Interactive Analysis of Multidimensional Data. for Adverse Event Detection

Transcription

1 University of Peradeniya, Sri Lanka, December 21 st 2009 Interactive Analysis of Multidimensional Data for Adverse Event Detection Artur Dubrawski Director, The Auton Lab School of Computer Science Carnegie Mellon University Agenda 1. Introduction to the Carnegie Mellon Auton Lab 2. First technical topic Interactive Analysis of Multidimensional Data for Adverse Event Detection 3. Second technical topic T-Cube Web Interface for Real-time Biosurveillance in Sri Lanka and Tamil Nadu 4. Discussion Slide 2 of 51

2 CMU Auton Lab: Research and Applications Founded in 1993 Central topic of research: scalable, self-adaptive analytic systems with real life impact ~20 people 2 regular+3 affiliated faculty, 1 post-doc, 6 programmers and analysts, 8 PhD students; plus a few interns; led by Artur Dubrawski and Jeff Schneider Working on ~10 sponsored projects Current and past funding from NSF, DARPA, DHS/HSARPA, ONR, AFRL, NASA, USDA, FDA, CDC, DoE, a few Fortune 100 companies, and a number smaller corporate & academic sponsors and partners Deliverables Algorithms for fast and scalable statistical machine learning Software for embedding in production systems Software available for download (currently 100+ new downloads per month) Slide 3 of 51 Statistical Machine Learning Learning: improving performance at some task through experience Statistical Machine Learning: building probabilistic models from data (and/or from human expertise) Predictive tasks E.g. predicting threat level of a shipment based on the record of spectral and contextual information about a number of previously processed and evaluated shipments. Descriptive tasks E.g. explaining (in probabilistic terms) the kind and extent of relationships between data elements; also: how strange is the current shipment, given historical data. Auton Lab s main contributions: Scalable versions of Machine Learning algorithms New data structures which enable efficient implementations Progress driven by practical applications & real-world deployments Slide 4 of 51

3 Astrophysics Interactive analytics Learning Locomotion United Nations CTBTO Fleet prognostics Bio-surveillance Food safety Nuclear threat detection Slide 5 of 51 Safety of agriculture Real-time Monitoring of Emergency Department Patient Data during 2002 Winter Olympics We have built an early warning system for outbreaks of diseases (with U.Pitt.)... Slide 6 of 51 February 5, 2002

4 Rapid Detection of Emerging Patterns of Adverse Events Why Do We Care? Highly motivational example: Adverse bio-events Terrorist acts (Anthrax, Smallpox,...) 100 kg of anthrax released in DC may kill 1 3 million people (WHO) For the life of me, I cannot understand why terrorists have not attacked our food supply because it is so easy to do (Resignation speech by departing U.S. Health Secretary Tommy Thompson, December 3, 2004) Naturally occurring events, including emerging threats (SARS, Avian Influenza, West Nile Virus, Mad Cow Disease, Foot-and- Mouth Disease, E.coli, Salmonella,...) Lower-bound estimate of death toll of Avian Influenza pandemic: 2 7 million lives Unintentionally introduced events (such as accidental contamination of food at processing plant) Each year in the US: 76 million cases of food borne diseases (5,000+ of them fatal) Slide 7 of 51 Rapid Detection of Emerging Patterns of Adverse Events Why Do We Care? May also be useful in: Environmental monitoring (water/air quality) Pharmaceuticals (adverse effects of drugs on patients) Law enforcement (correlating independently acquired tips) Work safety (new patterns of hazardous behavior) Maintenance operations (patterns of equipment failures) Logistics (systematic patterns of supply shortages) FMCG market performance (out-of-spec batches)... Slide 8 of 51

5 Rapid Detection of Emerging Patterns of Adverse Events Why Do We Care? Detection based on definitive diagnoses may involve too much latency: A 2-day gain in detection time during an incident of inhalational Anthrax release could reduce fatalities by a factor of six (DARPA) Improvements of even an hour over current detection capabilities could reduce economic impact of a bioterrorist anthrax attack by hundreds of millions of dollars (Wagner et al. 2004) US agricultural industry: $1T of economic activity, $60B in food exports, root cause identification is apparently very difficult: 82% of food-borne disease cases are of unknown origin. How we can help: Exploit available early signals Build algorithms and supporting data structures for rapid detection Slide 9 of 51 What kind of data carries early signals? Post-Diagnosis Data Specific, available after several weeks Lab Reports Mortality Reports Pre-Diagnosis Data Non-specific, available relatively soon ER & OPD Visits Pharmacy OTC Sales School Absenteeism Ambulance Response Logs Veterinarian Data Consumer Complaints Slide 10 of 51

