Medical heterogeneous data processing method and device based on deep learning and storage medium

Abstract

This application relates to the field of medical data processing technology, and discloses a method, apparatus, and storage medium for processing heterogeneous medical data based on deep learning. The method uses the same detection target as a unified processing unit, collecting next-generation sequencing data, quantitative PCR data, digital PCR data, and sample quality control data to construct platform feature inputs. It utilizes a deep learning model for branch coding to extract signal features from each detection platform and constructs platform observation capability features based on sample quality control data. On this basis, it generates compatible intervals corresponding to each detection platform; further, it estimates the potential true state of the detection target based on the compatible intervals, and performs a joint judgment based on consistency residuals, platform conflict degree, and observation sufficiency, outputting detection results and confidence information. This application can improve the detection accuracy and stability of low-abundance and weak-signal samples.

Description

Technical Field

[0001] This application relates to the field of medical data processing technology, specifically to a method, apparatus, and storage medium for processing heterogeneous medical data based on deep learning. Background Technology

[0002] With the development of molecular diagnostic technology, detection platforms such as next-generation sequencing, quantitative PCR, and digital PCR are widely used in medical testing. However, differences in detection principles, signal expression, and sensitivity range among these platforms can lead to inconsistent results for the same target, especially in low-abundance and boundary signal scenarios.

[0003] In practical applications, the processing of detection results from multiple platforms typically employs fixed threshold judgments or simple rule-based fusion methods. For example, undetected results are directly treated as negative, or any positive result is considered positive, or a majority voting strategy is used. However, these methods fail to effectively analyze the actual observation capabilities of different platforms under the current sample conditions, and they also fail to model the uncertainties corresponding to weak signals and undetected results. This leads to the easy missed detection of low-abundance targets, and makes it difficult to obtain stable and reliable judgment results when there are conflicts in the results from multiple platforms.

[0004] Furthermore, existing methods typically lack joint modeling of quality control factors such as sample input volume, nucleic acid quality, and library complexity, making it impossible to dynamically assess the effective observation capabilities of each platform based on detection conditions. At the same time, they lack a unified expression mechanism to describe and fuse the detection results of different platforms, making it difficult to perform consistency analysis and joint inference among heterogeneous data.

[0005] Therefore, there is an urgent need for a method that can uniformly model detection data from multiple platforms in order to achieve reasonable characterization of low-abundance targets and undetected signals, and on this basis, to complete the reliable fusion and judgment of detection results from multiple platforms. Summary of the Invention

[0006] To address the technical problems mentioned in the background section, the purpose of this application is to provide a method, apparatus, and storage medium for processing heterogeneous medical data based on deep learning.

[0007] According to the first aspect of this application, a method for processing heterogeneous medical data based on deep learning is provided, comprising the following steps: S1 uses the same detection target as a unified processing unit to collect next-generation sequencing data, real-time PCR data, digital PCR data and sample quality control data, and constructs the platform feature inputs corresponding to each detection platform. S2, use a deep learning model to branch and encode the feature inputs of each platform, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data; S3, Based on the signal characteristics and observation capability characteristics of each detection platform, generate a compatible range for the true abundance of the detection target for each detection platform; S4. Estimate the potential true state of the target based on each of the compatible intervals, and make a joint judgment based on the consistency residual between the potential true state and each of the compatible intervals, the degree of conflict between the detection platforms and the sufficiency of observation, and output the detection result and confidence information of the target.

[0008] According to a second aspect of this application, a medical heterogeneous data processing device based on deep learning is provided, the device comprising: The data acquisition module is used to collect next-generation sequencing data, quantitative PCR data, digital PCR data, and sample quality control data, using the same detection target as a unified processing unit, and to construct the platform feature inputs corresponding to each detection platform. The branch coding module is used to perform branch coding on the feature inputs of each platform using a deep learning model, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data. The interval generation module is used to generate compatible intervals for the true abundance of the detection target for each detection platform based on the signal characteristics and observation capability characteristics of each platform. The joint determination module is used to estimate the potential true state of the detected target based on each of the compatibility intervals, and to make a joint determination based on the consistency residual between the potential true state and each of the compatibility intervals, the degree of conflict between the detection platforms and the observation sufficiency, and output the detection result and confidence information of the detected target.

