Medicine industry credit investigation analysis method based on large model

Abstract

The invention discloses a pharmaceutical industry credit investigation analysis method based on a large model, and relates to the technical field of credit investigation analysis, and the method comprises the steps: obtaining pharmaceutical enterprise exclusive data, and carrying out the compliance processing; constructing a digital twinborn body of the exclusive data of the pharmaceutical enterprise; constructing a multi-modal risk scene, and carrying out multi-modal risk deduction based on the digital twinborn body; based on a risk deduction result, respectively calculating toughness values of the multi-modal risk; and calculating the comprehensive credit score of the pharmaceutical enterprise based on the toughness value of the multi-modal risk. The invention can provide a large-model-based credit investigation analysis method for the pharmaceutical industry, and credit investigation analysis is changed from historical backtracking to future prediction by constructing a pharmaceutical enterprise exclusive digital twinborn and combining with multi-modal risk deduction; research and development, compliance, supply chain and finance four-dimensional toughness indexes are constructed for pharmaceutical industry characteristics, and each index adopts industry exclusive computational logic, so that the credit score fitting degree of an innovative pharmaceutical enterprise is greatly improved, and the credit investigation pain point of difficult financing of a research and development type pharmaceutical enterprise is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of credit analysis technology, and in particular to a credit analysis method for the pharmaceutical industry based on a large model. Background Technology

[0002] As a technology-intensive and R&D-driven industry, the pharmaceutical industry has significant unique characteristics in its corporate credit analysis: on the one hand, the core value of pharmaceutical companies depends on their R&D pipelines, which have a long R&D cycle of 5-10 years and a high failure rate, making it difficult for traditional credit analysis to reflect the true value and potential risks of the companies; on the other hand, the industry is highly susceptible to policy regulation.

[0003] Current credit analysis methods in the pharmaceutical industry have the following key shortcomings:

[0004] Static assessments are lagging and unable to address dynamic risks: Existing methods mostly rely on the company's financial statements and qualification documents from the past 1-3 years for static scoring, which cannot predict the impact of sudden risks such as clinical trial failures of core drugs or price reductions in centralized procurement on the company's credit. This results in credit assessments lagging behind actual business changes, making it easy for financial institutions or investors to miss the opportunity for risk warnings.

[0005] Insufficient multi-source data processing capabilities and prominent compliance and integration challenges: Pharmaceutical industry data is scattered across multiple channels such as the official website of drug regulatory authorities, CDE database, and enterprise ERP systems, and includes patient clinical data and undisclosed R&D pipeline data. Existing methods lack an integrated solution for compliance desensitization and multimodal fusion.

[0006] Risk simulations lack industry adaptability and have low predictive accuracy: Existing risk simulations mostly use stress tests from general industries and have not built specific risk scenarios for the pharmaceutical industry.

[0007] The lack of resilience assessment and the single dimension of credit scoring, with existing credit reporting focusing only on general financial indicators such as debt-to-asset ratio and revenue growth rate, and failing to build an industry-specific resilience indicator system, leads to distorted scoring of innovative pharmaceutical companies with high R&D investment and low short-term profitability, failing to match their true credit level.

[0008] In summary, there is an urgent need for a credit analysis method that integrates large-scale modeling technology, is adapted to the characteristics of the pharmaceutical industry, and possesses dynamic extrapolation and probabilistic prediction capabilities. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a credit analysis method for the pharmaceutical industry based on a large-scale model. The technical solution adopted is as follows:

[0010] The credit analysis method for the pharmaceutical industry based on a large model includes the following steps:

[0011] Step 1: Obtain pharmaceutical companies' proprietary data and process it for compliance.

[0012] Step 2: Construct a digital twin of the pharmaceutical company's specific data;

[0013] Step 3: Construct multimodal risk scenarios and perform multimodal risk simulation based on digital twins;

[0014] Step 4: Based on the risk simulation results, calculate the resilience value of the multimodal risks respectively;

[0015] Step 5: Calculate the comprehensive credit score of pharmaceutical companies based on the resilience value of multimodal risk;

[0016] Step 6: Use the Prophet time series model combined with the language model to generate the credit change path for the next 12 months, output the probability distribution of credit score intervals through the Bayesian probability model, set the trigger conditions for credit downgrade, and output dynamic credit scores and risk warning reports.

[0017] Optionally, in step 1, the pharmaceutical company-specific data includes structured data, unstructured data, and semi-structured data; structured data includes GMP certification records, GSP certification records, CDE clinical trial data, drug registration approvals, medical insurance payment data, enterprise R&D pipeline data, and CDE clinical trial data; unstructured data includes medical insurance access policy texts, PubMed research papers, and adverse drug reaction feedback data; semi-structured data includes drug instructions and API procurement contracts.

[0018] The compliance processing method is as follows: use time series interpolation to complete the missing data, use hash encryption and access control to de-identify the controlled data, and uniformly convert it into time axis coordinates to achieve standardization;

[0019] Optionally, in step 1, the data processed for compliance is input into a graph neural network to construct a relational graph of enterprises, drugs, patents, API suppliers, and policies. The relational graph achieves cross-data source entity alignment, and virtual supplementary data is generated using a large biomedical model to complete data augmentation. A unified multimodal data graph of pharmaceutical enterprises is output.

[0020] Optionally, step 2, constructing a digital twin of the pharmaceutical company's specific data, includes the following sub-steps:

[0021] Step 21: Construct compliance qualification model, R&D pipeline model, financial resilience model, and supply chain model based on the characteristics of pharmaceutical companies;

[0022] Step 22: Based on historical decision-making cases of pharmaceutical companies over the past 10 years, fine-tune the R&D decision-making model using BioGPT, fine-tune the financial decision-making model using FinBERT, and fine-tune the compliance response model using the large risk decision-making model. Output the decision-making basis through the thinking chain of the large model.

[0023] Step 23: Connect with real-time policy channels of drug regulatory authorities and medical insurance departments, as well as drug bidding market data, to generate policy impact assessment and market impact analysis, and output a digital twin of pharmaceutical companies with behavioral simulation capabilities.

