Modeling Temporal Biomedical Data by SRL

Sriraam Natarajan, Irene Ong, David Haight, David Page and Vítor Santos Costa

September 2008

---

Abstract

Many biomedical applications involve temporal data, for example time-series gene expression experiments or longitudinal clinical data. From the statistical modeling side of SRL, dynamic Bayesian networks are a natural fit for such data, but they normally cannot incorporate additional types of relational information, such as the interaction between genes. In this paper, we examine the construction of logical DBNs from rules, either learned by relational learning methods or human-provided; in some cases, this construction requires combining rules. In some cases using such DBNs requires improved inference algorithms; we propose efficient inference within such DBNs by a Rao-Blackwellized particle filter. This work is motivated by two very different biomedical applications. One is incorporating the rich available background knowledge about genes and proteins when modeling time-series gene expression data. The other is modeling longitudinal nursing home data to estimate health status and health trajectory of individual patients and of groups of patients at specific facilities. This paper focuses on general principles rather than specific representations; we believe the lessons here are applicable for modeling time-series data within SRL.

Bibtex

@InProceedings{natarajan-strebio08,
  author =    {S. Natarajan and I. Ong and D. Haight and D. Page and V. Santos Costa},
  title =     {{Modeling Temporal Biomedical Data by SRL}},
  booktitle = {Proceedings of the Workshop on Statistical and Relational Learning in 
               Bioinformatics (StReBio 2008)},
  editor =    {J. Ramon and F. Costa and C. Florêncio and J. Kok},
  month =     {September},
  year =      {2008},
  address =   {Antwerp, Belgium},
}

Download Paper

PDF file