Semantic Processing of Twitter Traffic for Epidemic Surveillance David S. Hale, Alla Keselman, Thomas C. Rindflesch Lister Hill National Center for Biomedical Communications Specialized Information Services

Pandemic Preparedness Early detection is critical to effective response “ The truth is out there, it is just not indexed well” (A bumper sticker; NLM parking lot) Disaster information traffic: delays, loss, overload Outbreaks data requires fast collection / dissemination Collection – from disease to syndromic surveillance Dissemination – from formal announcements to informal channels Government agencies are entering web 2.0 innovations E.g., CDC on Twitter

Early detection is critical to effective response

“ The truth is out there, it is just not indexed well”

(A bumper sticker; NLM parking lot)

Disaster information traffic: delays, loss, overload

Outbreaks data requires fast collection / dissemination

Collection – from disease to syndromic surveillance

Dissemination – from formal announcements to informal channels

Government agencies are entering web 2.0 innovations

E.g., CDC on Twitter

Monitoring the Internet for Syndromic Surveillance Current methods: keywords analysis News Aggregator and visualization tools (e.g., HealthMap) Web searches - queries Google Trends; Google Flu Trends Brownstein et al. – peak for “food poisoning” preceded peak for “salmonella”, “peanut butter”, “recall” Requires massive amounts of data Ambiguous as to searchers’ precise information needs

Current methods: keywords analysis

News

Aggregator and visualization tools (e.g., HealthMap)

Web searches - queries

Google Trends; Google Flu Trends

Brownstein et al. – peak for “food poisoning” preceded peak for “salmonella”, “peanut butter”, “recall”

Requires massive amounts of data

Ambiguous as to searchers’ precise information needs

Distribution of “swine flu” google query

The Future of Syndromic Surveillance Social media, chatter Blogs , “Microblogging” More real-time data Monitor sentiment as well as events NLP analysis Requires less data / lower computational intensity More informative “ swine flu” and “travel” VS. “how fast swine flu travels” and “is it safe to travel during a swine flu epidemic”

Social media, chatter

Blogs , “Microblogging”

More real-time data

Monitor sentiment as well as events

NLP analysis

Requires less data / lower computational intensity

More informative

“ swine flu” and “travel” VS. “how fast swine flu travels” and “is it safe to travel during a swine flu epidemic”

Twitter Micro-blogging service SMS gateway enables posting from mobile devices Users post without breaking context or setting JIT (just-in-time) blogging API promotes community development of user experience and interaction 4-5 million users (Nov 2008) 17 million visitors (April 2009)

Micro-blogging service

SMS gateway enables posting from mobile devices

Users post without breaking context or setting

JIT (just-in-time) blogging

API promotes community development of user experience and interaction

4-5 million users (Nov 2008)

17 million visitors (April 2009)

Tweet Characteristics Format: [username] [text] [date time client] Length: Text limited to 140 characters Char Set: Not limited to ISO 8859-1 Western (Latin) Grammaticality: Variable Hashtags (#): Denote topics Primarily utilized by experienced users

Format: [username] [text] [date time client]

Length: Text limited to 140 characters

Char Set: Not limited to ISO 8859-1 Western (Latin)

Grammaticality: Variable

Hashtags (#): Denote topics

Primarily utilized by experienced users

Tweet Content Some provide (purported) information Authority not determined Majority express opinions Often with humor or sarcasm Value for syndromic surveillance Source for assessing public sentiment Observation of information trending As a guide for government action

Some provide (purported) information

Authority not determined

Majority express opinions

Often with humor or sarcasm

Value for syndromic surveillance

Source for assessing public sentiment

Observation of information trending

As a guide for government action

Tweets: Examples CDC tips for preventing the flu: wash hands often and stay home when sick Oklahoma health officials say swine flu headed to state, public needs to take precautions Napolitano says “not a pandemic” yet I bet this whole swine flu scare really has Kermit the Frog rethinking his relationship What’s next? Three-toed sloth flu?

CDC tips for preventing the flu: wash hands often and stay home when sick

Oklahoma health officials say swine flu headed to state, public needs to take precautions

Napolitano says “not a pandemic” yet

I bet this whole swine flu scare really has Kermit the Frog rethinking his relationship

What’s next? Three-toed sloth flu?