[0024] Optionally, in step 3, the multimodal risk scenarios include R&D risk, policy risk, supply chain risk, and market risk; the multimodal risk simulation includes the following steps:

[0025] The time step is set according to the risk type. First, the direct impact of the risk scenario on the current state of the enterprise is analyzed through GPT-4o. Then, the enterprise decision is output by calling the digital twin. The enterprise status is updated by applying the decision and the scenario impact. Finally, the pharmaceutical-specific health indicators are calculated, including compliance health score, R&D pipeline survival rate and cash flow support. The termination conditions of the pharmaceutical-specific health indicators are set. The simulation stops when the calculated pharmaceutical-specific health indicators meet the termination conditions.

[0026] Optional, multimodal risk resilience values ​​include R&D resilience R1, compliance resilience R2, supply chain resilience R3, and financial resilience R4;

[0027] The formula for calculating the R&D resilience R1 is: Where S i P is the target scarcity score for the i-th backup R&D pipeline. i M is the clinical trial success rate of the i-th backup R&D pipeline. i S0 is the projected market size of the backup R&D pipeline; P0 is the target scarcity score of the core R&D pipeline; P0 is the clinical trial success rate of the core R&D pipeline; and M0 is the projected market size of the core R&D pipeline.

[0028] The formula for calculating compliance resilience R2 is: Where L is the percentage of revenue loss due to the penalty, which is obtained by dividing the expected revenue loss amount caused by this compliance penalty by the total revenue of the company in the previous full fiscal year, and T is the penalty rectification period;

[0029] The formula for calculating supply chain resilience R3 is: Where v is the time from discovering the core supplier's supply disruption to using alternative suppliers and resuming normal procurement, C is the switching cost percentage, which is obtained by dividing the additional cost of switching alternative suppliers by the company's total supplier procurement costs in the previous quarter, and α is the standardization coefficient.

[0030] The formula for calculating financial resilience R4 is: Where M c It is the number of months that a company's cash flow can sustain. a This is the industry average number of months that cash flow can sustain.

[0031] Optionally, if R4 > 1, then take R4 = 1.

[0032] Optionally, the formula for calculating the comprehensive credit score S of pharmaceutical companies in step 5 is:

[0033] S = S0 × [(R1 × β + R2 × γ + R3 × δ + R4 × ε) · ω], where S0 is the basic credit score of the enterprise, β is the R&D resilience weight coefficient, γ is the compliance resilience weight coefficient, δ is the supply chain resilience weight coefficient, ε is the financial resilience weight coefficient, and ω is the environmental adaptability coefficient.

[0034] Optionally, step 6 is also included, which uses a timeline view to display the curves showing the changes in historical credit scores over the past 12 months and predicted credit scores over the next 12 months.

[0035] Optionally, credit rating can be conducted based on the comprehensive credit score of pharmaceutical companies. The mapping rules between the comprehensive credit score and the credit rating are as follows: 90-100 points are AAA level, corresponding to extremely low risk; 80-89 points are AA level, corresponding to low risk; 70-79 points are A level, corresponding to medium-low risk; 60-69 points are BBB level, corresponding to medium risk; and less than 60 points are BB level, corresponding to high risk.

[0036] In summary, the present invention has at least one of the following beneficial technical effects:

[0037] This invention provides a credit analysis method for the pharmaceutical industry based on a large model. By constructing a digital twin specifically for pharmaceutical companies and combining it with multimodal risk extrapolation, credit analysis shifts from historical retrospection to future prediction.

[0038] Based on the characteristics of the pharmaceutical industry, we have constructed a four-dimensional resilience index encompassing R&D, compliance, supply chain, and finance. Each index employs industry-specific calculation logic, significantly improving the credit scoring accuracy for innovative pharmaceutical companies and effectively addressing the credit assessment pain point of financing difficulties for R&D-oriented pharmaceutical companies.

[0039] By optimizing the data-to-model dual-loop system, the system's adaptability to changes in pharmaceutical industry policies and technological advancements is significantly shortened, ensuring that credit analysis stays aligned with industry development trends in the long term and extending the lifecycle and application value of the technical solution. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the pharmaceutical industry credit analysis method based on a large model, as described in this invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings.

[0042] This invention discloses a credit analysis method for the pharmaceutical industry based on a large model.

[0043] Reference Figure 1 Example 1, a credit analysis method for the pharmaceutical industry based on a large model, includes the following steps:

[0044] Step 1: Obtain pharmaceutical companies' proprietary data and process it for compliance.

[0045] Step 2: Construct a digital twin of the pharmaceutical company's specific data;

[0046] Step 3: Construct multimodal risk scenarios and perform multimodal risk simulation based on digital twins;

[0047] Step 4: Based on the risk simulation results, calculate the resilience value of the multimodal risks respectively;

[0048] Step 5: Calculate the comprehensive credit score of pharmaceutical companies based on the resilience value of multimodal risk;

[0049] Step 6: Use the Prophet time series model combined with the language model to generate the credit change path for the next 12 months, output the probability distribution of credit score intervals through the Bayesian probability model, set the trigger conditions for credit downgrade, and output dynamic credit scores and risk warning reports.

[0050] In Example 2, step 1, the pharmaceutical company's proprietary data includes structured data, unstructured data, and semi-structured data. Structured data includes GMP certification records, GSP certification records, CDE clinical trial data, drug registration approvals, medical insurance payment data, company R&D pipeline data, and CDE clinical trial data. Unstructured data includes medical insurance access policy texts, PubMed research papers, and adverse drug reaction feedback data. Semi-structured data includes drug instructions and API procurement contracts.

[0051] The compliance processing method is as follows: use time series interpolation to complete the missing data, use hash encryption and access control to de-identify the controlled data, and uniformly convert it into time axis coordinates to achieve standardization;

[0052] In Example 3, in step 1, the data processed for compliance is input into a graph neural network to construct a relational graph of enterprises, drugs, patents, API suppliers and policies. The relational graph achieves cross-data source entity alignment. Virtual supplementary data is generated using a large biomedical model to complete data augmentation, and a unified multimodal data graph of pharmaceutical enterprises is output.

