A diabetic retinopathy early screening and primary intelligent follow-up management system

Abstract

This invention discloses an early screening and intelligent follow-up management system and method for diabetic retinopathy (DR), belonging to the field of artificial intelligence in medicine and primary care chronic disease management. The system consists of a fundus image acquisition terminal, a lightweight DR AI diagnostic module, a multimodal physiological monitoring unit, an intelligent follow-up and early warning engine, a hierarchical diagnosis and treatment scheduling platform, and a cloud-edge collaborative server. The DR AI diagnostic module is built based on EfficientNetV2-B0, with a file size ≤15MB after INT8 quantization and a single-image inference time ≤50ms, allowing it to run offline on mobile phones and tablets. After training on 10386 labeled fundus images, it achieves an accuracy of 93.7%, an AUC of 0.952, and a sensitivity of 91.3%, effectively identifying early lesions. The system connects to devices such as blood glucose meters and blood pressure monitors to achieve real-time monitoring of physiological indicators, pushing early warnings within 10 seconds of abnormalities, improving follow-up compliance, and establishing a closed loop of "primary screening at the grassroots level—higher-level diagnosis and treatment—community follow-up," thereby improving the efficiency of primary care chronic disease screening and management.

Description

Technical Field This invention belongs to the fields of artificial intelligence medical applications, primary chronic disease health management, smart elderly care and digital medical technology. Specifically, it relates to a collaborative management system and implementation method for intelligent screening of diabetic retinopathy, multimodal physiological data monitoring, personalized follow-up intervention and hierarchical diagnosis and treatment based on lightweight deep learning. It is applicable to the routine management of chronic diseases in township health centers, village clinics, community health service centers and elderly diabetic groups. Background Technology Diabetic retinopathy (DR) is one of the most common chronic microvascular complications of diabetes, with an overall prevalence of 24.7% to 37.5% in the diabetic population, and the risk of blindness is 25 times higher than in the general population. Early diabetic retinopathy often presents with no obvious symptoms and minimal vision loss. However, once it progresses to the proliferative phase, it can lead to serious consequences such as retinal hemorrhage and retinal detachment, resulting in irreversible vision damage. Clinical data shows that early screening and standardized intervention can prevent more than 90% of severe vision loss. The following practical problems exist in current primary healthcare and elderly health management:

[0001] Adherence to follow-up care for chronic diseases is low at the grassroots level, and the continuity of health management is insufficient.

[0002] The lack of real-time health monitoring and abnormal alert mechanisms means that abnormal indicators cannot be detected in a timely manner.

[0003] Traditional fundus cameras are expensive and complex to operate, making them difficult to popularize in village-level medical centers; existing AI diagnostic models are bulky and resource-intensive, and cannot run offline on low-computing-power terminals such as mobile phones and tablets.

