• Multiscale Health Networks
• Recognized leader in data-drive patient care
• Partnership between CODONiS and Providence Health & Services
• Clark & Parsia
• Experts in Semantic Technologies
• Developers of Stardog the leading semantic graph database

• Working with Multiscale Health Networks
• Assist health care professionals to manage critical care situations and increase positive outcomes for patients
• Build a generic, automated, algorithmic, evidence-based system for monitoring patients
• First use case for this approach is Sepsis

• One of the oldest and most elusive medical conditions
• Introduced by Hippocrates in the 4th century BC as organic decay
• In 1992, defined as systemic inflammatory response to severe infection
• Can continue even after infection is gone
• Includes fatal complications such as organ dysfunction

• 10th leading cause of death in the US
• #2 in (non-coronary) intensive care units
• Mortality rate up to 30%, even in industrialized countries
• Up to 70% for patient with septic shock
• Rate significantly higher when untreated for more than seven days
• Occurs in 1-2% of hospitalizations, and as much as 25% of ICU bed utilization
• Most expensive condition in hospital stays in 2011
• Over $20 billion in direct costs for 1.1M hospitalizations
• Indirect costs topped $90B
• Cost has quadrupled since 1997; average 11.5% yearly increase
• Most costly condition billed to medicare

• Symptoms of Sepsis are quite varied
• 22 diagnostic criteria affecting all parts of the body
• Difficult to diagnose, can look like other conditions
• Prompt diagnosis is key
• Antibiotics should be administered within an hour of diagnosis
• Each hour delay increase mortality rate by 6%
• Bottom line: a deadly, difficult to diagnose disease, which escalates in severity very quickly
• How do you help get patients the care they need, when they need it?

• Research suggests automated monitoring can improve early diagnostic and therapeutic intervention
• Goal is to quickly, and automatically, identify at-risk patients
• Not an attempt to replace human diagnosis
• Provide a safety net for clinical judgement
• Patient Scores
• Aggregated, summarizing ~24 clinical measures
• Drill down and see individual component measures
• Includes information to help workflow and coordination

• Based on thousands of real-world cases of Sepsis
• Mining of several hundred million clinical data points
• Encode our domain knowledge of Sepsis
• Diagnostic concepts such as fever or hypothermia
• Pathology such as proinflammatory response
• Treatment programs for infection control, general care
• Combined with real-world predictive models

• Sets of objects (nodes) connected to other objects by links (edges)
• Useful, intuitive way to think about and model many problems and domains

• Semantics are a natural complement
• Provide formal, declarative definitions of nodes, edges and their relationships
• Using a high-level language
• Create graphs with meaning
• Specifically, computer understandable meaning
• Which is encoded in the graph

• They're based on standards
• Semantics, query language, format, protocol are all W3C standards; just like HTML
• Lets you consider them as more than just a graph
• So we can use the appropriate abstractions
• Natural fit for information integration and other data problems
  

• It's a fact of life, non-programmers exist
• And they can make valuable contributions to a project/codebase
• A common example is the business logic
• Except they can't write code
• But are often our domain experts

• Encode our business logic using a formal semantics
•  Via Semantic Graphs
• So we encode it in the graph
• Done with a high-level language
• So no programming required
• Frees it from the codebase; frees it from programmers

• Perform complex information processing tasks without writing code
• More directly capture expertise
• By letting the experts author the business logic
  
• Also yields a more maintainable codebase

• Making implicit information explicit
• Implicit in the data, the schema, or both
• Represent domain knowledge in a formal, declarative model
• Called an ontology
• Like UML, but with formal semantics
• Based on the W3C specification OWL, Web Ontology Language
• Reasoners consume ontologies to derive new information
• Answer queries, find inconsistencies
• Dynamic, executed on the fly

• A reasoner can tell you why it inferred some piece of information
• Called an explanation
• Gives insight into how the system uses the data and rules
• Provides traceability
• Explanations are simply proofs for why something happened
• Pinpoints the exact data/rules responsible
• Can reference other proofs

