1 / 33

Analytics for Situational Awareness in Healthcare

Analytics for Situational Awareness in Healthcare. Anupam Joshi CSEE Department, UMBC http://ebiquity.umbc.edu/. October 2010. Situational Awareness. Applies both to people and to systems.

prema
Download Presentation

Analytics for Situational Awareness in Healthcare

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Analytics for Situational Awareness in Healthcare Anupam Joshi CSEE Department, UMBC http://ebiquity.umbc.edu/ October 2010

  2. Situational Awareness Applies both to people and to systems • SA is a field of study concerned with perception of the environment critical to decision-makers in complex, dynamic areas. Wikipedia Being aware of what is happening around you to understand how infor-mation, events, and your own actions will impact your goals and objectives, both now and in the near future

  3. Situational Awareness (SA) • Common theme in many scenarios as we become increasingly instrumented and interconnected Hot conflicts, homeland security, cyber-security, cyber-physical systems, disaster relief, health-care, IT services, network operations & management … • Moving from “react and respond” to “predict and effect” • Spans CSEE and IS – sensors, wireless networks, embedded systems, streaming data, analytics, image processing, AI, human-computer interfaces, cloud computing, … • Highly distributed, dynamic and interconnected systems

  4. Some Key Research Challenges • Analysis and integration of unprecedented volumes of rich media and text to infer context • Representation of complex, uncertain, and evolving information, e.g., human networks • Fusion into a common operational picture for SA • Mining of patterns and trends to predict events • Planning under uncertainty • Intuitive presentation to the responder in the field and in the command center to enable decision making

  5. Some recent UMBC research • DHS: Unified Incident Command & Decision Support (recommended for funding, awaiting contract) • NIST: Image Analysis for Bio and Clinical Informatics (current) • NIST: Personalized Medicine (current) • NSF: Situational Awareness and equipment supplement for I/UCRC (current) • ONR: Relief Social Media (sub. from LMCO, current) • NSF: Platys: from position to context (current) • DARPA: PbWAN: policy based network control (STTR with Shared Spectrum Co., 2006-10) • UMMS: Operating Room of the Future (2006-09) • DARPA: Traumapod robotic surgery (2005-06)

  6. Underlying Technologies we work on …. • Agent based Systems • MANET management and security • Ontologies for Semantic Interoperability • Semantic Policy Languages (especially for Security/Privacy/Trust) • Social Media Analytics (NER, Community Detection, …) • Streaming Analytics on KBs • Reasoning with uncertain/incomplete information

  7. Operating Room Of the Future • ORswill be pervasive computing environments Devices, sensors, tags, trainers, PDAs, monitors will discover one another and interoperate • Components will require access to a context model to manage resources effectively Includes relevant information on people, roles, activities, events, workflow, devices, … • Intelligent components willrecognizeevents and activities Even in the presence of noisy, incomplete or contradictory data

  8. Continuous Queries Physiological Data Patient Monitor Stream Processor (TelegraphCQ) Medicines Assert facts Context Aware Agent Assert facts Tools RFID System Video Clipper Database Events Staff Medical Encounter Record Patient History Staff Medical Supplies System Architecture Trend Analyzer Rule Base

  9. Event Detection - Level 3 Medical Encounter Record Video Clipper Rule Base Event Detection - Level 2 Events Staff Assert facts Assert facts Patient History Trend Analyzer Low-Level Event Processor Events Database Physiological Data Medical Supplies Event Detection - Level 1 Stream Processor (TelegraphCQ) RFID System Patient Monitor Continuous Queries Medicines Tools Staff

  10. Simulations and Results • The Human Patient Simulator (HPS) from METI • Designed to react like a human • Used for training resident doctors • Responds to medical treatment • Physiological data sets from HPS

  11. Scenario and Patient Profile • HPS can run patient profiles • Data logs from simulations used to evaluate the system • Significant events for a blunt trauma multiple injuries profile include hypovolemia, tension pneumothorax, decompression and fluid infusions • Provides data for Medical Encounter Record • Ran 30 simulations on 7 profiles measuring false positives & negatives and latency in detecting events Patient Profile

  12. Tables are everywhere and yet they are ignored!! • 154 million high quality relational tables on the web • Key domains like Medical, Bio-technology store information in spread sheets and tables • Traditional text processing techniques do not work well with tables • Key public policy decisions is often based on the information encoded in tables. • Lack of systems that can understand and infer the intended meaning of tables 14 nations (including US) share datasets publicly. Most of it is in spreadsheets and CSV

  13. Evidence based medicine The idea behind Evidence-based Medicine is to judge the efficacy of treatments or tests by meta-analyses or reviews of clinical trials. Key information in such trials is encoded in tables. However, the rate at which meta-analyses are published remains very low … hampers effective health care treatment … Figure: Evidence-Based Medicine - the Essential Role of Systematic Reviews, and the Need for Automated Text Mining Tools, IHI 2010

  14. Given a table, we … http://dbpedia.org/class/yago/NationalBasketballAssociationTeams dbprop:team http://dbpedia.org/resource/Allen_Iverson Map numbers as values of properties Generate a machine – understandable representation

  15. Class label prediction and Entity Linking Possible Classes for the column - dbpedia-owl:Place dbpedia-owl:City yago:WomenArtist yago:LivingPeople yago:NationalBasketballAssociationTeams dbpedia-owl:PopulatedPlace dbpedia-owl:Film… …. ….. 1. Chicago Bulls 2. Chicago 3. Judy Chicago 1. Philadelphia 2. Philadelphia 76ers 3. Philadelphia (film) • Generate a set of possible classes • Query the Knowledgebase • Score the classes Re-query KB with predicted class label as additional evidence An SVM-Rank classifier ranks the result set A second SVM classifier decides whether to link to the top-ranked instance or not • A machine learning based approach for entity linking

