MODELING SHOULDER STRENGTH USING A SUPPORT VECTOR MACHINE

1
MODELING SHOULDER STRENGTH USING A SUPPORT VECTOR
MACHINE

• Richard E. Hughes, Ph.D.
• Orthopaedic Surgery

2
OBJECTIVES

• To develop a metric of shoulder function that
  represents a continuum from pathologic to healthy
• To develop a strength-based test for rotator cuff
  tears that is superior to MRI or ultrasound

3
ISOMETRIC ER STRENGTH
4
NORMATIVE SHOULDER STRENGTH

• Healthy volunteers (n120)
• Age 20-78 years
• Dominant and non-dominant sides
• Isometric strength
• Abduction/adduction
• Internal/external rotation
• Flexion/extension

Hughes, R.E. et al. (1999) AJSM 27651-657
5
ROTATOR CUFF TEAR (RCT) PATIENT STRENGTH

• Same protocol as normative data
• Full-thickness RCT
• Measurements
• Pre-op
• 6 months post-op
• 12 months post-op
• Intraoperative tear size measurement
• n37

6
METHODS

• Use isometric shoulder strength measurements for
  asymptomatic shoulders and symptomatic cuff tear
  shoulders
• Model data using a least-squares support vector
  machine (SVM)

7
2D EXAMPLE
x(x1, x2)
x2
EXTERNAL ROTATION STRENGTH (Nm) _at_ NEUTRAL
x1
ABDUCTION STRENGTH (Nm) _at_ 0o ABDUCTION
8
STRENGTH DATA

• Abduction _at_ 30o, 60o, 90o
• Adduction_at_ 30o, 60o, 90o
• External rotation
• 30o IR and 15o abduction
• 0o IR and 90o abduction
• 0o IR and 15o abduction
• 30o ER and 90o abduction
• Internal rotation
• 0o IR and 15o abduction
• 30o ER and 90o abduction
• 30o ER and 15o abduction
• 60o ER and 90o abduction

9
STRENGTH DATA
x1, x2, x3

• Abduction _at_ 30o, 60o, 90o
• Adduction _at_ 30o, 60o, 90o
• External rotation
• 30o IR and 15o abduction
• 0o IR and 90o abduction
• 0o IR and 15o abduction
• 30o ER and 90o abduction
• Internal rotation
• 0o IR and 15o abduction
• 30o ER and 90o abduction
• 30o ER and 15o abduction
• 60o ER and 90o abduction

x4, x5, x6
x7
x8
x9
x(x1 , , x14)
x10
x11
x12
x13
x14
10
SVM MODEL
(no tear data points)
(tear data points)
11
SVM ADVANTAGES

• Can model highly nonlinear relationships in high
  dimensional spaces
• More intuitive than competing machine learning
  methods (i.e. artificial neural networks)
• Very computationally efficient
• Can rigorously represent expert knowledge in
  model formulation

12
SVM APPLICATIONS IN MEDICINE

• Breast tumor identification from ultrasound
  images
• Microarray gene expression classification
• Nosocomial infection detection
• Bioinformatics
• EEG analysis

13
SVM MODELING STEPS

• Identify and prepare data set
• Predict healthy shoulder strength from regression
  models (gender, age, body mass)
• Train SVM
• Test (evaluate) SVM performance

14
ROC CURVE
15
ROC RESULTS
16
SHOULDER METRIC
(no tear data points)
Distance from hyperplane
(tear data points)
17
RESULTS
18
DISCUSSION

• Developed a simple unifying metric for healthy
  shoulder strength based on healthy-pathologic
  continuum
• Exceeded some but not all US studies of detecting
  cuff tears
• Did not exceed diagnostic ability of MRI to
  detect cuff tears

19
LIMITATIONS

• Asymptomatic people assumed to represent intact
  rotator cuff case
• Data based on Mayo Clinic data normative data
  from rural Minnesota volunteers
• Did not use separate training and testing data
  sets for evaluation

20
ACKNOWLEDGEMENTS

• Aaron Silver
• Matt Lungren
• Oleg Svintsitski
• Shawn ODriscoll, M.D., Ph.D.
• Mike Rock, M.D.
• Kai-Nan An, Ph.D.
• Marj Johnson, P.T.