Database Reference
In-Depth Information
delivered to the patient. Standard approaches are not well-equipped
to cope with this bias in the data, especially as it is hard to quantify
precisely. Furthermore, most studies in medical informatics are retro-
spective. Well-done prospective studies are hard to do, and are often
done on small populations, limiting the statistical significance of any
derived results.
2.2.5 Evaluation and Interpretation Challenges.
Data
mining results consist of models and predictions that need to be inter-
preted by domain experts. Many modeling techniques produce models
that are not easily interpretable. For example, the weights of a neural
network may be dicult to grasp for a domain expert. But for such a
model to be adopted for clinical use, it needs to be validated with existing
medical knowledge. It becomes imperative to track provenance meta-
data describing the process used to derive any results from data mining
to help domain expert interpret these results. Furthermore, the prove-
nance of the data sets, and analysis decisions used during the modeling
are also required by the experts to evaluate the validity of the results.
This imposes several additional requirements on the selected models and
analysis.
2.2.6 Generic Systems Challenges.
Beyond analytical chal-
lenges, sensor data mining also comes with a set of systems challenges
- that apply to medical informatics applications. The mining of sen-
sor data typically requires more than conventional data management
(database or data warehousing) technologies for the following reasons:
The temporal aspect of the data produced by sensors sometimes
generate large amounts of data that can overwhelm a relational
database system. For example, a large population monitoring so-
lution requiring the real-time analysis of physiological readings,
activity sensor readings and social media interactions, cannot be
supported with relational database technologies alone.
Sensor mining applications often have real-time requirements. A
conventional store and analyze paradigm with the use of relational
database technologies may not be appropriate for such time sensi-
tive applications.
The unstructured nature of some of the data produced by sensors
coupled with the real-time requirements imposes requirements on
the programming and analysis models used by developers of sensor
data mining applications.