6 Potential utility of non-specific data Detection Response delay Deaths from Anthrax Number Count Dead Sales of cough and cold medicines Visits to emergency rooms for cough or fever Days After the Release Days after Release Slide 11 of 51 Auton Lab s algorithms and data structures for rapid detection of emerging patterns What s Strange About Recent Events (WSARE) Monitoring Emergency Rooms chief complaints for unusually high counts of patients from specific sub-populations who report with similar symptoms Fast Spatial Scan Monitoring volumes of non-prescription drug sales and/or ER data for spatio-temporal over-densities Multi-Stream Temporal Monitor Combining evidence from multiple streams of time series for detection of anomalies and specific patterns of interest Tip Monitor Monitoring food consumer complaints for small sets of similar reports which may be attributed to a common underlying cause T-Cube In-memory data structure enabling huge speedups in time series extraction and aggregation Slide 12 of 51

7 WSARE: What s Strange About Recent Events? Representative Surveillance Data Date Time Hospital ICD9 Prodrome Gender Age Home Many Location more 6/1/03 9: Fever M 20s NE 6/1/03 9: Diarrhea F 40s SE : : : : : : : : : Standard d Approach WSARE Approach Select in advance which subpopulations to monitor (e.g., each county, zip) Do not pay close attention to effect of multiple testing Monitor hundreds of thousands of subpopulations Pay close attention to effect of multiple testing Evaluation Detailed comparison on 2,000 simulated scenarios and Western PA ER Data WSARE2. 0 WSARE2. 5 WSARE3. 0 Standard Method Search over 100,000s of subpopulations For each subpopulation, use as good of a model as can be created to predict expected counts The model will have a form of a multi-component rule, e.g. There is a surprisingly large number of children with respiratory problems today Compute p value, taking into account multiple testing All Historical Data Today s Environment What should be happening today? Today s Cases What s strange about today, considering its environment? And how significant is this? Significance WSARE is designed to detect small clusters of illness in healthcare workers, age groups, workplaces... Israeli Center for Disease Control evaluated WSARE retrospectively using an unusual outbreak of influenza type B that occurred in an elementary school in central Israel. It detected the outbreak on the second day from its onset. Retrospective analysis of Walkerton case: alerted one day ahead of the issuance of boil water advisory (at the expected rate of 2 false positives per year). Slide 13 of 51 Fast Spatial Scan OTC drug sales Method (Expectation-Based Variant of FSS) Time series of counts for each zip code (at least 3 months of historical data). Multivariate Bayesian variant of this method analyzes multiple streams of data to maximize detection power while enabling disambiguation among possible causes of outbreaks Evaluation Current version typically turns multi-day analysis into 20 minutes for daily counts from over 20,000 drugstores nationwide Searches all rectangular regions. Results are exact (not approximations). This increase could be due to an outbreak, or due to chance. Which regions of increase are significant? Search all rectangular regions on the grid. Solution: We learn the expected count for each area from historical data. Then we find regions where the recent counts are significantly higher than expected, accounting for anticipated spatial and temporal variations Significance Traditional spatial scan is very expensive, especially with randomization tests of significance A few hours difference may actually matter. Retrospective analysis of Walkerton case: alerted two days ahead of the issuance of boil water advisory. Algorithm Search space # of regions Search time Time / region Likelihood ratio SaTScan Circles centered at datapoints 150 billion 16 hours 400 ns exhaustive Axis-aligned rectangles 1.1 trillion 45 days 3600 ns FSS Axis-aligned rectangles 1.1 trillion 81 minutes 4.4 ns ER dataset (600,000 records), 1000 replicas; for SaTScan: M=17,000 distinct spatial locations; for exhaustive/fast: 256 x 256 grid. Slide 14 of 51

