Vol: 48 Page: 39

Anna Bartkowiak
Jerzy Liebhart

Random forests supplemented by parallel fences plots: a study in spirometric diagnostics

Language: English

medical diagnosis
random forests
parallel fences plot

We try to assess, what truly can be obtained when using only the spirometric measurements for the diagnosis of pulmonary disorders. Using spirometric data recorded for 202 patients and the Random Forests methodology, we obtained an algorithm permitting for a 100% correct classification of the learning sample, and for a 94% correct classification of the 125 'not-healthy' patients included into the validation sample.
We propose, as a routine continuation, to verify the results - especially those for misclassified patients - by an independent multivariate graphical procedure called by us the parallel fences plot. By inspecting the patient's data vectors displayed in parallel fences plots it becomes evident if they are typical or not-typical for the 'healthy' or 'not-healthy' state of the patient.
The conclusion is, that the spirogram can be used for a relatively safe diagnosis of majority of patients. Only those that do not pass the parallel fences plot test should undergo further pulmonary scrutinizing test.