Multi-modal data fusion and machine learning method for constructing psoriasis biological agent curative effect prediction model

Abstract

The invention relates to the technical field of disease auxiliary identification, and discloses a multi-modal data fusion and machine learning method for constructing a psoriasis biological agent curative effect prediction model. The invention aims to solve the technical problems that the curative effect is difficult to predict due to individual difference, drug selection depends on trial and error, the decision-making period is long and the cost is high in the current psoriasis biological agent treatment. According to the invention, a multi-stage machine learning framework is constructed through deep fusion of multi-modal data, and full-process intelligent decision-making from precise etiological inference to individualized medication suggestions is realized. The core of the method is to construct a prediction model with continuous learning and optimization capabilities by combining an integrated deep learning architecture and a dynamic adjustment mechanism. By establishing an individualized and dynamically adjustable diagnosis and treatment path, the accuracy of prediction of the curative effect of the biological preparation is remarkably improved, the time for obtaining an optimal treatment scheme by a patient is effectively shortened, and a powerful technical support is provided for precise medical treatment of psoriasis.

Description

Technical Field

[0001] This invention relates to the field of disease-aided identification technology, and more specifically, to a multimodal data fusion and machine learning method for constructing a predictive model of the efficacy of psoriasis biological agents. Background Technology

[0002] Psoriasis is a chronic inflammatory skin disease mediated by the immune system, driven by both genetic and environmental factors. It presents with diverse clinical manifestations and a protracted, recurring course, severely impacting patients' quality of life and mental and physical health. In recent years, with in-depth research into the pathogenesis of this disease, a series of biologics targeting key molecules in specific inflammatory pathways have emerged, bringing revolutionary breakthroughs to the treatment of moderate to severe psoriasis.

[0003] These biologics, such as monoclonal antibodies targeting tumor necrosis factor-α or members of the interleukin family, can rapidly and significantly improve skin lesions by precisely blocking key aspects of disease progression, demonstrating efficacy unmatched by traditional therapies. However, clinical practice shows that even drugs targeting the same target exhibit significant differences in efficacy among different psoriasis patients. Some patients achieve near-perfect skin lesion clearance, while others respond little or no. This individualized difference means that drug selection still largely relies on the clinician's experience and the patient's trial-and-error approach.

[0004] Currently, before administering biologics, doctors typically make decisions based on a patient's clinical phenotype, disease severity, and past medication history. While this approach has some guiding value, it fails to fully utilize the deeper biological information inherent in the patient's body. The lack of an effective tool to comprehensively analyze multidimensional patient data and quantitatively predict the efficacy of specific drugs before treatment begins creates unavoidable decision-making blind spots in clinical practice. This not only potentially delays patients from receiving optimal treatment but also leads to a potential waste of expensive medical resources, constituting a widespread problem that urgently needs to be addressed in the field of precision medicine for psoriasis. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a multimodal data fusion and machine learning method for constructing a predictive model of the efficacy of psoriasis biologics, which solves the technical pain points in the current treatment of psoriasis biologics, such as the difficulty in predicting efficacy due to individual differences, the reliance on trial and error in drug selection, and the long decision-making cycle and high cost.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a multimodal data fusion and machine learning method for constructing a predictive model for the efficacy of psoriasis biological agents, which includes the following steps: Primary model construction: Based on complete training data, a primary etiology identification model including fusion and calibration layers is constructed. Based on the etiology probability distribution, individual patient characteristics, and mapping relationship of different drug efficacy in the complete training data, a primary efficacy prediction model is constructed. A safety knowledge base including adverse drug reactions, allergy information, and drug interactions is constructed. Specialized reinforcement: Using simulated limited samples formed by information masking of another batch of complete training data, mixed with original limited data cases, the image analysis capability of the primary etiology identification model is specifically reinforced and trained to obtain a final etiology identification model. The statistical etiology probability and model parameters on which the primary efficacy prediction model depends are fine-tuned regionally to obtain a final efficacy prediction model. Personalized diagnosis and recommendations: The final etiology identification model is used to identify the patient's skin lesion images to obtain a candidate set of etiologies. Abnormal indicators in the patient's limited physiological test data are extracted and compared with the biological judgment indicators corresponding to the etiologies in the candidate set to generate a personalized etiology diagnosis report. The final efficacy prediction model generates a candidate drug set by analyzing the etiology diagnosis report, and then conducts a safety review of the candidate drug set through the safety knowledge base. The medication recommendations are dynamically adjusted in combination with the etiology probability to generate a treatment recommendation plan. Treatment efficacy tracking: Receives the final medication plan determined by the doctor and the patient's efficacy feedback data. If the efficacy feedback data is lower than expected, treatment adjustment suggestions are generated through reverse reasoning analysis.

