A clinical scientific research project management system based on AI technology

Abstract

The application discloses a kind of clinical scientific research project management systems based on AI technology, comprising: project declaration and intelligent duplicate checking module, automatically extract project feature label and historical project knowledge base to carry out multidimensional similarity calculation, generate duplicate declaration probability value and intercept duplicate declaration;Subject identity AI dual-mode authentication module, based on three-dimensional grid generation network and geometric consistency detection, live body identity verification is generated, and visit index coding is generated;Subject data AI automatic acquisition and structured input module, value-unit-mean triplets are extracted from multi-source heterogeneous data by BERT model, and are automatically mapped to EDC system;Project progress visualization and intelligent early warning module, in the form of cockpit, milestone, enrollment progress, quality indicators are aggregated and displayed, and early warning is pushed;Personnel resource intelligent scheduling module, based on genetic algorithm to generate optimal scheduling scheme;The application realizes the intelligent support of clinical research project whole process, significantly improves management efficiency, data quality and traceability credibility.

Description

Technical Field

[0001] This invention relates to the fields of medical informatics and artificial intelligence, specifically to a clinical research project management system based on AI technology. Background Technology

[0002] Clinical research project management is a core management activity in the implementation of investigator-initiated investigations (IITs) and registration-type clinical trials in medical institutions. Currently, clinical research project management still suffers from the following deficiencies in several technical aspects:

[0003] 1. The lack of an automated screening mechanism for duplicate project applications.

[0004] When initiating new research projects, clinicians typically rely on personal experience or manual review of historical records to determine whether proposed projects overlap with already approved ones. However, due to the vast amount of project files accumulated by medical institutions over the years, covering different departments, disease areas, and research types, manual retrieval is inefficient and prone to omissions. Researchers may unknowingly submit duplicate applications for projects with similar technical approaches and overlapping research objectives, resulting in wasted ethical review resources, conflicts in participant recruitment, and inefficient allocation of research funding. Existing project management tools only number and register approved projects and lack the capability to perform semantic understanding and similarity comparison of project content during the application stage.

[0005] 2. The reliance on manual verification of participant identities poses a security vulnerability.

[0006] In clinical trial recruitment, informed consent signing, and periodic visits, accurate verification of participant identity is the first line of defense for ensuring the authenticity of trial data and participant safety. Currently, the mainstream operating model still relies on Clinical Research Coordinators (CRCs) to visually compare participants' ID photos with their facial features on-site. This method cannot effectively detect non-liveness attacks such as photo manipulation, pre-recorded video playback, and 3D masking. With the increasing adoption of remote / decentralized clinical trials (DCT), the scenarios where participants complete informed consent and visit confirmation via mobile devices are becoming more frequent, further exacerbating the risk of remote identity forgery. Existing information systems generally lack the capability for liveness detection and 3D geometric consistency verification for medical clinical research scenarios.

[0007] 3. Clinical data collection from test subjects relies on manual transcription, making it difficult to balance efficiency and accuracy.

[0008] Researchers conducting clinical studies need to extract participants' medical data from existing hospital information systems as research data sources. Currently, there is a data pathway gap between EMR, LIS, PACS, and other systems and the Electronic Data Acquisition (EDC / eCRF) system for clinical research. Researchers or CRCs must manually search, copy and paste, and transcribe paper documents to transfer test results, imaging reports, medical records, etc., scattered across different systems to the research database. Data entry for a single participant's visit takes an average of 2 to 3 hours. The manual transcription process not only consumes a large amount of highly skilled human resources but is also prone to errors, omissions, and unit conversion discrepancies, directly affecting the completeness and reliability of the research data.

[0009] 4. The project implementation progress lacks real-time visual tracking methods.

[0010] Clinical research project leaders (usually clinicians) often have to juggle multiple responsibilities, including clinical practice, teaching, and research. During project execution, they often lack real-time access to critical management information such as enrollment progress, data quality control status, funding status, and protocol deviations. They must frequently contact scattered locations, including the institutional office, GCP pharmacy, finance department, and CRC team, via phone, WeChat, or in-person inquiries. This model leads to delayed management information, slow decision-making, and a significant increase in the non-medical workload of the project leader.

[0011] 5. Lack of algorithmic support for personnel scheduling and task allocation in multi-project parallel operation.

[0012] Clinical research institutions commonly face the pressure of high turnover among research coordinators (CRCs) and clinical monitors (CRAs), as well as staff shortages due to multiple concurrent projects. Current staff scheduling often relies on manual spreadsheets created by department secretaries or project managers, depending primarily on personal experience and lacking the ability to comprehensively optimize for multiple constraints such as staff skill sets, real-time workload, and project priorities. The resulting scheduling often exhibits a structural imbalance, with experienced staff overloaded and new staff idle, or insufficient staff for urgent projects and redundant staff for routine projects, impacting overall project execution efficiency.

[0013] 6. Project operation logs are stored in a scattered manner, limiting the reliability of audit and traceability.

[0014] Clinical trial regulatory agencies impose stringent requirements on the authenticity, completeness, and traceability of research data. Existing project management systems often use a centralized local database storage model for operational logs, which exposes log records to the possibility of unauthorized alteration, deletion, or overwriting. In multi-center collaborative scenarios, each center's logs are stored independently, failing to form a unified and credible chain of evidence across centers. This necessitates significant manual effort for comparing and verifying logs from multiple sources when responding to regulatory audits.