[0053] By adopting the above technical solutions, pharmaceutical companies' proprietary data is divided into three categories based on its structural characteristics. Structured data focuses on the core operational and compliance dimensions of the pharmaceutical industry, covering GMP certification records, GSP certification records, CDE clinical trial data, drug registration approvals, medical insurance reimbursement data, and corporate R&D pipeline data. Unstructured data revolves around policy, research, and feedback scenarios, including medical insurance access policy texts, PubMed research papers, and adverse drug reaction feedback data. Semi-structured data covers drug technical specifications and supply chain cooperation information, including drug instructions and API procurement contracts. These three types of data are collected from official platforms, research databases, and corporate disclosure channels, respectively, forming a data foundation covering the entire chain of pharmaceutical company compliance, R&D, operations, and marketing.

[0054] To address data quality and security requirements, a three-layer processing scheme is adopted. The first layer is data completion, which uses time series interpolation to fill in missing information such as stage progress and investment amount in the R&D pipeline data based on the time distribution patterns of similar data from similar pharmaceutical companies. The second layer is privacy protection, which uses hash encryption to irreversibly process patient clinical data and unpublished R&D pipeline data, coupled with a hierarchical access control mechanism, granting access to corresponding data only to authorized users. The third layer is data standardization, which converts non-quantifiable time information such as clinical trial stages and drug approval milestones into a unified time axis coordinate system in months, eliminating differences in time representation from different data sources and laying the foundation for subsequent cross-data comparisons.

[0055] After compliance processing, various types of data are input into a graph neural network. Using enterprises, pharmaceuticals, patents, API suppliers, and policies as core nodes, the network leverages its node feature learning capabilities to identify differentiated representations of the same entity across different data sources. For example, it automatically associates the generic and brand names of the same drug, or the full and abbreviations of the same company, across different databases, achieving precise matching and association of entities across data sources and breaking down data silos.

[0056] By introducing a large-scale biomedical model, virtual supplementary data that conforms to the laws of the pharmaceutical industry is generated based on the characteristics of acquired real data. For example, by referencing the success rate of clinical trials and market size prediction logic of drugs with the same target, the probability of the stage progress of undisclosed R&D pipelines is generated; based on the historical impact trend of centralized procurement policies on similar drugs, revenue change data after a specific drug is included in centralized procurement is generated. By integrating virtual and real data, the data dimensions are enriched, and the data coverage and completeness are improved.

[0057] Based on multimodal data mapping, a digital twin of a pharmaceutical company is constructed from both static attributes and dynamic behaviors. At the static level, four models are built: compliance qualifications, R&D pipeline, financial resilience, and supply chain, quantifying the company's basic status. At the dynamic level, BioGPT, FinBERT, and a large-scale risk decision-making model are used to fine-tune the R&D, financial, and compliance decision-making models, enabling the twin to simulate the company's decision-making in different scenarios. Simultaneously, it connects to real-time policy and market data to ensure consistency between the twin and the real company's environmental interaction logic.

[0058] Step 3 constructs four major scenarios: R&D risk, policy risk, supply chain risk, and market risk, and uses a digital twin to perform risk simulation. A time step is set according to the risk impact cycle. First, the direct impact of the risk on the enterprise's status is analyzed. Then, the digital twin outputs decisions, updates the enterprise's status, and calculates specific indicators such as compliance health score, R&D pipeline survival rate, and the number of months cash flow can sustain operations. The simulation stops when the termination condition is met. Step 4 calculates a four-dimensional resilience value based on the simulation results. Step 5 combines the resilience value with the basic credit score and environmental adaptability coefficient to form a comprehensive credit score. Step 6 uses a Prophet time series model and a language model to collaboratively generate credit change paths, and combines a Bayesian probability model to output probability distributions, ultimately forming a dynamic score and early warning report, completing the closed loop of data, simulation, evaluation, and prediction.

[0059] Example 4, step 2, constructing a digital twin of pharmaceutical company-specific data includes the following sub-steps:

[0060] Step 21: Construct compliance qualification model, R&D pipeline model, financial resilience model, and supply chain model based on the characteristics of pharmaceutical companies;

[0061] Step 22: Based on historical decision-making cases of pharmaceutical companies over the past 10 years, fine-tune the R&D decision-making model using BioGPT, fine-tune the financial decision-making model using FinBERT, and fine-tune the compliance response model using the large risk decision-making model. Output the decision-making basis through the thinking chain of the large model.

[0062] Step 23: Connect with real-time policy channels of drug regulatory authorities and medical insurance departments, as well as drug bidding market data, to generate policy impact assessment and market impact analysis, and output a digital twin of pharmaceutical companies with behavioral simulation capabilities.

[0063] By adopting the above technical solution, focusing on the core static attributes that distinguish pharmaceutical companies from general industries, key dimensions are broken down and transformed into quantifiable models:

[0064] Focusing on the mandatory compliance requirements of the pharmaceutical industry, this study takes the core qualifications for drug production and operation (GMP / GSP certification, drug registration certificate) as the core dimension, and builds a model by quantifying the validity of qualifications and compliance history (surprise inspection results, penalties) to adapt to the characteristic of the pharmaceutical industry where compliance is survival.

[0065] R&D Pipeline Model: Targeting the core value carrier of pharmaceutical companies (R&D pipeline), a quantitative model is constructed to reflect the potential value of the pipeline, using key parameters such as the clinical stage of the pipeline, target characteristics, clinical trial success rate, and market size forecast. This model matches the value assessment logic of R&D-driven companies.

[0066] Financial resilience model: Breaking through the limitations of general financial models that only focus on revenue and profit, it focuses on quantifying the sustainability of R&D investment and cash flow structure, and is adapted to the financial characteristics of pharmaceutical companies with high R&D investment and long capital recovery cycle.

[0067] Supply chain model: Taking API (active pharmaceutical ingredient) as the core node, it quantifies the dependence on core API suppliers, supplier compliance risks, and alternative supplier substitution capabilities, which is in line with the supply chain characteristics of the pharmaceutical industry where the stability of API supply directly affects production compliance.

[0068] Given the specialized nature of decision-making scenarios in the pharmaceutical industry, a proprietary large-scale model combined with historical case fine-tuning is used to train decision-making capabilities.