[0008] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the primary etiology identification model in the primary model construction step specifically includes: A physiological etiology identification module using a hybrid reasoning architecture identifies abnormal data in the test data and compares and reasons with a pre-set etiology knowledge base containing biological judgment indicators to generate a physiological inference report. An image etiology identification module of a deep learning network integrating an attention mechanism extracts visual features of skin lesions and generates an initial etiology set containing possible pathogenic factors and their corresponding probabilities based on an image database annotated by experts and associated with visual features and pathogenic factors. It also generates a morphological analysis report and a visual heatmap for explaining the basis of the judgment. The outputs of the physiological etiology identification module and the image etiology identification module are input into the fusion and calibration layer; the difference between the predicted result output by the fusion and calibration layer and the standard answer is calculated, and the difference is used as an error signal to synchronously adjust the parameters of the fusion and calibration layer, the physiological etiology identification module and the image etiology identification module through the backpropagation algorithm.

[0009] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the process of constructing the primary efficacy prediction model in the primary model construction step specifically includes: by analyzing the complete training data, establishing an efficacy knowledge structure that associates pathogenic factors with candidate drugs, wherein the efficacy knowledge structure provides a drug list sorted by priority according to statistical efficacy data for each pathogenic factor, and determines a quantitative expected effect range for each drug in the drug list.

[0010] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the specific process of the specialized reinforcement training in the specialized reinforcement step is as follows: guiding the primary etiology identification model to autonomously explore and extract microscopic visual features that have a stable statistical correlation with the etiology of the standard answer and are difficult to detect with the naked eye from the images of the simulated limited samples and original limited data cases; and iteratively optimizing the extraction process according to the standard answer until the model's ability to identify microscopic visual patterns reaches a preset high-precision standard.

[0011] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the process of obtaining the etiology candidate set in the individualized diagnosis and suggestion step specifically includes: Identify standard macroscopic features for preliminary diagnosis of disease type, specific features strongly associated with a single etiology, and common features that can correspond to multiple etiologies; significantly increase the probability of the pathogenic factors corresponding to the specific features; for multiple pathogenic factors corresponding to the common features, increase the probability of each with a lower probability to ensure that when multiple common features point to the same pathogenic factor, the cumulative probability increase can make it the primary pathogenic factor. The features that can be directly observed by the naked eye are identified during the analysis process and displayed side by side with standard images in a pre-set image library. The images are enhanced to highlight the subtle color differences between the features that the model learns autonomously and that are difficult to be detected by the naked eye and the normal skin areas, so as to achieve a visual presentation.

[0012] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the process of analyzing the patient's limited physiological test data in the individualized diagnosis and recommendation step specifically includes: Abnormal indicators in the detection data are classified into specific indicators that are strongly associated with a single cause and common indicators that can correspond to multiple causes; the probability of the pathogenic factors corresponding to the specific indicators is significantly increased; for the multiple pathogenic factors corresponding to the common indicators, each is increased with a lower probability. The identified abnormal indicators are cross-referenced with a pre-set knowledge base that associates physiological indicators with multiple diseases, and other disease names that may cause the abnormal indicators are retrieved and used as diagnostic information to be identified.

[0013] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the process of generating the individualized etiology diagnosis report in the individualized diagnosis and suggestion step specifically includes: excluding etiologies from the candidate set of pathogenic factors after probability adjustment, those that are not identified with any associated features in image analysis, and whose all associated physiological indicators show normality; and normalizing the probability values ​​of each pathogenic factor retained after exclusion in an equal proportion to form a probability distribution of pathogenic factors.

[0014] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the dynamic adjustment of medication recommendations in the individualized diagnosis and recommendation step specifically includes: providing a priority list containing multiple candidate drugs for the most probable primary pathogenic factor in the etiological diagnosis report, and labeling the corresponding expected efficacy range for the drugs in the priority list; for secondary pathogenic factors with lower probabilities, selecting a drug with the lowest known allergy risk from its corresponding candidate drugs as the recommended drug, and proportionally reducing the standard dose based on their respective probabilities to form the recommended dosage.