[0009] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0010] Compared with existing technologies, this application extracts signal features by branching and encoding data from different detection platforms, and constructs platform observation capability features by combining sample quality control data. On this basis, a compatible interval is introduced to characterize the true abundance of the detection target, so that undetected and weak signals are no longer simply mapped to fixed negative results, thereby effectively avoiding the problem of missed detection of low abundance targets. Furthermore, by fusing the compatible intervals to estimate the potential true state, and making joint judgments based on consistency residuals, the degree of conflict between platforms, and the sufficiency of observations, adaptive fusion and conflict resolution of multi-platform information are achieved, thereby significantly improving the accuracy, stability, and reliability of detection results under complex noise and heterogeneous data conditions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating a deep learning-based medical heterogeneous data processing method provided in this application embodiment; Figure 2 A schematic diagram of the network structure of the deep learning model provided in the embodiments of this application; Figure 3 A schematic diagram of the structure of a medical heterogeneous data processing device 200 based on deep learning provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device 300 provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] The technical solution of this embodiment can be applied to medical detection data analysis systems, molecular diagnosis information processing platforms, gene detection report generation systems, or detection devices with data processing capabilities (such as analysis workstations supporting gene sequencers, data processing modules in laboratory information management systems). The execution entity of the solution can be an independent server, cloud computing platform, personal computer, embedded device, or a processor integrated inside a detection instrument. The data sources include but are not limited to: FASTQ or BAM files output by second-generation sequencers (such as Illumina platforms, BGI platforms), amplification curve data (such as SDS files) output by fluorescence quantitative PCR instruments, droplet analysis data (such as CSV files of the QX200 system) output by digital PCR instruments, and sample quality control records (such as nucleic acid concentration, purity ratio, etc.) recorded in laboratory information systems. The above data is transmitted to the execution entity through wired or wireless networks, or manually imported by the user. The execution entity processes the heterogeneous data according to the steps described in this embodiment and finally outputs the detection results and confidence information for clinical doctors or inspection personnel to refer to for decision-making.

[0015] As Figure 1 shown, an embodiment of this application discloses a method for processing heterogeneous medical data based on deep learning. The method includes the following steps: S1, taking the same detection target as a unified processing unit, collecting second-generation sequencing data, fluorescence quantitative PCR data, digital PCR data, and sample quality control data, and constructing platform feature inputs corresponding to each detection platform; The detection target can be a specific gene mutation site, fusion gene, copy number variation, pathogen nucleic acid sequence, etc. For example, in the detection scenario of the lung cancer EGFR T790M drug resistance mutation, taking this mutation site as the unified processing unit, starting from the same patient sample (such as plasma or tissue), the original detection data of the second-generation sequencing (NGS) platform, fluorescence quantitative PCR (qPCR) platform, and digital PCR (dPCR) platform are collected in parallel, and the quality control-related data of this sample is recorded at the same time. It should be noted that there are significant differences in the formats, dimensions, and information dimensions of the original data output by different detection platforms. Therefore, it is necessary to construct platform feature inputs corresponding to each detection platform that can be directly processed by the deep learning model.

[0016] The following explains the specific construction process of each platform feature input: (1) For the second-generation sequencing platform, the original data is usually sequencing reads (reads) in FASTQ format or a BAM file obtained after alignment. In this embodiment, the following set of statistical features that can reflect the target signal intensity and reliability are extracted from it and organized into a feature vector with a fixed dimension: Coverage depth characteristics: Calculate the average sequencing depth, minimum depth, and standard deviation of depth within a certain range (e.g., ±50bp) before and after the target site.

[0017] Mutation supporting read characteristics: Extract the number of reads supporting the target mutation, the proportion of mutated reads to the total reads (i.e., allele frequency), and the number of forward and reverse reads supporting the mutation.

[0018] UMI family statistical characteristics: If UMIs (Unique Molecular Identifiers) were used during library construction, the total number of UMI families corresponding to the target site, the number of UMI families supporting the mutation, and the consistency ratio of UMI families are statistically analyzed. UMI information can effectively distinguish between true mutations and errors introduced by PCR amplification or sequencing.

[0019] Alignment quality characteristics include the mean alignment quality (MAPQ) score and base quality score of the target locus, as well as whether the locus is located in a high-GC-content region or a repetitive region. These characteristics are used to assess the reliability of the sequencing data.

[0020] The above features are concatenated into a one-dimensional feature vector in a preset order (such as coverage depth features, variant support read features, UMI family statistical features, and alignment quality features), which is the feature input for the second-generation sequencing platform.

[0021] (2) For quantitative real-time PCR platforms, the raw data is usually a real-time amplification curve, i.e., the fluorescence signal value for each cycle. This embodiment extracts the following temporal features and derived features from the amplification curve: Timing characteristics of the amplification curve: Using the amplification cycle number as the time axis, the original fluorescence signal value or the baseline-corrected ΔRn value is collected to form a time series with a length equal to the amplification cycle number (usually 40 or 45). This series directly reflects the dynamic process of the amplification reaction.

[0022] Cycle threshold (Ct) related characteristics: Calculate the Ct value (i.e., the number of cycles when the amplification curve exceeds the threshold), the confidence interval of the Ct value, and the peak value of the first derivative obtained by differentiating the amplification curve (i.e., the number of cycles corresponding to the maximum amplification rate). A lower Ct value usually indicates a higher initial template amount.

[0023] Amplification curve morphology characteristics: Extract the second derivative characteristics of the amplification curve (used to determine whether it is a typical S-shaped curve), the fluorescence value at the plateau phase of the curve endpoint, and the slope of the exponential phase of the curve. These characteristics help distinguish between specific amplification and non-specific amplification or primer dimers.

[0024] For time-series features, the normalized original fluorescence signal sequence can be directly used as input; for derived features, they can be concatenated into a short vector. The above time-series features and derived features are combined to form a multi-dimensional time-series feature input tensor (e.g., 40 cycles × 3 channels (original fluorescence, first derivative, second derivative)), which is used as the feature input for the real-time PCR platform.