8 Safety of Food Supply and Agriculture Common denominator: To quickly detect emerging patterns and rank them according to the expected risk, so that investigative resources can be allocated efficiently Import elexnet (FDA) Food Sample Test Results (microbes, toxins) Farm Fork CZDSS (USDA) Slaughterhouses (condemnations & results of lab tests for pathogens, residues and microbes causing zoonotic diseases) PHRBIS PHIS (USDA) Multiple Streams (condemnations, lab tests, inspection results, recalls, ) CCMS II (USDA) Consumer Complaints Slide 15 of 51 Efficient representation of data in support of Event Detection The next part of the talk claims that adverse events can be detected and analyzed at interactive speeds when using smart computer science to represent and process data Slide 16 of 51

9 How can we benefit from representing data efficiently? Some relevant tasks in which it can help: 1. Early detection of emerging crises 2. Characterization and interpretation of detected patterns/processes 3. Tracking evolution of adverse events Slide 17 of 51 How can we benefit from representing data efficiently? Some relevant tasks in which it can help: 1. Early detection of emerging crises 2. Characterization and interpretation of detected patterns/processes 3. Tracking evolution of adverse events This can be achieved via: Interactive navigation through data Human-initiated iti t drill-downs d and summarizations Highly responsive visualizations Execution of data-intensive statistical algorithms Slide 18 of 51

10 Example data Fragment of a multidimensional record of disease cases: sign_joint_pain sign_muscle_pain sign_red_eyes sign_light_colored d_stool sign_diarrhea sign_rash sign_vomiting sign_dark_urine sign_abdominal_p pain sign_chills sign_other symptom_loss_of f_appetite symptom_fever symptom_fatigue symptom_nausea symptom_headach he symptom_other date location disease age_group gender 16 Dec 06 Colombo Dengue_fever Female Dec 06 Colombo Dengue_fever Above_45 Female Dec 06 Colombo Dengue_fever Male Dec 06 Colombo Dysentery Female Dec 06 Kandy Dengue_fever Male Dec 06 Kandy Dysentery 0 1 Male Dec 06 Kandy Dysentery Female Dec 06 Matale Viral_Hepatitis Above_45 Female Dec 06 Nuwara_Eliya Viral_Hepatitis Above_45 Male Dec 06 Galle Dengue_fever Male Dec 06 Hambantota Typhus_fever Above_45 Female Dec 06 Matara Dengue_fever Above_45 Male Dec 06 Matara Leptospirosis Female Dec 06 Vavuniya Dengue_fever Female Slide 19 of 51 Plausible idea: Replace raw data with sufficient statistics Pre-compute key statistics about data ahead of its extensive analyses in order to amortize the bulk of the costs of future computations Example: Using Contingency Tables to represent categorical data Mining categorical data is very often all about counting (co-)occurrences Precomputing counts makes the future costs of analyses independent on the data size Slide 20 of 51

11 Plausible idea: Replace raw data with sufficient statistics Pre-compute key statistics about data ahead of its extensive analyses in order to amortize the bulk of the costs of future computations Example: Using Contingency Tables to represent categorical data Mining categorical data is very often all about counting (co-)occurrences Precomputing counts makes the future costs of analyses independent on the data size E[ P(ArrackDrinker SriLankan) ] = = NumberOf(SriLankan ArrackDrinkers)/NumberOf(SriLankans) Slide 21 of 51 Plausible idea: Replace raw data with sufficient statistics Pre-compute key statistics about data ahead of its extensive analyses in order to amortize the bulk of the costs of future computations Example: Using Contingency Tables to represent categorical data Mining categorical data is very often all about counting (co-)occurrences Precomputing counts makes the future costs of analyses independent on the data size E[ P(ArrackDrinker SriLankan) ] = = NumberOf(SriLankan ArrackDrinkers)/NumberOf(SriLankans) rds N=7 Recor Raw data SriLankan? ArrackDrinker? CricketFan? Contingency SriLankan=0 0 rackdrinker=0 Arr 1 1 rackdrinker=1 Arr 2 2 table M=3 Attributes Slide 22 of 51 SriLankan=1 0 0