[0015] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the reverse reasoning analysis process in the efficacy tracking step specifically includes: performing correlation analysis between the changes in the patient's latest physiological test indicators and skin lesion image features and the expected mechanism of action of the drug targeting the etiology in the current medication regimen; The changes were consistent with the mechanism of action of drugs targeting the primary cause, but the efficacy was significantly lower than expected, indicating a single cause with insufficient therapeutic effect; If the changes are only reflected in indicators related to the mechanism of action of some drugs, while other pathological indicators do not improve, it can be considered that multiple pathogenic factors work together and are not fully covered. When the changes are highly consistent with the mechanism of action of a drug targeting a secondary cause, but unrelated to the mechanism of action of a drug targeting the primary cause, it indicates a deviation in the initial diagnosis of the cause.

[0016] As a preferred embodiment of the multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to the present invention, the treatment adjustment suggestions in the efficacy tracking step specifically include: When the cause is determined to be a single etiology but the treatment effect is insufficient, the primary option of increasing the current drug dosage is generated, as well as the alternative option of replacing it with an alternative drug targeting the same etiology. If it is determined that multiple pathogenic factors are acting together and not fully covered, it is recommended to increase the dosage of drugs that were abandoned by the doctor, have low doses, and target secondary pathogenic factors in the treatment plan; When there is a deviation in the initial diagnosis of the cause, the secondary cause that is consistent with the current actual treatment effect should be adjusted to the primary cause, and the recommended dosage of the corresponding drug should be increased to the standard dose.

[0017] The beneficial effects of this invention are as follows: By deeply fusing multimodal data to construct a multi-stage machine learning framework, it achieves intelligent decision-making throughout the entire process, from accurate etiological inference to personalized medication recommendations. Its core lies in constructing a predictive model with continuous learning and optimization capabilities by combining an integrated deep learning architecture with a dynamic adjustment mechanism. By establishing a personalized and dynamically adjustable treatment pathway, it significantly improves the accuracy of predicting the efficacy of biologics, effectively shortens the time for patients to obtain the optimal treatment plan, and provides strong technical support for achieving precision medicine for psoriasis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a multimodal data fusion and machine learning method for constructing a predictive model for the efficacy of psoriasis biologics.

[0020] Figure 2 Build a flowchart for the primary model.

[0021] Figure 3 A flowchart for special reinforcement.

[0022] Figure 4 A flowchart for individualized diagnosis, recommendations, and efficacy tracking. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments. Example

[0026] Reference Figures 1-4 This is one embodiment of the present invention, which provides a multimodal data fusion and machine learning method for constructing a predictive model for the efficacy of psoriasis biological agents, including the following steps: Primary model construction: Based on complete training data, a primary etiology identification model including fusion and calibration layers is constructed. Based on the etiology probability distribution, individual patient characteristics, and mapping relationship of different drug efficacy in the complete training data, a primary efficacy prediction model is constructed. A safety knowledge base including adverse drug reactions, allergy information, and drug interactions is constructed. The primary etiology identification model includes: a physiological etiology identification module that adopts a hybrid reasoning architecture. After identifying abnormal data in the test data, it compares and reasons with a pre-set etiology knowledge base containing biological judgment indicators to generate a physiological inference report. An image etiology identification module of a deep learning network integrating an attention mechanism extracts visual features of skin lesions and generates an initial etiology set containing possible pathogenic factors and their corresponding probabilities based on an image database annotated by experts and associated with visual features and pathogenic factors. It also generates a morphological analysis report and a visual heatmap for explaining the basis of the judgment. The outputs of the physiological etiology identification module and the image etiology identification module are input into the fusion and calibration layer; the difference between the predicted result output by the fusion and calibration layer and the standard answer is calculated, and the difference is used as an error signal to synchronously adjust the parameters of the fusion and calibration layer, the physiological etiology identification module and the image etiology identification module through the backpropagation algorithm; A preliminary efficacy prediction model is constructed. By analyzing complete training data, an efficacy knowledge structure is established that associates pathogenic factors with candidate drugs. For each pathogenic factor, the efficacy knowledge structure provides a drug list prioritized according to statistical efficacy data, and determines a quantitative expected effect range for each drug in the drug list.

[0027] First, a batch of high-quality, anonymized retrospective cases were collected and compiled from multiple large medical centers to form a complete training dataset. This dataset includes standardized clinical records, high-resolution images of skin lesions taken under fixed conditions, transcriptome sequencing data of skin lesion tissue, key gene loci data related to psoriasis, and efficacy data at key time points after using different biologics.

