Biosurveillance: Machine Learning And Disease Surveillance by Kass-Hout Di Tada

Slideshow Transcript

1. Slide 1: MACHINE LEARNING AND DISEASE SURVEILLANCE Taha Kass-Hout, MD, MS Nicolás di Tada October 2008
2. Slide 2: Image source: http://www.birds.cornell.edu/crows/images/deadcrow.jpg Image source: http://farm3.static.flickr.com/2029/2239605500_6ef2fd2295.jpg?v=0
3. Slide 3: LATE DETECTION – RESPONSE Opportunity for control CASES DAY
4. Slide 4: EARLY DETECTION AND RESPONSE Opportunity for control CASES DAY
5. Slide 5: INFORMATION SOURCES  Event-based – ad-hoc unstructured reports issued by formal or informal sources  Indicator-based – (number of cases, rates, proportion of strains…)
6. Slide 6: PUBLIC HEALTH MEASURES  Representativeness  Completeness  Predictive Value  Timeliness
7. Slide 7: PUBLIC HEALTH MEASURES Specificity / 50 Malaria Reliability notifications (5%) Urge frequent reporting: Weekly  daily  immediately 1000 Malaria infections (100%) Get as close to the bottom of the pyramid as possible Sensitivity / Timeliness • Main attributes o Representativeness o Completeness o Predictive value positive
8. Slide 8: PUBLIC HEALTH MEASURES Health care hotline • Main attributes o Timeliness Signal as early as possible Analyze and interpret Time Automated analysis/ thresholds
9. Slide 9: THE PROBLEM SPACE  Current systems design, analysis and evaluation has been geared towards specific data sources and detection algorithms – not humans  We have systems in place for those threats we have been faced with before
10. Slide 10: PUBLIC HEALTH – TWO PERSPECTIVES  Case management  Individual cases of notifiable diseases  Relationship networks (contact tracing)  Population surveillance  Larger risk patterns
11. Slide 11: CASE MANAGEMENT  Questions/problems:  Is a case due to recent transmission?  If so, does the case share any feature with other, recent cases?  Ways it's being done:  Investigations/interviews  Meeting with other investigators
12. Slide 12: POPULATION SURVEILLANCE  Questions/problems:  Are more cases happening than expected?  Does an excess suggest ongoing transmission in a specific region?  Way it's being done:  Semi-automated routine temporal and space-time statistical analysis
13. Slide 13: WHY LOCATION MATTERS – CASE MANAGEMENT  If you are studying a case of a certain disease that was just declared  It is harder to picture the situation by looking at something as this..
14. Slide 14: WHY LOCATION MATTERS – CASE MANAGEMENT
15. Slide 15: WHY LOCATION MATTERS – CASE MANAGEMENT  Than by looking at this..
16. Slide 16: WHY LOCATION MATTERS – CASE MANAGEMENT
17. Slide 17: WHY LOCATION MATTERS – POP SURVEILLANCE  If you are studying the spatial distribution of a set of disease clusters  This would seem more difficult..
18. Slide 18: WHY LOCATION MATTERS – POP SURVEILLANCE
19. Slide 19: WHY LOCATION MATTERS – POP SURVEILLANCE  Than this..
20. Slide 20: WHY LOCATION MATTERS – POP SURVEILLANCE
21. Slide 21: MODERN DISEASE SURVEILLANCE  In the past two decades, much disease surveillance research has focused on developing analytical methods for automatically detecting anomalous patterns in data  Modern methods can achieve timely detection of anomalies by incorporating temporal, spatial, and multivariate information
22. Slide 22: MODERN DISEASE SURVEILLANCE 9/20, 15213, cough/cold, … 9/21, 15207, antifever, … 9/22, 15213, CC = cough, ... 1,000,000 more records… Huge mass of data Detection algorithm Too many “What are we supposed alerts to do with this?”
23. Slide 23: MODERN DISEASE SURVEILLANCE 9/20, 15213, cough/cold, … 9/21, 15207, antifever, … 9/22, 15213, CC = cough, ... 1,000,000 more records… Huge mass of data Feedback loop
24. Slide 24: ADVANTAGES OF MACHINE LEARNING P(malaria) = 22% P(influenza) = 13% P(other ILI) = 33%
25. Slide 25: MACHINE LEARNING TECHNIQUES  Classifiers  Clustering  Bayesian Statistics  Neural Networks  Genetic Algorithms
26. Slide 26: HOW TO REPRESENT A DOCUMENT? “I had a flu last month. […] I had flu a flu early this week.” fever “This morning I woke up with fever, I might have a flu.”
27. Slide 27: CLASSIFIERS – PROBLEM DEFINITION  Map items to vectors (Feature extraction)  Normalize those vectors  Train the classifier  Measure the results with new information  Feedback the classifier  Separate classes in feature space
28. Slide 28: CLASSIFIERS - SVM
29. Slide 29: SVM – MARGIN MAXIMIZATION  Support vectors define the separator
30. Slide 30: SVM – NON LINEAR? Map to higher-dimension space Φ: x → φ(x)
31. Slide 31: SVM – FILTERING OR CLASSIFYING Training Training Document Document Positives Document 1 Document 2 Classifier Document 3 Negatives
32. Slide 32: CLUSTERING – PROBLEM DEFINITION  Map items to vectors (Feature extraction)  Normalization  Agglomerative and Partitional
33. Slide 33: CLUSTERING - AGGLOMERATIVE
34. Slide 34: CLUSTERING - PARTITIONAL
35. Slide 35: BAYESIAN STATISTICS Probability of Probability of flu Probability of fever once flu is disease A (flu) once (prior or confirmed marginal) symptoms B (fever) are observed P(B | A).P(A) P(A | B) = P(B) Probability of fever (prior or marginal)
36. Slide 36: NEURAL NETWORKS  Given a set of stimulus, train a system to produce a given output
37. Slide 37: NEURAL NETWORKS - STRUCTURE Weight {I0,I1,……In} Input Layer […] I H n = ￥ (Ii .w in ) i= 0 ﾠ Hidden Layer […] Output Layer {O0,O1,……On}