[0015] In summary, there is an urgent need in this field for a clinical research project management system that is designed with clinical researchers as the core users and deeply integrates artificial intelligence technology. This system would provide intelligent support for the entire process of project application, subject identity verification, data collection, progress tracking, personnel scheduling, and auditing and tracing, thereby addressing the aforementioned long-standing technical pain points. Summary of the Invention

[0016] This invention provides a clinical research project management system based on AI technology, which aims to achieve: automatic plagiarism detection and novelty assessment during the project application stage; liveness detection of subjects and automatic data collection during the project implementation stage; visual tracking and intelligent early warning of project progress; and intelligent allocation of personnel resources and optimized task assignment.

[0017] To achieve the above objectives, the present invention provides the following technical solution: a clinical research project management system based on AI technology, the system comprising:

[0018] The project application and intelligent plagiarism detection module is used to automatically extract project feature tags when researchers initiate new project applications, perform multi-dimensional similarity calculations with the historical project knowledge base, generate a duplicate application probability value, and intercept the application and push similar historical projects when the probability value exceeds a preset threshold.

[0019] The AI ​​dual-modal authentication module for subject identity is used to collect video streams during the clinical trial process. It performs liveness verification based on a 3D mesh generation network and geometric consistency detection. After successful verification, a visit index code is generated.

[0020] The AI-automated data collection and structured data entry module for subject data is used to access subject diagnosis and treatment data from multiple heterogeneous data sources. It extracts numerical-unit-meaning triples and diagnostic event sequences through the BERT language model and automatically maps them to the EDC / eCRF system to complete data entry.

[0021] The project progress visualization and intelligent early warning module is used to aggregate and display project milestone status, team entry progress, and quality indicators from the perspective of the project leader, and automatically push early warning information when preset rules are triggered.

[0022] The intelligent personnel resource scheduling module is used to collect researchers' skill tags, available time, and task load, and generate the optimal schedule and task allocation scheme based on a genetic algorithm.

[0023] Preferably, in the project application and intelligent plagiarism detection module, the formula for calculating the probability value R of duplicate applications is:

[0024]

[0025] Where M represents the number of data source types for the project. Here, N represents the coefficient corresponding to the m-th type of data source, and N is the number of project keywords. The weight of the project tag corresponding to the nth keyword. Let be the cosine similarity between the nth keyword and the corresponding project tag.

[0026] Preferably, the subject identity AI dual-modal authentication module includes:

[0027] The 3D facial network reconstruction unit is used to input video frames into a 3D mesh generation network, generate a 3D facial network hypothesis and project it onto a 2D plane to obtain projection key points and projection texture features.

[0028] A two-dimensional biometric extraction unit is used to extract biometric key points and local microtexture features from video frames;

[0029] The geometric consistency detection unit is used to compare the alignment error between the projected key points and the biometric key points, as well as the structural similarity index and cosine similarity between the projected texture features and the local microtexture features, to comprehensively determine whether it is a live, real human face.

[0030] Preferably, the subject data AI automatic collection and structured input module includes:

[0031] The multi-source heterogeneous data access unit supports HL7 / FHIR protocol interface with EMR, LIS, and PACS systems, and is compatible with unstructured data formats such as PDF, DICOM, and paper scans.

[0032] The unstructured data parsing unit converts images into text based on OCR technology, performs named entity recognition based on the BERT pre-trained language model, and extracts the value-unit-meaning triplet.

[0033] The automatic pre-filling and mapping unit automatically maps and inputs triples and diagnostic event sequences according to the CRF page field definitions of the EDC system.

[0034] Preferably, the project progress visualization and intelligent early warning module includes:

[0035] The project dashboard unit, built on Vue3's composable API and ECharts, visualizes multi-dimensional project progress indicators in the form of Gantt charts, funnel charts, and dashboards.

[0036] The intelligent early warning rule engine allows researchers to customize early warning rules. Once a rule is triggered, a work order is automatically generated and assigned to the responsible role.

[0037] Preferably, the intelligent personnel resource scheduling module includes:

[0038] The personnel capability profiling unit collects researchers' professional skill tags, available time slices, and current task workload;

[0039] The project priority assessment unit dynamically generates a project priority ranking based on project type, urgency of enrollment, and importance of the sponsor.

[0040] The intelligent scheduling engine generates the optimal schedule and task allocation scheme based on genetic algorithms under the triple constraints of task-capability matching, load balancing, and project priority.

[0041] Preferably, the system also includes a blockchain evidence storage and auditing module, which is used to calculate the hash value of key operation logs throughout the project lifecycle and upload them to the consortium blockchain, supporting authorized parties to retrieve evidence storage records for hash comparison and verification.

[0042] Preferably, the system adopts a front-end and back-end separation architecture. The front-end is built on Vue3 + Pinia, and the back-end is built on Java11 modular isolation + SpringBoot2.5 to build a microservice cluster. The caching layer uses Redis, the persistence layer uses MySQL8.0, and the AI ​​algorithm layer communicates with the back-end through gRPC.

[0043] The beneficial effects of this invention are:

[0044] (I) Beneficial Effects of Project Application and Intelligent Plagiarism Detection Module

[0045] Significantly reduces duplicate reporting rate: Through deep semantic understanding and multidimensional similarity calculation based on the BERT model, the system automatically identifies the overlap between the duplicate reporting and historical projects in terms of disease domain, intervention measures, and research design. The accuracy rate of duplicate reporting interception reaches 92.7%, which is about 65 percentage points higher than the manual retrieval mode, effectively avoiding the waste of ethical review resources and conflicts in subject recruitment.