[0028] Establish a structured etiological knowledge base defined by dermatopathologists, classifying psoriasis into multiple known causative factors and associating each factor with a set of explicit biological indicators. Employ a hybrid reasoning architecture to immediately compare and infer abnormal indicators in patient test data against this etiological knowledge base, thereby generating a preliminary physiological inference report.

[0029] Establish a dedicated image database where experts directly correlate various subtle visual features in a massive number of skin lesion images with the ultimately identified causative factors for each patient. Construct a deep learning network integrating an attention mechanism. By extracting the visual features of the patient's skin lesions and learning from this image database, the network can independently generate an initial etiology set containing different causative factors and their corresponding probabilities, forming a morphological analysis report and outputting a visual heatmap highlighting the image regions on which the diagnosis was based.

[0030] Physiological inference reports and morphological analysis reports are simultaneously input into a fusion and calibration layer. Using the standard answers ultimately determined by the expert team for each case in the complete training data, the difference between the predicted results output by the fusion layer and the standard answers ultimately determined by the expert team for each case in the complete training data is calculated. This difference is used as an error signal and, through the backpropagation algorithm, all internal parameters of the fusion layer, physiological module, and image module are adjusted synchronously and finely until the model can accurately reproduce the expert's diagnostic logic. in, This represents the value of the loss function, which is the objective that the model needs to minimize; This indicates the total number of predefined categories of pathogenic factors; Represents the first in the batch One patient sample; Representing the One category of pathogenic factors; Represents an indicator variable, for the first... If the patient's standard answer is number [number] If there is a pathogenic factor, its value is 1; otherwise, it is 0. The fusion and calibration layer of the model predicts the first... The patient was the first The probability of each pathogenic factor.

[0031] From the complete training data, all patient groups meeting the single etiological factor defined in the etiology knowledge base were selected. Efficacy data for each candidate drug used by these patients were compiled to form the efficacy distribution of different drugs under this specific etiology. Key indicators such as the median and quartiles were calculated. The median assesses the relative efficacy of different drugs under this etiology, thus establishing an objective priority ranking. The interval formed by the first and third quartiles is defined as the quantified expected efficacy range of the drug, reflecting the potential efficacy range for most patients. By repeating this process for all etiological factors and related drugs, a data-driven, refined efficacy knowledge structure is ultimately established.

[0032] From the established efficacy knowledge structure, a definitive list of candidate drugs relevant to the system is extracted. This list is then systematically queried from authoritative pharmacopoeias, official drug instructions, and the latest pharmacovigilance reports. Adverse reactions associated with these specific drugs, allergy information linked to specific gene loci or physiological markers, and potential interactions between them are precisely extracted and structured for storage. By constructing a specialized cross-query matrix, the safety profile of any two drug combinations from the candidate drug list can be quickly retrieved, thus forming an efficient and non-redundant safety assurance system.

[0033] Specialized reinforcement: Using simulated limited samples formed by information masking of another batch of complete training data, mixed with original limited data cases, the image analysis capability of the primary etiology identification model is specifically reinforced and trained to obtain a final etiology identification model. The statistical etiology probability and model parameters on which the primary efficacy prediction model depends are fine-tuned regionally to obtain a final efficacy prediction model. Specialized intensive training guides the primary etiology identification model to autonomously explore and extract microscopic visual features that are statistically stable and imperceptible to the naked eye and have a stable correlation with the etiology of the standard answer from images of simulated limited samples and original limited data cases; and iteratively optimizes the extraction process based on the standard answer until the model's ability to identify microscopic visual patterns reaches a preset high-precision standard.

[0034] This process involves setting data masking rules within the program to artificially and randomly mask high-dimensional physiological data, such as transcriptome data and partial gene locus information, from complete case studies, generating a large number of simulated limited samples. These simulated samples are then mixed with the original limited data cases to form a reinforced training set. During training, the pre-trained primary etiology identification model is guided to identify pathogenic factors based solely on the input images and minimal basic information. Since the standard answer for each case (i.e., the final etiology determined by the expert team) is known, the model's prediction bias under insufficient information can be accurately calculated. This bias serves as a driving force, using backpropagation to focus on correcting the network parameters of the image analysis component. The model autonomously explores and extracts microscopic visual features in images that are difficult to detect with the naked eye but have a stable statistical correlation with the standard etiology (e.g., subtle changes in scale sheen or specific textures of inflammatory infiltration at the edge of skin lesions). Iterative optimization continues until the model achieves a preset high-precision accuracy on an independent test set based solely on images. At this point, a highly robust final etiology identification model is complete.