38. Slide 38: NEURAL NETWORK - APPLICATION Event?
39. Slide 39: GENETIC ALGORITHM - BASICS  Define the model that you want to optimize  Create the fitness function  Evolve the gene pool testing against the fitness function.  Select the best individual
40. Slide 40: GENETIC ALGORITHM – MODEL  Model the transmission process using a set of parameters:  Onset time between an infection and illness  Latency period  Incubation period  Symptomatic period  Infectious period (Onset, Latency, Incubation, Symptomatic , Infectious) ( 2 days, 3 days, 1 day, 4 days, 3 days)
41. Slide 41: GENETIC ALGORITHM – MODEL FITNESS Fitness = 1/Area
42. Slide 42: GENETIC ALGORITHM – PROCESS 1. Create an initial population of candidates 2. Use operators to generate new candidates (mating and mutation) 3. Discard worst individuals or select best individuals in generation 4. Repeat from 2 until you find a candidate that satisfies the solution searched
43. Slide 43: GENETIC ALGORITHM - PROCESS (5,3,4,6,2) (2,4,6,3,5) (4,3,6,5,2) (2,3,4,6,5) (4,3,6,2,5) (4,5,6,3,5) (3,4,5,2,6) (3,4,4,6,2) (5,3,2,6,5) (3,5,4,6,2) (4,5,3,6,2) (5,4,2,3,6) (4,6,3,2,5) (3,4,2,6,5) (3,6,5,1,4) (5,3,2,6,5) (3,4,4,6,2)
44. Slide 44: RESULTS – IMPROVED SURVEILLANCE
45. Slide 45: Q&A
46. Slide 46: THANK YOU! Taha Kass-Hout, MD, MS http://www.instedd.org kasshout@instedd.org http://taha.instedd.org Nicolás di Tada http://www.manas.com.ar nditada@manas.com.ar http://weblogs.manas.com.ar/ndt/
47. Slide 47: BACKUP SLIDES
48. Slide 48: REFERENCES  Izadi, M. and Buckeridge, D., Decision Theoretic Analysis of Improving Epidemic Detection, AMIA 2007, Symposium Proceedings 2007  EpiNorth-Based material (http://www.epinorth.org):  Mereckiene, J., Outbreak Investigation Operational Aspects. Jurmala, Latvia, 2006  Bagdonaite, J., and Mereckiene, J., Outbreak Investigation Methodological aspects. Jurmala, Latvia, 2006  Epidemic Intelligence: Signals from surveillance systems, Anne Mazick, Statens Serum Institut, Denmark, EpiTrain III, Jurmala, August 2006  Daniel Neil, Incorporating Learning into Disease Surveillance Systems
49. Slide 49: REFERENCES  Algorithms  Complex Event Processing Over Uncertain Data in Wasserkrug (2008)  Outbreak detection through automated surveillance A review of the determinants of detection in Buckeridge (2007)  Approaches to the evaluation of outbreak detection methods in Watkins (2006)  Algorithms for rapid outbreak detection a research synthesis Buckeridge (2004)  Data mining in bioinformatics using Weka in Frank (2004)
50. Slide 50: REFERENCES  Automating Laboratory Reporting  Automatic Electronic Laboratory-Based Reporting in Panackal (2002)  Benefits and Barriers to Electronic Laboratory Results Reporting for Notifiable Diseases in Nguyen (2007)  Using EMR Data for Disease Surveillance  Using Electronic Medical Records to Enhance Detection and Reporting of Vaccine Adverse Events in Hinrichsen (2007)  Electronic Medical Record Support for PH in Klompas (2007)  A knowledgebase to support notifiable disease surveillance in Doyle (2005)  Automated Detection of Tuberculosis Using Electronic Medical Record Data in Calderwood (2007)  Misc Readings  Breakthrough in modeling emerging disease hotspots in Jones (2008)  Use of data mining techniques to investigate disease risk classification as a proxy for compromised biosecurity of cattle herds in Wales in Ortiz-Pelaez (2008)
51. Slide 51: RELATED PROJECTS  InSTEDD RNA (or Event Evolution): Collaborative Analytics and Environment for Linking Early Health-Related Event Detection to an Effective Response ( http://taha.instedd.org/2008/09/collaborative-analytics-and-environment.html )  ALPACA \"ALPACA Light Parsing And Classifying Application (ALPACA) is a classifying tool designed for use in community-oriented software as well as in Academia. The application consists of two parts: a parsing tool for transforming raw documents into readable data, and a classifying tool for categorizing documents into user-provided classes. The application provides a user-friendly interface and a Plug-in functionality to provide a simple way to add more parsers/classifiers to the application.\" http://2008.hfoss.org/ALPACA  Surveillance Project An Open Source R-package disease surveillance framework for \"...the development and the evaluation of outbreak detection algorithms in univariate and multivariate routine collected public health surveillance data.\" http://surveillance.r-forge.r-project.org/  Weka An open source \"...collection of machine learning algorithms for data mining tasks. The algorithms can either be applied directly to a dataset or called from your own Java code. Weka contains tools for data pre-processing, classification, regression, clustering, association rules, and visualization. It is also well-suited for developing new machine learning schemes.\" http://www.cs.waikato.ac.nz/~ml/weka/