NLP Analysis Unified Medical Language System (UMLS) Medical concepts in semantic types (or classes) MetaMap Identifies UMLS concepts in text SemRep Identifies semantic relations between concepts Rifampin for tuberculosis Rifampin [Pharmacologic Substance] TREATS Tuberculosis [Disease or Syndrome]

Unified Medical Language System (UMLS)

Medical concepts in semantic types (or classes)

MetaMap

Identifies UMLS concepts in text

SemRep

Identifies semantic relations between concepts

Rifampin for tuberculosis

Rifampin [Pharmacologic Substance]

TREATS

Tuberculosis [Disease or Syndrome]

Monitoring Twitter with NLP Processed 1300 Twitter posts Known to be about swine flue Sent during 1 hour on Monday, April 27, 2009 Preprocessed, to accommodate format Ran MetaMap and SemRep Extracted semantic concepts and relationships Defined a semantic schema for influenza epidemic

Processed 1300 Twitter posts

Known to be about swine flue

Sent during 1 hour on Monday, April 27, 2009

Preprocessed, to accommodate format

Ran MetaMap and SemRep

Extracted semantic concepts and relationships

Defined a semantic schema for influenza epidemic

Schema: UMLS Semantic Types Focus output In the area of interest And with the components in that area Schema for influenza epidemic Disease or Syndrome Sign or Symptom Geographic Area Mammal Health Care Organization Medical Device

Focus output

In the area of interest

And with the components in that area

Schema for influenza epidemic

Disease or Syndrome

Sign or Symptom

Geographic Area

Mammal

Health Care Organization

Medical Device

MetaMap and SemRep Output Tweet Texas confirms third case of swine flu Concepts extracted Texas [Geographic Area] Third [Quantitative Concept] Family suidae [Mammal] Influenza [Disease or Syndrome] Relationship Influenza PROCESS_OF Family suidae

Tweet

Texas confirms third case of swine flu

Concepts extracted

Texas [Geographic Area]

Third [Quantitative Concept]

Family suidae [Mammal]

Influenza [Disease or Syndrome]

Relationship

Influenza PROCESS_OF Family suidae

Results: Most Frequent Concepts 371 Family suidae [Mammal] 324 Influenza [Disease or Syndrome] 115 Not [Functional Concept] 113 Mexico [Geographic Area] 89 Centers for Disease Control and Prevention (U.S.) [Health Care Related Organization] 71 Case unit dose [Quantitative Concept] 54 Time [Temporal Concept] 53 Pandemics [Phenomenon or Process]

371 Family suidae [Mammal]

324 Influenza [Disease or Syndrome]

115 Not [Functional Concept]

113 Mexico [Geographic Area]

89 Centers for Disease Control and Prevention (U.S.) [Health Care Related Organization]

71 Case unit dose [Quantitative Concept]

54 Time [Temporal Concept]

53 Pandemics [Phenomenon or Process]

Results: Filtered through Schema Disease or Syndrome: Influenza Sign or Symptom: Coughing Geographic Area: Mexico Mammal: Family suidae Health Care Organization: Centers for Disease Control and Prevention (U.S.) Medical Device: Mask

Disease or Syndrome: Influenza

Sign or Symptom: Coughing

Geographic Area: Mexico

Mammal: Family suidae

Health Care Organization: Centers for Disease Control and Prevention (U.S.)

Medical Device: Mask

Results: PROCESS_OF Relation Influenza PROCESS_OF Family suidae Influenza PROCESS_OF Farmer, unspecified Influenza PROCESS_OF Hispanics Influenza PROCESS_OF Mexican Influenza in Birds PROCESS_OF Human Influenza-like symptoms PROCESS_OF Passenger Flu symptoms PROCESS_OF Family suidae Swine influenza PROCESS_OF Family suidae

Influenza PROCESS_OF Family suidae

Influenza PROCESS_OF Farmer, unspecified

Influenza PROCESS_OF Hispanics

Influenza PROCESS_OF Mexican

Influenza in Birds PROCESS_OF Human

Influenza-like symptoms PROCESS_OF Passenger

Flu symptoms PROCESS_OF Family suidae

Swine influenza PROCESS_OF Family suidae

Next Steps Twitter access Further testing for effectiveness Refine filters (frequency, semantic types) User control Implement proof-of-concept Preprocessing for tweet format NLP Final filtering Output format Graphs

Twitter access

Further testing for effectiveness

Refine filters (frequency, semantic types)

User control

Implement proof-of-concept

Preprocessing for tweet format

NLP

Final filtering

Output format

Graphs

Opportunities Biosurveillance Monitoring of wide-spread sentiment Targeted information provision Respond to misinformation trends Potential for evaluating authenticity Semantic comparison to trusted source

Biosurveillance

Monitoring of wide-spread sentiment

Targeted information provision

Respond to misinformation trends

Potential for evaluating authenticity

Semantic comparison to trusted source

Conclusion Exploiting the Internet for disaster preparedness Assessing public sentiment and events Leveraging social media, e.g. Twitter Using semantic NLP Useful to CDC and other government agencies Proof-of-concept experiment suggests the viability of this approach

Exploiting the Internet for disaster preparedness

Assessing public sentiment and events

Leveraging social media, e.g. Twitter

Using semantic NLP

Useful to CDC and other government agencies

Proof-of-concept experiment suggests the viability of this approach