[0025] (3) For digital PCR platforms, the raw data typically includes the endpoint fluorescence signal value for each reaction unit (droplet or chip well) and the positive / negative result for each channel. This embodiment extracts the following statistical features from these data: Positive droplet counting characteristics: The number of droplets identified as positive in the statistical detection channel, and the proportion of positive droplets to the total number of effective droplets. This proportion is directly converted into the absolute copy number concentration of the target molecule.

[0026] Cluster separation characteristics: Calculate the difference in fluorescence signal intensity between positive and negative droplet clusters (e.g., the Euclidean distance between the centers of the two clusters) and the degree of dispersion within the clusters (e.g., standard deviation). Higher cluster separation indicates a clearer distinction between positive and negative clusters, and more reliable results.

[0027] Repeatability characteristics: If a sample is subjected to multiple wells or repeated testing, the coefficient of variation (CV) and intraclass correlation coefficient (ICC) of positive droplet counts between each repeated test are calculated. Higher repeatability indicates more stable test results.

[0028] Background noise characteristics: The mean and standard deviation of fluorescence intensity of negative droplet clusters were extracted to assess the background noise level of the system.

[0029] After standardizing the above features, they are concatenated into a fixed-dimensional statistical feature vector (e.g., a vector containing 10 values), which serves as the feature input for the digital PCR platform.

[0030] (4) Sample quality control data include: the initial input amount of the sample (e.g., the number of nanograms of extracted nucleic acid), nucleic acid purity (e.g., the A260 / A280 ratio), fragmentation degree (e.g., DNA integrity index), and library complexity after library construction (e.g., the estimated number of unique molecules). It should be understood that these data, after processing in the subsequent step S2, form a quantitative value of the observation capability of each platform under the current sample conditions. For example, when the sample input amount is lower than the platform's recommended threshold, the observation capability characteristic values ​​of NGS and dPCR will be suppressed; when the nucleic acid purity is poor (indicating the possible presence of inhibitors), the observation capability characteristic values ​​of qPCR and dPCR will be suppressed.

[0031] In this embodiment, the raw data from next-generation sequencing, quantitative PCR, and digital PCR, as well as the sample quality control data, are transformed into structured platform feature inputs that can be processed by deep learning models for subsequent branch coding, signal extraction, and uncertainty modeling.

[0032] S2, use a deep learning model to branch and encode the feature inputs of each platform, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data; In this step, a deep learning model is used to branch and encode the feature inputs of each platform as described above, in order to extract the signal features corresponding to each detection platform that can characterize the essence of the detection signal. At the same time, the observation capability features of each detection platform under the current sample conditions are independently constructed based on the sample quality control data, which serves as the basis for subsequent uncertainty modeling.

[0033] It should be noted that the signal characteristics of different detection platforms vary significantly. Therefore, each platform uses its own independent branch coding network for processing. The structure of each branch network can be designed specifically according to the characteristics of its input data (time series, statistics, high-dimensional sparsity, etc.), which will not be elaborated further.

[0034] In some embodiments, such as Figure 2 As shown, a deep learning model is used to perform branch encoding on the feature inputs of each platform to extract the signal features corresponding to each detection platform, including: S21, input the platform features corresponding to the second-generation sequencing data into the second-generation sequencing branch coding network to extract second-generation sequencing signal features that characterize the sequencing support and sequencing reliability of the target detection. Deep learning models include second-generation sequencing branch coding networks. These networks can be composed of multiple fully connected layers or one-dimensional convolutional layers stacked together. The input is the platform's feature vector (including coverage depth, variant support reads, UMI family statistics, and alignment quality, etc.).

[0035] This network learns to map the original statistical features into two core representations through nonlinear transformation: sequencing support strength, which reflects the proportion of mutated reads in the total reads after background noise calibration; and sequencing reliability, which integrates alignment quality score, UMI family consistency, and base quality information. The network outputs a low-dimensional vector, which represents the second-generation sequencing signal features and is used for subsequent interval generation.

[0036] Additionally, the platform features corresponding to the quantitative PCR data are input into the quantitative PCR branch coding network to extract quantitative PCR signal features that characterize the amplification response process and amplification intensity changes of the detection target. The deep learning model also includes a branch coding network for quantitative PCR. Since qPCR data contains temporal features of the amplification curve, this branch network preferably employs a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN) to capture the dynamic changes in the amplification curve with cycle number. The input consists of the temporal signal of the amplification curve and Ct-related derived features.

[0037] This network uses time-series modeling to extract features that characterize the amplification response process (such as the curve start point, exponential slope, and plateau height) and amplification intensity changes (such as the confidence interval of the Ct value and the peak value of the first derivative). The output vector represents the features of the quantitative PCR signal, which can effectively distinguish between specific amplification and non-specific amplification or primer dimers.

