be used, such as using other relevant variables to predict the value of the missing variable, or more advanced statistical methods such as multiple imputation.

[0035] Understandably, the missing data mask is a binary matrix or vector with the same dimensions as the original data. A value of 1 indicates that the data point is original, while a value of 0 indicates that the data point was generated by an imputation algorithm. This missing data mask serves as additional information and is input into the subsequent network along with the imputed data. The advantage of this is that the model can not only utilize the imputed data but also clearly identify which data is real and which is speculative. This allows the model to learn the information that the missing data itself may carry. For example, frequent missing data may be related to patient compliance or certain specific clinical conditions, and the model can learn to utilize this pattern.

[0036] S02: For each modality of data at each time point in the data sequence, a pre-trained modality quality assessment network is used for processing. The modality quality assessment network receives the modality data itself and its associated metadata as input, including but not limited to image resolution, sequencing depth, or data integrity markers. Through the processing of the modality quality assessment network, a modality quality score is generated for each modality of data at each time point to quantify its data reliability and information content at that time point.

[0037] Specifically, a modal quality score can be dynamically generated for each piece of input data. The modal quality score is a normalized value between 0 and 1. The higher the value, the better the data quality, the richer the information, and the higher the reliability.

[0038] A modality quality assessment network can be a lightweight neural network whose structure varies depending on the type of modality being assessed; for example, it could be a small multilayer perceptron or a shallow convolutional neural network. This modality quality assessment network needs to be pre-trained, and the training data can be a batch of quality-labeled data obtained through manual annotation or automatic evaluation; for example, radiologists can be asked to score a batch of ultrasound images for aspects such as sharpness and artifact levels, which can be used as training labels.

[0039] The input to a modal quality assessment network is multifaceted, including the modal data itself or its simplified representation; for example, for a CT image, a thumbnail of the image can be directly input.

[0040] Another part is the metadata associated with the data; for example, for medical imaging data, the associated metadata may include: image resolution, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), slice thickness parameters in the scanning protocol, scanning equipment model and manufacturing year, and markers indicating the presence of artifacts. The SNR can be obtained by calculating the ratio of the signal standard deviation to the noise standard deviation over a uniform region of the image. Artifacts include motion artifacts, metal artifacts, etc., which can be detected by preprocessing algorithms or extracted from radiology reports. For genomics data, the associated metadata may include: sequencing depth, alignment rate, Q30 ratio, and data integrity markers. Sequencing depth is the average number of times each base in the genome is sequenced; higher depth indicates greater reliability. The alignment rate indicates the success rate of sequencing reads. The percentage of data aligned to the reference genome, specifically the Q30 ratio (the percentage of bases with a quality value greater than or equal to 30 in the sequencing data), is a key indicator of sequencing quality. Data integrity markers include whether certain regions of key genes were not successfully covered in targeted sequencing. For clinical time-series test data, associated metadata may include: data integrity markers, laboratory-provided sample status markers, and the time interval since the last test. Too long or too short a time interval may mean different reference values ​​for the data. Among these, the data integrity marker can be the percentage of items actually completed in this test within a preset set of indicators, and the laboratory-provided sample status markers include whether the sample has hemolysis, lipemia, etc., and the time interval since the last test.

[0041] Understandably, the modal quality assessment network processes these inputs, and its output layer uses the Sigmoid activation function to constrain the output values ​​between 0 and 1, forming the final modal quality score. Through this step, each data point in the data sequence is assigned an independent quality score.

[0042] S03: For each modal data at each time point in the data sequence, the modality-specific encoder that matches the data type of the modality is used to process and convert it into a high-dimensional modal feature vector; different types of modal data correspond to dedicated encoders with different structures in order to extract deep semantic features within each modality.

[0043] Understandably, the heterogeneous raw data, after preprocessing and quality assessment, is transformed into a unified, information-condensed, high-dimensional vector representation, namely, a modality feature vector. To maximize the extraction of information unique to each data type, a modality-specific encoder can be used. Specifically, for medical imaging data, a convolutional neural network can be used as its modality-specific encoder; for genomics data, a graph neural network or an autoencoder can be used as its modality-specific encoder; and for clinical time-series test index data, a recurrent neural network or a one-dimensional convolutional network can be used as its modality-specific encoder.

