HuiMei Technology Leverages AI to Tackle Pulmonary Embolism Prevention Challenges in VTE Management

Since the late 1980s, surgeons such as Caprini have been dedicated to developing a highly detailed, individualized risk assessment model for venous thromboembolism (VTE), continuously optimizing its accuracy and clinical utility. Nevertheless, more than four decades later, pulmonary embolism remains a significant cause of unexpected death among hospitalized patients.

The high mortality rate stems from the low proportion of patients receiving VTE prophylaxis, as well as physicians’ subjective and objective neglect of VTE risk assessment models. Under normal circumstances, if healthcare professionals could accurately assess patients’ VTE risk, they would be able to promptly identify high-risk individuals and provide appropriate preventive treatment. However, in practice, the implementation of this scoring system has encountered numerous challenges.

“It’s too time-consuming.” When asked why various VTE assessment scales have failed to achieve their intended preventive and therapeutic effects, a physician at a Grade 3A hospital in Beijing replied without hesitation, “In large hospitals, it is already a miracle if ward beds are not overbooked. Even with the best intentions to prevent VTE, doctors and nurses cannot thoroughly assess every patient’s condition. Moreover, the issue of VTE is not merely due to negligence or non-standardized practices by healthcare professionals; at times, it may also stem from physicians’—particularly young physicians’—insufficient familiarity with clinical guidelines and lack of adequate evidence-based judgment.”

The challenges of VTE prevention and management are not confined to large hospitals; in fact, primary healthcare institutions may face even more severe issues, as many physicians in these settings lack awareness of VTE prevention and the ability to identify at-risk patients.

With numerous challenges arising, fortunately, those with determination are unafraid to confront them and actively seek solutions.

VTE is a preventable condition; we cannot wait until thrombosis occurs to address the problem. If early screening and prevention of VTE are effectively implemented, the major challenges in VTE prevention and control will be readily resolved. The key lies in integrating sufficient patient information, enabling automated data extraction and providing rational clinical judgments, thereby streamlining clinical workflows.

In fact, hospitals already possess automated systems for capturing patient information; however, basic information systems can only extract fundamental data such as the patient’s name, age, and gender. For more critical and diverse descriptive clinical information, manual entry by physicians is often required due to variations in documentation habits and differing levels of standardization among clinicians. This approach fails to address practical clinical challenges, offers little relief in reducing physicians’ workload, and does not resolve issues of misassessment or omission associated with manual completion of assessment forms.

Automatically parsing information from patient diagnosis and treatment processes is highly challenging. “For example, to enable a computer to determine whether a patient has experienced a cerebral infarction within the past month—a risk factor for venous thromboembolism (VTE)—the system must not only help the machine understand the temporal constraint of the event (i.e., it should have occurred recently, within the specified one-month window, excluding events that happened three months or six months ago), but also assess the patient’s disease status, such as its severity and classification. This often requires synthesizing information from medical history, MRI reports, and other sources to reach a conclusion, which traditional information systems are incapable of handling,” explained Zhang Qi, CEO of Beijing Huimei Cloud Technology Co., Ltd. (hereinafter referred to as Huimei Technology), in an interview with VCBeat.

The advent of deep learning has brought a turning point to big data processing. Emerging natural language processing (NLP) technologies enable semantic understanding of medical record data, structuring and standardizing descriptive natural language. Deep learning algorithms can further integrate personalized medical record data into rule bases, and construct VTE knowledge graphs by combining clinical guidelines and various assessment scales. This approach enhances the accuracy of final machine learning models while addressing two major challenges faced by physicians: “difficulty in data entry” and “difficulty in data analysis.”

However, NLP has its limitations. If developers are unfamiliar with the diagnostic workflow for venous thromboembolism (VTE), lack access to high-quality and comprehensive data on patients’ medical history and medication records, or fail to construct a well-designed knowledge graph, the accuracy of the generated results will be significantly compromised. Building an effective NLP system is, in fact, far from straightforward.

Applying NLP to VTE prevention and control is no easy task. Take Huimei Technology, which offers a VTE quality control product, as an example; the entire product development process was fraught with twists and hardships.