[0046] Improve application review efficiency: Reduce the average time for plagiarism checking per project from 2-3 working days of traditional manual review to 1.2 seconds, enabling plagiarism checking and real-time feedback upon application. Researchers no longer need to wait for manual search results from the institution's office, significantly shortening the preparation cycle before project launch.

[0047] Knowledge asset accumulation and reuse: The system automatically builds a historical project knowledge base and tag vector index to form a unique research knowledge graph for the institution, supporting researchers to conduct multi-dimensional searches and statistical analyses by disease area, intervention measures, and PI dimension, providing data decision support for scientific research planning.

[0048] (II) Beneficial Effects of the Subject Identity AI Dual-Modal Authentication Module

[0049] Completely eliminate the risk of non-living identity forgery: The first to introduce a joint detection mechanism of 3D mesh reconstruction and 2D texture features into the clinical research identity verification scenario, the comprehensive interception success rate of attack methods such as photo copying, pre-recorded video playback, and 3D masks is no less than 96.5%, completely solving the technical pain point that traditional visual comparison and ordinary face recognition cannot effectively resist deepfakes.

[0050] Supporting remote / decentralized clinical trial models: Through mobile H5 / mini-program adaptation and lightweight 3D reconstruction algorithms, subjects can complete high-security identity authentication in an off-site environment, providing key infrastructure support for the promotion of DCT models and expanding the implementation boundaries of clinical research.

[0051] Achieve full-process identity traceability during visits: After each successful authentication, a globally unique visit index code is generated and strongly linked to all subsequent data collection operations, forming a three-dimensional trusted link of subject identity, visit time, and data operation, which meets the strict traceability requirements of regulatory agencies for the subject identity verification process.

[0052] (III) Beneficial effects of the AI-automated data collection and structured data entry module for test subjects

[0053] Significantly reduced data entry time: The average data entry time per subject per visit has been reduced from 2-3 hours in the traditional manual mode to 15-25 minutes, with a time reduction ratio of over 85%. This frees up highly skilled CRC human resources from inefficient and repetitive transcription work, allowing them to focus on higher-value tasks such as subject management and quality control.

[0054] Significantly improve data quality: Through automated collection and semantic mapping mechanisms, problems such as typos, omissions, and unit conversion errors that are easily introduced in the manual transcription process are completely eliminated. The entry error rate is reduced from about 3.7% to less than 0.2%, and the integrity and reliability of the clinical research data source are fundamentally guaranteed.

[0055] Breaking down silos in medical information systems: It connects to heterogeneous systems such as EMR, LIS, and PACS within hospitals using a non-invasive adapter model, compatible with both structured interfaces and unstructured files. It enables the smooth flow of clinical data to research data without modifying existing clinical information systems, significantly reducing the cost of information technology transformation for medical institutions.

[0056] (iv) Beneficial effects of the project progress visualization and intelligent early warning module

[0057] Achieve real-time visibility into project management status: Project leaders no longer need to inquire with multiple departments through discrete means such as phone calls and WeChat. They can grasp all dimensions of management information in real time through the project dashboard, including team entry progress, quality indicators, budget execution, and plan deviations. The delay in obtaining management information is reduced from an average of 2.3 days to real time, and the decision response speed is improved by more than 90%.

[0058] Reduce researchers' non-medical workload: Clinicians undertake multiple tasks such as diagnosis, teaching and research. This module reduces their workload in managing ongoing projects by about 67%, effectively alleviating researchers' professional burnout and ensuring that the quality of clinical medical services is not excessively squeezed by scientific research management tasks.

[0059] Building a proactive quality management system: Through a configurable intelligent early warning rule engine, the quality control mode of reactive response after the fact is transformed into proactive early warning during the process. The average detection time of risk events such as plan deviation, delayed SAE reporting, and data questioning backlog is reduced by 9.6 days, and the intervention success rate is improved by 41%.

[0060] (v) Beneficial effects of the intelligent personnel resource scheduling module

[0061] Achieving Pareto optimization of human resource allocation: Based on the global optimization capability of genetic algorithms under the triple constraints of task-capability matching, load balancing, and project priority, compared with the manual experience-based scheduling mode, the overall task coverage rate is increased by 22%, the variance of man-hours is reduced by 34%, and the number of times there is insufficient manpower in emergency projects is reduced by 67%.

[0062] Reduce the impact of staff turnover: By accumulating skills tag profiles and historical performance data, newly hired CRCs / CRAs can be quickly matched with suitable projects, shortening the job adaptation period by about 40%, effectively mitigating the project execution continuity risks brought about by the high turnover in the clinical research industry.

[0063] Supports institutional-level resource coordination: Breaks down scheduling barriers between departments and project teams, enabling pooled scheduling of human resources in multi-project parallel scenarios, increasing the overall CRC resource utilization rate of the institution by 28.5%, and increasing the number of projects that can be undertaken with the same staffing by approximately 35%.

[0064] (vi) Beneficial effects of blockchain-based evidence storage and traceability module

[0065] Construct an immutable chain of evidence: Hash key operation logs such as subject authentication, data modification, and project status changes on the blockchain. Any unauthorized tampering with the logs can be detected instantly through on-chain hash comparison, completely solving the security risks of administrators being able to bypass audits and delete or modify records under the traditional centralized log storage model.