[0044] Specifically, for medical image data, the modality-specific encoder can be a deep convolutional neural network. For example, mature architectures such as ResNet-50, VGG-19, or EfficientNet, which are pre-trained on large natural image datasets, can be used. The network can be fine-tuned on medical image data through transfer learning. The input is pre-processed image slices or 3D volume data. Through its multiple layers of convolutional layers, pooling layers, and activation functions, the convolutional neural network can automatically learn hierarchical features in the image, from low-level edge and texture features to high-level semantic features related to lesion morphology. The output of the last few fully connected layers or global average pooling layers of the network can be used as a fixed-dimensional high-dimensional modality feature vector, such as 512-dimensional or 1024-dimensional.

[0045] For genomic data, the choice of encoder depends on the specific form of the data. If the data is gene mutation information, it can be represented as a multi-hot vector and then encoded using a multilayer perceptron. If the data is a gene expression profile, an autoencoder can be used. An autoencoder consists of an encoder and a decoder, and its input is reconstructed through training. The output of its bottleneck layer is a condensed representation of the original high-dimensional expression profile data, i.e., a modal feature vector. Furthermore, if information about gene interactions can be obtained, such as protein-protein interaction networks obtained from public resources like the STRING database, a graph neural network can be constructed. In this graph neural network, each gene is a node, and the interactions between genes are edges. Gene expression levels or mutation states are used as the initial features of the nodes. By propagating and aggregating neighbor node information on the graph, the graph neural network can learn a feature representation containing the topological structure of the biological network. This representation is better at capturing the patterns of gene synergy than a simple vector.

[0046] For clinical time-series laboratory test data, since it is inherently a time-ordered sequence, it is well-suited for models capable of processing sequential data. For example, recurrent neural networks, especially their variants Long Short-Term Memory (LSTM) networks or gated recurrent units, can be used. Taking a vector sequence of multiple indicators (such as cancer antigen 125, human epididymal protein 4, etc.) within a time window (e.g., a quarter) as input, the LSTM network or gated recurrent unit, through its internal gating mechanism including input gate, forget gate, and output gate, can effectively capture the dynamic trends and dependencies of these indicators over time. After processing the sequence, the hidden state of the last time step can serve as the modal feature vector of the clinical laboratory test indicators within that time period. Alternatively, a one-dimensional convolutional network can be used; by sliding the convolution kernel along the time dimension, local patterns and trends in the indicator sequence can also be extracted.

[0047] Understandably, the missing mask can be used as an additional feature and input into the modality-specific encoder along with the corresponding modality data. Specifically, the missing mask and the original data can be concatenated along the feature dimension and then fed into the encoder together. In this way, the encoder can utilize the information of whether the data is missing while learning the feature representation.

[0048] S04: Construct a multimodal data fusion model based on a temporal attention mechanism, and input modal feature vectors from multiple time points into the multimodal data fusion model; within the multimodal data fusion model, when calculating the contribution of each modal feature vector to the fusion result at different time points through its temporal attention mechanism, use the corresponding modal quality score as a dynamic adjustment factor to modulate the attention weight; finally, generate a unified fusion feature vector for each time point in the data sequence, and form a unified fusion feature vector sequence.

[0049] Understandably, for an object to be evaluated, assuming we have M modalities and observe T time points, we obtain M×T modal feature vectors, which constitute the input of the fusion model.

[0050] It is understandable that the multimodal data fusion model based on the temporal attention mechanism can be specifically implemented as a multi-head attention module based on the Transformer architecture. The core of Transformer is the self-attention mechanism, which can calculate the importance of each element in the sequence to all other elements. In the scenario of this application, we regard all modal feature vectors of different modalities and different time points as a large sequence. The specific fusion process includes the following steps.

[0051] All modal feature vectors are concatenated into a long sequence. For example, image feature vectors, gene feature vectors, and clinical indicator feature vectors are concatenated into a long sequence. In order for the model to distinguish the source of each vector, position encoding and modality encoding can be added to it. Position encoding tells the time point corresponding to the vector, and modality encoding tells the data modality to which the vector belongs.