  16. Relation identification and RDF representation Rel ‘A’ • Query the knowledge base • Generate a set of possible relations • Rank the relations Rel ‘A’ Rel ‘A’, ‘C’ Rel ‘A’, ‘B’, ‘C’ Rel ‘A’, ‘B’ @prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> . @prefix dbpedia: <http://dbpedia.org/resource/> . @prefix dbpedia-owl: <http://dbpedia.org/ontology/> . @prefix yago: <http://dbpedia.org/class/yago/> . "Name"@en is rdfs:label of dbpedia-owl:BasketballPlayer . "Team"@en is rdfs:label of yago:NationalBasketballAssociationTeams . "Michael Jordan"@en is rdfs:label of dbpedia:Michael Jordan . dbpedia:Michael Jordan a dbpedia-owl:BasketballPlayer . "Chicago Bulls"@en is rdfs:label of dbpedia:Chicago Bulls . dbpedia:Chicago Bulls a yago:NationalBasketballAssociationTeams . • “Linked” RDF representation of data

  17. Results Class prediction for column: 76.92 % Entity Linking for table cells: 66. 12 % Examples of class label prediction results:Column – NationalityPrediction – MilitaryConflict Column – Birth PlacePrediction – PopulatedPlace Category wise Entity Linking results * The number in the brackets indicates # excluding columns that contained numbers

  18. Medical Table – Challenges Clues often hidden in captions More Numbers; less strings ! Numbers expressed in a pattern [e.g. 24 – 3.2 is Mean – Std. deviation] Row headers (in addition to column headers)

  19. Proposed approach Number preprocessing [Identify patterns such as Mean – S.D.] Information Extraction techniques(OMC -> Omeprazole, Metronidazole, Clarithromycin) Incorporate caption and text surrounding table as additional evidence Medical Tables Graphical Model based framework for table interpretation Background knowledge of Medical Domain Knowledge Source

  20. A Relational Learning based framework R33 R11 R12 R13 R21 R22 R23 R31 R32 C2 C3 C1 Function that captures the interaction between the column headers and row values Alternative – Markov Logic Network based system

  21. Problem Current epidemiologic studies of large cohorts does not take into account individual’s genetic, proteomic and metabolic characteristics. For complex diseases, state-of-the-art clinical-genomic based studies do not provide simple means to disseminate new findings into clinical practice. Hence gap continues to grow between the knowledge (clinical research, therapeutic guidelines, etc.) accumulated and it’s implementation at the patient’s bedside.

  22. Our focus -> Type 2 diabetes -> Chronic disease that comprises 90% of people with diabetes around the world Genetic Vs Non-Genetic Risk Factors

  23. Proposed Solution Develop a Web-based Clinical Decision Support System that will integrate genomic, metabolic associations and data mining correlative evidence gathered by computational algorithms for prediction and knowledge discovery and will be invoked on demand at the point of care. Study data – GENEVA Diabetes Study data obtained from dbGAP (NCBI-NIH database)

  24. Approaches Identification of dominant phenotypes in controls vs cases. (Primary - Body mass index, family history, cholesterol, physical activity, high blood pressure. Secondary – fat intake, cereal fiber intake, glycemic load, etc. Identification of Type II diabetes genes and their corresponding risk alleles. Correlating risk SNPs with the phenotypes and other environmental factors. Assign genomic risk score to a patient based on presence of risk SNPs in his/her genotype. Parameters to be used: - % of occurrence in cases vs controls. - Intensity variation in alleles A against allele B using scattered plots of .CEL files. Build a prediction model based on clinical and genomic datasets. Identify groups of patients based on their clinical and genomic behaviors.

  25. Key Results Type II Diabetes risk SNPs identified in the dataset under study: TCF7L2, FTO, MCR4, TSPAN8, VEGFA, BCL11A, HHEX, CDAKL1, MTNR1B. Findings: -Risk alleles always present in higher % of cases than controls. -TCF7L2 gene prominently present in cases than controls. -Individuals with FTO (fat mass and obesity association gene) weigh 2kgs more than an average person. -Individuals with TCF7L2 are strongly associated with family inheritance. Patients with higher genomic score (more no. of risk SNPs present) have higher chances of occurrence of T2D and vice versa. Genomic score along with phenotype features increases prediction accuracy by ~2%.

  26. Laparoscopic Cholecystectomy (Lap Chole) • Most common laparoscopic procedure performed • Cholecystectomy – surgical removal of gallbladder • First-choice treatment for: • gallstones and • inflammation of gall bladder • 4-5 incisions of 0.5-1.5 cm in diameter • CO2 – used to inflate abdominal cavity

  27. Complications in Lap Chole Hemorrhage Injury to common bile duct – connects gallbladder and liver Bile Leakage – Dangerous infection Stray burns from electrocauter Injury to bowel or vascular structures Abdominal peritoneal adhesions 5 – 20% conversion to open cholecystectomy

  28. Our Approach

  29. Region of Interest (ROI) Patch-based Template Matching Relative distance estimation Reduced Region of Interest (ROI) Cystic artery course the neck of the gallbladder It lies towards the lower right side of the neck

  30. SVM classifier training Training data – 900 images (470 positive, 430 negative) Test data – 213 images (135 positive, 78 negative) Linear kernel

  31. Results • 26.7 % increase in accuracy

  32. http://ebiquity.umbc.edu/

More Related