Son Doan, National Institute of Informatics, Tokyo, Japan Mike Conway*, Univ. of California, San Diego, San Diego, United States Nigel Collier, National Institute of Informatics, Tokyo, Japan Track: Research Presentation Topic: Web 2.0 approaches for behaviour change, public health and biosurveillance Presentation Type: Rapid-Fire Presentation Submission Type: Single Presentation Building: Joseph B. Martin Conference Center at Harvard Medical School Room: C-Rotunda Room Date: 2012-09-15 04:00 PM – 04:45 PM Last modified: 2012-09-12

If you are the presenter of this abstract (or if you cite this abstract in a talk or on a poster), please show the QR code in your slide or poster (QR code contains this URL).

Background
In psychology, positive affect (PA) is defined as feelings that reflect a level of pleasurable engagement with the environment, such as happiness, joy, and contentment while negative affect (NA) is defined as negative emotion such as anxiety, depression, and hostility. The relationship between PA/NA and physical health has been extensively studied in health psychology. It has been shown that in laboratory data NA is an indicator for increasing illness and mobility while PA is often associated with decreasing illness.

Social media have proven to be a useful resource to track the Influenza-Like Illness (ILI) rate. Previous studies showed that Twitter can be used to track the ILI rate with a correlation coefficient over 0.90 in comparison to laboratory data (Culotta (2009), Lampos and Cristianini (2010)). In this study, we investigate effectiveness of PA and NA in calculating the ILI rate using Twitter data. To the best of our knowledge, this is the first study on using PA and NA in calculating the ILI rate.

Methods
We created a word list of PA and NA from psychology resources such as text books, literature and the Internet. We ended up with the list of 966 words in which the PA list consists of 509 words, e.g.,“happy”, “keen”, “joyful”, and the NA list consists of 457 words, e.g., “anger”, “angry”, “nervous”.

We investigated the effectiveness of PA and NA by applying their word lists after ILI-related keyword filtering methods. Three filtering methods were used: 1) Culotta (2009) with “flu”, “cough”, “headache” and “sore throat”; 2) Signorini et al. (2011) with “h1n1”, “swine”, “flu”, “influenza”; 3) Chew and Eysenbach (2011) with “h1n1”, “swineflu” and “swine flu”. Then we calculated the Pearson correlation coefficients to the laboratory data from the US Outpatient Influenza-Like Illness Surveillance Network (ILINet) in the US. We compared the correlation scores before and after applying PA and NA word lists.

Twitter data was collected within 36 weeks for the US 2009 influenza season from 30th August 2009 to 8th May 2010. The data contains over 587 million tweets from 23 million users. The correlation coefficients were calculated using R package.

Results
Before PA and NA lists were applied, Cullota (2009) achieved the best result with 0.9485 Pearson correlation coefficient, followed by Signorini et al. (2011) with 0.9470 and Chew and Eysenbach (2011) with 0.9448, respectively. When applying NA and PA lists, the results are as follows: 0.9548 (NA) and 0.9483 (PA) for Culotta (2009); 0.9586 (NA) and 0.9532 (PA) for Signorini et al. (2011); and 0.9467 (NA) and 0.9444 (PA) for Chew and Eysenbach (2011). All correlation coefficients were reported with all p-value < 2.2e-16.

Conclusions
The results show that NA helped to increase the correlation with the ILI rate slightly for all three methods with the best ILI rate of 0.9586 (an improvement of +1.16%) while PA did not show much effectiveness. This suggests that in practice NA can be integrated with keyword-filtering methods to calculate the ILI rate.