[0069] Model selection and adaptation: Select industry-adaptive large models for different decision-making scenarios: BioGPT has professional knowledge in the biopharmaceutical field and is suitable for handling R&D decisions such as R&D pipeline adjustments and clinical trial advancements; FinBERT excels in financial text and financial logic analysis and is suitable for financial decisions such as centralized procurement price reductions and cash flow management; the risk decision-making large model focuses on compliance risk response logic and can handle scenarios such as unannounced inspection rectification and policy compliance adjustments;

[0070] Historical case training: Using real decision-making cases of pharmaceutical companies over the past 10 years (such as pipeline switching after clinical failure, capacity adjustment after centralized procurement price reduction, and rectification plan after unannounced inspection) as training data, the model is fine-tuned to learn the decision-making logic of pharmaceutical companies in specific scenarios, so as to avoid the decision-making of general models from deviating from the actual situation of the industry.

[0071] Thought Chain Decision Output: Through the large model thought chain technology, the decision-making process is broken down into a traceable chain of scenario analysis, parameter input, solution derivation, and result prediction, so that the decision output by the twin not only includes the conclusion, but also has clear reasoning basis.

[0072] To address the pharmaceutical industry's strong reliance on policy and high market volatility, a real-time interaction capability between the twin and the external environment is constructed.

[0073] Real-time data integration: Targeted integration with policy release channels of drug regulatory authorities (National Medical Products Administration) (new regulations, approval information), medical insurance department policy channels, and drug bidding market data (winning bid prices, procurement volume) to ensure that the twin can obtain key external data affecting the company's operations in real time;

[0074] Quantitative impact assessment: By analyzing real-time policy and market data through large models, we quantify their direct impact on enterprises, such as the impact of centralized procurement price reductions on revenue and the increase in compliance costs due to new policies and regulations; and indirect impacts, such as the impact of competitors winning bids on market share. We generate structured policy impact assessment and market impact analysis reports.

[0075] Behavioral simulation capability activation: By integrating real-time environmental impact data into the twin decision-making model, the twin can adjust its decision-making logic based on the current environmental state, realizing dynamic behavioral simulation of environmental perception, decision response, and state updates, rather than just fixed decision output based on static data.

[0076] Example 5, in step 3, the multimodal risk scenarios include R&D risk, policy risk, supply chain risk, and market risk; the multimodal risk simulation includes the following steps:

[0077] The time step is set according to the risk type. First, the direct impact of the risk scenario on the current state of the enterprise is analyzed through GPT-4o. Then, the enterprise decision is output by calling the digital twin. The enterprise status is updated by applying the decision and the scenario impact. Finally, the pharmaceutical-specific health indicators are calculated, including compliance health score, R&D pipeline survival rate and cash flow support. The termination conditions of the pharmaceutical-specific health indicators are set. The simulation stops when the calculated pharmaceutical-specific health indicators meet the termination conditions.

[0078] Example 6: The resilience values ​​for multimodal risks include R&D resilience R1, compliance resilience R2, supply chain resilience R3, and financial resilience R4.

[0079] The formula for calculating the R&D resilience R1 is: Where S i P is the target scarcity score for the i-th backup R&D pipeline. i M is the clinical trial success rate of the i-th backup R&D pipeline. i S0 is the projected market size of the backup R&D pipeline; P0 is the target scarcity score of the core R&D pipeline; P0 is the clinical trial success rate of the core R&D pipeline; and M0 is the projected market size of the core R&D pipeline.

[0080] The formula for calculating compliance resilience R2 is: Where L is the percentage of revenue loss due to the penalty, which is obtained by dividing the expected revenue loss amount caused by this compliance penalty by the total revenue of the company in the previous full fiscal year, and T is the penalty rectification period;

[0081] The formula for calculating supply chain resilience R3 is: Where v is the time from discovering the core supplier's supply disruption to using alternative suppliers and resuming normal procurement, C is the switching cost percentage, which is obtained by dividing the additional cost of switching alternative suppliers by the company's total supplier procurement costs in the previous quarter, and α is the standardization coefficient.

[0082] The formula for calculating financial resilience R4 is: Where M c It is the number of months that a company's cash flow can sustain. a This is the industry average number of months that cash flow can sustain.

[0083] Example 7: If R4 > 1, then take R4 = 1.

[0084] By adopting the above technical solutions, and based on the risk exposure characteristics of core business operations in pharmaceutical companies, risks are classified into four specific types to align with the industry's risk transmission patterns:

[0085] Focusing on the core value creation process of pharmaceutical companies, R&D risks directly determine the long-term survival ability of enterprises. Therefore, they are identified as the primary risk type, covering scenarios such as failure in the clinical stage of core drugs, patent expiration, and the first launch of competing products. These risks have a long-term and high-impact nature on enterprises, and are the core risks that distinguish the pharmaceutical industry from general industries.

[0086] Given the pharmaceutical industry's strong policy and regulatory nature, policy adjustments directly impact corporate revenue and compliance costs. Therefore, scenarios such as centralized procurement price reductions, failure to meet expectations in medical insurance access negotiations, and rectification following unannounced inspections are included. These risks are characterized by their suddenness and mandatory nature, and should be prioritized in the simulation to assess companies' policy response capabilities.

[0087] Given the API-dependent nature of pharmaceutical production, the stability of API supply is directly related to the compliance and continuity of drug production. Therefore, scenarios such as core API supply disruptions, API price surges, and cold chain logistics interruptions are included. These risks have an immediate and rapid impact on enterprises, and can quickly trigger production disruptions and compliance risks.

[0088] Given the clinical demand-driven and feedback-sensitive nature of pharmaceutical products, market risks directly impact product sales and brand value. Therefore, scenarios such as adverse drug reaction feedback, a sharp drop in hospital procurement, and price reductions by competitors are included. These risks are characterized by rapid spread and wide impact, requiring a quick assessment of their impact on the company's short-term operations.

[0089] Based on the differences in the evolution cycle and impact speed of different types of risks, differentiated time steps are set to ensure that the simulation closely matches the actual development pattern of risks:

[0090] Short-term risk time step logic: For short-term risks such as adverse drug reaction feedback and cold chain logistics disruptions, the impact can be seen within a few days and the pace of impact is fast. Therefore, the time step is set by day - which can accurately capture the daily changes in the company's status during the risk diffusion process and avoid missing key impact nodes due to the step being too large.

[0091] In response to long-term risks such as clinical trial failures of core drugs, patent expiration, and the long-term impact of centralized procurement, the impact of these risks needs to be released gradually over several months. Therefore, setting a time step of one month can balance the accuracy and efficiency of the projection and avoid redundancy caused by too small a step.