[0038] Furthermore, the platform features corresponding to the digital PCR data are input into the digital PCR branch coding network to extract digital PCR signal features that characterize the absolute quantitative response of the detection target and the stability of detection repeatability. The deep learning model also includes a digital PCR branch coding network. This network can be composed of fully connected networks, with inputs including statistical features such as positive droplet counts, cluster separation, and repeatability consistency. The network learns to map these features to absolute quantitative response representations (such as the maximum likelihood estimate of copy number concentration) and repeatability stability representations (such as the stability score of the coefficient of variation between repeats after nonlinear transformation). The output vector is the digital PCR signal feature.

[0039] S22, perform feature mapping on the branch encoding results corresponding to each detection platform to obtain the low-dimensional signal representation vector corresponding to each detection platform, and determine the low-dimensional signal representation vector corresponding to each detection platform as the signal feature corresponding to each detection platform.

[0040] Since the original feature vectors output by each branch of the encoding network may have different dimensions (e.g., the NGS branch outputs 16 dimensions and the qPCR branch outputs 32 dimensions), they need to be mapped to a unified low-dimensional space in order to facilitate joint processing by the subsequent interval generation network.

[0041] Specifically, one or more fully connected mapping layers (called feature mapping heads) are set for each platform to compress the high-dimensional feature vectors output by the branch coding network to a pre-defined uniform dimension (for example, each platform is mapped to a 16-dimensional or 32-dimensional low-dimensional signal representation vector). The low-dimensional signal representation vectors obtained after feature mapping are the signal features corresponding to each detection platform.

[0042] Furthermore, platform observation capability characteristics for each detection platform are constructed based on the sample quality control data. These characteristics are independent of the aforementioned signal characteristics, and their construction process is as follows: The collected sample quality control data (such as sample input volume, nucleic acid purity, library complexity, PCR inhibitor index, etc.) are input into a small, fully connected network (this network can partially share parameters or be set individually for each platform). This network outputs one or more scalar values ​​to quantify the effective observation capability of each detection platform for the target signal under the current sample conditions.

[0043] For example, for NGS platforms, the output is an effective sequencing depth factor between 0 and 1. This factor is close to 1 when the sample input volume is high and the library complexity is appropriate, and close to 0 when the sample degradation is severe or the input volume is insufficient. For qPCR and dPCR platforms, the output is an amplification efficiency factor. The higher the nucleic acid purity and the fewer the inhibitors, the closer this factor is to 1.

[0044] S3, Based on the signal characteristics and observation capability characteristics of each detection platform, generate a compatible range for the true abundance of the detection target for each detection platform; In this step, each detection platform outputs a numerical range that reflects the uncertainty of its detection results. This range represents the set of all possible values ​​for the true abundance of the detected target (such as mutation abundance and copy number) under the current observed signal and observation capabilities of the platform, thus effectively characterizing weak signals and undetected situations.

[0045] In some embodiments, based on the signal characteristics corresponding to each detection platform and the observation capability characteristics of the platform, a compatible range for the true abundance of the detection target for each detection platform is generated, including: S31, the signal features and observation capability features corresponding to each detection platform are jointly input into the interval generation network to obtain the lower bound parameters and upper bound parameters of the interval corresponding to each detection platform; An interval generation network is set up for each detection platform. This interval generation network can be a lightweight fully connected network (e.g., containing 1-2 hidden layers, 64 neurons per layer). Its input layer receives two parts: the previously obtained low-dimensional signal representation vector of the platform (e.g., 16-dimensional), and the platform's observation capability features constructed simultaneously (e.g., 1-3 scalars). These two parts are concatenated or weighted and fused before being input into the interval generation network. The output layer of this network contains two neurons, corresponding to the lower bound parameters of the intervals. and the upper bound parameter of the interval .

[0046] The network is trained through a regression task, learning a mapping from joint features to reasonable interval boundaries. Specifically, when the input signal features are strong and the platform's observation capability features are high, the range of the network's output interval is relatively narrow, indicating that the detection result has high certainty. and The output range will be relatively close; conversely, when the signal is weak or the observation capability is low, the output range will be relatively wide. It will decrease This will increase to cover various true abundance values ​​that may arise due to noise or poor sample quality. For example, for a weak positive sample, if its observation capability is poor, the interval generation network outputs a wide interval from 0 (covering the possibility of non-detection) to a certain low positive number, thereby avoiding false negatives caused by directly classifying it as negative.

[0047] S32, based on the lower bound parameter and upper bound parameter of the interval corresponding to each detection platform, generate a compatible interval for the true abundance of the detection target for each detection platform, and perform interval constraint processing on the compatible interval to obtain the target compatible interval.

[0048] Obtaining the lower bound parameter of the interval and the upper bound parameter of the interval Then, directly construct the compatibility range for each platform. However, the original parameters of the neural network output may contain unreasonable conditions (e.g., If the interval boundary exceeds the physically permissible range (e.g., the abundance is negative), then the generated compatible interval needs to be constrained. Specific constraint processing includes one or more of the following operations: Order constraint: Apply the following transformation,

[0049] This is to ensure that the lower bound is not greater than the upper bound.