[0052] In a standard self-attention mechanism, each input vector generates a query vector, a key vector, and a value vector. By calculating the dot product of a query vector with all key vectors and then passing it through the Softmax function, the attention weights of the query vector with respect to all value vectors can be obtained.

[0053] Understandably, after calculating the dot product of the query vector and the key vector, but before applying the Softmax function, the original attention score matrix can be multiplied element-wise with the modulation matrix. This modulation matrix can be composed of modality quality scores. Specifically, if the attention score score(i, j) represents the attention of vector i to vector j, and the modality quality score of vector j is q(j), then the modulated attention score'(i, j) is equal to score(i, j) multiplied by q(j).

[0054] The modulated attention score is input into the Softmax function for normalization to obtain the final attention weight. Then, this weight is used to sum all the value vectors to obtain the output vector at that position.

[0055] Through the above modulation process, if the original data quality of a modal feature vector (e.g., j) is very low, its corresponding modal quality score q(j) will be very small, for example, close to 0. Then, when calculating the attention weight, no matter how high the original attention of other vectors is to it, after multiplying by this small quality score, its final attention weight will be significantly suppressed. Conversely, if the data quality of a modal feature vector is very high, its quality score q(j) is close to 1, and its influence will be maintained or even amplified. By dynamically controlling the flow of information according to the data quality, the focus on high-quality information and the ignoring of low-quality information are achieved.

[0056] Multi-head attention mechanisms further enhance this process by allowing multiple independent attention computation processes (i.e., heads) to run in parallel. Each head learns to focus on different aspects of the input sequence, and finally, the outputs of all heads are concatenated and subjected to a linear transformation to obtain the final fused representation.

[0057] After processing by the quality-aware adaptive temporal fusion model, for each time point in the input sequence, a highly condensed unified fusion feature vector can be output. This unified fusion feature vector integrates the effective information of all modalities of that time point and its context time points that has been quality-filtered.

[0058] S05: Input the unified fusion feature vector sequence or a single vector aggregated from the unified fusion feature vector sequence into the artificial intelligence risk assessment model. Through the processing of the artificial intelligence risk assessment model, calculate and output the risk assessment value used to characterize the current or specified time point of the subject being assessed as having ovarian cancer.

[0059] Understandably, an AI risk assessment model can be a relatively simple classifier or regressor, such as a multilayer perceptron containing several fully connected layers and activation functions, where the last layer uses a sigmoid activation function to output a probability value between 0 and 1, which is the quantified risk assessment value.

[0060] Understandably, for input processing, the entire unified fused feature vector sequence can be used. This sequence can be fed into another recurrent neural network (such as a long short-term memory network) or a one-dimensional convolutional network, and the output of its last time step can be used to make the final risk prediction. This approach can better utilize the dynamic information of the entire time series. Alternatively, the sequence can be aggregated first. For example, only the unified fused feature vector of the last time point in the sequence can be taken, because it has already incorporated historical information. Or, the vectors of the entire sequence can be subjected to time-dimensional average pooling or max pooling to obtain a single aggregated vector, and then this aggregated vector can be input into a multilayer perceptron for prediction.

[0061] Understandably, the output risk assessment value is a quantitative numerical value, such as 0.75. This value can be interpreted as the probability or risk level of the subject being assessed having ovarian cancer at the current point in time or at a specified future point in time. Clinicians can use this value and preset thresholds to decide on the next clinical action. For example, if the risk value is greater than 0.5, it can be considered as high risk.

[0062] In other embodiments of this application, the following steps may also be included: inputting a unified fusion feature vector sequence into a hierarchical time series prediction model in chronological order; the hierarchical time series prediction model includes a bottom-level time series network for capturing short-term feature patterns and a top-level time series network for learning long-term risk evolution trends; and generating a prediction result representing the risk evolution trend of the object to be evaluated at one or more future time points through processing by the hierarchical time series prediction model.

[0063] The underlying temporal network can be a relatively small gated recurrent unit or temporal convolutional network that receives a unified fused feature vector sequence and learns and encodes feature fluctuations and local patterns within a short time window; for example, it can capture the rapid upward trend of a certain fused feature dimension within two or three consecutive time points.