12 Plausible idea: Replace raw data with sufficient statistics Pre-compute key statistics about data ahead of its extensive analyses in order to amortize the bulk of the costs of future computations Example: Using Contingency Tables to represent categorical data Mining categorical data is very often all about counting (co-)occurrences: Complaint: Several modeling techniques rely on that (Decision trees, Bayesian networks, ) Contingency E[ P(Flyfisher Canadian) Tables can ] = reach enormous sizes (numbers of cells) if the underlying = NumberOf(Canadian data is highly Flyfishers)/NumberOf(Canadians) dimensional and if the involved It needs variables to be done can in multivariate assume data many spaces different values rds N=7 Recor Raw data SriLankan? ArrackDrinker? CricketFan? SriLankan=0 0 rackdrinker=0 Arr 1 1 rackdrinker=1 Arr 2 2 Contingency table M=3 Attributes Slide 23 of 51 SriLankan=1 0 0 All-Dimensional-Tree: A smarter representation Replace contingency table with a tree It represents the same counts of co-occurrences A1=* A2=* C=8 Vary nodes contain queries in which specific attributes are instantiated Vary A1 Vary A2 Data A1 A A1=1 A2=1 C=0 A1=1 A2=* C=1 Vary A2 A1=1 A2=2 C=1 A1=2 A2=1 C=2 A1=2 A2=* C=3 Vary A2 A1=2 A2=2 C=1 A1=3 A2=1 C=1 A1=3 A2=* C=4 Vary A2 A1=3 A2=2 C=3 A1=* A2=1 C=3 A1=* A2=2 C=5 Count nodes store queries and the corresponding counts of the records of data matching them Slide 24 of 51

13 AD-Tree: A smarter representation Take advantage of sparseness and redundancies in data Do not store anything with zero counts and do not store sub-trees of Most Common Values A1=* A2=* C=8 Data A1 A NULL A1=1 A2=* C=1 Vary A2 mcv->2 NULL (mcv) NULL (mcv) Vary A1 mcv->3 A1=2 A2=* C=3 Vary A2 mcv->1 A1=2 A2=2 C=1 NULL (mcv) A1=* A2=1 C=3 Vary A2 mcv->2 NULL (mcv) Such tree consumes less memory than the equivalent Contingency Table (the higher dimensionality of data the more attainable savings) Yet we still can cheaply re-compute all of the removed counts Slide 25 of 51 The idea can be extended T-Cube Efficient representation of multi-dimensional time series of counts If we have a set of additive time series (e.g. series of counts of events) annotated with multiple categorical descriptor variables: Technical details: Sabhnani M., Moore A. and Dubrawski A. T-Cube: A Data Structure for Fast Extraction of Time Series from Large Datasets. Technical Report CMU-ML , Carnegie Mellon University, Slide 26 of 51

14 T-Cube: Evaluation in a controlled environment (circa 2007) Commercial datacube tools evaluated: TimesTen (Oracle), ANTS Data Server (ANTs Software), extremedb (McObject), TimeSeries DataBlade (IBM, designed for time series) Memory Complex Query Response Time Tool MB 6.8s Tool MB 7.6s Tool 3 1+ GB 3.5s T-Cube 236 MB 22ms T-Cube 845 MB 5ms All tested on 12 million transactional records of synthetic data with 3 attributes (arities of 1,000; 10 and 5); on a Windows XP machine with 2GB RAM and 2.4GHz CPU. Format of data used in the example Date Gender Place Complaint Count 1/1/2006 M 100 GI 4 1/1/2006 M 300 Resp 3 1/1/2006 F 300 Fever 11 1/1/2006 M 200 Resp 3 1/1/2006 F 400 Fever 2 1/2/2006 M 200 GI 1 1/2/2006 F 400 GI 4 1/2/2006 M 300 Resp 2 1/2/2006 F 300 Fever 5 1/2/2006 M 200 GI 6 1/3/2006 M 200 GI 2 1/3/2006 F 300 Resp 1 1/3/2006 M 100 Fever 4 1/3/2006 F 300 GI 2 1/3/2006 F 400 GI 3 Slide 27 of 51 T-Cube Applications: Detection of adverse events in public health Rapid and reliable detection of emerging threats to public health Usage of multiple sources of early-warning data (ER visits, OTC drug sales, ) Autonomous learning of probabilistic models of detection Multivariate frequentist and Bayesian approaches Required scalability to large quantities of spatio-temporal temporal data Other requirements: Rapid detection of even small clusters of illness, sensitivity to signals supported by little evidence, management of false positives Timely detections may results in huge savings (lives, dollars): processing time is a factor Slide 28 of 51