[0066] Significantly reduces audit preparation time: In multi-center clinical trial regulatory audit scenarios, auditors no longer need to spend weeks collecting scattered operation logs from various centers and conducting manual comparisons. Through the system's verification function, consistency verification of cross-center logs can be completed in minutes, improving audit efficiency by more than 90%.

[0067] Enhancing the credibility of clinical research data: Blockchain-based evidence records are jointly maintained by all nodes of the consortium blockchain, preventing any single institution from unilaterally altering historical data. This provides sponsors, regulatory agencies, and ethics committees with highly credible proof of data authenticity, thereby increasing the trust of all parties in the research results.

[0068] (vii) Beneficial effects of the overall system architecture

[0069] High availability and high scalability: Adopting a microservice architecture and containerized deployment, each module can be upgraded independently and scaled elastically. The failure of a single module does not affect the overall availability of the system. It supports an average of 150,000 requests per day, with a system availability of 99.97%, meeting the concurrent usage needs of a large tertiary hospital's GCP center.

[0070] Rapid deployment and low-intrusion transformation: Supports both hybrid cloud deployment and pure private deployment modes, without replacing the existing core information systems of medical institutions such as EDC, HIS, and LIS. Low-coupling integration is achieved through the adapter layer, and the average implementation cycle is controlled within 4 weeks.

[0071] Excellent user experience: The front-end application built on Vue3+Pinia supports mobile adaptation and dark mode. The response speed and interface friendliness have been tested by various users, including subjects, researchers, and CRCs. The system usability score (SUS) reached 84.6 points, which is at the excellent level in the industry.

[0072] In summary, this invention, through the deep integration of cutting-edge technologies such as artificial intelligence, 3D vision, natural language processing, operations research, and blockchain, provides a comprehensive and intelligent technical solution to address six long-standing pain points in clinical research project management: duplicate applications, identity fraud, inefficient data transcription, opaque progress tracking, unbalanced scheduling, and fragile traceability and evidence preservation. Compared to existing clinical research information systems, this invention represents a paradigm shift from an "electronic recording tool" to an "intelligent decision-making center," significantly improving the quality and operational efficiency of clinical research project implementation. It strongly supports researchers in initiating high-quality clinical research and registration trials, possessing significant industrial application value and promising prospects for widespread adoption. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0074] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0075] Figure 2 This is a flowchart of the project application and intelligent plagiarism detection module of this invention. Detailed Implementation

[0076] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Example 1: System Overall Architecture

[0078] like Figure 1 As shown, the AI-based clinical research project management system of this invention adopts a microservice architecture with a front-end and back-end separation. The front-end is a single-page application built using Vue 3's composite API and the Pinia state management library, while the back-end is a microservice cluster built using a Java 11 modular isolation environment and the Spring Boot 2.5 framework. The system's caching layer uses Redis 6.2, and the persistence layer uses MySQL 8.0. The AI ​​algorithm layer communicates with the back-end business microservices at high performance via the gRPC protocol.

[0079] The system is divided into six core microservice modules: project application and intelligent plagiarism detection microservice, subject identity AI dual-modal authentication microservice, subject data AI automatic collection and structured input microservice, project progress visualization and intelligent early warning microservice, personnel resource intelligent scheduling microservice, and blockchain evidence storage and audit traceability microservice. Each microservice is deployed and evolves independently, and service discovery and invocation are performed through the registry center (Nacos).

[0080] Example 2: Specific Implementation of the Project Application and Intelligent Plagiarism Detection Module

[0081] according to Figure 2 As shown, this embodiment details the implementation of the project application and intelligent plagiarism detection module.

[0082] 2.1 Automatic Extraction of Project Feature Tags

[0083] When researchers fill out a new project application form (including fields such as project name, research objective, inclusion / exclusion criteria, intervention measures, and primary endpoint) on the system frontend, the frontend asynchronously transmits the filled content to the backend every 3 seconds using anti-shake technology. The backend then calls a BERT-based pre-trained Chinese medical model (such as BioBERT-zh) to perform named entity recognition and relation extraction on the text, automatically extracting project feature tags in the following dimensions:

[0084] (1) Disease domain labels: such as "non-small cell lung cancer" and "type 2 diabetes";

[0085] (2) Intervention label: such as "PD-1 inhibitor", "lifestyle intervention";

[0086] (3) Research type label: such as "randomized controlled trial", "single-arm trial";

[0087] (4) Target population labels: such as "first-line treatment" and "EGFR mutation positive";

[0088] (5) Primary endpoint labels: such as "progression-free survival" and "adverse event rate".

[0089] The tags mentioned above are stored in the project_tags table in MySQL as key-value pairs, and are also synchronously written to the Redis cache to support real-time deduplication.

[0090] 2.2 Construction of Historical Project Knowledge Base

[0091] The system periodically performs offline batch processing on historical projects that have been approved, using the same BERT model as in Section 2.1 to extract feature labels for each historical project and construct a high-dimensional vector index. The vector index is stored using the Faiss library, supporting millisecond-level similarity retrieval for hundreds of millions of vectors. Each historical project stores its set of tag vectors using the project ID as the key.

[0092] 2.3 Multidimensional Similarity Calculation and Repetition Probability Value Generation

[0093] When a new project enters the application stage, the system calculates the probability value R of duplicate applications with historical projects in real time. The specific calculation formula is as follows:

[0094]

[0095] in:

[0096] M represents the number of data source types for the project. In this example, M=5, which correspond to five data source types: project name, research objective, inclusion and exclusion criteria, intervention measures, and primary endpoint indicators.

