Table 2.

Table comparing different types of clinical data on some points important to clinical decision support systems.

Clinical decision support issues	Electronic health record free-text/unstructured data (eg, clinical notes)	Registry/trial data (eg, case record forms case record forms and questionnaires)	Structured data/electronic health record (eg, lab values and smoking status)
Context completeness	Excellent: contextual information can be included.	Poor: context is essentially absent as a priori interpretation is an integral part of recording data in case record forms.	Depends on implementation. Context may be lost because of predetermined categorization.
Machine readability	Poor: information is mostly useful for case-specific usage by humans. May require text mining/text retrieval to convert to a machine-readable format.	Good: data are uniformly formatted and can be parsed by computers.	Excellent: data can be parsed or directly used by computers.
Translatability (between institutions)	Poor: free text contains jargon-specific, ambiguous abbreviations (eg, PCI: percutaneous coronary intervention/prophylactic cranial irradiation).	Excellent: trial data are usually collected using a standardized protocol, allowing for interoperability between institutions.	Good: lab values can be converted using reference values. Structured data, such as smoking and hypertensive status, can be reformatted for interoperability.
Noise resistance	Very poor: These type of data are very sensitive to interobserver noise (eg, personal abbreviations, spelling mistakes, and personal focus in recording certain types of information).	Excellent: data are recorded in a standardized way, designed to prevent noise.	Good: data are often machine-derived or recorded in a standardized way. However, bias because of differences in information-recording habits among health care professionals may arise.
Availability for reuse/general applicability	Excellent: these type of data are readily available, contain a lot of context (see Context completeness), and can thus be repurposed for a variety of applications.	Limited: trials are designed and conducted for one specific research question.	Excellent: these type of data are readily available and can thus be used for a plethora of purposes.
Design flexibility	Excellent: study design can be revisited if unanticipated bias effects arise. In this sense, bias could be corrected by altering the data selection.	Poor: study design is hit-or-miss. Bias cannot be corrected after the data recording process.	Excellent: study design can be revisited if unanticipated bias effects arise. In this sense, bias could be corrected by altering the data selection.