Sensitivity - Diagnosis Example

If you want to determine which test is going to provide the best information about the presence of a fault or disease, select the node for the fault or disease and choose Sensitivity to Findings.  Use the summary list of sensitivities at the end of the report generated to identify possible varying nodes for which a finding will provide the most information about the query node.  If you want more detailed information of how these varying nodes can effect the query node, look up each of them in the first part of the report.

If you want a single number which best describes the degree of sensitivity of one node to another, it is recommended that you use the numbers provided in the first column of the summary report at the end.  For continuous nodes, or nodes with state numbers defined, this will be the variance reduction, otherwise it will be the mutual information (i.e. entropy reduction).

When you do a complete sensitivity report (i.e. only the query node selected), then in the report Netica also shows the sensitivity of the query node to a finding at the query node itself.  Of course, the minimum and maximum beliefs for each state will be 0 and 1 respectively, and the maximum reductions in variance and entropy will be 100%.  This node is included in the report for completeness, and to quickly see what the maximum of each sensitivity value is (for example, what the full variance and entropy is).

When the sensitivities are calculated, all findings currently entered into the network will be taken into account, which can effect the sensitivities significantly.

If you are trying to find the next best observation to make a diagnosis, you will probably want to combine the cost of each possible observation with its expected value as indicated by the sensitivity to that observation (finding).

If you want to consider pairs of observations, or multiple observations, the results can be quite different than if you consider observations one at a time.  To do a pair of observations, you must enter each possible finding of the first observation, and do a sensitivity analysis on the second observation, then average the results (weighted by the probability of the finding for the first observation) to find an expected value.

You will probably need Netica API to automate this. The sensitivity measures available from Netica API, as of version 3.10, are mutual information (i.e. entropy reduction), and RMS change of real (i.e. variance reduction of real).