[0097] The coefficients corresponding to the m-th type of data source are determined by institutional experts through the Analytic Hierarchy Process (AHP). In this embodiment, the default value is C=[0.15,0.25,0.20,0.30,0.10].

[0098] N represents the number of keywords extracted from the current project;

[0099] The item tag weight corresponding to the nth keyword is calculated using the TF-IDF algorithm;

[0100] Let be the cosine similarity between the nth keyword and the corresponding tags of historical projects, and take the average similarity with the tags of the top-3 most similar historical projects.

[0101] The system sets a duplicate submission threshold T=0.65. When R ≥ 0.65, the system determines there is a risk of duplicate submission, displays a pop-up notification window, and shows a list of up to five similar historical projects ranked from highest to lowest similarity, including project number, project name, project initiation date, PI name, and similarity score. Researchers can click to view the details of historical projects or apply for "mandatory submission" and fill in an explanation of the rationale for duplicate submission. Mandatory submission records will be marked and entered into the institution's office manual review queue.

[0102] Example 3: Specific Implementation of the Subject Identity AI Dual-Modal Authentication Module

[0103] This embodiment details the implementation of the subject identity AI dual-modal authentication module.

[0104] 3.1 Hardware Environment and Interaction Process

[0105] This module is suitable for two types of use cases:

[0106] (1) In-hospital scenario: Subjects complete identity authentication in the GCP center visitation room using a fixed high-speed document scanner terminal, which has a built-in RGB camera and an infrared camera;

[0107] (2) Remote / decentralized scenario: Subjects complete authentication by calling the front camera of their mobile device through WeChat mini program or H5 page.

[0108] The authentication process is as follows:

[0109] Step 1: The subject places their ID card in the terminal's card reader area or takes a picture of the national emblem side and the portrait side of the ID card. The system extracts the document information (name, ID number, validity period) and the portrait image through OCR.

[0110] Step 2: The system prompts the subject to complete the specified action according to the voice command (such as nodding, blinking, opening the mouth, shaking the head), while simultaneously collecting a continuous video stream with a frame rate of no less than 25fps and a resolution of no less than 720p;

[0111] Step 3: The system inputs the video stream frame by frame into the AI ​​dual-modal authentication model and outputs the liveness detection results and face similarity in real time;

[0112] Step 4: After successful authentication, the system generates a globally unique Visit Index Code (VIC), with the encoding format "PROJID_VISITNO_YYYYMMDD_HHMMSS_6-bit random number". This code serves as the identity token for all operations during this visit.

[0113] 3.2 Three-dimensional facial network reconstruction and geometric consistency detection

[0114] The core algorithm of this module includes two sub-modules: a 3D mesh generation network and a geometric consistency detection module.

[0115] The 3D facial reconstruction unit employs a 3D Morphable Model (3DMM) variational autoencoder implemented using the PyTorch3d framework. A single-frame RGB image is input into a pre-trained ResNet50 backbone network to extract a 512-dimensional facial feature vector. This vector is then used to regress 200 identity coefficients, 64 expression coefficients, and illumination coefficients from the 3DMM through fully connected layers. A 3D facial mesh (approximately 35,000 vertices) is reconstructed, and based on a weak perspective projection model, the mesh is rendered onto a 2D plane to obtain the coordinates of projection keypoints (e.g., 68 keypoints such as the corners of the eyes, nose, and mouth) and a 256-dimensional projection texture feature vector.

[0116] Two-dimensional biometric extraction unit: The MediaPipe framework is used to extract 468 dense 3D facial key points from the original video frame in real time, and downsampled to 68 key points for alignment with the 3D projection. At the same time, Local Binary Pattern (LBP) and multi-scale Gaussian filtering are used to extract local micro-texture feature vectors (128-dimensional) of the facial region.

[0117] Geometric consistency detection unit: Calculates the normalized mean Euclidean distance (NMED) between projected keypoints and 2D biometric keypoints. The calculation formula is as follows:

[0118]

[0119] in, The Euclidean distance between the outer corners of the eyes is used for scale normalization. Geometric alignment is considered passed when NMED < 0.08.

[0120] Simultaneously, the cosine similarity (CS) and structural similarity index (SSIM) between the projected texture features and the local micro-texture features are calculated. If CS ≥ 0.75 and SSIM ≥ 0.65, the texture consistency is considered acceptable.

[0121] Comprehensive judgment rule: If both geometric alignment and texture consistency pass, and the accuracy of live action command response is ≥80%, then it is judged as a live real face and passes identity authentication.

[0122] 3.3 Verification of Anti-Attack Capabilities

[0123] Tests have shown that this module achieves the following success rates in intercepting the following attack methods: photo re-enhancing attack (99.7%), pre-recorded video playback attack (99.2%), 3D mask attack (96.5%), and screen re-enhancing attack (98.9%), meeting the high-security requirements of clinical research scenarios.

[0124] Example 4: Specific Implementation of the Subject Data AI Automatic Acquisition and Structured Input Module

[0125] This embodiment details the implementation of the AI-based automatic data collection and structured data entry module for test subjects.

[0126] 4.1 Access to multi-source heterogeneous data

[0127] The system has a built-in medical data adapter and supports the following access methods:

[0128] (1) Standard protocol access: Connect to hospital information systems such as EMR, LIS, and RIS through HL7 v2.x or FHIR R4 protocol to subscribe to relevant diagnosis and treatment events of subjects;

[0129] (2) Direct database connection: With authorization, the test results and examination reports of the designated subjects are directly connected to the hospital's business database through a JDBC read-only account on a regular basis.

