Twitter Improves Seasonal Influenza Prediction
1 like772 views
The document discusses the use of Twitter data to improve predictions of seasonal influenza outbreaks. It presents an analysis indicating a strong correlation between flu-related tweets and CDC data on influenza-like illnesses, demonstrating that Twitter can serve as a timely and effective tool for epidemic tracking. The study concludes that incorporating social media data enhances public health preparedness and provides valuable insights into flu activity trends based on demographics and regions.
![Twitter improves Seasonal Influenza Prediction
Harshavardhan Achrekar
Avinash Gandhe
Ross Lazarus
[2]
[3]
Ssu-Hsin Yu
[2]
Benyuan Liu
[1]
HEALTHINF 2012
International Conference on Health Informatics
Vilamoura, Algarve, Portugal
[1] Computer Science Department, University of Massachusetts Lowell
[2] Scientific Systems Company Inc , Woburn, MA
[3] Department of Population Medicine - Harvard Medical School
1
Monday, January 30, 2012
[1]](https://image.slidesharecdn.com/healthinf2012-140226103007-phpapp01/85/Twitter-Improves-Seasonal-Influenza-Prediction-1-320.jpg)
![Twitter Data Set Analysis
In our Twitter dataset [2010-2011]
22 % users are from USA,
46 % users are outside USA
32 % users have not published their location details.
Status posting times (tweet timestamp in GMT) are converted to the local timezone
of the individual profile. Day light saving are applied within required time frame.
Number of Tweets (in %)
15%
10%
5%
0%
CA NY TX FL IL OH PA GA MI VA MA NC WA MO NJ AZ MD TN DC IN OR WI MN CO AR NV LA SC KY AL CT OK UT IA KS MS NE RI HI NM WV ME ID NH MT DE ND WY AK VT SD PR
State-wise Distribution of USA users on Twitter for flu postings
12
Monday, January 30, 2012](https://image.slidesharecdn.com/healthinf2012-140226103007-phpapp01/85/Twitter-Improves-Seasonal-Influenza-Prediction-12-320.jpg)
Twitter Improves Seasonal Influenza Prediction
- 1. Twitter improves Seasonal Influenza Prediction Harshavardhan Achrekar Avinash Gandhe Ross Lazarus [2] [3] Ssu-Hsin Yu [2] Benyuan Liu [1] HEALTHINF 2012 International Conference on Health Informatics Vilamoura, Algarve, Portugal [1] Computer Science Department, University of Massachusetts Lowell [2] Scientific Systems Company Inc , Woburn, MA [3] Department of Population Medicine - Harvard Medical School 1 Monday, January 30, 2012 [1]
- 2. Talk Outline Background Related Work Our Approach Twitter Dataset Description Twitter Dataset Analysis Extended Analysis (Text Mining, Age-wise, US Regions) Conclusions 2 Monday, January 30, 2012
- 3. Seasonal flu Influenza (flu) is contagious respiratory illness caused by influenza viruses. Seasonal - wave occurrence pattern. 5 to 20 % of population gets flu ≈ 200,000 people are hospitalized from flu related complications. 36,000 people die from flu every year in USA, worldwide death toll is 250,000 to 500,000. Epidemiologists want to use early detection of disease outbreak to reduce number of people affected. CDC collects Influenza-like Illnes(ILI) from its surveillance network and publishes weekly (typically 1-2 weeks delay) CDC stands for Center for Disease Control 3 Monday, January 30, 2012
- 4. Emerging Flu Epidemics Spanish Flu SARS, 2002-2003 H1N1 (swine flu), 2009-2010 ? 4 Monday, January 30, 2012
- 5. Related Work :- Google Flu Trends Over the counter drug sale, patients visit log for flu shots, tapping telephone advice lines. Certain Web Search terms are good Indicators of flu activity. Google Trend uses Aggregated search data on flu indicators. Estimate current flu activity around the world in real time. From example :- Google Flu Trend detects increased flu activity two weeks before CDC. Link:- www.google.com/flutrends 5 Monday, January 30, 2012
- 6. OSN - Novel Data Source for Detection and Prediction Online Social Networks (OSN) has emerged as popular platform for people to make connections, share information and interact. Facebook:~ 750 million users , Twitter :~ 200 million users Billions of pieces of information being posted and shared on the web every week. Applications: Real-world outcome of box-office revenues for movie Large scale fire emergencies and Earth-quake detection and reporting Online Service Downtime and disruptions of content providers People’s mood Live Traffic updates 6 Monday, January 30, 2012
