Condition Monitoring for Evidencebased Patient Care

1
Condition Monitoring for Evidence-based Patient
Care
Samuel H. Huang, Associate Professor and
Director http//www.eng.uc.edu/icams
2
Outline

• Background
• Evidence-based Patient Care and Condition-based
  Maintenance
• Research Thrust
• Knowledge-based Modeling
• Dimensionality Reduction
• Missing Data Imputation
• Summary

3
Background

• Intelligent Systems Laboratory established in
  September 1998 at the University of Toledo,
  relocated to the University of Cincinnati in
  September 2001
• Vision computer systems as a tool to expand
  human perception and computation ability so human
  can make better decisions
• Research Focus productivity improvement for
  manufacturing and service systems
• Current Team Member 4 Ph.D. and 3 M.S. students
• Alumni 3 Post-Docs, 5 Ph.D. and 26 M.S. students

4
Evidence-based Patient Care (EBPC)

• Promotes the collection, interpretation, and
  integration of applicable patient-reported,
  clinician-observed, and research-derived evidence
  to improve the quality of clinical judgments and
  facilitate cost-effective healthcare

5
Condition-based Maintenance (CBM)

• Monitor production equipment using various
  sensors
• Real-time diagnosis of impending failures
• Maintenance actions take place in a timely
  fashion and on an as-needed basis.

6
Summary of CBM Application Results
7
Challenges in Applying CBM to EBPC

• Unavailability of analytical models for patient
  diagnosis/prognosis
• Large number of parameters in patient data
• Gaps (missing values) in patient data

8
Knowledge-based Modeling

• An expert-centered data-driven modeling approach
  that integrates knowledge embedded in data and
  those possessed by experts
• IF-THEN linguistic rules is used as a unified
  knowledge representation scheme because of their
  compatibility to expert heuristic reasoning and
  their adaptability to actual data
• General modeling framework with enabling tools

9
Diagnosis Case Study

• Pima Indian Diabetes dataset from the University
  of California at Irvine Machine Learning
  Repository
• Eight (8) parameters, 768 female patients at
  least 21 years old of Pima Indian heritage either
  healthy or diabetic
• Two parameters are significant features F2
  (plasma glucose concentration) and F6 (body mass
  index)
• Four rules are extracted with an diagnosis
  accuracy of 75 (78 is best achievable)
• IF F2 is low and F6 is low THEN 86.4 Healthy
• IF F2 is low and F6 is high THEN 67.3 Healthy
• IF F2 is high and F6 is low THEN 36.9 Diabetic
• IF F2 is high and F6 is high THEN 84.3 Diabetic

10
Dimensionality Reduction

• Directly reduce the dimension of the input space
  by identifying parameters that are necessary and
  sufficient for model building

11
SF-36 Mental/Physical Score Study

• The Centers for Medicare and Medicaid Services
  use a set of 54 variables to survey their
  subscribed patients
• SF-36 Physical score are highly dependent on the
  following questions
• getting in or out of chairs (Q12d)
• walking (Q12e)
• Shortness of breath when walking less than one
  block (Q14c)
• Shortness of breath when climbing one flight of
  stairs (Q14d)
• Interference of low back pain (Q36)
• SF-36 Mental score are highly dependent on the
  following questions
• Two Weeks of Depression Question (Q38)
• Depression Much of the Time Question (Q39)
• Depression Most of the Time Question (Q40)

12
Missing Data Imputation

• Missing data mechanism Missing Completely At
  Random (MCAR), Missing At Random (MAR), Not
  Missing At Random (NMAR)
• Single imputation impute a single value for each
  missing element
• Multiple imputation each missing element is
  replaced by a vector of possible values (NORM)
• Iterative imputation estimate the values of
  missing elements by conducting systematic
  analysis until convergence
• Principal component analysis (PCA)
• Clustering (SC)

13
Guidelines for Selecting Imputation Methods
14
Summary

• Knowledge-based modeling is an expert centered
  data driven modeling approach that could
  facilitate evidence-based patient care
• Dimensionality reduction improves modeling
  efficiency and model effectiveness
• The ability to deal with missing data will allow
  the extraction of useful information from
  real-world messy data
• Technical approaches may need to be adjusted to
  meet the needs of healthcare applications