Technologies and Solutions for Trend Detection 
        in Public Literature for Biomarker Discovery

                      Bernd Wachmann
             Siemens Corporate Research,NJ, USA


Research for new biomarkers usually begins with a literature 
review to identify the mechanisms of action and to define a set 
of biomarkers that can jointly be used as a panel to characterize 
the type and stage of a disease. However, the manual search for 
biomarkers is an increasingly difficult task, since the number 
of publications is steadily increasing in volume and broadening 
in terms of complexity and diversity. The PubMed database of
publications in biomedical science lists more than 6 million 
articles from the last 10 years. Currently more than 600k 
publications are added to the knowledge base every year, making 
a manual search for information a time consuming task. Even for a 
single disease, like lung cancer, several thousand related 
publications are published every year (i.e., in 2007, more than 
300 per month on average for lung cancer).

To address this challenging task, we have developed a system that 
can identify structural and longitudinal patterns in the biomedical 
literature data that support the understanding of trends and 
relationships between diseases and biomarkers over time. We believe 
that the information of time is important, since it helps in 
tracking
    o when a biomarker has been discovered and how important it
      has become for the understanding of the disease over time,
    o if a biomarker has been “replaced” or complemented by another, 
      more informative biomarker,
    o at what time we can see an emerging biomarker that will become 
      relevant for a disease on a broader basis.

The solution addresses the above challenges and covers several 
functionalities that automate a large part of the previously manual 
search process and can therefore significantly reduce the time of 
searching for biomarkers and increase the quality of the results.

The system uses biomedical entity recognition and dictionaries of 
medical terms to emphasize the key concepts and their relations in a 
document but in principle all words and phrases are considered as 
possible trends. This way even very current biomarkers not recognized 
by the detection modules can be discovered. A hybrid text clustering 
and time series analysis engine detects topics evolving over time 
and ranks trends by their past and predicted future relevance to a 
certain disease.

To make the retrieved information accessible to the user in a 
convenient way, the system visualizes the trends of publications with 
respect to diseases and biomarkers. Narrowing the results down to 
specific elements allows the user to understand the relationships 
between them both statically and over time.