[0050] Range constraints: truncate the interval boundaries, for example, by... The abundance is limited to not less than 0 (because the abundance cannot be negative). The limit is set to no more than a preset maximum abundance threshold. (like (or 100%)

[0051] Smoothing constraint: Temporal smoothing is applied to the boundaries of consecutive samples or adjacent detection windows to avoid drastic jumps caused by noise. For example, an exponential moving average can be used.

[0052] in, This is a smoothing factor.

[0053] After the above interval constraint processing, the final and reasonable target compatible interval is obtained. This interval represents the uncertainty expression given by each detection platform for the true abundance of the detected target.

[0054] Taking a digital PCR platform as an example, if it detects a target copy number concentration of... If the sample has good repeatability and good cluster separation, the generated compatible interval is [4.5, 5.5]. If the sample has inhibitors that cause a decrease in amplification efficiency, the generated interval will be broadened to [2, 8]. If the sample is completely undetectable but the quality control shows that the sample input is low, the generated interval will be [0, 3], instead of simply being judged as negative.

[0055] S4. Estimate the potential true state of the target based on each of the compatible intervals, and make a joint judgment based on the consistency residual between the potential true state and each of the compatible intervals, the degree of conflict between the detection platforms and the sufficiency of observation, and output the detection result and confidence information of the target.

[0056] Within the compatibility range of all testing platforms (where subscript) Indicates the first A testing platform, After determining the total number of platforms participating in the fusion, this step, through interval fusion, consistency assessment, and joint inference, arrives at a final, more reliable detection conclusion. It should be understood that this process effectively integrates information from multiple platforms, especially when there are discrepancies in results across platforms or when some platforms lack sufficient observation capabilities, still providing stable and interpretable judgment results. Specifically: In some embodiments, estimating the potential true state of the detected target based on each of the compatibility intervals includes: S41, perform interval fusion processing on the compatible intervals corresponding to each detection platform to obtain a joint interval representation that characterizes the true abundance range of the detection target, and estimate the potential true state of the detection target based on the joint interval representation.

[0057] Because different platforms may produce different observation results for the same detection target, directly comparing the raw outputs (e.g., positive / negative) of each platform can easily lead to information loss. Therefore, this embodiment adopts an interval fusion strategy to merge compatible intervals from all platforms into a unified joint interval. This joint interval can more comprehensively reflect the possible range of the true abundance of the detection target. One or more of the following fusion methods can be selected based on the actual data characteristics: (1) Intersection Fusion. When the detection results of various platforms are relatively consistent, the overlapping area of ​​all intervals is the true abundance range commonly recognized by all platforms. For example, if the NGS platform gives an interval [0.02, 0.08] (representing a mutation abundance of 2%-8%), and the dPCR platform gives an interval [0.03, 0.07], then the intersection is [0.03, 0.07]. A non-empty intersection means that the evidence from each platform supports each other, and the joint interval converges. If the intervals from each platform do not intersect, then the intersection is empty, indicating a conflict between platforms, in which case other fusion strategies are needed.

[0058] (2) Weighted average fusion. When certain platforms have stronger observation capabilities (e.g., sample quality control data shows that the platform performs more stably under the current sample conditions), these platforms can be assigned higher weights. This weight value is the platform observation capability feature constructed in step S2. (The value range is usually from 0 to 1; a larger value indicates stronger observation capability). Then, a weighted average is calculated for the lower and upper bounds of each platform interval to obtain the lower and upper bounds of the joint interval. The specific calculation formula is as follows:

[0059] In this way, platforms with strong observation capabilities contribute more to the joint interval, while the impact of platforms with weak observation capabilities (such as those with poor sample quality leading to unreliable data) is effectively suppressed.

[0060] Obtain joint interval representation Then, the potential true state of the detected target is estimated. The potential true state refers to the optimal estimate of the true abundance of the detected target after synthesizing evidence from all platforms. A simple and effective method is to take the midpoint of the joint interval, i.e.:

[0061] Of course, a weighted expectation method can also be used, that is, first calculate the midpoint of each platform interval. Then, a weighted average is calculated using observation capabilities as the weights:

[0062] It should be understood that this potential real state It is a continuous numerical estimate that includes both the detection signal strength information of each platform and the differences in the observation capabilities of each platform.

[0063] In some embodiments, a joint determination is made based on the consistency residuals between the potential true state and each of the compatible intervals, the degree of conflict between detection platforms, and the sufficiency of observations, and the detection result and confidence information of the target are output, including: S42, calculate the consistency residual between the potential true state and each of the compatible intervals, and determine the degree of conflict between the detection platforms based on the interval overlap relationship between each of the compatible intervals, while combining the platform observation capability characteristics corresponding to each detection platform to determine the observation sufficiency; To quantitatively evaluate the quality of multi-platform detection results, this embodiment constructs three independent judgment factors: consistency residual, conflict degree, and observation sufficiency. These three factors can characterize the reliability of the current detection from different dimensions, as detailed below: Consistent residuals: used to measure the potential true state Compatibility range with each platform The degree of deviation between them. In short, if a platform provides an interval that includes the estimated true state... This indicates that the platform's results are consistent with the final estimate, with zero residuals; if It falls on the left side of the interval (i.e. This indicates that the platform believes the true abundance should be higher, resulting in a positive residual; conversely, if... It falls on the right side of the interval (i.e.) This indicates that the platform believes the actual abundance should be lower, and the residual is also positive. Specifically, the first... The consistency residual of each platform is defined as the Euclidean distance from a point to an interval:

[0064] The overall consistency residuals are obtained by summing the residuals of all platforms according to their observation capabilities. . The smaller the value, the more consistent each platform interval is with the estimated state, indicating a higher degree of agreement between the fusion results and the original platform observation evidence.