[0035] The actual frequency of each causative factor in all cases within a target region is statistically analyzed to obtain a regional causal probability distribution that reflects the specific patterns of the disease in that region. This regional probability distribution replaces the baseline probability, based on global data statistics, relied upon in the primary efficacy prediction model. The model parameters are incrementally trained using case data from the target region at a small learning rate, ensuring that the model does not forget global knowledge and can fully absorb and learn the disease manifestation patterns and drug response patterns specific to that region. This process is repeated for other different geographical regions, ultimately resulting in a series of final efficacy prediction models tailored to different regions.

[0036] Personalized diagnosis and recommendations: The final etiology identification model is used to identify the patient's skin lesion images to obtain a candidate set of etiologies. Abnormal indicators in the patient's limited physiological test data are extracted and compared with the biological judgment indicators corresponding to the etiologies in the candidate set to generate a personalized etiology diagnosis report. The final efficacy prediction model generates a candidate drug set by analyzing the etiology diagnosis report. The candidate drug set is then subject to safety review through a safety knowledge base. The medication recommendations are dynamically adjusted based on the etiology probability to generate a treatment recommendation plan. Acquire a candidate set of causes, identify standard macroscopic features for preliminary diagnosis, specific features strongly associated with a single cause, and common features that can correspond to multiple causes; significantly upregulate the probability of pathogenic factors corresponding to specific features; for multiple pathogenic factors corresponding to common features, upregulate them with a lower probability to ensure that when multiple common features point to the same pathogenic factor, the cumulative probability increase can make it the primary pathogenic factor. The features that can be directly observed by the naked eye are identified during the analysis process and displayed side by side with standard images in a pre-set image library; the images are enhanced to highlight the subtle color differences between the features that the model learns autonomously and that are difficult to be perceived by the naked eye and the normal skin areas, so as to achieve a visual presentation. The patient's limited physiological test data were analyzed, and the abnormal indicators in the test data were divided into specific indicators that are strongly associated with a single etiology and common indicators that can correspond to multiple etiologies. The probability of the pathogenic factors corresponding to the specific indicators was significantly increased; for the multiple pathogenic factors corresponding to the common indicators, each was increased with a lower probability. The identified abnormal indicators are cross-referenced with a pre-set knowledge base that associates physiological indicators with multiple diseases, and other disease names that may cause the abnormal indicators are retrieved and used as information for differential diagnosis. A personalized etiology diagnosis report is generated. Etiology factors that are excluded from the candidate set of pathogenic factors after probability adjustment, that are not identified with any related features in image analysis, and that all related physiological indicators show normality are excluded. The probability values ​​of each pathogenic factor that is retained after exclusion are normalized proportionally to form a probability distribution of pathogenic factors. The medication recommendations are dynamically adjusted. For the primary pathogenic factors with the highest probability in the etiological diagnosis report, a priority list containing multiple candidate drugs is provided, and the expected efficacy range of each drug in the priority list is marked. For secondary pathogenic factors with lower probability, the drug with the lowest known allergy risk is selected from the corresponding candidate drugs as the recommended drug, and the standard dose is reduced proportionally based on their respective probabilities to form the recommended dosage.

[0037] When a new patient visits, the system first analyzes their skin lesion images, identifying standard macroscopic features for preliminary diagnosis, such as plaque morphology and scale color, as well as microscopic visual features related to specific causes, distinguishing between specific and common features. The probability of specific features corresponding to a particular cause is significantly increased; the probabilities of multiple causes corresponding to common features are increased by a relatively low preset margin. Features that are directly observable to the naked eye are displayed side-by-side with standard images from a preset image library on the interface. For features learned by the model but imperceptible to the naked eye, the subtle color differences between these features and normal skin areas are highlighted by adjusting contrast and saturation, providing effective visualization.

[0038]

[0039] in, This indicates that after a new feature is identified, the first... The updated probability of the first pathogenic factor; then represents the probability before identifying this new feature, the probability of the first pathogenic factor. The probability of each pathogenic factor; These are the weighting coefficients; Let be an indicator function, if the currently identified feature matches the first... If a pathogenic factor is associated with a value of 1, the value is 0 otherwise.

