Explore chapters and articles related to this topic
Machine learning and public health
Published in Sridhar Venkatapuram, Alex Broadbent, The Routledge Handbook of Philosophy of Public Health, 2023
Many of the high-profile applications of ML in healthcare can be subsumed under the categories of image-based diagnostic systems and risk-prediction models (see Topol 2019 for a review). In the former case, there are by now dozens of studies reporting on ML algorithms achieving “expert-level” accuracy when detecting diseases by examining medical images. The envisioned role of these algorithms is usually to support clinicians by providing a secondary opinion (Grote and Berens, forthcoming). In the latter case, algorithms are used to monitor patients for the purpose of facilitating timely intervention. While all these developments hold great promise for the application of ML algorithms within clinical settings, there is no such counterpart yet for public health. Indeed, precursors to ML applications in public health, such as Google Flu Trends, have proved somewhat inaccurate—and were abandoned. The basic idea of Google Flu Trends was to predict the influenza activity for different countries by aggregating search queries, which then were compared to a historical baseline of influenza activity in the given region. Inaccuracy resulted from many flu-related searches by people who were not ill but prompted to search by some other “exposure” to the population in question, such as a news story (Butler 2013).
Measuring population health and disease
Published in Kevin McCracken, David R. Phillips, Global Health, 2017
Kevin McCracken, David R. Phillips
Dedicated web query-based real-time surveillance systems are also valuable. One of the earliest developed was Google Flu Trends in 2008, using influenza-related search terms to estimate flu activity ('nowcasting'). Over time the system grew to cover twenty-nine countries. Later the company added another system, Google Dengue Trends. Unfortunately the open access publishing of current estimates was terminated in August 2015, though flu and dengue fever signal data based on search history is still being produced and provided directly to Google partner institutions who specialize in infectious disease research. Google Flu Trends generally produced results compatible with more traditional CDC influenza surveillance data, and several days faster. The need to see new digital surveillance tools as complementing traditional laboratory-based diagnostic surveillance rather than as a replacement was emphasized by several atypical US flu seasons in which the flu tracking algorithm delivered inaccurate results. Web-query approaches have also been implemented using other internet search engines (e.g. Yahoo, Bing, Baidu).
The 1918 Influenza A Pandemic
Published in Patricia G. Melloy, Viruses and Society, 2023
Seasonal influenza still occurs every year, primarily in the winter months in temperate zones of the world. The Centers for Disease Control and Prevention (CDC) have an influenza surveillance system that can help the agency predict the trajectory of each flu season. The CDC gathers local and national data weekly, working with clinical laboratories, public health officials, and healthcare providers in hospitals and outpatient facilities. The CDC not only looks at the number of cases and deaths, but also characterizes the virus itself, keeping an eye out for new versions of the influenza A virus (CDC 2021d). Companies like Google have also gotten involved in influenza surveillance. Google created a website called “Google Flu Trends,” but retired the website in 2015 when it was found that it could not reliably be used to predict flu cases (Brown 2018). The overprediction of flu cases by Google Flu Trends may have been due to “big data hubris” and changes in the Google algorithm that occurred over time during the flu season, adjustments that may improve Google searches but also affect the strength of the forecasting (Lazer et al. 2014). To try to get ahead of the game, the CDC tries to predict the patterns of season influenza, including how long the season will last and how bad it will be, in a program known as FluSight. The program directors also invite academic and private industry researchers to participate in their annual challenge to forecast the flu season, allowing the CDC to develop external research partners for their efforts (CDC 2021a). Effective modeling approaches tend to consider many different kinds of data, including social media posts, web search data, and information such as weather and humidity, analyzed using either a “statistical” or “mechanistic” model (Ali and Cowling 2021). We also must not forget about surveilling wild birds to figure out what subtypes of influenza A are circulating among the birds, in addition to conducting whole genome analysis. Scientists have called for greater coordination and standardization of practices among countries for monitoring of birds long term (Machalaba et al. 2015). Aside from U.S. flu surveillance efforts, over 100 research centers around the world communicate with the WHO, so that the WHO can make recommendations on the composition of the seasonal flu vaccine (Krammer et al. 2018).
Forecasting tuberculosis using diabetes-related google trends data
Published in Pathogens and Global Health, 2020
Leonie Frauenfeld, Dominik Nann, Zita Sulyok, You-Shan Feng, Mihály Sulyok
The first attempt – Google Flu Trends (GFT) – used exclusively online data related to several highly correlated search terms. Despite the initial high expectations, GFT’s failure to accurately predict the 2013 epidemic led to the cancellation of that project [15] and a similar project, Google Dengue Trends. A possible reason for the failure was the belief that ‘Big Data’ can fully replace traditional surveillance. Using several highly correlated terms likely substantially overfitted the model [15].
Use and interest of electronic nicotine delivery systems (ENDS): Assessing the validity of Google Trends
Published in The American Journal of Drug and Alcohol Abuse, 2021
Abhishek Ghosh, Simranjit Kaur, Fazle Roub
Our study results should be interpreted with the following cautions- (a) the results need replications from other countries; the internet penetration in India is nearly 50% and more than 90% of the internet users use Google as search engine; both of these figures for the US are nearly 90%; however, Google analytics may not be suitable in countries where internet penetration is low and the use of other search engines are more prevalent, (b) Google Trends data could never replace the routine community-based tobacco surveillance, (c) Google Trends can only act as indicators of change and demographic profile of the persons carrying out searchers and patterns of ENDS use cannot be discerned. This has implications for our research. As an example when we correlated RSV with the NATS data, it might not be very accurate as the former included searches made by any age groups but the later provided data only for adults, (d) Google search returns what it believes the ‘searcher’ is most interested in but what the person finally does with the list of topics displayed by Google cannot be determined by the Google Trends analysis; nonetheless, the fact that people are searching about the topics, suggests a general level of interest, (e) people can search information while they are away from their states or even country; hence, the geo-spatial Google Trends may not always reflect the behavior and interest of a particular state; this becomes an issue when we correlate state-wise RSV and prevalence of ENDS use and finally (e) Google analytics is just a “proxy” of the actual population-level behavior and do not represent actual behavior (e.g. actual use of ENDS). This limitation has also been observed in Google Flu Trends that has otherwise continued to perform remarkably well in estimating the occurrence of influenza-like illness, gave an erroneously high national flu-peak in 2013–14 (35). Moreover, previous authors showed that the performance of different models of Google Flu Trends analysis may differ according to the changes in health seeking behavior and the onset of the condition of interest (36). These observations are also relevant for Google Trends-based ENDS monitoring; future researchers may find more accurate (than ours) models for Google Trends analytics for estimating ENDS use.