15 T-Cube Applications: Real-time surveillance of public health in Sri Lanka and Tamil Nadu Bayesian Spatial Scan Time series of counts for each analyzed region Multivariate Bayesian Spatial Scan analyzes streams of data to maximize detection power while enabling disambiguation among possible causes of outbreaks Bivariate Temporal Scan Temporal Scan alerts of unusual increases in counts of events of interest (e.g. patients reporting recently with Dengue fever) which cannot be explained by the changes target counts (e.g. daily counts of positive results of microbial tests for Salmonella) Jan-01 Apr-01 Jul-01 Oct-01 Jan-02 Apr-02 Jul-02 Oct-02 Jan-03 Apr-03 Jul-03 current time window Oct-03 baseline counts (e.g. daily counts of negative results of tests for Salmonella) in baselines (e.g. total number of patients). TCube runs a large number of such tests quickly. T-Cube Web Interface Within seconds of loading the data, the analysts found an emerging Leptospirosis event in the South-Western and Central provinces of Sri Lanka. Spatial scan algorithm tracks probabilities of Leptospirosis outbreak anywhere in the nation for all days in the past data (red line, and red dots in the map). On 8/20/2008, it went over 97%. Blue bars in the time series plot and blue circles in the map depict the temporal and spatial distribution of the volume of patients diagnosed with the disease Jan-01 Apr-01 Jul-01 Oct-01 Jan-02 Apr-02 Jul-02 Oct-02 Jan-03 Apr-03 Jul-03 Oct-03 Jan-04 Jan-04 Apr-04 Apr-04 Jul-04 Jul-04 Oct-04 Oct-04 Jan-05 Jan-05 Apr-05 Apr-05 Jul-05 Jul-05 Oct-05 Oct-05 Jan-06 Jan-06 Apr-06 Apr-06 Jul-06 Jul-06 Oct-06 Oct-06 Slide 29 of 51 T-Cube Applications: Crises in food safety Interactive visualization of multiple data streams USDA Food Sampling Data CDC PulseNet Human Illness Data About 150,000 records related to Salmonella across 3 years Daily, transactional temporal resolution Spatial resolution by unique establishment Key attributes: test result (positive, negative), serotype, PFGE pattern, antibiotic resistance pattern, product type, establishment production profile Certain statistics such as moving average are computed/updated on-the-fly Color coding is used to show multiple categories of data About 100,000 records related to Salmonella across 3 years Daily, transactional temporal resolution Spatial resolution by state Key attributes: serotype, PFGE pattern, outbreak ID Controls such as time window sliders make visualization interactive Slide 30 of 51

16 T-Cube Applications: Crises in food safety Supporting trace-back investigations Est. A Human health data contained a cluster of human cases of salmonellosis for which microbiological tests identified an identical PFGE pattern FSIS CDC Est. A After correlating that data against records of microbial sampling of food at food factories, Establishment A was found to report positives with a similar PFGE immediately prior to the emergence of human cluster It was located in the same geographic area as the human cases Slide 31 of 51 T-Cube Applications: Detection and monitoring of logistic crises in fleets of equipment Systematic failures in maintenance and logistic support, operational processes or hardware/software quality can limit availability, incur serious costs, reduce personnel safety and mission capability, if not identified quickly Idea: Exhaustively search records of equipment maintenance to provide early identification and characterization of emerging crises Note: Crisis mapping is not always geo-spatial! Slide 32 of 51