[0130] (3) Unstructured file access: Supports uploading of PDF, DICOM, JPEG, PNG and other file formats, as well as scanning of paper documents.

[0131] 4.2 Unstructured Data Parsing and Semantic Extraction

[0132] The core of this module is a BERT-based medical text parsing engine, and the processing flow is as follows:

[0133] (1) OCR recognition: For unstructured data such as scanned documents and screenshots, the PaddleOCR engine is called to perform text recognition and output a text layer containing position coordinates;

[0134] (2) Text cleaning: Remove noise such as headers and footers, table lines, and special symbols, and perform traditional and simplified Chinese conversion and full-width and half-width normalization;

[0135] (3) Named entity recognition: Load the BERT-BiLSTM-CRF model finely tuned based on CMeEE (Chinese Medical Named Entity Dataset) to recognize the following entity categories: laboratory indicator name (LAB), value (VALUE), unit (UNIT), diagnosis name (DIAG), surgery name (OP), and drug name (DRUG);

[0136] (4) Triple extraction: A combination of dependency parsing and predefined pattern matching is used to construct "numerical value-unit-meaning" triples. For example, extract (6.8, 10^9 / L, white blood cell count) from the text "white blood cell count 6.8×10^9 / L".

[0137] (5) Time-sequencing: Sort the triplet according to the examination / test execution time to generate a personalized diagnosis and treatment event sequence for the subject.

[0138] 4.3 Automatic pre-filling and mapping

[0139] The system comes pre-installed with an EDC system adapter, supporting API integration with mainstream commercial EDC systems (such as Medidata Rave and Veeva VaultCDMS) and open-source EDC systems (such as OpenClinica).

[0140] The automatic mapping logic is as follows:

[0141] (1) Field fingerprint matching: Extract the tag text of the CRF page field, calculate its Jaccard similarity and edit distance with the extracted entity name, and take the weighted comprehensive similarity;

[0142] (2) Automatic unit conversion: For indicators with multiple units (such as blood glucose: mg / dL and mmol / L), the system has built-in common medical indicator unit conversion formulas to automatically convert the source data units to the CRF target units;

[0143] (3) Diagnostic event sequence mapping: For repetitive structured field groups such as "past medical history" and "adverse events", the diagnosis name, occurrence time and outcome status are automatically filled into the corresponding rows in date order;

[0144] (4) Manual verification mechanism: For automatic mapping results with a confidence level of less than 0.90, the system marks them with a highlighted background on the CRF page and prompts the CRC or researcher to manually confirm them.

[0145] Actual measurements showed that the average data entry time per subject per visit was reduced from 2-3 hours in the traditional manual mode to 15-25 minutes, and the data entry error rate was reduced from about 3.7% to below 0.2%.

[0146] Example 5: Specific Implementation of the Project Progress Visualization and Intelligent Early Warning Module

[0147] This embodiment details the implementation of the project progress visualization and intelligent early warning module.

[0148] 5.1 Project Cockpit Unit

[0149] A project-level dashboard is built using Vue 3's composable API and ECharts 5. The main views include:

[0150] (1) Milestone Gantt Chart: The horizontal axis is the time axis, and the vertical axis is the key milestone nodes of the project (institutional establishment, ethics review, contract signing, first case enrollment, last case exit, database locking, summary report). Different colored blocks represent the deviation between the planned time and the actual time.

[0151] (2) Enrollment progress funnel chart: showing the number of cases screened, the number of cases successfully enrolled, the number of cases completed visits, the number of dropouts, and the conversion rate at each stage;

[0152] (3) Quality indicator dashboard: Displays quality KPIs such as data questioning rate, scheme deviation rate, SAE report timeliness rate, and CRF completion rate in the form of radar chart;

[0153] (4) Funding execution dashboard: Displays the total project budget, amount already paid, amount to be paid, and execution percentage of each item (subject subsidy, examination fee, CRC working hours fee).

[0154] Project leaders can switch between "Projects I am in charge of", "Projects I am involved in", and "All projects" perspectives with one click, and support multi-level drill-down to the single subject visit calendar.

[0155] 5.2 Intelligent Early Warning Rule Engine

[0156] The system has a built-in rule engine that uses the Drools framework to define, store, and execute rules. A rule consists of three elements: trigger condition, alert level, and action.

[0157] The triggering conditions support logical combinations of the following dimensions:

[0158] (1) Time dimension: such as "≤30 days until the expiration of the ethics approval document";

[0159] (2) Progress dimension: such as "actual number of enrolled cases / planned number of enrolled cases < 80% and more than half the time has passed";

[0160] (3) Quality dimension: such as "the same subject has raised questions ≥ 3 times";

[0161] (4) Security dimension: such as “unexpected SAE reporting timeout”.

[0162] The warning levels are divided into three levels: blue (alert), orange (attention), and red (emergency).

[0163] The actions taken include:

[0164] (1) Work orders are generated within the system and automatically assigned to preset responsible roles;

[0165] (2) Push warning cards to project groups via WeChat / DingTalk robots;

[0166] (3) Trigger an email notification to the designated personnel;

[0167] (4) Displayed on the top of the project cockpit.

[0168] The rules support hot reloading, and administrators can add, modify, enable / disable rules through the visual rule editor without restarting the service.

[0169] Example 6: Specific Implementation of the Intelligent Personnel Resource Scheduling Module

[0170] This embodiment details the implementation of the intelligent personnel resource scheduling module.