[0092] Leveraging GPT-4o's deep understanding of pharmaceutical industry texts and scenarios, abstract risk scenarios are transformed into quantifiable impact parameters: by analyzing the key characteristics of risk scenarios and the current state of enterprises, direct impact results such as the amount of revenue reduction, the need to adjust R&D investment, and the amount of compliance cost increase are output, avoiding the subjective bias and inefficiency of traditional manual assessment.

[0093] Based on the constructed digital twin with pharmaceutical-specific decision-making capabilities, the system calls upon its trained R&D, financial, and compliance decision-making models. The twin can output response decisions tailored to the actual situation of pharmaceutical companies based on the impact results output by GPT-4o and its own static attributes, ensuring the industry adaptability of the decisions.

[0094] Enterprise status update principle: Following the dynamic interaction logic of risk impact, enterprise decision-making, and status feedback, the risk impact output by GPT-4o and the response decision output by the twin are used together as the input for status update; for example, after the impact of centralized procurement price reduction and decision are superimposed, the new status of the enterprise is: the revenue decline narrows to 8% and the production capacity of non-centralized procurement products increases by 30%, which truly simulates the dynamic operation process of the enterprise under risk.

[0095] The compliance and health score, R&D pipeline survival rate, and the number of months that cash flow can sustain the business are selected as the criteria for determining the termination of the simulation. These three factors are the core bottom line for the survival of pharmaceutical companies: the compliance and health score reflects whether the company meets the compliance requirements for production and operation, the R&D pipeline survival rate reflects whether the core value carrier is still viable, and the number of months that cash flow can sustain the business reflects short-term operating capacity. When any indicator reaches the termination condition, such as the number of months that cash flow can sustain the business is less than 3 months or the compliance and health score is less than 60 points, the simulation will automatically stop to avoid invalid simulations and improve the efficiency and relevance of the simulation.

[0096] To address the core dimensions of the pharmaceutical industry's resilience, a four-dimensional resilience value calculation formula was designed, encompassing R&D, compliance, supply chain, and finance. The R&D resilience calculation formula revolves around the industry's fundamental nature that the R&D pipeline is the core value carrier for pharmaceutical companies. The core logic is based on the replacement capability of alternative pipelines after the failure of the core pipeline. The numerator (total value of alternative R&D pipelines) is designed based on the following principles: target scarcity score, clinical trial success rate, and product market size forecast are selected as core parameters for the value of alternative pipelines. Target scarcity score reflects the pipeline's technological barriers; for example, first-in-class targets score higher than me-too targets. Clinical trial success rate reflects the pipeline's feasibility. Market size forecast reflects the pipeline's commercialization potential. The product of these three factors quantifies the potential value of a single alternative pipeline, and summing them yields the total replacement capability of all alternative pipelines.

[0097] The design principle of the denominator (value of the core R&D pipeline): The value of the core pipeline is calculated using the same parameters as the alternative pipelines (scarcity of targets, success rate, market size). The core pipeline is the main source of the company's current revenue or future value, and its value is the benchmark for measuring the replacement capability of alternative pipelines.

[0098] The R&D resilience value is "total alternative value / core value". The closer the ratio is to 1, the stronger the R&D resilience is, indicating that alternative pipelines can completely replace the value of the core pipeline after its failure. A ratio of 0 indicates that there are no effective alternative pipelines and the R&D resilience is extremely weak. This design directly quantifies the ability of pharmaceutical companies to cope with the highest risk of "core pipeline failure", which is in line with the R&D-driven characteristics of the industry.

[0099] In response to the industry requirement that "compliance is a prerequisite for the survival of pharmaceutical companies," a formula is designed based on the logic of "the degree of impact of compliance penalties on business operations and the recovery period":

[0100] The calculation method is based on the estimated revenue loss amount resulting from this compliance penalty divided by the company's total revenue in the previous full fiscal year. Compliance penalties (such as production stoppage due to unannounced inspections or restrictions on bidding) directly result in the inability of pharmaceutical companies to sell their core products. The revenue loss ratio can quantify the impact of the penalty on the company's revenue (e.g., a loss ratio of 20% means that the penalty affects one-fifth of the revenue).

[0101] Using the time (in months) from the issuance of the penalty notice to the completion of rectification and the resumption of normal operations as a parameter, divide by 12 (converted to an annual cycle); the longer the rectification cycle, the longer the business interruption time, and the higher the possibility of compliance risks being transmitted to the financial and R&D stages;

[0102] The compliance resilience value is obtained by subtracting the impact coefficient (L×T / 12) from 1. The larger the impact coefficient, the stronger the overall impact of the penalty on the enterprise, and the closer the compliance resilience value is to 0. When the impact coefficient is 0 (no penalty), the compliance resilience value is 1, which means that the enterprise's current compliance status is good. This design quantifies the enterprise's ability to recover from compliance penalties, which is in line with the characteristics of strong compliance supervision in the pharmaceutical industry.

[0103] Based on the industry characteristic that APIs are a core element of pharmaceutical production, the formula is designed logically around the recovery efficiency and cost of enterprises after a core API supply disruption:

[0104] The parameter is the time (in months) from the discovery of a core supplier's supply disruption to the activation of alternative suppliers and the resumption of normal procurement; the shorter the switching speed, the stronger the company's response capability to API supply disruptions and the shorter the production interruption time.

[0105] The additional costs are calculated as the ratio of the additional costs incurred in switching alternative suppliers to the company's total supplier procurement costs in the previous quarter. Additional costs include alternative supplier qualification review fees, production line process adjustment fees, temporary transportation premiums, etc. The lower the percentage, the smaller the economic cost for the company to switch its supply chain.

[0106] “V×C” reflects the combined “time-cost” cost of switching the supply chain, and its reciprocal reflects the supply chain recovery efficiency. The introduction of a standardization coefficient is to eliminate parameter differences caused by different enterprise sizes and API types, and to uniformly map the supply chain resilience value to the 0-1 range, so as to make the resilience values ​​of different enterprises comparable. This design quantifies the ability of enterprises to cope with API supply chain risks, which is in line with the pharmaceutical industry’s dependence on API supply stability.