[0065] Conflict level: This quantifies the degree of difference in results between different testing platforms. The core manifestation of conflict is the overlap between the compatible intervals of each platform. If all intervals have a large overlapping area, it indicates that the results of each platform corroborate each other, and the degree of conflict is low; conversely, if multiple intervals are separated from each other or do not overlap at all, it indicates that there is a significant discrepancy between the platforms, and the degree of conflict is high.

[0066] This embodiment quantifies conflict by calculating the overlap between every two intervals. Two intervals and The overlap is defined as the length of their intersection divided by the length of their union, as shown in the following formula:

[0067] When the two intervals completely overlap When two intervals do not overlap at all (i.e., the upper bound of one interval is less than the lower bound of the other interval), Overall level of conflict Defined as the average of the complements of the overlap of all intervals:

[0068] Obviously, The closer the result is to 0, the more consistent the results are across platforms; The closer the value is to 1, the more severe the conflict between platforms. For example, if three platforms provide intervals [0, 0.1], [0.4-0.6], and [0.9-1.0] respectively, with almost no overlap, then the degree of conflict is... Close to 1.

[0069] Observational sufficiency: This reflects whether the overall observational capabilities of all detection platforms are sufficient under the current sample conditions. This factor is based on the observational capability characteristics of each platform. Calculation. Observation adequacy can be defined as the average or minimum of the observation capability characteristics of each platform. When using the average, observation adequacy is:

[0070] When using the minimum value, the observation sufficiency is:

[0071] The higher the value, the better the observation capability of each platform under the current sample conditions, and the higher the reliability of the detection results; The lower the value, the more likely it is that at least one or more platforms have limited observation capabilities due to insufficient sample input, poor nucleic acid quality, or excessive library complexity, and the uncertainty of the test results should increase accordingly.

[0072] S43, based on the consistency residual, the degree of conflict and the sufficiency of observation, a fusion judgment is made, and the detection result and confidence information of the detection target are output.

[0073] Consistency residuals Degree of conflict and observation sufficiency Three factors are input into a fusion decision network for joint reasoning. This fusion decision network can be a pre-trained logistic regression classifier or a small multilayer perceptron. The network learns a non-linear mapping from these three factors to the final decision and confidence information. Its mathematical form can be expressed as:

[0074] in For having trainable parameters The determination model.

[0075] To facilitate understanding, the following is an example of the rule determination logic: Rule 1: If observation sufficiency Below the preset sufficiency threshold (e.g.) Regardless of the consistency residuals and the degree of conflict, the output result is "uncertain," and the confidence information is directly set to... This is because the sample quality itself cannot guarantee the reliability of the test.

[0076] Rule 2: If the degree of conflict Higher than the preset conflict threshold (e.g.) ), and consistent residuals If the value is large, the output will be "Pending Review" or "Uncertain", and the confidence information will be set to [value missing]. This indicates that the results from different platforms are significantly inconsistent and require manual review or retesting.

[0077] Rule 3: If the residuals are consistent Smaller (e.g.) ), level of conflict Smaller (e.g.) ), and the sufficiency of observations Larger (e.g.) The output will be either positive or negative (based on the potential true state). Whether the mutation abundance exceeds a preset clinical threshold (e.g., mutation abundance > 0.5% is considered positive), the confidence information can be calculated as follows: A value close to 1 indicates a high level of confidence.

[0078] In actual training, the fusion decision network uses pre-labeled clinical sample data for end-to-end learning and automatically optimizes the decision boundaries to achieve optimal performance in different scenarios.

[0079] Ultimately, the method outputs the detection result (e.g., "positive," "negative," or "uncertain") and the corresponding confidence information (a confidence score between 0 and 1). For example, when the consistency residual is small, the degree of conflict is low, and the observation adequacy is high, the decision result has high confidence; conversely, when the degree of conflict is high or the observation adequacy is low, even if a platform shows a positive result, the final decision result may be "uncertain" or a "positive" result with low confidence.

[0080] It should be understood that this approach can effectively avoid the misjudgment problem of simple majority voting or fixed threshold rules in complex scenarios, and significantly improve the accuracy, stability and reliability of low abundance targets and weak signal samples detection.

