Abstract:
Effective solution of medical diagnostics tasks requires the use of complex probabilistic
models which allow one to adequately describe real data and permit the use of analytical methods
of the supervised learning classification. Choosing a model of a mixture of normal distributions
solves the posed problems but leads to the curse of dimensionality. The transition to the model of a
mixture of probabilistic principal component analyzers allows one to formally set the task of
choosing its structural parameters. The solution is proposed to search by combining the application
of information criteria for the formation of initial approximations followed by refinement of the
resulting estimates. Using the example of experiments to diagnose liver diseases and
to predict the
chemical composition of urinary stones, the capabilities of the described data analysis procedures
are demonstrated. The proposed solutions give a source of improving the accuracy of classification,
impetus to experts in the subject area to clarify the essence of the processes.
Keywords:medical diagnostics, mixture of probabilistic principal component analyzers, model selection criterion, cross validation.
Fulltext:
PDF file (115 kB)