Infectious diseases prevention and control using an integrated health big data system in China

The advent of big data platforms and advancements in machine learning algorithms have allowed researchers to distil insights from large amounts of information, which has the potential to create increasingly effective infectious diseases surveillance and control methods [1]. The term big data refers to complex datasets that are too voluminous and sophisticated to be handled by traditional analytics [2, 3] and public health agencies are increasingly reliant on big data to improve infectious disease screening. There is a wide variety of data structures that are relevant to public health concerns, including electronic health records (EHRs), clinical databases at the state and local level, social media, and reports from public health agencies such as the Centers for Disease Control and Prevention (CDC) [4]. The ease of storing, manipulating, and analyzing these wide ranging data types at scale has potential to empower health organization and public health officials to preempt the spread of infectious diseases and respond to and manage outbreaks in a timely manner [5].

While researchers have recognized the promises of big data in enhancing infectious disease monitoring and control, the efficacy of big data in preparing public health officials at both the state and national levels to respond to infectious diseases outbreak is largely contingent on the existence of successful partnerships between different stakeholders, such as government health agencies, departments in infectious disease control, and private and public hospitals and clinics. A concerted effort in storing, transferring, manipulating, and analyzing data efficiently for infectious disease control requires the support of computational and organizational infrastructures and the streamlining of processes between different stakeholders.

The overall objective of this study is to demonstrate how building an integrated, big data health infrastructure and platform at the district-level can significantly improve public health officials’ ability to leverage insights from big data for infectious disease prevention and control. To achieve this objective, we compared the effectiveness of a new tool with an integrated big data platform located in Yinzhou, a major district in the city of Ningbo in Zhejiang province in China [6], with a traditional infectious disease surveillance method—which relies on reporting from laboratories and healthcare facilities (e.g., doctors’ diagnosis after face-to-face consultation)—in case identification for infectious disease control and prevention. For this study, we focused on three types of infectious diseases categories: (a) acute infectious diseases (i.e., dengue fever), (b) chronic infectious diseases (i.e., pulmonary tuberculosis (TB)), and (c) gaps in vaccination uptake among migrant children.

The big data system in Yinzhou was established in 2016 as part of the district’s effort to accelerate and strengthen infectious disease control by pooling together a diverse set of medical and clinical data that health organizations can access for decision-making in a timely manner. To effectively screen for infectious diseases, we developed an overarching infectious disease screening framework with five key guidelines (see Fig. 2) that guide the detection of infectious diseases. These guidelines were derived from best practices and findings from existing research on the use of various forms of big data systems in complementing traditional infectious disease surveillance methods [14‐16]. For the purpose of this study, we programmed specific clinical criteria (see Table 1) after consultations with public health and medical experts involved in the big data system and screened for (a) dengue fever among the outpatients, (b) TB among university students, and (c) migrant children with insufficient vaccination coverage. To screen for TB among individuals, we extracted any key unstructured texts in medical imaging diagnosis abstracts for patients with diagnosis of acute respiratory infection as well as cold and pneumonia with the following keywords: shadow; pleural thickening; pleural effusion; lung infection; infective lesion; pulmonary nodules. If the patients were diagnosed with the above symptoms for more than 2 times across 14 days in a month, or more than 5 times in a month, they would be identified as a potential suspect patient. When a student is listed as a suspect case based on the above criteria, we also applied the same screening criteria on students’ social networks in their classes and dormitories to identify potential spread of TB.

For dengue fever, we screened for patients above 15 years old, with white blood cell count (WBC) < 4.5 * 10⁹/L, or reduced by 10% compared with the most recent medical record; Platelet count (PLT) < 125 * 10⁹/L; PLT reduced by 10% compared with the most recent medical record; or In the past 5 days with any following diagnoses: fever, infectious fever, viral infection, upper respiratory tract infection, acute pharyngitis, cold, and thrombocytopenia.

For screening of migrant children with incomplete immunization, we first matched the name list of children under 15 years old visiting medical institutions in Yinzhou with the name list of children who have been covered by local immunization program by matching the ID number of children, children’s name and birthday, children’s family name, township of residence, and parents’ names and children’s’ birthday. If the children’s records don’t match with any of the cases in the dataset of Yinzhou Immunization program, the children will be identified as a potential case with incomplete immunization.

We then compared the effectiveness of big data screening done with existing traditional methods of detecting TB, dengue, and the number of migrant children without adequate vaccinations.

Our study has indicated the potential of having a coordinated health big data system in drawing upon different health sources to identify individuals who contracted dengue fever and TB, thereby preempting their further spread, as well as engaging in preventive infectious disease control by identifying high-risk populations who require vaccinations. We conducted our study in Yinzhou district in Zhejiang province in China, which was suitable as it has successfully implemented a health big data system where the health data of individuals (e.g., daily clinic data produced by local health providers, EHRs, annual medical check-up data, immunization records, chronic and infectious diseases reporting data, mortality surveillance data, students’ registry data) are uploaded into a secured centralized EHR on a daily basis. This allows public health officials to have timely access to the latest data in tracking and mitigating the spread of infectious diseases.