[0081] Reference Figure 3As shown in the illustration, this application also provides a deep learning-based medical heterogeneous data processing device 200, the device comprising: The data acquisition module 10 is used to collect next-generation sequencing data, quantitative PCR data, digital PCR data and sample quality control data with the same detection target as a unified processing unit, and to construct the platform feature inputs corresponding to each detection platform. The branch coding module 20 is used to perform branch coding on the feature inputs of each platform using a deep learning model, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data. The interval generation module 30 is used to generate a compatible interval for the true abundance of the detection target for each detection platform based on the signal characteristics and observation capability characteristics of each platform. The joint determination module 40 is used to estimate the potential true state of the detection target based on each of the compatibility intervals, and to make a joint determination based on the consistency residual between the potential true state and each of the compatibility intervals, the degree of conflict between detection platforms and the sufficiency of observation, and output the detection result and confidence information of the detection target.

[0082] In some embodiments, the branch coding module 20 is specifically used to implement: The platform features corresponding to the second-generation sequencing data are input into the second-generation sequencing branch coding network to extract second-generation sequencing signal features that characterize the sequencing support and sequencing reliability of the target detection. Additionally, the platform features corresponding to the quantitative PCR data are input into the quantitative PCR branch coding network to extract quantitative PCR signal features that characterize the amplification response process and amplification intensity changes of the detection target. Furthermore, the platform features corresponding to the digital PCR data are input into the digital PCR branch coding network to extract digital PCR signal features that characterize the absolute quantitative response of the detection target and the stability of detection repeatability. Feature mapping is performed on the branch encoding results corresponding to each detection platform to obtain the low-dimensional signal representation vector corresponding to each detection platform, and the low-dimensional signal representation vector corresponding to each detection platform is determined as the signal feature corresponding to each detection platform.

[0083] In some embodiments, the interval generation module 30 is specifically used to implement: The signal features and observation capability features of each detection platform are jointly input into the interval generation network to obtain the lower bound parameters and upper bound parameters of the interval for each detection platform. Based on the lower bound parameter and upper bound parameter of the interval corresponding to each detection platform, a compatible interval for the true abundance of the detection target is generated for each detection platform, and interval constraint processing is performed on the compatible interval to obtain the target compatible interval.

[0084] In some embodiments, the joint determination module 40 is specifically used to implement: The compatible intervals corresponding to each detection platform are subjected to interval fusion processing to obtain a joint interval representation that characterizes the true abundance range of the detection target, and the potential true state of the detection target is estimated based on the joint interval representation.

[0085] In some embodiments, the joint determination module 40 is further configured to implement: Calculate the consistency residual between the potential true state and each of the compatible intervals, and determine the degree of conflict between the detection platforms based on the interval overlap relationship between each of the compatible intervals. At the same time, determine the observation sufficiency by combining the platform observation capability characteristics corresponding to each detection platform. Based on the consistency residual, the degree of conflict, and the sufficiency of observation, a fusion judgment is made, and the detection result and confidence information of the detection target are output.

[0086] like Figure 4 As shown, this application embodiment also provides a storage medium storing a computer program thereon, which, when executed by processor 301, implements the method as described in any of the preceding claims.

[0087] The storage medium 302 can be any physical medium capable of storing computer programs, including but not limited to: USB flash drives, portable hard drives, read-only memory (ROM, such as CD-ROM, DVD-ROM), random access memory (RAM), magnetic disks (such as floppy disks, hard disks), optical disks (such as optical discs, laser disks), flash memory (such as NAND Flash, NOR Flash), solid-state drives (SSDs), secure digital cards (SD cards), and other non-transitory computer-readable storage media. The storage medium can be the memory 302 inside an electronic device 300, which also includes a processor 301, a communication interface 303, input / output devices 304, and a power management module 305.

[0088] It should be noted that the storage medium does not include transient propagation signals, such as carrier waves or electromagnetic signals. In this embodiment, the storage medium specifically refers to a non-transitory computer-readable storage medium capable of stably storing program code, unaffected by power outages (for non-volatile media) or allowing temporary loading (for volatile media, such as RAM). In specific application scenarios, computer programs can be loaded into the random access memory (RAM) of an electronic device, read and executed by the central processing unit (CPU) or graphics processing unit (GPU), thereby achieving the corresponding functions.

[0089] The above description is the preferred embodiment of this application. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A method for processing heterogeneous medical data based on deep learning, characterized in that, Includes the following steps: S1 uses the same detection target as a unified processing unit to collect next-generation sequencing data, real-time PCR data, digital PCR data and sample quality control data, and constructs the platform feature inputs corresponding to each detection platform. S2, use a deep learning model to branch and encode the feature inputs of each platform, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data; S3, Based on the signal characteristics and observation capability characteristics of each detection platform, generate a compatible range for the true abundance of the detection target for each detection platform; S4. Estimate the potential true state of the target based on each of the compatible intervals, and make a joint judgment based on the consistency residual between the potential true state and each of the compatible intervals, the degree of conflict between the detection platforms and the sufficiency of observation, and output the detection result and confidence information of the target.