[0171] 6.1 Personnel Capability Profile Construction

[0172] The system collects the following three types of data to construct a capability profile for CRC / CRA:

[0173] (1) Static attributes: employee ID, name, department, date of employment, and professional title;

[0174] (2) Skill tags: These are derived from questionnaire assessment and historical task completion quality calculations, covering more than 20 binary tags such as Phase I ward experience, oncology project experience, vaccine project experience, image transmission operation, and English CET-6.

[0175] (3) Dynamic load: The employee’s occupied working hours (including fixed meetings, training and assigned tasks) for the next 14 days are obtained through calendar integration, and the available remaining working hours are calculated.

[0176] 6.2 Project Priority Assessment

[0177] The system dynamically calculates the project execution priority P from three dimensions:

[0178]

[0179] in:

[0180] U represents the enrollment urgency, calculated as (planned enrollment number - number of enrolled cases) / remaining days;

[0181] I represents the importance of the sponsor, which is pre-set by the organization (1-5 points).

[0182] T is the project type coefficient, with 1.2 for clinical trials of registered drugs and 1.0 for IIT studies;

[0183] α, β, and γ are weighting coefficients. In this embodiment, α = 0.5, β = 0.3, and γ = 0.2.

[0184] 6.3 Intelligent Scheduling Engine

[0185] The core of this module is an intelligent scheduling engine based on genetic algorithms, with constraints including:

[0186] (1) Hard constraints: No one can be assigned two tasks in the same time slot; the personnel must possess the skill tags required for the task; personnel can work continuously for no more than 4 hours;

[0187] (2) Soft constraints: tasks with high project priority are assigned first; personnel workload is balanced (the variance of working hours occupied by each person is minimized); personnel skills are matched with task difficulty.

[0188] The encoding scheme uses integer encoding, with the chromosome length equal to the total number of tasks to be assigned, and each gene bit representing a person's ID. The fitness function is defined as:

[0189] Fitness=w1⋅Coverage−w2⋅Variance−w3⋅PriorityPenalty

[0190] Where Coverage is the task coverage rate, Variance is the variance of human employee hours, and PriorityPenalty is the penalty for low-priority tasks occupying high-priority resources.

[0191] Algorithm parameter settings: population size 200, crossover probability 0.8, mutation probability 0.1, maximum number of iterations 500. Verified in actual projects, this algorithm can output a weekly scheduling plan containing 30 tasks and 15 CRCs within 15 seconds, improving overall task coverage by 22% and reducing human work hour variance by 34% compared to manual scheduling, while reducing the number of times manpower is insufficient in urgent projects by 67%.

[0192] Example 7: Specific Implementation of the Blockchain Evidence Storage and Investigation Tracing Module

[0193] This embodiment details the implementation of the blockchain-based evidence storage and traceability module.

[0194] 7.1 Scope of Evidence Preservation

[0195] The system stores the following key operation logs on the blockchain:

[0196] (1) Subject authentication record: authentication time, VIC code, face comparison similarity score, liveness detection result;

[0197] (2) Data collection and modification records: CRF field value before modification, value after modification, modifier, modification time, and reason for modification;

[0198] (3) Project status change record: Project phase transition, milestone completion confirmation;

[0199] (4) Warning work order processing record: the entire lifecycle of work order generation, assignment, response, and closure;

[0200] (5) Personnel scheduling change record: scheduling table generation time and adjustment record.

[0201] 7.2 Implementation of Evidence Preservation Technology

[0202] The system uses the FISCO BCOS 2.0 consortium blockchain framework to build the evidence storage nodes. Each medical institution deploys its own local blockchain node, forming a multi-center consortium blockchain.

[0203] Evidence preservation process:

[0204] Step 1: The system generates an operation log and calculates the SHA-256 hash value of the entire log text;

[0205] Step 2: Package the hash value, operation timestamp, operation user ID, project ID, and organization ID into structured evidence storage data;

[0206] Step 3: Call the smart contract to write the evidence data to the blockchain. After the transaction is recorded on the chain, the transaction ID and block height are returned.

[0207] Step 4: Write the transaction ID and block height back to the local MySQL evidence storage table.

[0208] 7.3 Verification and Inspection

[0209] Regulatory personnel or authorized inspectors can enter a unique identifier (such as an operation record ID) of the log to be verified through the system's "Inspection and Tracing" interface. The system will automatically extract the original log content from the local database and recalculate the hash value, while simultaneously pulling the evidence storage hash value from the blockchain node. If the two values ​​match, the interface will display "Verification Passed" along with the evidence storage time, block height, and transaction ID; if they do not match, it will display "Verification Failed" and the possible time point of tampering.

[0210] This module effectively solves the technical problems of scattered storage of multi-center logs and their susceptibility to unauthorized tampering, and greatly improves the credibility and audit efficiency of clinical research data in regulatory verification.

[0211] Example 8: System Deployment and Performance Indicators

[0212] This system supports both hybrid cloud deployment and pure private deployment modes. In a specific implementation example of this invention, the system was deployed at the GCP center of a top-tier tertiary hospital, with the following core hardware configuration:

[0213] Application server: 4 virtual machines, each with an 8-core CPU and 32GB RAM;

[0214] Database server: 2 physical machines for dual-machine hot standby, each with a 16-core CPU, 64GB RAM, and SSD disk array;

[0215] AI inference server: 2 GPU servers, each equipped with an NVIDIA Tesla T4 graphics card (16GB of video memory).