[0064] Top-level temporal networks are larger-scale and more powerful temporal models, such as a deep long short-term memory network or a Transformer encoder. Their input is the sequence of hidden states output by the bottom-level temporal networks at each time step. Since the bottom-level networks have already encoded and abstracted short-term patterns, the top-level networks can focus more on learning the long-term dependencies and macro-evolutionary trends between these short-term patterns. For example, it can identify that patient risk has shown a periodic fluctuation over the past three years, or has entered a sustained upward trend after a key event.

[0065] This hierarchical structure enables more effective modeling of complex temporal dynamics, and the final output is a sequence, which is a prediction of the risk value at one or more future time points (e.g., the next 6 months, 1 year, 2 years). This sequence constitutes the prediction result of risk evolution trend.

[0066] It is understandable that the risk evolution trend prediction result is as follows: for at least one preset time point in the future, output the risk probability distribution, and based on the preset statistical method, calculate and output the confidence interval corresponding to the risk probability distribution together to characterize the uncertainty of the prediction result.

[0067] To make the prediction results more clinically valuable, this application provides rich information to allow the output layer to predict the parameters of a probability distribution, such as the mean and variance of a Gaussian distribution. In this way, for a future point in time, we do not get a point estimate, but a complete risk probability distribution.

[0068] Understandably, to calculate confidence intervals, a statistical method compatible with deep learning models, Monte Carlo Dropout, can be used. In standard model training, the Dropout layer is activated only during the training phase to prevent overfitting. However, when using Dropout for prediction, the Dropout layer remains activated even during the inference phase. Then, for the same input sample, forward propagation is repeated multiple times (e.g., 100 times). Due to the randomness of Dropout, each forward propagation yields a slightly different prediction, resulting in an empirical distribution of 100 predicted values. Based on this empirical distribution, its mean can be calculated as the final predicted value, and its standard deviation or specific quantiles can be calculated to construct a 95% confidence interval, such as the 2.5% and 97.5% quantiles. This confidence interval visually demonstrates the uncertainty of the model's prediction. A narrow confidence interval indicates that the model is very confident in the prediction, while a wide confidence interval warns the doctor that the current prediction has significant uncertainty and requires careful interpretation.

[0069] Understandably, after generating risk assessment values ​​or risk evolution trend prediction results, the process also includes performing multi-level interpretability analysis, which can include feature-level attribution analysis, modality-level attribution analysis, and time-series attribution analysis.

[0070] Specifically, feature-level attribution analysis uses a pre-defined attribution algorithm to calculate the contribution of each original input feature to the final risk assessment value or risk evolution trend prediction result. For example, an attribution algorithm that can be used is SHAP (SHapley Additive exPlanations). SHAP can fairly allocate the model's predicted value to each input feature. By calculating the SHAP value of each feature, we can quantitatively know whether a specific indicator, such as "the value of cancer antigen 125 in the first quarter of 2025 is 300 U / mL", has an increasing or decreasing effect on the final high-risk prediction result, and the magnitude of that effect.

[0071] Modal-level attribution analysis aggregates the contribution values ​​of all original input features originating from the same modality to obtain the overall contribution of each modality to the risk assessment value or risk evolution trend prediction result. It then combines this with the analysis of the attention weight distribution of the time-series attention mechanism to cross-validate the modal contributions. The aggregation method can be simple summation or averaging of absolute values; for example, summing the SHAP values ​​of all image-related original features yields the overall contribution of the image modality. Simultaneously, the attention weight matrix of the multimodal data fusion model can be visualized. This matrix visually demonstrates how much attention is focused on the image modality, genetic modality, and clinical indicator modality when making decisions. Comparing the contribution of the SHAP aggregation with the attention weights reveals that if both indicate the image modality is the primary influencing factor, the conclusion is more credible.

[0072] Temporal attribution analysis analyzes hierarchical temporal prediction models to identify and highlight one or more historical time points that decisively influence the prediction results of risk evolution trends. Specifically, temporal attribution analysis works by calculating the gradient of the risk evolution trend prediction results with respect to the unified fused feature vector sequence, or by analyzing the temporal attention weights within the hierarchical temporal prediction model to identify key time points that significantly impact the prediction results. If the hierarchical temporal prediction model itself includes an attention mechanism, its temporal attention weights can be directly extracted and analyzed; the historical time point with the highest weight is the key time point with the greatest impact on future predictions. If there is no built-in attention mechanism, gradient-based attribution methods can be used. For example, the gradient of the final prediction result relative to the unified fused feature vector at each time point in the input sequence can be calculated. The magnitude of the gradient value reflects the sensitivity of the input at that time point to the final result; time points with large gradient values ​​are the key influencing time points.