17 Variety of structures for efficient representation of data They can provide large speedups of data access with respect to previously considered efficient alternatives: AD-Trees: useful to represent multidimensional categorical data 1-4 orders of magnitude savings in processing time required by computationally intensive data mining processes (attainable if the data is at least partially correlated) Example: 1,580,000 Galaxies, 27 binary attributes per galaxy, time to build AD-Tree: 4 minutes, tree memory: 2 Megs, time required to execute 50,000 iterations of Bayesian network structure search: less than 2 minutes (previously 1.3 days) Kd-Trees: useful to represent multivariate numeric data up to ~8-D Gaussian mixture density modeling and clustering: Speedups of 8 to 1000 times vs. an efficient, but not kd-tree based implementation [Moore 1999] K-means clustering: Speedups of 150+ times on real-world astrophysical data [Pelleg & Moore 1999] Non-parametric multi-resolution regression: Speedups of times [Deng & Moore 1995] Spatial Scan Statistic: 50 times speedups in analyzing nationwide OTC pharmacy sales [Neill 2006] Metric Trees: useful to represent highly multivariate numeric data Mining data up to 10,000 (yes! ten thousand!) dimensions reveals 2.5 to 2,000-fold speedups w.r.t. otherwise efficient classical approaches to k-means clustering, grouping attributes and non-parametric anomaly detection [Moore 2000]. Slide 33 of 51 Summary of benefits Efficient representations of data: Support interactive visualizations Support rapid drill-downs, summarizations (roll-ups) Enable comprehensive monitoring of highly dimensional data for indications and tracking of emerging or unexpected patterns Combined benefits improve attainable situational awareness and they can make monitoring of adverse events more effective A simple variant of T-Cube Web Interface is available for evaluation at: Acknowledgements of partial support: USDA (award ), CDC (R01-PH000028), NSF (IIS , IIS ), USDoD (FA C-7264), IDRC (Canada, ). Slide 34 of 51

18 And now, to the next set of slides Slide 35 of 51 Originally presented at the 2009 ISDS 8 th Annual Conference, Miami, Florida, USA T-Cube Web Interface for Real-time Biosurveillance in Sri Lanka and Tamil Nadu Maheshkumar Sabhnani 1, Artur Dubrawski 1 and Nuwan Waidyanatha 2 1 The Auton Lab Carnegie Mellon University i 2 LIRNEAsia Colombo, Sri Lanka Slide 36 of 51

19 The Need for Real-Time Biosurveillance Slide 37 of 51 Source: DARPA T-Cube Web Interface for Real-time Biosurveillance in Sri Lanka and Tamil Nadu Objective: To present an on-going effort to establish a system for real-time disease surveillance in a developing country Agenda: 1. Current disease reporting system in Sri Lanka 2. Concept of the Real-Time Biosurveillance Program 3. TCWI: An interactive surveillance component of RTBP a. Short demo (video) b. Examples of evaluations performed so far 4. Conclusion Slide 38 of 51

20 Sri Lanka Disease Notification: Current System Black arrows represent the current manual paper/postal system for health data collection and reporting It can take up to 10 days for Epidemiology Unit to receive data It only records cases of a few reportable diseases and no syndrome or diagnosis information No direct alerting capability; alerts distributed by phone throughout organizational hierarchy (Epi Unit Regional Epis Medical Officers of Health Hospitals, Providers, etc ) Slide 39 of 51 SL Disease Notification: Opportunity for Improvement Red lines: RTBP mobile phone communication system for heath data collection and reporting Many developing countries, including Sri Lanka, have pretty good cell phone coverage RTPB leverages that to mitigate reporting latencies RTBP adds automated alerting and advanced analytic capabilities It also aims at deepening and widening the scope of what is being reported Slide 40 of 51