[0216] Blockchain node server: A consensus cluster consisting of 3 virtual machines.

[0217] After three consecutive months of stress testing in the production environment, the key performance indicators of the system are as follows:

[0218]

[0219] Example 9: User Experience Optimization Design

[0220] To enhance the user experience for clinical researchers, the system has undergone refined interaction design in several areas:

[0221] (1) Intelligent filling of application forms: Based on the project keywords entered by the researcher, the system suggests historical project templates and supports one-click application of CRF design and visit plan;

[0222] (2) Real-time feedback of subject certification results: During the certification process, the front-end interface displays the status of the action in real time with green / red boxes to avoid subjects waiting in vain;

[0223] (3) Visualization of data collection confidence: The confidence level of the data fields automatically collected by AI is marked with three colored dots of "high / medium / low". Clicking on the dots will allow you to view the specific evidence.

[0224] (4) Mobile adaptation: The core card of the project cockpit, the processing of early warning work orders, the review of the subject authentication results and other functions have been adapted to the mobile H5, which supports researchers to keep track of the project dynamics anytime and anywhere;

[0225] (5) Dark mode: The front end fully implements the dark theme to reduce visual fatigue during long-term use at night.

[0226] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A clinical research project management system based on AI technology, characterized in that, The system includes: The project application and intelligent plagiarism detection module is used to automatically extract project feature tags when researchers initiate new project applications, perform multi-dimensional similarity calculations with the historical project knowledge base, generate a duplicate application probability value, and intercept the application and push similar historical projects when the probability value exceeds a preset threshold. The AI ​​dual-modal authentication module for subject identity is used to collect video streams during the clinical trial process. It performs liveness verification based on a 3D mesh generation network and geometric consistency detection. After successful verification, a visit index code is generated. The AI-automated data collection and structured data entry module for subject data is used to access subject diagnosis and treatment data from multiple heterogeneous data sources. It extracts numerical-unit-meaning triples and diagnostic event sequences through the BERT language model and automatically maps them to the EDC / eCRF system to complete data entry. The project progress visualization and intelligent early warning module is used to aggregate and display project milestone status, team entry progress, and quality indicators from the perspective of the project leader, and automatically push early warning information when preset rules are triggered. The intelligent personnel resource scheduling module is used to collect researchers' skill tags, available time, and task load, and generate the optimal schedule and task allocation scheme based on a genetic algorithm.

2. The system according to claim 1, characterized in that, In the project application and intelligent plagiarism detection module, the formula for calculating the probability value R of duplicate applications is: Where M represents the number of data source types for the project. Here, N represents the coefficient corresponding to the m-th type of data source, and N is the number of project keywords. The weight of the project tag corresponding to the nth keyword. Let be the cosine similarity between the nth keyword and the corresponding project tag.

3. The system according to claim 1, characterized in that, The subject identity AI dual-modal authentication module includes: The 3D facial network reconstruction unit is used to input video frames into a 3D mesh generation network, generate a 3D facial network hypothesis and project it onto a 2D plane to obtain projection key points and projection texture features. A two-dimensional biometric extraction unit is used to extract biometric key points and local microtexture features from video frames; The geometric consistency detection unit is used to compare the alignment error between the projected key points and the biometric key points, as well as the structural similarity index and cosine similarity between the projected texture features and the local microtexture features, to comprehensively determine whether it is a live, real human face.

4. The system according to claim 1, characterized in that, The AI-automated data collection and structured input module for subject data includes: The multi-source heterogeneous data access unit supports HL7 / FHIR protocol interface with EMR, LIS, and PACS systems, and is compatible with unstructured data formats such as PDF, DICOM, and paper scans. The unstructured data parsing unit converts images into text based on OCR technology, performs named entity recognition based on the BERT pre-trained language model, and extracts the value-unit-meaning triplet. The automatic pre-filling and mapping unit automatically maps and inputs triples and diagnostic event sequences according to the CRF page field definitions of the EDC system.

5. The system according to claim 1, characterized in that, The project progress visualization and intelligent early warning module includes: The project dashboard unit, built on Vue3's composable API and ECharts, visualizes multi-dimensional project progress indicators in the form of Gantt charts, funnel charts, and dashboards. The intelligent early warning rule engine allows researchers to customize early warning rules. Once a rule is triggered, a work order is automatically generated and assigned to the responsible role.

6. The system according to claim 1, characterized in that, The intelligent personnel resource scheduling module includes: The personnel capability profiling unit collects researchers' professional skill tags, available time slices, and current task workload; The project priority assessment unit dynamically generates a project priority ranking based on project type, urgency of inclusion, and importance of the sponsor. The intelligent scheduling engine generates the optimal schedule and task allocation scheme based on genetic algorithms under the triple constraints of task-capability matching, load balancing, and project priority.

7. The system according to claim 1, characterized in that, The system also includes a blockchain evidence storage and auditing module, which is used to calculate the hash value of key operation logs throughout the project lifecycle and upload them to the consortium blockchain, supporting authorized parties to retrieve evidence records for hash comparison and verification.

8. The system according to claim 1, characterized in that, The system adopts a front-end and back-end separation architecture. The front-end is built on Vue3 + Pinia, and the back-end is built on Java11 modular isolation + SpringBoot2.5 to build a microservice cluster. The caching layer uses Redis, the persistence layer uses MySQL8.0, and the AI ​​algorithm layer communicates with the back-end through gRPC.

Images