[0107] Considering the financial characteristics of pharmaceutical companies, namely "high R&D investment and long capital recovery cycle," a formula is designed based on the logic of the company's cash flow risk resistance capability relative to the industry level:

[0108] The ratio of current available working capital to average monthly operating costs is used for calculation. Available working capital reflects the actual amount of funds a company can use to cope with risks, while average monthly operating costs reflect the company's capital needs to maintain basic operations. The ratio of the two directly reflects the company's ability to survive in the short term (on a monthly basis) without new revenue. The average M_c of pharmaceutical companies of similar size and in the same sub-sector is selected as the benchmark. The financial characteristics of sub-sectors in the pharmaceutical industry vary greatly. Using the industry average as a comparison benchmark can avoid the absolute bias of data from a single company.

[0109] Financial resilience is the ratio of a company's own financial resilience to the industry average. The closer the ratio is to 1, the better the company's cash flow resilience is compared to the industry average. A ratio greater than 1 indicates that the company is better than the industry average, while a ratio less than 1 indicates that the company is weaker than the industry average. This design quantifies financial resilience through relative values, which is more in line with the characteristics of large differences in the pharmaceutical industry's sub-sectors and ensures the horizontal comparability of resilience values.

[0110] The financial resilience value reflects a company's ability to withstand cash flow risks relative to the industry average. When a company's cash flow supports the company for significantly more months than the industry average (e.g., M_c = 12, M_avg = 6, resulting in a resilience value of 2), its financial resilience is already at the top level in the industry. At this point, further increasing the resilience value (e.g., 2, 3) has limited effect on improving the overall credit assessment of the company (the difference in the impact of top-tier financial resilience and super-top-tier financial resilience on credit risk is minimal). Therefore, setting an upper limit of 1 can avoid excessive interference from extreme values ​​in the assessment results. Example 8: The formula for calculating the comprehensive credit score S of the pharmaceutical company in step 5 is:

[0111] S = S0 × [(R1 × β + R2 × γ + R3 × δ + R4 × ε) · ω], where S0 is the basic credit score of the enterprise, β is the R&D resilience weight coefficient, γ is the compliance resilience weight coefficient, δ is the supply chain resilience weight coefficient, ε is the financial resilience weight coefficient, and ω is the environmental adaptability coefficient.

[0112] Example 9 also includes step 6, which uses a timeline view to display the change curves of historical credit scores over the past 12 months and predicted credit scores over the next 12 months.

[0113] Example 10: Credit rating is conducted based on the comprehensive credit score of pharmaceutical companies. The mapping rule between the comprehensive credit score and the credit rating is as follows: 90-100 points are AAA level, corresponding to extremely low risk; 80-89 points are AA level, corresponding to low risk; 70-79 points are A level, corresponding to medium-low risk; 60-69 points are BBB level, corresponding to medium risk; and less than 60 points are BB level, corresponding to high risk.

[0114] By adopting the above technical solution, the basic credit score serves as the core benchmark of the scoring system, focusing on the historical stability and compliance bottom line of pharmaceutical companies' operations. It integrates key indicators adapted to the industry from traditional credit reporting—covering financial health (revenue growth, assets and liabilities, R&D investment ratio), compliance foundation (GMP / GSP effectiveness, unannounced inspection records), and operational stability (supply chain qualification rate, R&D investment fluctuations). Its design logic is to ensure that the score is supported by historical data, avoid subjective predictions that deviate from the company's past operating performance, and provide a reliable benchmark for subsequent dynamic adjustments.

[0115] The weighting coefficients for R&D resilience, compliance resilience, supply chain resilience, and financial resilience are set based on the priority differences in the impact of risks in the pharmaceutical industry: R&D is the core value source of pharmaceutical companies, so R&D resilience has the highest weight, ensuring the key impact of the R&D pipeline's ability to withstand risks on credit; compliance is a prerequisite for the operation of pharmaceutical companies, so compliance resilience has the next highest weight, matching the industry's characteristic that compliance failure means business interruption; the supply chain is directly related to production continuity, and finance is related to the bottom line of the company's survival. The weights of the two are set according to the degree of impact, forming a weighting logic of core value - operating prerequisite - operational guarantee - survival foundation, so that the score can accurately reflect the unique risk resistance capabilities of pharmaceutical companies.

[0116] The environmental adaptability coefficient is designed to address the characteristics of the pharmaceutical industry, which is highly dependent on policies and subject to high market volatility. By quantifying the response of enterprises to changes in the external environment, the comprehensive score is dynamically adjusted. For enterprises that respond positively to policies and achieve good results, the coefficient is adjusted upward to reflect their environmental adaptability; for enterprises that respond lagging behind or achieve poor results, the coefficient is adjusted downward to reflect potential risks. This overcomes the limitations of traditional static scoring in adapting to dynamic changes in the external environment and makes the score more closely match the real-time operating scenarios of pharmaceutical companies.

[0117] The timeline view places historical credit scores from the past 12 months and predicted credit scores for the next 12 months on the same time dimension, connecting them with a linear curve to form a complete link between the past trajectory of credit changes, the current state, and future trends. Its design logic leverages the continuity of time to intuitively present the fluctuation patterns of credit scores with key events, helping users quickly identify the driving factors behind credit score changes.

[0118] To address the phased characteristics of credit changes in the pharmaceutical industry, the timeline view transforms abstract score changes into intuitive trend features through curve slope variations and key node annotations. The principle behind this is to reduce the cost of data interpretation for users, allowing them to grasp the long-term trend and short-term fluctuations of credit scores without complex calculations, thus providing visual support for rapid decision-making.

[0119] The credit score ranges are determined based on the probability and impact of risks in the pharmaceutical industry: Companies scoring 90-100 points demonstrate excellent performance in R&D, compliance, supply chain, and finance, with an extremely low probability of risk, and are therefore rated AAA; those scoring 80-89 points indicate no risk in core dimensions but minor fluctuations in certain areas, with a low probability of risk, and are rated AA; those scoring 70-79 points indicate minor risks in certain dimensions that can be quickly addressed, with a low to medium probability of risk, and are rated A; those scoring 60-69 points indicate significant risk in one dimension requiring close monitoring, with a medium probability of risk, and are rated BBB; and those scoring below 60 points indicate multiple overlapping risks or severe risk in a single dimension, with a high probability of risk, and are rated BB. This range classification logic is highly correlated with the actual risk performance of pharmaceutical companies, ensuring that the rating directly reflects the risk level.

