How big data safeguards food safety

Law enforcement officers from the Market Supervision Administration of Shushan District, Hefei, carry newly configured rapid food safety testing instruments to conduct spot checks on various foods in supermarkets and shopping malls within their jurisdiction, with results produced on site. Guangming Photo/Visual China

Recently, the national on-site meeting on the creation of national food safety demonstration cities and the creation of agricultural product quality and safety counties was held in Chengdu, Sichuan. In the Internet era, how can we eat more safely? Experts at the meeting offered suggestions, proposing that attention be paid to using modern information technology to innovate regulatory methods and improve regulatory effectiveness.

As one of the most popular information technologies in recent years, how can big data technology improve the effectiveness of food safety regulation? In Chengdu, which is in the process of creating a food safety demonstration city, we have seen one possibility.

Breaking down information silos

Unlike many food safety regulators, Wang Lishan does not need to go to farmers' markets, supermarkets or restaurants. Most of the time he sits at his desk, and the objects of his "inspection" are piles of data. "Data analysis can improve the efficiency of food safety regulation," he said.

Wang Lishan is the director of the Chengdu Food Safety Risk Monitoring Data Center, which is building the core technical support system for "Smart Food Safety." "Smart regulation, data first" is their slogan.

Walking into the data center, the most eye-catching thing is a black electronic screen, on which the data is constantly refreshing. At this moment, the data on the screen shows a total of 674,136 testing batches, a total of 8,469,618 test items, a total of 15,619 non-compliant samples, with 4 non-compliant samples pending processing and 3 being processed...

This data is not all of it. The data that Wang Lishan and his colleagues face comes not only from various food testing institutions, but also from local agriculture, quality inspection, health, and industry and commerce departments. The data includes food testing information, food safety public opinion information, enterprise industrial and commercial registration and credit data, foodborne disease case data, and environmental data, among others.

"Chengdu has nearly 200,000 food production and business entities. From farm to table, every link is scattered with massive amounts of data, and fragmentation is severe," Wang Lishan said. What they need to do is integrate this data, establish a basic database, break down information silos, and achieve data interconnection and resource sharing.

At present, the database has accumulated 70,000 pieces of government department shared data and 8.6 million pieces of inspection and testing data, connected 300,000 pieces of basic producer data and 70 million pieces of global commodity data, simultaneously monitored public opinion data from multiple online media outlets, and maintained an average monthly growth rate of 200,000 pieces.

"But big data does not equal a large volume of data; rather, it must generate value where 1 plus 1 is far greater than 2," Wang Lishan said.

They achieved this. From January to August this year, through big data mining and analysis, the problem detection rate of Chengdu's food safety supervisory spot checks increased from 2.36% to 10.71%, and regulatory effectiveness improved substantially.

So how exactly did they do it?

Achieving targeted strikes

At four in the morning, the Chengdu Agricultural Products Central Wholesale Market is already bustling. Trucks loaded with all kinds of vegetables drive into the market and split into two lines. One line of trucks is equipped with magnetic cards, while the other is not. The magnetic cards contain the vegetable vendors' incoming goods filing and registration information and safety certificates. For trucks without magnetic cards, the goods are sent to the rapid screening testing laboratory, and only after passing the test can they enter the market. The testing data is automatically uploaded to the Chengdu Food Safety Risk Monitoring Data Center dozens of kilometers away.

Here, rapid testing can produce results in 24 minutes, with a daily rapid testing volume of more than 600 batches. After the testing information is uploaded to the data center, its effect is quickly reflected in the next link.

For those trucks equipped with magnetic cards, although they have safety certificates, according to regulations a certain proportion still needs to undergo spot checks. This creates a difficult problem: specifically, which vegetables should be selected for spot checks to be most likely to reveal problems?

Wang Lishan said that through data analysis they found that non-compliant foods are highly unevenly distributed geographically and seasonally, and that foods prone to safety problems differ under different seasons and different climate conditions. Moreover, enterprises with a single shareholder structure, no corporate shareholders, and relatively low paid-in capital are more likely to have non-compliant foods detected. This inspired them to integrate various types of inspection data, enterprise registration information and credit data, among others, to establish a machine learning model. This work was completed jointly with a team from the University of Electronic Science and Technology of China.

"Through a certain algorithm, the machine learning model can quantitatively assess the possibility that an item to be tested will fail inspection, thereby achieving targeted spot checks aimed at high-risk foods," said Zhou Tao, chief expert of the data center and a professor at the University of Electronic Science and Technology of China. The wholesale market's information platform is connected with the data center, and the high-risk food information derived from data analysis is directly transmitted to the wholesale market.

"We receive high-risk product data from them every day, and the data will make our safety regulation more effective," Zheng Keke, deputy general manager of the wholesale market, told reporters.

The data center also integrated all case data on foodborne gastrointestinal diseases from the health department. Wang Lishan said that by monitoring changes in the number of cases infected by different strains with the seasons and weather, and by analyzing 8.6 million pieces of testing data, they obtained the correlation between foods and detected strains, thereby establishing an automatic decision-making channel from relevant cases to targeted spot checks.

This kind of targeted spot checking based on big data analysis ultimately enabled them to achieve the regulatory efficiency described above, allowing more precise strikes against non-compliant foods.

Terminating rumors in advance

Discovering food safety problems more efficiently is only one of the "magic powers" of big data. What once puzzled Zhou Tao, Wang Lishan and others was whether the various food safety rumors circulating online every day and the panic they bring to the public could also be monitored in a timely manner, or even terminated in advance.

They monitored more than 3 billion pieces of public opinion data accumulated from over a hundred well-known websites, forums, and tieba, among others, of which nearly 8 million pieces were related to food safety. "After obtaining public opinion information, through back-end algorithms, we extract key semantics and automatically cluster public opinion events, and we can judge the trend of public opinion dissemination," Zhou Tao said.

Next is a more critical step: dividing all the mined public opinion data into positive sentiment and negative sentiment, and analyzing the intensity of positive and negative sentiment. "Through this analysis, we can assess the risk level of a certain food safety public opinion event and whether the regulatory authorities should actively intervene, breaking rumors and eliminating safety hazards through timely testing or strict law enforcement," Zhou Tao introduced.

Not long ago, they monitored at the earliest opportunity the online rumor in the Chengdu area that "cabbage is dipped in formaldehyde to keep it fresh." At the time, this public opinion event had not yet spread widely, but after analysis they believed the risk level was very high and promptly notified the regulatory authorities to conduct large-scale spot checks on cabbage in the market. "In the end, no formaldehyde cabbage was found, and the authoritative testing report was also made public to a notary institution, terminating the rumor in advance," Wang Lishan said. They also promptly discovered 30 food safety hazards through data analysis, including aluminum in flour products and malachite green in aquatic products, and promptly carried out targeted rectification to eliminate safety risks.

In fact, many regulatory authorities are no strangers to this "moving the threshold forward" approach of grasping risks from the earliest signs. "But the methods are mostly rather backward, the amount of data is small, and they rely on people to specially follow certain websites and report regularly. Now the data and information are growing larger and larger, and relying purely on people is not very realistic," Zhou Tao said.

Data is an important resource

What excites Wang Lishan even more is that data analysis can also actively discover possible defects in current standards and optimize inspection and testing standards.

Their database contains a large amount of testing information on water-injected meat. After analysis, they found that the moisture content testing data for pork clustered around 63% and between 73% and 77%, while the national standard is 77%. "Many merchants inject water into pork to just around the standard line and pass the test. But in fact, pork with a moisture content of 76% is already seriously water-injected meat," Wang Lishan said.