[0040] in, The value of is set to 15 when the identified feature is a specific feature, to ensure that the calculated result of the lowest probability cause is higher than the value of the highest initial probability cause plus a single-level specific feature when both physiological and imaging levels are specific features. When the feature is a common feature, the value is set to 9, but it needs to be divided by the number of all common causes. Even if only two causes are common, the average is 4.5, and it still requires the probability of 4 or more common features to be higher than the probability increase of a single specific feature.

[0041] After an initial diagnosis, the doctor will conduct physiological tests on the patient based on the results. The limited physiological data obtained from these tests (such as complete blood count, specific inflammatory factors, etc.) are first excluded from those within the normal range. For the remaining abnormal data, the associated pathogenic factors are adjusted using the same probability upregulation rules as for image features. However, a search is still needed in multiple disease knowledge bases to screen for other possible causes of abnormal data. This screening is then conducted through communication between the doctor and patient, as well as a review of the patient's medical records.

[0042] After two rounds of probability adjustment, the pathogenic factors that were not identified in image analysis and whose related physiological indicators were all normal were removed from the candidate set of pathogenic factors. The probability values ​​of the remaining pathogenic factors were normalized proportionally so that their sum equaled 100%, forming the final pathogenic factor probability distribution, which serves as the core of the diagnostic report.

[0043]

[0044] in, Indicates the first The final probability obtained after normalization of each pathogenic factor; Indicates the first The probability values ​​of each pathogenic factor after screening and before normalization. This represents the collection of all pathogenic factors that were retained after the final screening. Representative set Any one of the pathogenic factors.

[0045] For one or more of the most probable causative factors reported, select the best few medications to avoid situations where patients experience allergic reactions to a single drug, requiring repeated prescriptions and skin tests. For other less probable causative factors, their potential should be preserved for subsequent efficacy evaluation; selecting the safest medications avoids wasting resources due to excessive skin testing.

[0046]

[0047] Among them, table This is for the first Recommended dosage for each secondary pathogenic factor; Indicates that for the first The standard therapeutic dose of the drug corresponding to each pathogenic factor; This indicates that in the individualized etiological diagnosis report, the first... The final normalized probability of each minor pathogenic factor.

[0048] Treatment efficacy tracking: Receive the final medication plan determined by the doctor and the patient's efficacy feedback data. If the efficacy feedback data is lower than expected, treatment adjustment suggestions will be generated through reverse reasoning analysis. Reverse reasoning analysis correlates changes in the patient's latest physiological test indicators and skin lesion image features with the expected mechanism of action of the drugs in the current medication regimen targeting the etiology. The changes were consistent with the mechanism of action of drugs targeting the primary cause, but the efficacy was significantly lower than expected, indicating a single cause with insufficient therapeutic effect; If the changes are only reflected in indicators related to the mechanism of action of some drugs, while other pathological indicators do not improve, it can be identified as the combined effect of multiple pathogenic factors and incomplete coverage. When the changes closely match the mechanism of action of a drug targeting a secondary cause, but are unrelated to the mechanism of action of a drug targeting the primary cause, it indicates a deviation in the initial diagnosis of the cause. Treatment adjustment recommendations specifically include: When the cause is determined to be a single etiology but the treatment effect is insufficient, the primary option of increasing the current drug dosage is generated, as well as the alternative option of replacing it with an alternative drug targeting the same etiology. If it is determined that multiple pathogenic factors are acting together and not fully covered, it is recommended to increase the dosage of drugs that were abandoned by the doctor, have low doses, and target secondary pathogenic factors in the treatment plan; When there is a deviation in the initial diagnosis of the cause, the secondary cause that is consistent with the current actual treatment effect should be adjusted to the primary cause, and the recommended dosage of the corresponding drug should be increased to the standard dose.

[0049] At the end of a treatment cycle, the actual therapeutic effect is first compared with the initially predicted range. If the actual effect meets or exceeds expectations, the system will recommend maintaining the current treatment plan.

[0050] When the actual therapeutic effect is lower than expected, we will conduct an in-depth correlation analysis between the changes in the patient's latest physiological test indicators and skin lesion image characteristics and the expected mechanism of action of each drug in the current medication regimen targeting the etiology.

[0051] When the latest physiological changes (e.g., a decrease in the level of a key inflammatory factor) are perfectly consistent with the mechanism of action of a drug targeting the primary cause, but the magnitude of the change is significantly lower than expected, the system will classify it as a single cause with insufficient efficacy. It will then suggest increasing the dosage of the current drug and provide the physician with an alternative drug that also targets the primary cause.