Driven by the combined forces of four years of foundational work in algorithm R&D and knowledge graph construction, the involvement of clinical experts, and policy support, Huimei Technology has rapidly advanced the development of its VTE quality control system. Zhang Qi told VCBeat, “The entire system was initiated in May 2019, and within just four months, Huimei Technology developed an intelligent VTE prevention and treatment system with an accuracy rate exceeding 97%. This success is attributable not only to Huimei Technology’s diligent efforts but also to the strong support from partner hospitals.”

Reflecting on the four-month development project, Zhang Qi couldn’t help but feel emotional: “When we launched the first-generation product in August, its accuracy was below 80%. At that time, we also faced numerous challenges.”

Integrating CDSS systems with clinical data is the first major hurdle every company encounters. Fortunately, Huimei Technology has become highly proficient in handling the steps for CDSS integration. The real challenge lies in standardizing medical record information. Since medical records largely consist of unstructured text descriptions, individual physicians have their own documentation habits, resulting in low levels of terminology standardization and consistency, as well as pronounced multi-source heterogeneity. Huimei Technology needs to “explain” these diverse pieces of information to machines in a universal format, enabling computers to accurately understand their semantics. Zhang Qi stated, “Even though our team has been deeply embedded in clinical practice for several years, incorporating hundreds of thousands of real medical records from tertiary hospitals to train AI models and achieving strong recognition performance, we still needed to iterate and validate the models against real patient records from our partner hospitals. This entire process took as long as three months before the recognition accuracy stabilized at over 97%.”

It is worth noting that Huimei Technology’s platform already possesses robust generalization capabilities. Even when physicians’ reports do not strictly adhere to standards, Huimei AI can accurately extract patient information and generate precise results.

In practice, even with the support of a VTE quality control system, physicians may inevitably overlook VTE risk assessments for some patients amid their busy workflows. To address this issue, the intelligent alerts feature of the Huimei VTE Quality Control System proves invaluable.

“Our current goal is to advance the timing of screening, assisting physicians in addressing relevant issues as soon as patients are admitted, thereby shifting quality control measures upstream. The first 24 hours after admission constitute a critical assessment window; if any oversights occur, our Clinical Decision Support System (CDSS) embedded within the electronic medical record will issue reminders to physicians within this 24-hour period, alerting them that the patient’s VTE risk assessment upon admission has not yet been completed.”

To ensure timely data statistics and feedback, Huimei Technology has also established a statistical management platform within the hospital to record VTE prevention and treatment measures for each patient. This allows patient information to be clearly presented to the attending physicians, making any missed screenings immediately apparent.

Taking China-Japan Friendship Hospital as an example, Huimei Technology’s post-event big data statistical analysis platform is displayed on the large screen at the nurses’ station, enabling nurses to instantly and intuitively access patients’ bed information as well as the status of VTE assessment and preventive measure implementation.

Furthermore, Huimei Technology is also striving to implement VTE prevention and control measures in primary care hospitals.

Unlike tertiary hospitals, primary healthcare institutions have a low level of internal informatization, and physicians often lack the ability to assess venous thromboembolism (VTE). In this context, Huimei Technology’s VTE quality control system primarily serves to standardize practices and provide physician education. Through this process, physicians’ capabilities in VTE prevention and treatment will gradually improve.

On October 13, 2018, the launch ceremony for the “National Project on Capacity Building for Prevention and Control of Pulmonary Embolism and Deep Vein Thrombosis,” officially approved by the Bureau of Medical Administration and Hospital Supervision of the National Health Commission, along with the press conference for the release of the “Chinese Guidelines for Prevention and Treatment of Thrombotic Diseases,” was held at China-Japan Friendship Hospital in Beijing, providing a strong impetus to the prevention and control of venous thromboembolism (VTE). This also served as a significant opportunity for companies such as Huimei Technology to develop VTE quality control systems.

Many hospitals have also responded to policy requirements by establishing in-hospital VTE prevention and control systems through measures such as forming prevention and treatment management teams, implementing supervisory mechanisms for VTE prevention and treatment management, advancing disciplinary development, and conducting health education.

However, advancing in-hospital informatization and raising physicians’ awareness of venous thromboembolism (VTE) prevention and control will both be protracted, arduous endeavors requiring concerted efforts from the government, enterprises, and hospitals.