- 7. Our Approach {“i am down with flu”, “got flu.”} msg exchange between users provide early, robust predictions. OSN represent a previously untapped data source for detecting onset of an epidemic and predicting its spread. Twitter/Facebook mobile users tweet/ posts updates with their geo-location updates. helps in carrying out refined analysis. User demographics like age, gender, location, affiliated networks.,etc can be inferred from data. snapshot of current epidemic condition and preview on what to expect next on daily or hourly bases. sought to develop model that estimates number of physician visits per week related to ILI as reported by CDC. 7 Monday, January 30, 2012 OSN stands for Online Social Network
- 8. System Architecture of SNEFT Data Collection Engine downloader ILI Data Internet crawler OSN Data ILI stands for Influenza-Like Illness 8 Monday, January 30, 2012 ARMA Model ILI Prediction Novelty Detector Flu Warning
- 9. OSN Data Collection Design of the Twitter data collection engine / Crawler 9 Monday, January 30, 2012
- 10. Twitter Data Set Real Time Response Stream fetches entries relevant to searched keyword having the tweets in reverse-time order. Data collection active from October 18, 2009 until present. 2009-2010: 4.7 million tweets from 1.5 million unique users 2010-2011: 4.5 million tweets from 1.9 million unique users 10 Monday, January 30, 2012
- 11. Spatio Temporal Database for Twitter Data Set Crawler uses Streaming Real time Search Application Programming Interface (API) to fetch data at regular time intervals. A tweet has the Twitter User Name, the Post with status id Time stamp attached with each post. From Twitter’s username we can get profile details attached to every user which include number of followers, number of friends, his/her profile creation date, location {public or private from the profile page or mobile client} with status updates count User’s current location is passed as an input to Google’s location based web services to get geo-location codes (i.e., latitude and longitude) along with the country, state, city with a certain accuracy scale. 11 Monday, January 30, 2012
- 12. Twitter Data Set Analysis In our Twitter dataset [2010-2011] 22 % users are from USA, 46 % users are outside USA 32 % users have not published their location details. Status posting times (tweet timestamp in GMT) are converted to the local timezone of the individual profile. Day light saving are applied within required time frame. Number of Tweets (in %) 15% 10% 5% 0% CA NY TX FL IL OH PA GA MI VA MA NC WA MO NJ AZ MD TN DC IN OR WI MN CO AR NV LA SC KY AL CT OK UT IA KS MS NE RI HI NM WV ME ID NH MT DE ND WY AK VT SD PR State-wise Distribution of USA users on Twitter for flu postings 12 Monday, January 30, 2012
- 13. Twitter Data Set Analysis 18% 7% 16% 6% Number of Tweets (in %) Number of Tweets (in %) 14% 5% 4% 3% 2% 12% 10% 8% 6% 4% 1% 2% 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 0 Hour of the Day SUNDAY MONDAY TUESDAY WEDNESDAY THURSDAY FRIDAY SATURDAY Hourly Twitter usage pattern in USA Average daily Twitter usage within a week The hourly activity patterns observed at different hours of the day are much to our expectations, with high traffic volumes being witnessed from late morning to early afternoon and less tweet posted from midnight to early morning, reflecting people’s work and rest hours within a day. Average daily usage pattern within a week suggests a trend on OSN sites with more people discussing about flu on weekdays than on weekends. 13 Monday, January 30, 2012
- 14. Twitter Data Set Cleaning “I got flu shot”, “got stomach flu”, “flu season”...do not indicate real flu events Need to classify the tweets into positive or negative categories. 25,000 tweets classified by Amazon Mechanical Turk as training set Use trained SVM to classify all other tweets Significant improvement of correlation between the Twitter data and CDC data. Classifier Class Precision Recall F-value J48 decision tree Yes 0.801 0.791 0.796 No 0.813 0.704 0.755 Yes 0.725 0.829 0.773 No 0.813 0.704 0.755 Yes 0.807 0.822 0.814 No 0.829 0.814 0.822 Naive Bayesian Support Vector Machine Text Classification 10 fold cross validation results 14 Monday, January 30, 2012
