• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
Health-e-Child CaseReasoner
 

Health-e-Child CaseReasoner

on

  • 397 views

Presentation in Darmstadt for LWA 2009

Presentation in Darmstadt for LWA 2009

Statistics

Views

Total Views
397
Views on SlideShare
397
Embed Views
0

Actions

Likes
0
Downloads
4
Comments
0

0 Embeds 0

No embeds

Accessibility

Upload Details

Uploaded via as Microsoft PowerPoint

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

    Health-e-Child CaseReasoner Health-e-Child CaseReasoner Presentation Transcript

    • HeC CaseReasoner: Neighborhood Graph for Clinical Case Retrieval and Decision Support LWA 2009 Darmstadt, Germany 21 th – 23 th September 2009 Gabor Rendes Alexey Tsymbal Siemens AG, Erlangen, Germany
    • Overview
      • Health-e-Child CaseReasoner
      • Neighborhood Graphs
      • Clustering algorithms
      • Learning Discriminative Distance Functions
      • Framework and Developement
    •  
    • Motivation
        • Case-Based Reasoning (CBR) system -> reasoning by similarity
        • Challenges:
          • complex clinical data
          • lack of transparency and explanation
        • Very generic
          • basically works for any type of data (all you need is a distance function)‏
          • has been applied to gene expression, 3D anatomical meshes, clinical data
        • User defined or learned distances
    • Motivation
        • Task:
          • Display inter-patient proximity for the given context with the purpose of decision making and possibly knowledge discovery
        • Visualization using neighborhood graphs
          • Overview + explanation
        • Potential clinical use cases
          • Diagnosis: have similar patients been diagnosed as diseased?
          • Treatment selection: did similar patients profit from treatment (surgery, chemotherapy, etc.)
    • Neighborhood Graphs
        • Node-link entity-relationship representation
        • Three basic types
          • Relative Neighborhood Graphs (RNG): A and B are connected if: (no patient C is closer to both A and B than AB) [Toussaint, 1980]
          • In a nearest neighbor graphs, each case is connected with one or a set of its nearest neighbors
          • A threshold graph is simply defined as a graph where two vertices are connected with an edge if the distance between the two corresponding cases is less than a certain threshold
        • The graphs nicely represent patient clusterings, are adaptive to the distance function
        • Node-link entity-relationship representation
        • Three basic types
          • Relative Neighborhood Graphs (RNG): A and B are connected if: (no patient C is closer to both A and B than AB) [Toussaint, 1980]
          • In a nearest neighbor graphs, each case is connected with one or a set of its nearest neighbors
          • A threshold graph is simply defined as a graph where two vertices are connected with an edge if the distance between the two corresponding cases is less than a certain threshold
        • The graphs nicely represent patient clusterings, are adaptive to the distance function
    • Neighborhood Graphs functionality
        • Node colouring to represent numeric and nominal attributes
        • Node filtering according to values of nominal and numeric attributes in patient record
        • Edge colouring and filtering according to the underlying distance
        • Patient clustering (Newman's on the graph and “semantic” in the original space)‏
        • Mesh visualization and transformation (within node, e.g. meshes corresponding to the pulmonary trunk)‏
        • Reconfigurable tooltips displaying patient records and images
        • Nearest Neighborhood classification and regression performance visualization for each node, for a selected class attribute and a certain similarity context
    • GUI for NGraphs: plain nodes Toolbar Statusbar Graph panel
    • GUI for NGraphs: mesh visualization Besides clinical data and patient similarities, graphs are nicely suitable for displaying meshes corresponding to patients. Same operations can be used as for usual graphs; also meshes can be scaled and rotated
    • Clustering
        • Two algorithms applied:
          • Girvan-Newman's algorithm for graphs
          • “ Semantic clustering” for general data
        • Both algorithms produce a hierarchical, top-down clustering tree
        • Intuitive, interactive navigation among the clusters
        • Newman's clustering is clustering on the graph
          • Starting from the fill weighted graph, it iteratively removes edges according to the edge betweenness criterion (it removes “key links”), so that the graph is split into disconnected clusters
          • Popular in the social networks community
    • Semantic Clustering
        • Goal: provide a top-down hierarchical clustering with semantic splits (rules like “age>10”)
        • Intra-cluster variances (square of the distance from centroid) re considered before and after the split
        • Almost “every” possible split is considered for every existing cluster
        • At every step the cluster with the currently highest “gain” (difference) will be selected for a real split