[0120] Different credit ratings correspond to specific risk levels. The principle behind this is to transform abstract credit scores into risk labels that users can directly apply: banks can quickly set credit granting strategies based on the rating (AAA ratings grant long-term, large-amount credit lines, BB ratings suspend new credit lines); investors can use this to assess the risk-reward ratio (A ratings correspond to low to medium risk, suitable for stable investments, BB ratings correspond to high risk, requiring caution); and regulatory agencies can classify and monitor according to ratings (BB-rated companies are listed as key regulatory targets). Through rating mapping, the decision-making process for different users is simplified, enabling credit results to directly adapt to the decision-making needs of multiple scenarios such as finance, investment, and regulation, thus solving the problem that while comprehensive credit scores are accurate, they require complex interpretation.

[0121] The following specific embodiments illustrate the implementation principle of the present invention:

[0122] Innovative pharmaceutical company A is mainly engaged in the research and development and production of targeted cancer drugs. Its core pipeline (Drug X, a targeted drug for lung cancer) is in Phase III clinical trials, and two alternative pipelines (Drug Y and Drug Z, both targeted drugs for gastric cancer) are in Phase II clinical trials. Its core API relies on supplier B (accounting for 80%), and its GMP certification is valid until 2026.

[0123] Step 1: Data Collection and Compliance Processing

[0124] Collect three types of data:

[0125] Structured data: GMP certification records (valid), Phase III clinical trial data of drug X (CDE), medical insurance reimbursement data (medical insurance reimbursement of 120 million yuan for oncology drugs in 20XX), and the progress of drug X / Y / Z R&D pipelines;

[0126] Unstructured data: 20XX medical insurance access negotiation text for oncology drug price reduction, PubMed-indexed papers related to drug X's target, and adverse reaction feedback for drug X (no major feedback in the past 3 months);

[0127] Semi-structured data: Drug X's instruction manual, API procurement contract with supplier B.

[0128] Compliance handling:

[0129] Use time series interpolation to complete the missing R&D investment data for drug Y in Q3 20XX;

[0130] Hash-encrypted patient data from the clinical trials of drug X is authorized for review only.

[0131] The "Start date of Phase III clinical trials for drug X (May 2023)" and "Approval and acceptance date by the drug regulatory authority (November 20XX)" will be uniformly converted into "monthly coordinates" (May 2023 and November 20XX, respectively).

[0132] Generate multimodal data maps:

[0133] Input a graph neural network, associate it with "Pharmaceutical Company AX Drug - Supplier B - 20XX Medical Insurance Access Policy", use a large biomedical model to generate "virtual data on revenue changes if Drug X is included in centralized procurement", and output a graph.

[0134] Step 2: Constructing a digital twin:

[0135] Static model:

[0136] Compliance qualification model: GMP valid (score 100), qualified for unannounced inspections in the past 3 years (score 100);

[0137] R&D pipeline model: Drug X (high target scarcity, success rate 65%, market size 5 billion yuan), Drug Y / Z (medium target scarcity, success rate 50% / 48%, market size 3 billion / 2.5 billion yuan);

[0138] Financial resilience model: R&D investment has fluctuated by 15% over the past 3 years;

[0139] Supply chain model: Supplier B has an 80% dependence rate and low compliance risk, with 2 alternative suppliers.

[0140] Dynamic decision-making model:

[0141] BioGPT was fine-tuned using "pharmaceutical companies' decision-making cases of failed Phase III clinical trials" over the past 10 years, and FinBERT was fine-tuned using "cases of responding to centralized procurement and price reductions" to ensure that the output decisions are traceable (e.g., "if the core pipeline fails, prioritize the advancement of drug Y").

[0142] Environmental interaction:

[0143] Connect with the drug regulatory authority's "2025 Oncology Drug Centralized Procurement Pre-notification" to generate an impact assessment of "If Drug X's price is reduced by 40%, revenue is expected to decrease by 18%", and output a twin.

[0144] Step 3: Risk Simulation

[0145] Risk scenario: "Failure of Phase III clinical trials for drug X + supplier B's supply disruption for 1 month";

[0146] Time step: Set monthly (risk impact cycle 6 months);

[0147] Deduction process:

[0148] GPT-4o Analysis of the Impact: The failure of drug X led to a 60% drop in future revenue expectations, and supplier B's supply disruption led to an 80% drop in monthly production capacity;

[0149] Twin decision: Suspend non-core R&D and activate alternative API supplier C;

[0150] Update status: Production capacity has recovered to 70% after one month, and the expected revenue decline has narrowed to 35%.

[0151] Calculate health indicators: compliance health score 90, R&D pipeline survival rate 80%, cash flow can support 8 months (not meeting termination conditions), and the simulation continues for 6 months.

[0152] Step 4: Calculate the toughness value:

[0153] R&D resilience: Total value of drug Y / Z / Value of drug X ≈ 0.8;

[0154] Compliance resilience: No recent penalties, resilience value 1;

[0155] Supply chain resilience: Switching to supplier C took 1 month and accounted for 5% of the cost, with a resilience value of 0.7;

[0156] Financial resilience: Cash flow support for 8 months / industry average 6 ≈ 1 (take 1).

[0157] Step 5: Comprehensive Credit Score

[0158] Basic credit score: 85;

[0159] Weighting coefficients: R&D resilience 0.3, compliance resilience 0.3, supply chain resilience 0.2, financial resilience 0.2;

[0160] Environmental adaptability coefficient: Actively respond to centralized procurement (1.1);

[0161] Overall credit score: 85×(0.8×0.3+1×0.3+0.7×0.2+1×0.2)×1.1≈82, corresponding to AA level (low risk).

[0162] Step 6: Forecasting and Early Warning

[0163] Credit change path: The timeline view shows the historical score (80-85) from September 20XX to August 2025 and the predicted score (78-83) from September 2025 to August 2026;

[0164] Probability distribution: The Bayesian model outputs "70% probability of a score of 78-83 in Q1 2026";

[0165] Downgrade trigger conditions: Cash flow support for less than 5 months or R&D resilience less than 0.5;

[0166] Warning report: "Pay attention to the clinical progress of drug X, and it is recommended to prepare one alternative API supplier."