[0052] The latest changes are only reflected in indicators related to the mechanism of action of some drugs (e.g., low-dose drugs targeting only secondary causes), while other indicators reflecting the primary pathological state do not improve. In such cases, it is determined that multiple pathogenic factors are acting together and not fully covered. In this situation, it is recommended that doctors increase the dosage of drugs that were previously abandoned or used at lower doses to target those secondary pathogenic factors that were not effectively covered.

[0053] The latest changes closely match the drug's mechanism of action targeting a secondary cause, but are almost unrelated to the drug's mechanism of action targeting the primary cause, indicating a deviation in the initial diagnosis. The first step is to adjust the secondary cause, which is currently consistent with the actual therapeutic effect, to the primary cause, and then increase the dosage of the corresponding drug to the standard dose.

[0054] In summary, this invention constructs a multi-stage machine learning framework by deeply integrating multimodal data, enabling intelligent decision-making throughout the entire process, from accurate etiological inference to personalized medication recommendations. Its core lies in building a predictive model with continuous learning and optimization capabilities by combining an integrated deep learning architecture with a dynamic adjustment mechanism. By establishing a personalized and dynamically adjustable treatment pathway, the accuracy of predicting the efficacy of biologics is significantly improved, effectively shortening the time for patients to obtain the optimal treatment plan, and providing strong technical support for achieving precision medicine for psoriasis.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multimodal data fusion and machine learning method for constructing a predictive model for the efficacy of psoriasis biological agents, characterized in that, Performed by a computer device, including the following steps: Primary model construction: Based on complete training data, a primary etiology identification model including fusion and calibration layers is constructed. Based on the etiology probability distribution, individual patient characteristics, and mapping relationship of different drug efficacy in the complete training data, a primary efficacy prediction model is constructed. A safety knowledge base including adverse drug reactions, allergy information, and drug interactions is constructed. Specialized reinforcement: Using simulated limited samples formed by information masking of another batch of complete training data, mixed with original limited data cases, the image analysis capability of the primary etiology identification model is specifically reinforced and trained to obtain a final etiology identification model. The statistical etiology probability and model parameters on which the primary efficacy prediction model depends are fine-tuned regionally to obtain a final efficacy prediction model. Personalized diagnosis and recommendations: The final etiology identification model is used to identify the patient's skin lesion images to obtain a candidate set of etiologies. Abnormal indicators in the patient's limited physiological test data are extracted and compared with the biological judgment indicators corresponding to the etiologies in the candidate set to generate a personalized etiology diagnosis report. The final efficacy prediction model generates a candidate drug set by analyzing the etiology diagnosis report, and then conducts a safety review of the candidate drug set through the safety knowledge base. The medication recommendations are dynamically adjusted in combination with the etiology probability to generate a treatment recommendation plan. Treatment efficacy tracking: Receives the final medication plan determined by the doctor and the patient's efficacy feedback data. If the efficacy feedback data is lower than expected, treatment adjustment suggestions are generated through reverse reasoning analysis.

2. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 1, characterized in that, The primary etiology identification model in the primary model construction step specifically includes: A physiological etiology identification module using a hybrid reasoning architecture identifies abnormal data in the test data and compares and reasons with a pre-set etiology knowledge base containing biological judgment indicators to generate a physiological inference report. An image etiology identification module of a deep learning network integrating an attention mechanism extracts visual features of skin lesions and generates an initial etiology set containing possible pathogenic factors and their corresponding probabilities based on an image database annotated by experts and associated with visual features and pathogenic factors. It also generates a morphological analysis report and a visual heatmap for explaining the basis of the judgment. The outputs of the physiological etiology identification module and the image etiology identification module are input into the fusion and calibration layer; the difference between the predicted result output by the fusion and calibration layer and the standard answer is calculated, and the difference is used as an error signal to synchronously adjust the parameters of the fusion and calibration layer, the physiological etiology identification module and the image etiology identification module through the backpropagation algorithm.

3. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 1, characterized in that, The process of constructing the primary efficacy prediction model in the primary model construction step specifically includes: by analyzing the complete training data, establishing an efficacy knowledge structure that associates pathogenic factors with candidate drugs. The efficacy knowledge structure provides a drug list for each pathogenic factor, which is prioritized according to statistical efficacy data, and determines a quantitative expected effect range for each drug in the drug list.

4. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 1, characterized in that, The specific process of the specialized reinforcement training in the specialized reinforcement step is as follows: guiding the primary etiology identification model to autonomously explore and extract microscopic visual features that have a stable statistical correlation with the etiology of the standard answer and are difficult to be detected by the naked eye from the images of the simulated limited sample and the original limited data case. The extraction process is iteratively optimized based on the standard answer until the model's ability to recognize microscopic visual patterns reaches a preset high-precision standard.

5. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 4, characterized in that, The process of obtaining the etiology candidate set in the personalized diagnosis and recommendation steps specifically includes: Identify standard macroscopic features for preliminary diagnosis of disease type, specific features strongly associated with a single etiology, and common features that can correspond to multiple etiologies; significantly increase the probability of the pathogenic factors corresponding to the specific features; for multiple pathogenic factors corresponding to the common features, increase the probability of each with a lower probability to ensure that when multiple common features point to the same pathogenic factor, the cumulative probability increase can make it the primary pathogenic factor. The features that can be directly observed by the naked eye are identified during the analysis process and displayed side by side with standard images in a pre-set image library. The images are enhanced to highlight the subtle color differences between the features that the model learns autonomously and that are difficult to be detected by the naked eye and the normal skin areas, so as to achieve a visual presentation.

6. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 5, characterized in that, The process of analyzing the patient's limiting physiological test data in the individualized diagnosis and recommendation steps specifically includes: Abnormal indicators in the detection data are classified into specific indicators that are strongly associated with a single cause and common indicators that can correspond to multiple causes; the probability of the pathogenic factors corresponding to the specific indicators is significantly increased; for the multiple pathogenic factors corresponding to the common indicators, each is increased with a lower probability. The identified abnormal indicators are cross-referenced with a pre-set knowledge base that associates physiological indicators with multiple diseases, and other disease names that may cause the abnormal indicators are retrieved and used as diagnostic information to be identified.

7. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 6, characterized in that, The process of generating the individualized etiology diagnosis report in the individualized diagnosis and suggestion steps specifically includes: excluding etiologies from the candidate set of etiologies after probability adjustment, those that are not identified with any associated features in image analysis, and whose associated physiological indicators are all normal; and normalizing the probability values ​​of each etiology factor that remains after exclusion to form a probability distribution of etiologies.

8. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 7, characterized in that, The dynamic adjustment of medication recommendations in the individualized diagnosis and recommendation steps specifically includes: providing a priority list containing multiple candidate drugs for the most probable primary pathogenic factor in the etiological diagnosis report, and marking the expected efficacy range of the drugs in the priority list; for secondary pathogenic factors with lower probabilities, selecting the drug with the lowest known allergy risk from its corresponding candidate drugs as the recommended drug, and adjusting the standard dose proportionally based on their respective probabilities to form the recommended dosage.

9. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 1, characterized in that, The reverse reasoning analysis process in the efficacy tracking step specifically includes: performing a correlation analysis between the changes in the patient's latest physiological test indicators and skin lesion image features and the expected mechanism of action of the drugs in the current medication regimen targeting the etiology; The changes were consistent with the mechanism of action of drugs targeting the primary cause, but the efficacy was significantly lower than expected, indicating a single cause with insufficient therapeutic effect; If the changes are only reflected in indicators related to the mechanism of action of some drugs, while other pathological indicators do not improve, it can be considered that multiple pathogenic factors work together and are not fully covered. When the changes are highly consistent with the mechanism of action of a drug targeting a secondary cause, but unrelated to the mechanism of action of a drug targeting the primary cause, it indicates a deviation in the initial diagnosis of the cause.

10. The multimodal data fusion and machine learning method for constructing a psoriasis biological agent efficacy prediction model according to claim 9, characterized in that, The treatment adjustment recommendations in the efficacy tracking steps specifically include: When the cause is determined to be a single etiology but the treatment effect is insufficient, the primary option of increasing the current drug dosage is generated, as well as the alternative option of replacing it with an alternative drug targeting the same etiology. If it is determined that multiple pathogenic factors are acting together and not fully covered, it is recommended to increase the dosage of drugs that were abandoned by the doctor, have low doses, and target secondary pathogenic factors in the treatment plan; When there is a deviation in the initial diagnosis of the cause, the secondary cause that is consistent with the current actual treatment effect should be adjusted to the primary cause, and the recommended dosage of the corresponding drug should be increased to the standard dose.

Images