[0073] It is understood that the embodiments of this application also include the following steps: integrating the results of feature-level attribution analysis, modality-level attribution analysis, and time-series attribution analysis to generate an integrated interpretability report; this interpretability report presents all analysis results in a clinician-friendly visualization manner, and the interpretability report includes at least a waterfall chart to show the contribution of key risk factors, where red bars represent factors that increase risk, green bars represent factors that decrease risk, and the length of the bars represents the size of the contribution; the interpretability report also includes a pie chart or bar chart to show the overall contribution of each modality, allowing doctors to see at a glance which type of examination this assessment mainly relies on; the interpretability report also includes a risk evolution curve, on which key impact time points identified by time-series attribution analysis are highlighted and accompanied by text descriptions, such as, "The system identified the follow-up in the third quarter of 2024 as a key turning point."

[0074] like Figure 2 As shown, this application also discloses an artificial intelligence system for early risk assessment of ovarian cancer, including multiple functional modules configured to perform the aforementioned artificial intelligence method for early risk assessment of ovarian cancer.

[0075] Understandably, the system can be a combination of hardware and software. At the hardware level, it may be deployed on a set of high-performance servers equipped with powerful graphics processors to accelerate the training and inference of neural networks. At the software level, the system is designed with a modular architecture, including at least a data acquisition and processing module, a modality quality assessment module, a feature representation module, an adaptive fusion module, and a risk assessment module.

[0076] Specifically, the data acquisition and processing module is responsible for interacting with various data systems in the hospital, performing functions such as data extraction, time series alignment, and missing value imputation; the modality quality assessment module contains one or more pre-trained modality quality assessment networks, receiving data and metadata, and outputting modality quality scores; the feature representation module contains a set of dedicated encoders for different modalities, responsible for converting raw data into high-dimensional modality feature vectors; the adaptive fusion module internally constructs a quality-aware Transformer fusion model, responsible for fusing modality feature vector sequences into a unified fusion feature vector sequence; the risk assessment module may contain a simple multilayer perceptron for static risk assessment, or a more complex hierarchical time series prediction model for dynamic trend prediction, and can calculate confidence intervals.

[0077] In other embodiments of this application, the system may further include an interpretability analysis module for performing multi-level attribution analysis and generating a final visualization report; in addition, there is a user interface module that provides clinicians with an interactive web interface or desktop application for inputting information about the object to be evaluated, viewing risk assessment results and interpretability reports.

[0078] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0079] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0082] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0083] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. An artificial intelligence method for early risk assessment of ovarian cancer, the method being implemented by a computer device, characterized in that, The following steps are performed by the computer device: Acquire longitudinal medical records of at least one subject to be evaluated, the longitudinal medical records containing multimodal time-series data collected at multiple time points; the multimodal time-series data includes at least two or more modal data selected from medical imaging data, genomic data, and clinical time-series test index data; and align the multimodal time-series data according to the collection timestamps to form a data sequence associated with the time axis; For each modality of data at each time point in the data sequence, a pre-trained modality quality assessment network is used for processing. The modality quality assessment network receives the modality data itself and its associated metadata as input, including but not limited to image resolution, sequencing depth, or data integrity markers. Through the processing of the modality quality assessment network, a modality quality score is generated for each modality of data at each time point to quantify its data reliability and information content at that time point. For each modal data point in the data sequence, the modal data is processed and converted into a high-dimensional modal feature vector by a modal-specific encoder that matches the modal data type; wherein, different types of modal data correspond to dedicated encoders with different structures. A multimodal data fusion model based on a temporal attention mechanism is constructed, and the modal feature vectors at multiple time points are input into the multimodal data fusion model. Within the multimodal data fusion model, when calculating the contribution of each modal feature vector to the fusion result at different time points through its temporal attention mechanism, the corresponding modal quality score is used as a dynamic adjustment factor to modulate the attention weights. Finally, a unified fusion feature vector is generated for each time point in the data sequence, and they are combined into a unified fusion feature vector sequence. The unified fusion feature vector sequence or a single vector aggregated from the unified fusion feature vector sequence is input into the artificial intelligence risk assessment model. Through the processing of the artificial intelligence risk assessment model, a risk assessment value for the current or specified time point of the subject to be assessed to be diagnosed with ovarian cancer is calculated and output.