[0167] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A credit analysis method for the pharmaceutical industry based on a large model, characterized in that, Includes the following steps: Step 1: Obtain pharmaceutical companies' proprietary data and process it for compliance. Step 2: Construct a digital twin of the pharmaceutical company's specific data; Step 3: Construct multimodal risk scenarios and perform multimodal risk simulation based on digital twins; Step 4: Based on the risk simulation results, calculate the resilience value of the multimodal risks respectively; Step 5: Calculate the comprehensive credit score of pharmaceutical companies based on the resilience value of multimodal risk; Step 6: Use the Prophet time series model combined with the language model to generate the credit change path for the next 12 months, output the probability distribution of credit score intervals through the Bayesian probability model, set the trigger conditions for credit downgrade, and output dynamic credit scores and risk warning reports.

2. The pharmaceutical industry credit analysis method based on a large model according to claim 1, characterized in that, In step 1, the pharmaceutical company-specific data includes structured data, unstructured data, and semi-structured data. Structured data includes GMP certification records, GSP certification records, CDE clinical trial data, drug registration approvals, medical insurance payment data, company R&D pipeline data, and CDE clinical trial data. Unstructured data includes medical insurance access policy texts, PubMed research papers, and adverse drug reaction feedback data. Semi-structured data includes drug instructions and API procurement contracts. The compliance processing method is as follows: use time series interpolation to complete the missing data, use hash encryption and access control to de-identify the controlled data, and uniformly convert it into time axis coordinates to achieve standardization; 3. The pharmaceutical industry credit analysis method based on a large model according to claim 1, characterized in that, In step 1, the data for compliance processing is input into a graph neural network to construct a relational graph of enterprises, drugs, patents, API suppliers, and policies. The relational graph achieves cross-data source entity alignment. Virtual supplementary data is generated using a large biomedical model to complete data augmentation, and a unified multimodal data graph of pharmaceutical enterprises is output.

4. The pharmaceutical industry credit analysis method based on a large model according to claim 1, characterized in that, Step 2, constructing a digital twin of pharmaceutical company-specific data, includes the following sub-steps: Step 21: Construct compliance qualification model, R&D pipeline model, financial resilience model, and supply chain model based on the characteristics of pharmaceutical companies; Step 22: Based on historical decision-making cases of pharmaceutical companies over the past 10 years, fine-tune the R&D decision-making model using BioGPT, fine-tune the financial decision-making model using FinBERT, and fine-tune the compliance response model using the large risk decision-making model. Output the decision-making basis through the thinking chain of the large model. Step 23: Connect with real-time policy channels of drug regulatory authorities and medical insurance departments, as well as drug bidding market data, to generate policy impact assessment and market impact analysis, and output a digital twin of pharmaceutical companies with behavioral simulation capabilities.

5. The pharmaceutical industry credit analysis method based on a large model according to claim 4, characterized in that, In step 3, the multimodal risk scenarios include R&D risk, policy risk, supply chain risk, and market risk; the multimodal risk simulation includes the following steps: The time step is set according to the risk type. First, the direct impact of the risk scenario on the current state of the enterprise is analyzed through GPT-4o. Then, the enterprise decision is output by calling the digital twin. The enterprise status is updated by applying the decision and the scenario impact. Finally, the pharmaceutical-specific health indicators are calculated, including compliance health score, R&D pipeline survival rate and cash flow support. The termination conditions of the pharmaceutical-specific health indicators are set. The simulation stops when the calculated pharmaceutical-specific health indicators meet the termination conditions.

6. The pharmaceutical industry credit analysis method based on a large model according to claim 5, characterized in that, The resilience values ​​for multimodal risks include R&D resilience R1, compliance resilience R2, supply chain resilience R3, and financial resilience R4. The formula for calculating the R&D resilience R1 is: Where S i P is the target scarcity score for the i-th backup R&D pipeline. i M is the clinical trial success rate of the i-th backup R&D pipeline. i S0 is the projected market size of the backup R&D pipeline; P0 is the target scarcity score of the core R&D pipeline; P0 is the clinical trial success rate of the core R&D pipeline; and M0 is the projected market size of the core R&D pipeline. The formula for calculating compliance resilience R2 is: Where L is the percentage of revenue loss due to the penalty, which is obtained by dividing the expected revenue loss amount caused by this compliance penalty by the total revenue of the company in the previous full fiscal year, and T is the penalty rectification period; The formula for calculating supply chain resilience R3 is: Where v is the time from discovering the core supplier's supply disruption to using alternative suppliers and resuming normal procurement, C is the switching cost percentage, which is obtained by dividing the additional cost of switching alternative suppliers by the company's total supplier procurement costs in the previous quarter, and α is the standardization coefficient. The formula for calculating financial resilience R4 is: Where M c It is the number of months that a company's cash flow can sustain. a This is the industry average number of months that cash flow can sustain.

7. The pharmaceutical industry credit analysis method based on a large model according to claim 6, characterized in that, If R4 > 1, then take R4 = 1.

8. The pharmaceutical industry credit analysis method based on a large model according to claim 7, characterized in that, The formula for calculating the comprehensive credit score S of pharmaceutical companies in step 5 is: S = S0 × [(R1 × β + R2 × γ + R3 × δ + R4 × ε) · ω], where S0 is the basic credit score of the enterprise, β is the R&D resilience weight coefficient, γ is the compliance resilience weight coefficient, δ is the supply chain resilience weight coefficient, ε is the financial resilience weight coefficient, and ω is the environmental adaptability coefficient.

9. The pharmaceutical industry credit analysis method based on a large model according to claim 8, characterized in that, It also includes step 6, which uses a timeline view to display the change curves of historical credit scores over the past 12 months and predicted credit scores over the next 12 months.

10. The pharmaceutical industry credit analysis method based on a large model according to claim 8, characterized in that, Credit rating is based on the comprehensive credit score of pharmaceutical companies. The mapping rule between the comprehensive credit score and the credit rating is as follows: 90-100 points is AAA level, corresponding to extremely low risk level; 80-89 points is AA level, corresponding to low risk level. 70-79 is classified as Grade A, corresponding to low to medium risk; 60-69 is classified as Grade BBB, corresponding to medium risk. A score below 60 is classified as BB, which corresponds to a high-risk level.

Images