Dose-response Explorer: An Open-source-code Matlab-based tool for modeling treatment outcome as a function of predictive factors

1
Dose-response ExplorerAn Open-source-code
Matlab-based tool for modeling treatment outcome
as a function of predictive factors
Gita Suneja Issam El Naqa, Patricia
Lindsay, Andrew Hope, James Alaly, Jeffrey
Bradley, Joseph O. Deasy
Supported by NIH grant R01 CA 85181
2
What is DREX?

• An open-source-code Matlab-based tool for
• Modeling tumor control probability (TCP) and
  normal tissue complication probability (NTCP)
• Evaluating robustness of models
• Graphing the results for purposes of outcomes
  analysis for practitioners, training for
  residents, and hypothesis-testing for further
  research

3
Motivation Objectives

• Motivation
• Cornerstone of treatment planning is the need to
  balance tumor control probability (TCP) with
  normal tissue complication probability (NTCP)
• Objective
• Physicians and scientists need a tool that is
  straightforward and flexible in the study of
  treatment parameters and clinical factors

4
Features

1. Analytical modeling of normal tissue complication
   probability (NTCP) and tumor control probability
   (TCP)
2. Combination of multiple dose-volume variables and
   clinical variables using multi-term logistic
   regression modeling
3. Manual selection or automated estimation of model
   parameters
4. Estimation of uncertainty in model parameters
5. Performance assessment of univariate and
   multivariate analysis
6. Capacity to graphically visualize NTCP or TCP
   prediction vs. selected model variable(s)

5
Basic Modules
Data Input
1
Radiobiological model?
2
TCP
NTCP
Model type?
Model type?
3
Analytical
Analytical
Multi-metric
Lyman-Kutcher-Burman (LKB) or Critical volume
Poisson or Linear quadratic
Logistic regression
4
Univariate/multivariate performance assessment
Graphical representation
5
Export output
6
Modeling Method I Analytical

• NTCP
• Lyman-Kutcher-Burman (LKB) Model (Lyman 1985,
  Kutcher and Burman 1989)
• Critical Volume Model (Niemierko and Goitein
  1993)
• TCP
• Poisson Statistics
• Linear-quadratic (LQ) Prediction

7
Modeling Method II Multimetric

• Logistic regression additive sigmoid model
• Two types of data exploration
• Manual
• Automated
• Determining Model Order by Leave-one-out-Cross-Va
  lidation (Ref. Multi-Variable Modeling of
  Radiotherapy Outcomes Determining Optimal Model
  Size, Deasy et al., poster SU-FF-T-376 )
• Model parameters estimated by forward selection
  on multiple bootstrap samples

8
Performance Assessment

• Spearmans Rank Correlation
• Area under the Receiver Operating Characteristic
  (ROC) curve
• Survival analysis using the Kaplan-Meier
  estimator

9
Univariate Graphical Representations
Graph/Plot Description/Function
Self-correlation Color-washed Spearmans cross-correlation image of selected variables and observed outcome
Scatter User selects abscissa and ordinate variables Provides user with visual cues about the discrimination ability of certain factors
Survival curves Use Kaplan-Meier estimates
10
Multivariate Graphical Representations
Graph/Plot Description/Function
Histogram Cumulative plot of observed response (bar graph) and model-predicted response (line graph)
Contour Demonstrates the effect of the model variables on shaping the predicted outcome
Octile Patients are uniformly binned into 8 groups Helps visualized goodness of fit of model
ROC Assess prediction power of model
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
Conclusions

• User-friendly software tool to analyze dose
  response effects of radiation
• Incorporates treatment and clinical factors, as
  well as biophysical models
• Various graphical representations
• Available in the near future on the web at
• radium.wustl.edu/DREX