• Accuracy of data limits its utility
• Custom code to validate can be complex and time consuming
• Easier to model integrity constraints logically
• No reason to write more code
• Use Semantic Graphs
• Can be authored by experts in a high-level language
• Constraints can be enforced by middleware, or as part of the system

• Things go wrong
• People fat-finger things when doing data entry
• Or simply do not know the rules
• Effectively communicating information about violations is helpful
• So, like with reasoning, we can get explanations
• The computer can tell us why a violation occurred
• Any offending data, or lack thereof, is pinpointed by the reasoner
• Can offer suggestions to fix

• We can express disease conditions
• If a patient's temperature is over 100F, then they have a High Fever
• If a patient's platelet count is less than 100,000/mm^3, then they have Thrombocytopenia
• Combined with domain knowledge
• High Fever is a general indicator of Sepsis
• Thrombocytopenia is an Organ Dysfunction
  
• We can start to infer useful diagnostics
• If a patient has an Organ Dysfunction and an infection, then they have Severe Sepsis
• If a patient has Sepsis and has either Hypotension or Hyperlactatemia, then they have Septic Shock

• System claims that a patient is at-risk for Septic Shock
• This is useful for alerting critical care staff
• But we want to know why before committing to treatment

• Reasoner tells us patient is at-risk for Septic Shock because:
• Patient diagnosed with Sepsis
• Because of a known infection
• Due to the results of the test X, performed by Y, at time Z
• They also have an organ dysfunction
• Organ dysfunction is a leading indicator of Septic Shock
• They were also diagnosed with hypotension
• When combined, these can indicate a high-risk of Septic Shock

• If basing patient care on data, need to be sure it's accurate and complete
• Use Integrity Constraints to safeguard existing data and validate incoming data
• Examples:
• Body temperature must be a non-negative number
• Patients must have at least one doctor
• Test results must have a lab result, a tech/doctor it was performed by, and the time it was performed

• We're encoding our knowledge about Sepsis
• With Semantic Graphs
• Computer can understand these definitions and assist with data analysis
• Using new, inferred knowledge about patient status
• Which is provably correct

• Leveraging the abstractions provided by Semantic Graphs
• Ask questions directly, "Who has Sepsis", instead of writing a program
• Domain specifics are handle by the reasoner
• No code required
• Programmers can focus on UX
• Doctors can define the business logic, i.e., Sepsis diagnosis and treatment
  
• Easier to apply this approach to new conditions
• Building a generic platform, not a tool specific to Sepsis

• Sepsis is a deadly, difficult to diagnose disease
• Mortality rate increases quickly when left untreated
• Common condition costing hospitals billions of dollars a year
• Getting at-risk patients to care as soon as possible
• This is the key to effective treatment
• Utilize semantic technologies to automate the identification of at-risk patients
• Reduce the costs of treating Sepsis
• Treat it before it starts
• Stop it before it's too late
• Prevent re-occurrence
• This approach matches or exceeds the best results presented in existing literature

if { 
    ?patient :temperature ?temp . 
    filter (?temp > 100) 
} 
then { 
    ?patient :has :HighFever 
} 

if { 
    ?patient :plateletCount ?count . 
    filter(?count < 100000) 
} 
then { 
    ?patient :has :Thrombocytopenia 
} 

:Thrombocytopenia a :OrganDysfunction
:HighFever a :GeneralIndicator 

if { 
    ?patient :has [ a :OrganDysfunction ] . 
    ?patient :has :Infection 
} 
then { 
    ?patient :has :SevereSepsis 
} 

if { 
    ?patient :hasCondition :Sepsis . 
    { ?patient :has :Hypotension } 
    union 
    { ?patient :has :Hyperlactamemia } 
} 
then { 
    ?patient :has :SepticShock 
}