In the big data system, all the screening criteria were set by clinical and public health workgroups and calibrated based on consultations with public health and medical professionals. Disease-like symptoms, such as acute upper and lower respiratory infection, cold and pneumonia, pleural thickening, pleural effusion, lung infection, infective lesion, and pulmonary nodules, were used to screen for TB in our big data system. As for dengue fever, we used symptomatic indicators such as fever, upper respiratory tract infection, acute pharyngitis, cold, erythra, and thrombocytopenia. Leveraging disease-like symptoms within the big data system for the screening of infectious diseases—a form of syndromic surveillance which often involves the usage of clinical data to scan for discernable symptoms to identify potential cases of infectious diseases before official diagnosis—was shown to be effective as early as the 1990s [22]. For instance, research has shown that reliance on pre-diagnostic signs and symptoms allows public health organizations to detect community wide influenza outbreaks [23]. This is because the systematic collection of continuous population health data (e.g., whenever people visit medical facilities) through the big data system ensures that the available data are the most up to date, thus bringing it closer to a “real-time” detection of outbreaks [24]. Moreover, they are arguably more accurate, as they do not rely on individuals’ self-reporting or participation through digital communication technologies (e.g., smartphones, web-search queries) [25], which suffer from problems such as non-representativeness, missing data, and mis- or under-reporting [10].

In addition to relying on individual-level disease symptoms, our screening included social network data of university students in an ethical and secure way. Only when a student was suspected or diagnosed with TB would the big data system scan the symptoms of their close contacts, which could be classmates or roommates in our records, for any medical anomalies. Social networks are crucial for understanding and tracking the spread of public health problems and diseases. This was most notably demonstrated by Christakis and Fowler, who modelled the spread of obesity in a large social network over three decades [26]—due to underlying tie-generative mechanisms in social networks. Some of these tie-generative mechanisms [27] are assortative mixing (i.e., people who are like-minded would be in closer proximity) and triadic closure—a situation where if node a is connected to node b and node b to node c, then node a is likely to be connected to node c, which explains the spread of diseases. The traditional method will initiate an investigation only when a student with TB was diagnosed, while the big data method can automatically screen for students with TB-like symptoms; the traditional method tracks the close contacts within a small scope, such as classmates or roommates, compared to the large scope captured by the big data method, such as students in the same grade; the traditional method investigates the TB-like symptoms of close contacts currently and in the past few weeks, and may be affected by recall bias, whereas the big data method analyses the 3 month period before and after detection.

Using our big data approach, we successfully identified six cases of patients who contracted dengue fever out of a pool of 3972 suspected cases in 2019, three cases of TB out of 288 suspected cases in a pool of 43,521 university students, and 240 children who were without immunization out of a total of 983,000 children. As such, using an integrated big data health system could significantly improve upon the efficiency, time-related costs, and misdiagnoses that come with reliance on manual contact-tracing alone.

There are several advantages to implementing an integrated big data system for infectious disease prediction and control. Through our three case studies, we have illustrated that reliance on big data is effective, timely, and may potentially result in significant cost savings (i.e., it is relatively expensive to send out public health officials to go door-to-door for screening). In addition to these advantages, because the data are located within a geographical boundary (e.g., Yinzhou), this approach allows public health officials working in that region to take practical action. As such, this is better than relying on other forms of “big data” epidemiology, such as the use of web-search queries and social media posts in predicting infectious diseases. There exist multiple confounders in the use of social media posts and web-search queries as proxies for the spread of infectious diseases. Even if researchers were able to identify and model the spread of diseases using user-generated content on social media platforms, public health researchers would not be able to take advantage of the information strategically and practically for their region.

This study has several limitations. First, we relied on aggregated data from Yinzhou CDC for this study and due to the data limitations, we were not able to conduct formal statistical testing to compare the efficacy of the integrated big data system and traditional infectious disease surveillance method beyond that of descriptive comparison. Second, for the three case studies, we did not rely on all possible data types, such as records from pharmacies, social media or web-search data. While social media and web-search data are meaningful in forecasting [28, 29], we have chosen to exclude them as the purpose of the study is to examine the effectiveness of the big data system in identifying potential cases, while social media and web-search queries are more effective in aggregated forecast. While our case studies were specific to dengue, TB, and immunization records, we are cognizant of the limitations of generalizability and scalability of this method to other infectious diseases, especially in the light of emerging pandemics such as COVID-19. Finally, we are mindful that individuals’ health data are highly sensitive and at-risk of malicious cyber-attacks—however we have put in place substantial cyber security measures to protect this sensitive data.

Despite these limitations, there are several areas where we are extending our research. First, the screening criteria used in the big data system were formulated based on consultations with public health experts and clinical practitioners. Moving forward, we plan to incorporate the implementation of artificial intelligence algorithms, such as deep learning, to refine the screening criteria to improve the detection rate. Second, even though our big data system was able to identify potential dengue cases, this information was not made available during face-to-face consultations. To improve the speed of detection, we aim to make the data available to doctors so that they can correctly diagnose patients. Third, we are expanding our big data system to include other data sources. In the medical area, we aim to include other forms of administrative health data such as insurance data and pharmacy transaction data [4]. We are also working on making our big data system interoperable with healthcare systems outside of Yinzhou. In that way, even if a person travels out of Yinzhou to receive treatment, his medical records would be available to the medical facility he visited to aid healthcare providers in making the right medical diagnosis. Besides health data, we are seeking to integrate non-medical data sources, such as individuals’ travel data as well as their mobile payments, to improve the efficacy of contact-tracing and epidemic surveillance. Finally, the screening criteria was developed in consultation with a panel of public health and medical experts so as to be sensitive to symptoms of infectious diseases in Yinzhou. Future research should incorporate both international and local contexts when developing screening criteria.

We have demonstrated through our three case studies that an integrated health big data approach is essential and effective for infectious disease control and management, and can potentially result in more lives saved and cost savings for the government and other decision makers. The ability of take advantage of health big data to improve infectious disease surveillance relies on a multi-party coordination and a concerted effort, as well as and the ability of government agencies, health organizations, hospitals and clinics, and private sectors to pull health data together and analyze them in a timely manner [10], provided that issues of data interoperability and privacy are adequately addressed [28].