2. The method for processing heterogeneous medical data based on deep learning according to claim 1, characterized in that, Using a deep learning model, branch encoding is performed on the feature inputs of each platform to extract the signal features corresponding to each detection platform, including: S21, input the platform features corresponding to the second-generation sequencing data into the second-generation sequencing branch coding network to extract second-generation sequencing signal features that characterize the sequencing support and sequencing reliability of the target detection. Additionally, the platform features corresponding to the quantitative PCR data are input into the quantitative PCR branch coding network to extract quantitative PCR signal features that characterize the amplification response process and amplification intensity changes of the detection target. Furthermore, the platform features corresponding to the digital PCR data are input into the digital PCR branch coding network to extract digital PCR signal features that characterize the absolute quantitative response of the detection target and the stability of detection repeatability. S22, perform feature mapping on the branch encoding results corresponding to each detection platform to obtain the low-dimensional signal representation vector corresponding to each detection platform, and determine the low-dimensional signal representation vector corresponding to each detection platform as the signal feature corresponding to each detection platform.

3. The method for processing heterogeneous medical data based on deep learning according to claim 1, characterized in that, Based on the signal characteristics and observation capability characteristics of each detection platform, a compatible range for the true abundance of the detection target is generated for each detection platform, including: S31, the signal features and observation capability features corresponding to each detection platform are jointly input into the interval generation network to obtain the lower bound parameters and upper bound parameters of the interval corresponding to each detection platform; S32, based on the lower bound parameter and upper bound parameter of the interval corresponding to each detection platform, generate a compatible interval for the true abundance of the detection target for each detection platform, and perform interval constraint processing on the compatible interval to obtain the target compatible interval.

4. The method for processing heterogeneous medical data based on deep learning according to claim 1, characterized in that, Estimate the potential true state of the detected target based on each of the aforementioned compatible intervals, including: S41, perform interval fusion processing on the compatible intervals corresponding to each detection platform to obtain a joint interval representation that characterizes the true abundance range of the detection target, and estimate the potential true state of the detection target based on the joint interval representation.

5. The method for processing heterogeneous medical data based on deep learning according to claim 4, characterized in that, Based on the consistency residuals between the potential true state and each of the compatible intervals, the degree of conflict between detection platforms, and the sufficiency of observations, a joint judgment is made, and the detection result and confidence information of the target are output, including: S42, calculate the consistency residual between the potential true state and each of the compatible intervals, and determine the degree of conflict between the detection platforms based on the interval overlap relationship between each of the compatible intervals, while combining the platform observation capability characteristics corresponding to each detection platform to determine the observation sufficiency; S43, based on the consistency residual, the degree of conflict and the sufficiency of observation, a fusion judgment is made, and the detection result and confidence information of the detection target are output.

6. A medical heterogeneous data processing device based on deep learning, characterized in that, The device includes: The data acquisition module is used to collect next-generation sequencing data, quantitative PCR data, digital PCR data, and sample quality control data, using the same detection target as a unified processing unit, and to construct the platform feature inputs corresponding to each detection platform. The branch coding module is used to perform branch coding on the feature inputs of each platform using a deep learning model, extract the signal features corresponding to each detection platform, and construct the platform observation capability features corresponding to each detection platform based on the sample quality control data. The interval generation module is used to generate compatible intervals for the true abundance of the detection target for each detection platform based on the signal characteristics and observation capability characteristics of each platform. The joint determination module is used to estimate the potential true state of the detected target based on each of the compatibility intervals, and to make a joint determination based on the consistency residual between the potential true state and each of the compatibility intervals, the degree of conflict between the detection platforms and the observation sufficiency, and output the detection result and confidence information of the detected target.

7. A medical heterogeneous data processing device based on deep learning according to claim 6, characterized in that, The branch coding module is specifically used to implement: The platform features corresponding to the second-generation sequencing data are input into the second-generation sequencing branch coding network to extract second-generation sequencing signal features that characterize the sequencing support and sequencing reliability of the target detection. Additionally, the platform features corresponding to the quantitative PCR data are input into the quantitative PCR branch coding network to extract quantitative PCR signal features that characterize the amplification response process and amplification intensity changes of the detection target. Furthermore, the platform features corresponding to the digital PCR data are input into the digital PCR branch coding network to extract digital PCR signal features that characterize the absolute quantitative response of the detection target and the stability of detection repeatability. Feature mapping is performed on the branch encoding results corresponding to each detection platform to obtain the low-dimensional signal representation vector corresponding to each detection platform, and the low-dimensional signal representation vector corresponding to each detection platform is determined as the signal feature corresponding to each detection platform.

8. A medical heterogeneous data processing device based on deep learning according to claim 6, characterized in that, The joint determination module is specifically used to implement: The compatible intervals corresponding to each detection platform are subjected to interval fusion processing to obtain a joint interval representation that characterizes the true abundance range of the detection target, and the potential true state of the detection target is estimated based on the joint interval representation.

9. A medical heterogeneous data processing device based on deep learning according to claim 8, characterized in that, The joint determination module is also specifically used to implement: Calculate the consistency residual between the potential true state and each of the compatible intervals, and determine the degree of conflict between the detection platforms based on the interval overlap relationship between each of the compatible intervals. At the same time, determine the observation sufficiency by combining the platform observation capability characteristics corresponding to each detection platform. Based on the consistency residual, the degree of conflict, and the sufficiency of observation, a fusion judgment is made, and the detection result and confidence information of the detection target are output.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.

Images