Transforming laboratory data to improve medical care for patients with chronic kidney disease.

Background: Reviewing laboratory data is a key method in dialysis research. However, data collection via a traditional prospective study is invariably costly in terms of money and time. Abstracting data from existing databases using multiple methods to identify underlying knowledge offers an alternative means of reducing the medical uncertainties.

Methods: Laboratory data covering the previous 1-year period before dialysis for 212 hemodialysis (HD) patients were adopted as the study data. The original 86 examination items (input variables), 1 assigned output distinguishing variable and 1,292 records were abstracted into a 13x931 data set (13 variables and 931 records). A transformed data set comprising a binary code of original values was then prepared. This work compared the data between 2 groups, 2-month-0 (records dated earlier than 2 months before HD started) and 2-month-1 (records dated within a period of 2 months). To avoid effects from missing values in the database, 4 analysis methods--namely t-test, ANOVA, Kruskal-Wallis H test and Spearman correlation--were applied to both the original scaling and the transformed data.

Results: Through appropriate data cleansing, 12 out of 86 examination items can be used as index factors to distinguish patient disease condition between 2-month-0 and 2-month-1. Of these, after data analysis and data format comparison, 4 items--serum potassium, serum calcium, red blood cell count and platelet count--rather than the traditional kidney function, were shown to be the most distinguishing variables. The transformed data provided more consistent results than the original scaling. However, original scaling yielded rich information and results that were consistent with current medical knowledge.

Conclusions: Reviewing different data formats via a multimethod approach can reduce uncertainties associated with real world databases, and thus achieve enhanced medical care.