- 15. Twitter Data Set Cleaning Retweet: A retweet is a post originally made by one user that is forwarded by another user. Syndrome elapsed time: An individual patient may have multiple encounters associated with a single episode of illness . To avoid duplication the first encounter for each patient within any single syndrome group is reported to CDC, but subsequent encounters with the same syndrome are not reported as new episodes until more than six weeks has elapsed since the most recent encounter in the same syndrome. We call it syndrome elapsed time. Remove retweets and tweets from the same user within a certain syndrome elapsed time, since they do not indicate new ILI cases. Retweet Syndrome Elapse Time Correlation coefficient RMSE errors No 0 week 0.8907 0.3796 No 1 week 0.8895 0.3818 No 2 week 0.8886 0.3834 No 3 week 0.886 0.3878 No 4 week 0.8814 0.3955 Correlation between Twitter dataset and CDC along with Root Mean Square Errors (RMSE). 15 Monday, January 30, 2012
- 16. Twitter Data Set Analysis Weekly plot of percentage of weighted ILI visits, positively classified Twitter dataset and predicted ILI rate using CDC and Twitter Number of Twitter users posting per week versus percentage of weighted ILI visits by CDC Data show strong correlation (Pearson correlation coefficient 0.8907) between Twitter data set and ILI rates from CDC, providing a strong base for accurate prediction of ILI rate. 16 Monday, January 30, 2012
- 17. Prediction Model Auto Regressive model with external input (Twitter data) where t indexes weeks y(t) : the percentage of physician visits due to ILI in week t u(t) : the number of unique Twitter users with flu related tweets in week t e(t) is a sequence of independent random variables c is a constant term to account for offset. m: previous CDC data in weeks n: previous Twitter data in weeks 17 Monday, January 30, 2012
- 18. Cross Validation Results n=0 n=2 n=3 0.5355 m=0 n=1 0.4814 0.4813 m=1 0.6331 0.4107 0.4147 0.4314 m=2 0.5395 0.3957 0.3986 0.4256 Root Mean Squared Errors from 10-fold cross validation Weekly plot of percentage of weighted ILI visits, positively classified Twitter dataset and predicted ILI rate using CDC and Twitter addition of Twitter data improves the prediction with past CDC data alone. use of Twitter data alone to predict the ILI rate (m=0) results in poor predictions. best result when m=2 , n=1: previous 2 week’s CDC data, current Twitter data 18 Monday, January 30, 2012
- 19. Regional and Age-based Flu Prediction Analysis 19 Monday, January 30, 2012
- 20. Regional and Age-based Flu Prediction Analysis ILI seems to peak later in the northeast (Region 1 and 2) than in the rest of the country by at least week. The Twitter reports also follow this trend. In Region 9 (CA, NV, AZ...), Region 4 (FL,etc.) and the northeast, the ILI rates seem to drop off fairly slowly in the weeks immediately following the peaks. This is also reflected in the Twitter reports. Approximately 20-25 weeks after the peak ILI, the northern regions have lower levels relative to the peaks in the southern regions. This is true for Twitter reports. 20 Monday, January 30, 2012
- 21. Regional and Age-based Flu Prediction Analysis Twitter data is a good indicator of ILI rates and can be used to effectively improve the regional prediction of current ILI rates. 21 Monday, January 30, 2012
- 22. Regional and Age-based Flu Prediction Analysis Prediction performance (root relative squared error) using Twitter in different age groups for different geographical regions within the US. The results indicates that for most of the regions, Twitter data best fits the age-groups of 5-24 yrs and 25-49 yrs, which correlates well with the fact that this likely is the most active age groups using Twitter 22 Monday, January 30, 2012
- 23. Conclusions Investigated use of a previously untapped data source, namely, messages posted on Twitter to track and predict influenza epidemic situation in real world. Results show that the number of flu related tweets are highly correlated with ILI activity in CDC data ( Pearson correlation coefficient 0.8907). Build logistic auto-regression models to predict number of ILI cases in a population as percentage of visits to physicians in successive weeks. Verified that Twitter data substantially improves our model’s accuracy in predicting ILI cases nationwide as well as region-wise. Twitter best fits age group 5-24 and 25-49 years, as these are most active age group communities on Twitter. Opportunity to significantly enhance public health preparedness among the masses for influenza epidemic and other large scale pandemic. Thank You Presenter: Harshavardhan Achrekar Personal Website :- www.hdachrekar.com 23 Monday, January 30, 2012
- 24. CDC’s Region-wise ILI data(left) and Twitter data(right) CDC’s Regional ILI data Twitter data 40 41 42 43 44 45 46 47 48 49 50 51 52 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 Weeks 24 Monday, January 30, 2012