2. The method according to claim 1, characterized in that, It also includes the following steps: The unified fused feature vector sequence is input into the hierarchical time series prediction model in chronological order; the hierarchical time series prediction model includes a bottom-level time series network for capturing short-term feature patterns and a top-level time series network for learning long-term risk evolution trends. The hierarchical time-series prediction model generates a prediction result characterizing the risk evolution trend of the object to be evaluated at one or more future time points.

3. The method according to claim 2, characterized in that, The modality-specific encoder specifically includes: For medical imaging data, convolutional neural networks are used as modality-specific encoders; for genomics data, graph neural networks or autoencoders are used as modality-specific encoders; and for clinical time-series test index data, recurrent neural networks or one-dimensional convolutional networks are used as modality-specific encoders.

4. The method according to claim 3, characterized in that, The multimodal time-series data is time-aligned according to the acquisition timestamps to form a data sequence associated with the time axis, specifically including: During time alignment, for missing data points in the data sequence, a preset interpolation algorithm is used to fill them in and generate a missing mask representing the missing data position. The missing mask, as an additional feature, is input into the modality-specific encoder along with the corresponding modal data for processing.

5. The method according to claim 2, characterized in that, After generating the risk assessment value or risk evolution trend prediction result, the method further includes a step of performing multi-level interpretability analysis, specifically including: A preset attribution algorithm is applied to calculate the contribution of the multimodal time series data to the final generated risk assessment value or risk evolution trend prediction result; The contribution values ​​of all original modal data originating from the same modality are aggregated to obtain the overall contribution of each modality data to the risk assessment value or risk evolution trend prediction result; and the contribution of each modality data is cross-validated by analyzing the attention weight distribution of the time-series attention mechanism. Analyze the hierarchical time series prediction model to identify and highlight one or more historical time points that have a decisive impact on the prediction results of the risk evolution trend.

6. The method according to claim 5, characterized in that, It also includes the following steps: The results of feature-level attribution analysis, modality-level attribution analysis, and time-series attribution analysis are integrated to generate a unified interpretability report. The interpretability report includes at least a waterfall chart to show the contribution of key risk factors, a pie chart or bar chart to show the overall contribution of each modality, and a visualization chart that highlights key impact time points on the risk evolution curve.

7. The method according to claim 5, characterized in that, The specific method of the time-series attribution analysis is as follows: By calculating the gradient of the risk evolution trend prediction result with respect to the unified fusion feature vector sequence, or by analyzing the time attention weights within the hierarchical time series prediction model, key influencing time points that have a significant impact on the risk evolution trend prediction result can be identified.

8. The method according to claim 2, characterized in that, The specific results of the risk evolution trend prediction are as follows: For at least one preset time point in the future, output a risk probability distribution, and calculate and output the confidence interval corresponding to the risk probability distribution based on a preset statistical method.

9. The method according to claim 1, characterized in that, The multimodal data fusion model based on the temporal attention mechanism is specifically implemented as a multi-head attention module.

10. An artificial intelligence system for early risk assessment of ovarian cancer, characterized in that, This includes a plurality of functional modules configured to perform the method as described in any one of claims 1 to 9, said functional modules comprising at least: The data acquisition and processing module is used to acquire and align multimodal time series data of at least one object to be evaluated at multiple time points. The modal quality assessment module is used to generate a modal quality score that quantifies the reliability and information content of each modal data point at each time point. The feature representation module is used to process each type of modal data into a modal feature vector; An adaptive fusion module is used to adaptively fuse the modality feature vector sequence into a unified fusion feature vector sequence based on a temporal attention mechanism and in combination with the modality quality score; The risk assessment module is used to process and output an early risk assessment value for the subject to be assessed to have ovarian cancer based on the unified fusion feature vector sequence.

Images