[0004] Fifth, screening, monitoring, follow-up, and referral are independent of each other and have not formed a standardized closed loop, making it easy for high-risk patients to be lost to follow-up or missed diagnosis. In summary, existing technologies cannot meet the needs of grassroots communities for low-cost, widespread, offline, and closed-loop management. Therefore, developing a diabetic retinopathy early screening and intelligent follow-up management system tailored to grassroots scenarios is of urgent practical significance. Summary of the Invention The purpose of this invention is to overcome the shortcomings of the existing technology and provide a practical, low-cost, AI-driven, closed-loop process management system for early screening and intelligent follow-up of diabetic retinopathy, suitable for promotion at the grassroots level. This system addresses practical problems such as insufficient medical resources at the grassroots level, inconvenient screening for elderly patients, low follow-up compliance, untimely referrals, and lack of remote monitoring. The present invention adopts the following technical solution: A management system for early screening and intelligent follow-up of diabetic retinopathy at the grassroots level is characterized by comprising: a fundus image acquisition terminal, a lightweight DR AI diagnostic module, a multimodal physiological monitoring unit, an intelligent follow-up and early warning engine, a hierarchical diagnosis and treatment scheduling platform, and a cloud-edge collaborative server. The DR AI diagnostic module is a five-class classification model built on a lightweight deep neural network, which can automatically identify normal, mild, moderate, severe, and proliferative diabetic retinopathy. The model is no larger than 15MB after INT8 quantization and compression, and can run offline on the CPU of a regular smartphone or tablet. The inference time for a single fundus image is no more than 50ms. The model has been trained and validated on no less than 10,000 clinically labeled fundus images, with a classification accuracy of 92%~95%, AUC≥0.94, sensitivity≥90%, and specificity≥88%, which can effectively reduce the missed diagnosis rate of high-risk patients. The fundus image acquisition terminal supports low-cost devices such as smartphones and portable fundus cameras. It has automatic image preprocessing, illumination correction, size normalization and quality verification functions, and can output standardized fundus images with a resolution of 224×224×3, which meets the input requirements of AI models. The multimodal physiological monitoring unit supports access to at least eight types of health monitoring terminals, including blood glucose meters, electronic blood pressure monitors, smart bracelets, smartwatches, body fat scales, and electrocardiogram monitoring devices. It can collect, analyze, and upload physiological indicators such as blood glucose, blood pressure, heart rate, activity level, and sleep in real time to form a continuous and dynamic personal health record. The intelligent follow-up and early warning engine has built-in abnormal judgment rules based on clinical guidelines. When blood glucose, blood pressure or AI screening results reach the risk threshold, it can push graded early warning information to the elderly's terminal, family members' WeChat and community doctor's terminal within 10 seconds, which can improve the compliance rate of elderly diabetic patients with standardized follow-up from less than 30% to more than 80%. The hierarchical diagnosis and treatment scheduling platform can automatically identify high-risk patients based on AI diagnosis results, generate standardized electronic referral forms, personalized follow-up plans and follow-up tasks, and connect with the ophthalmology diagnosis and treatment system of higher-level hospitals to realize functions such as appointment registration, green channel referral, and feedback of diagnosis and treatment results. The cloud-edge collaborative service terminal is deployed using a SaaS cloud architecture, providing three types of entry points: institutional management terminal, doctor workstation, and family member mini-program. It supports commercial models such as annual subscription by institution, billing by screening visits, and family health protection subscription, and can connect to the national basic public health service chronic disease management data reporting interface. This invention also discloses a method for early screening and intelligent follow-up management of diabetic retinopathy, comprising the following steps: 1. Acquire fundus images through a data acquisition terminal and perform standardized preprocessing; 2. A lightweight AI model completes DR classification diagnosis and risk assessment locally and offline; 3. Multimodal devices upload physiological indicators in real time, and the system dynamically updates the health profile; 4. Real-time tiered early warning and push notifications for abnormal indicators and high-risk screening results; 5. High-risk patients are automatically entered into the referral process, completing appointments and treatment at higher-level hospitals; 6. After patients return to the community, the system performs long-term follow-up, health reminders, and data tracking, forming a closed loop throughout the entire process. The present invention also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the above-described method. Compared with the prior art, the present invention has the following advantages: 1. Suitable for use at the grassroots level, requiring no professional equipment or dedicated physicians, and can be deployed to grassroots medical points.

[0005] 2. The AI ​​model is lightweight, supports offline operation, and has low deployment environment requirements.

[0006] 3. More accurate physiological monitoring: It connects to everyday home health devices, providing continuous and reliable data that suits the usage habits of the elderly.

[0007] 4. Improve follow-up compliance through multi-level reminders and automatic tracking mechanisms.

[0008] 5. A true closed loop for hierarchical medical services: screening, early warning, referral, follow-up visit, and follow-up are all connected to reduce loss to follow-up. Detailed Implementation

[0009] To make the technical solution of this invention clearer, more complete, and more practically applicable, the following detailed description is provided in conjunction with actual application scenarios. Example 1: System Deployment and Practical Application A township health center deployed the system of this invention, using a smartphone and an external simple fundus lens as the data acquisition terminal, and a medical tablet to run the AI ​​diagnostic module.

[0010] The operator takes fundus photos of elderly diabetic patients, and the system automatically preprocesses the images and provides the DR classification results within 42 milliseconds.

[0011] Meanwhile, patients upload their daily data using home blood glucose meters and blood pressure monitors, and the system automatically generates health trend charts.

[0012] When a patient experiences persistently high blood sugar or a moderate to high risk of renal impairment (DR), the system automatically sends an alert to the family and the family doctor.

[0013] For high-risk patients, the system automatically generates a referral form and connects to the green channel of the ophthalmology department of the county-level hospital to complete the registration and examination appointment.

[0014] After a patient completes treatment, the treatment data is automatically transmitted back to the community, and doctors use the platform to develop follow-up plans and remind patients to have regular check-ups. Actual pilot data shows: - The coverage rate of fundus screening increased from 13.7% to 81.2%. - Follow-up compliance improved from 27.3% to 82.6%. - Average waiting time for high-risk patients has been reduced by 70%. - Increased work efficiency of primary care physicians by more than 60% Example 2: AI Model Performance in Real-World Scenarios The AI ​​model of this invention uses publicly available datasets such as DDR, IDRide, and Messidor, as well as desensitized clinical data from partner hospitals, totaling 10,386 labeled fundus images.

[0015] The model training environment is a conventional deep learning server, and the deployment environment is a regular Android tablet and mobile phone.

