Personalized Event-Based Surveillance and Alerting Support for the Assessment of Risk

P ersonalized Ev en t-Based Surv eillance and Alerting Suppo rt for the Assess men t of Risk ∗ Av ar ´ e Stew ar 1 , Ricardo Lage 2 , Ernesto Diaz-Aviles 1 , P eter D olog 2 1 L3S R ese ar ch Center / LUH . Hannover, Germany – { stewart, diaz } @L3S.de 2 A alb or g U niversity. A alb or g, Denmark – { dolo g, ric ar dol } @cs.aau.dk Abstract In a t ypica l Ev ent-Based Surv eilla nce setting, a stre am of w eb do cuments is contin uously monitored for disea se repor ting . A structured representation of the disease re p o r ting even ts is extrac ted fro m the raw text, and the ev ents are then aggreg ated to pr o duce signals, which ar e intended to r epresent early warnings aga inst p otential public health threats. T o public hea lth officials, these warnings r e pr esent an ov erwhelming list o f “one-size- fits-all” information for r isk assess men t. T o reduce this o verload, tw o techn iq ues are pr op osed. First, filtering signals ac c ording to the user’s prefer- ences (e.g., lo cation, dis ease, s ymptoms, etc.) helps reduce the undesired noise. Second, re-r anking the filtered sig nals, accor ding to an individual’s feedback and annotation, allows a user-sp ecific, prio ritized ranking of the most relev ant warnings. W e int r o duce an approach that tak es into ac count this tw o-step pro cess of: 1) filtering and 2) re-ranking the res ults o f repo r ting sig nals. F or this, Collab ora tive Filtering and Personalization are common tec hniques us ed to supp ort users in dealing with the la rge amount of infor mation that they face. W e demonstrate the use of a mult i- int er est profile, which c o mpactly allows users to define numerous criteria for filtering . Profiles can b e decomp osed b y the system, so multiple interests can be automa tica lly genera ted from the comp osite profile and b e used to filter out undesired infor ma tion. A key re sult is tackling the problem of eq ua lly ra nked s ig nals, by exploiting the information within the underlying do cument and metadata provided b y the user, such as annotations and fav or ite items. This metadata a r e exploited to learn the user’s interests. Moreov e r , by comb ining m ultiple sources (e.g., annotations, fav or ite items, external W eb 2.0 a nd mult imedia data) a more ∗ In ternational Meeting on Emerging Diseases and Surveillance. IMED 2011 – PO STER SESSION – V ienna, Austria. F ebruary 4-7, 2011. 1 comprehensive computational view of the user can b e built to help r e-rank results accor ding to sy stem-learned user prefere nc e s . W e intro duce a metho d to allow users to explicitly build a profile to facilitate access to pers onalized health even ts, alerts and apply the profile for p e r sonalized ranking for different use rs or groups to supp or t filtering o f signa ls and ass o ciated do cument s . Keyw ords: bio s urveillance, epidemic intelligence, pe r sonalizatio n Ac kno wledgmen ts This w o rk is suppo rted by the Europ ean Comm unity’s Seven th F ramework Pr o- gram (FP7/2007 -2013 ) Me dic a l E cosystem: Personalized E ven t-Based Sur veil- lance grant num b er ICT 2 4 7829 . http://www.meco-pro ject.eu/. 2