    • GUI for Clustering
    • Distance Functions
        • Underlying distance functions for similarity
        • Two option is considered
          • Canonical distance: Euclidean metric, weighted
          • Learn a strong distance function for a given classification context, two techniques:
            • Learning from equivalence constraints in the product or difference space
            • The intrinsic Random Forest (RF) distance
    • Learning Discriminative Functions
        • Learning discriminative distance functions
          • Helps to combine the power of strong learners as AdaBoost, RF, SVM, with the transparency of case retrieval and nearest neighbor classification
          • Is especially useful for data sets with many irrelevant, weakly relevant and correlated features (e.g. images or clinical data)‏
    • Learning Discriminative Functions: two techniques
        • Learning from so called weak representation, equivalence constraints
          • Pairs of cases are considered, and the label displays whether they belong to the same class
          • Original space is transformed into a product or difference space for equivalence constraints
          • Any technique can be used to learn in the new space
          • Originates from imaging, no known works in other domains so far
        • Using intrinsic Random Forest distance
          • Two cases are more similar to each other if they fall into more of same leaves in a forest
          • Breiman predicted the power of this distance in his works, however it is still used not so often, mostly for clustering genetic data
    • Distance learning from equivalence constraints
        • Usually are represented using triplets (x 1 ,x 2 ,y), where x 1 and x 2 are points in the original space and y ∈ {+1, -1}
        • Originates and finds most applications so far from imaging
        • Was shown to work well for multidimensional data with many irrelevant and redundant features (as in imaging)
        • The following function is learnt directly which was shown to be optimal under iid [Mahamud and Hebert, 2003] :
    • Intrinsic RF distance
        • Is rather a “black horse”; has never been compared with learning from equivalence constraints so far
        • Not many applications yet, the most popular is clustering genetic data
          • See e.g. [Shi and Horvath, 2006]
        • But: usually good performance is reported
        • RF similarity between two cases x 1 and x 2: where K is number of trees and z ij is terminal position of case x i in tree j
    • Learning distance: experimental methodology
        • Two approaches to learning discriminative distance; learning from equivalence constraints and intrinsic RF distance compared
        • Equivalence constraints are learnt using AdaBoost and RF in product and difference spaces
        • Compared with simple learners; k-NN, AdaBoost and RF
        • 9 benchmark data sets; 1 SCR mesh data, 3 UCI clinical data sets, 4 gene expression data sets, 1 mass spectrometry data set
        • LOO for genetic and mesh data, 70/30% CV for the rest
    • CaseReasoner: the application
        • The basic philosophy: provide clinicians with a flexible and interactive tool
        • Explore and compare the patients' records
        • Various visualization techniques
          • User define similarity context
          • View basic statistics of the retrieved cases
          • Visualize them utilizing neighborhood graphs, treemaps and heatmaps
    • CaseReasoner: the workflow
    • CaseReasoner: the framework
        • Using a less strict variation of the Presentation-Abstraction-Control (PAC) pattern
        • Hierarchical structure of 'agents'
        • The core and the modules communicate with each other only through their Controller part
        • Flexible, easy module adding/removal, common libraries, simple workflow, clear APIs
    • Conclusion
        • Ngraphs are an alternative way to represent clinical data and interpatient similarity for knowledge discovery and decision support
        • Learning discriminative distances is a way to combine the power of string black box learners with the transparency of case retrieval and nearest neighbor classification
        • Decision support becomes transparent
        • Flexible architecture makes life easier
    • Ongoing / Future work
        • Comparison with other distance learning techinques
          • RCA (Relevant Component Analysis), etc.
        • One-class learning from positive or negative constraints only
        • Incrementalization of learning from equivalence constraints
          • Existing incrementalizations of boosting are far from being lossless and converge very slowly
          • Use incrementalization of RF instead? There exist lossless incrementalizations both for bagging and decision trees
        • Support of comparative constraints
          • Triplet (a,b,c): a is closer to b than to c
        • Application to other domains
          • Prediction of suitability to PPVI (mesh-based), VHD treatment
          • Brain tumour classification on microarray dara
    • Thank you! Any questions?