[0016] The actual performance of the model is as follows: Accuracy: 93.7% - AUC: 0.952 - Sensitivity: 91.3% - Specificity: 88.7% - Model size: 13MB - Single-card inference time: 42ms It fully meets the real needs of offline screening at the grassroots level. Example 3: Commercialization and Policy Implementation This system can provide SaaS services to primary healthcare institutions, with an annual technical service fee charged. It can provide health monitoring subscription services for children who live away from home; It can undertake public health projects such as chronic disease management for the elderly, prevention and treatment of eye diseases, and digital health from local health commissions.

[0017] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the foregoing technical solutions, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart follow-up management system for early screening and primary care of diabetic retinopathy, characterized in that, include: The system comprises a fundus image acquisition terminal, a lightweight DR AI diagnostic module, a multimodal physiological monitoring unit, an intelligent follow-up and early warning engine, a hierarchical diagnosis and treatment scheduling platform, and a cloud-edge collaborative server. The DR AI diagnostic module is a deep learning-based five-category diagnostic model for diabetic retinopathy. After INT8 quantization and compression, the model size is no more than 15MB, the inference time for a single fundus image is no more than 50ms, the classification accuracy is 92%~95%, the AUC value is no less than 0.94, and the sensitivity is no less than 90%. The system is used to construct a closed-loop management system for the entire hierarchical diagnosis and treatment process, from initial AI screening at the grassroots level to diagnosis at higher-level hospitals and long-term follow-up in the community.

2. The system according to claim 1, characterized in that, The DR AI diagnostic module is based on the EfficientNetV2-B0 lightweight network architecture. The model input is a standardized fundus image with a resolution of 224×224×3, and the output classification results are as follows: normal fundus, mild diabetic retinopathy, moderate diabetic retinopathy, severe diabetic retinopathy, and proliferative diabetic retinopathy.

3. The system according to claim 1, characterized in that, The fundus image acquisition terminal is compatible with smartphones and portable fundus camera acquisition devices. It has image preprocessing, normalization and quality verification functions and can automatically generate standardized fundus images adapted to AI model input.

4. The system according to claim 1, characterized in that, The multimodal physiological monitoring unit supports access to at least eight types of health monitoring terminals, including blood glucose meters, electronic blood pressure monitors, smart bracelets, smartwatches, body fat scales, and electrocardiogram monitoring devices. It can collect, analyze, and upload physiological characteristic data such as blood glucose, blood pressure, heart rate, activity level, and sleep in real time.

5. The system according to claim 1, characterized in that, The intelligent follow-up and early warning engine has a built-in strategy for identifying abnormal physiological indicators and risk classification alarms. When the monitoring data or AI screening results reach the high-risk threshold, the early warning information can be pushed to the elderly terminal, family monitoring terminal and community doctor terminal within 10 seconds, which can increase the compliance rate of elderly diabetic patients with standardized follow-up from less than 30% to more than 80%.

6. The system according to claim 1, characterized in that, The hierarchical diagnosis and treatment scheduling platform can automatically mark high-risk patients based on DR AI diagnosis results, generate standardized electronic referral forms, personalized follow-up plans and optimal treatment paths, and achieve green channel connection with the ophthalmology diagnosis and treatment system of higher-level hospitals.

7. The system according to claim 1, characterized in that, The cloud-edge collaborative service terminal is deployed using a SaaS architecture, which includes three levels of operation entry points: institutional management terminal, doctor workstation, and family member mini-program. It supports annual subscription for To B institutions, billing per screening session, and paid service model for To C family health protection, and is compatible with grassroots public health chronic disease management data reporting interface and policy project docking interface.

8. A method for early screening and intelligent follow-up management of diabetic retinopathy applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: Standardized fundus images are acquired and generated using a fundus image acquisition terminal; Standardized fundus images are input into a lightweight DR AI diagnostic module to complete offline hierarchical diagnosis and disease risk assessment; By connecting to external health monitoring equipment through a multimodal physiological monitoring unit, continuous collection and dynamic monitoring of patients' physiological indicators can be achieved; The intelligent follow-up and early warning engine performs real-time graded early warnings and push notifications for abnormal indicators and high-risk screening results. The hierarchical medical system automatically completes referral matching, appointment booking, and treatment process scheduling for high-risk patients through the hierarchical medical system scheduling platform. After patients complete their treatment at higher-level hospitals and return to the community, the system performs continuous follow-up, health intervention, and data tracking, forming a closed loop of the entire process of screening, diagnosis, referral, and follow-up.

9. The method according to claim 8, characterized in that, The DR AI diagnostic module completes model training, verification, and optimization based on no less than 10,000 professionally annotated fundus images. It can perform offline inference in the CPU environment of smartphones, tablets, and primary healthcare terminals, and the diagnosis time for a single fundus image is no more than 50ms.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the method for early screening and intelligent follow-up management of diabetic retinopathy as described in any one of claims 8 to 9.