21 RTBP: Actors, Processes, Data Flow Slide 41 of 51 Improved Detail and Resolution of Data Cur rrent distinct value count 17 9 case date location disease Min Bihalpola Dengue_fever Max Bingiriya Dysentery Bopitiya Encephalitis Dambadeniya Enteric_fever Dunakadeniya Food_poisoning Hettipola Human_rabies Horathapola Leptospirosis Kandanegedara Typhus_fever Kattimahana Viral_Hepatitis Katupotha Koshena Kuliyapitiya Minuwangatte Munamaldeniya Narammala Sandalankawa Wariyapola RTBP: More dimensions More values per dimension Not just reportable diseases Daily resolution and collection cycles Finer spatial resolution New distinct value count possible choices single single single single multiple multiple case date location disease gender age group sign symptom Min Ammanamulla Abdominal pain Female 0-1 Abdominal Pain Coated tongue Max Bogahapitiya Abscesses Male 1-5 Anal itching Crepitation Bopitiya Accident Unknown 6-14 Backache Cyanosis Deegalla Acne Body ache Dehydration Dunakadeniya Acute Diarrheal Disease Burning sensation in the stomach Delirium Havenegedara Acute Gastroenteritis Chest pain Deviation of the tongue Horathapola Allergy Cold Difficulty in breathing Horombawa Anemia Constipation Heart murmur Kadapathwehera Angina Cough Increase of respiratory rate Kandanegedara Anxiety Above_45 Deviation of the tongue Low Blood pressure Slide 42 of 51

22 T-Cube Web Interface in RTBP Efficient representation of data substantially speeds-up analyses and user interactions Implements statistical event detection techniques Enables data visualization along temporal and spatial dimensions at interactive speeds Automates alerts Slide 43 of 51 Video: Demo of the RTBP T-Cube Web Interface Slide 44 of 51

23 RTBP: TCWI Evaluations Performed So Far 17 Hospitals/Central Dispensaries 16 Sarvodaya Suwadana Centers Kurunegala District, North Western Province, Sri Lanka 24 Health Sub Centers; 4 Public Health Centers Thirupathur block, Sivaganga District, state of Tamil Nadu, India Data is being collected under the new schema only for a few months now Its amount & consistency do not warrant objective evaluations just yet Therefore, so far, evaluations rely on semi-synthetic data: Use actual historically reported counts per region and per reportable disease Other attributes in RTBP schema are then probabilistically synthesized We test the ability to detect known events, the usability in tracking of emerging events and in revealing systematic patterns of events. Slide 45 of 51 Example Evaluation: Leptospirosis: Deadly Pile-up during Q4_2008--Q1_ /30 4/17 8/1 Over 360 deaths Peak in Sept/08 originally noticed much later later RTBP would have picked up leading waves much earlier than Sept/08 10/30 4/17 8/1 Slide 46 of 51

24 Example Evaluation: Progression of Dengue Fever Outbreak (Spring 2009) 4/14 4/15 4/24 5/28 Spatial scan global score An elevated global score noted, centered in Kandy Situation in Kandy escalates, regions to the W and SW from it follow suit South begins to see increased counts South continues to escalate, ate, while the magnitudes in other regions subside Slide 47 of 51 Example Evaluation: Dengue Fever: Seasonal and Spatial Patterns Note apparent Aug 30,2007 seasonality of Dengue fever It seems to be moving South and West May 1,2007 May 21,2008 April 15,2009 May 28,2009 Slide 48 of 51

25 Summary and Conclusion 1. RTBP deepens and widens coverage of reported health data while significantly reducing latency of reporting 2. T-Cube Web Interface (TCWI) supports RTBP by enabling: a. Automated comprehensive searches for events of interest through large collections of data b. Interactive data navigation and visualization c. Automated explanation of detected patterns 3. A fielded RTBP-like system can qualitatively yet affordably improve ability to detect and mitigate bio-medical threats in countries with limited resources and infrastructure 4. The pilot studies are ongoing in Sri Lanka and India, but we are on the outlook for additional challenges. Slide 49 of 51 Acknowledgements This work has been supported in part by the International Development Research Centre of Canada (105130), and by CDC (R01-PH000028) and NSF (IIS ). We have also received support from numerous collaborators not specifically named above! Partners in RTBP: LIRNEAsia Indian Institute of Technology Madras, India National Center for Biological Sciences, India Sarvodaya Shramadana Society, Sri Lanka Respere Lanka (Private) Limited, Sri Lanka University of Alberta, Canada Carnegie Mellon University Auton Lab, USA Slide 50 of 51

26 Contact Carnegie Mellon University 5000 Forbes Avenue, NSH 3121 Pittsburgh, PA , USA Artur Dubrawski Director, Auton Lab Systems Scientist, Robotics Institute Adjunct Professor, Heinz School of Public Policy and Management Tel: Fax: Slide 51 of 51