The Classification of Risk Group for Liver Cirrhosis Using SPAN
유영애 ; 송기준
Journal of the Korea Society of Health Informatics and Statistics (보건정보통계학회지), Vol.40(2) : 13~21, 2015
Journal of the Korea Society of Health Informatics and Statistics (보건정보통계학회지)
Objectives: The statistical predictive methods have been used to find the risk factors related with diseases and to generate predictive probabilities of those diseases. Logistic regression is the most commonly used method
for predicting the probability of diseases in the medical fields. Also, data-driven methods, such as CART have been used to identify subjects at increased risk of diseases. However, both of regression and tree models
have their specific limitations in spite of their advantages. Recently, an alternative approach called by search partition analysis (SPAN) is suggested, which is based on direct non-hierarchical search algorithm to identify
subgroups at risk. SPAN searches subgroups among different Boolean combinations of risk factors.
Methods: SPAN was compared against the performance of the other 3 methods; logistic regression, polychotomous regression and quick unbiased efficient statistical trees. We applied these methods to the
real clinical data composed of 4,093 individuals who received the screening test in first and then visited Yonsei University Medical Center for check-up liver cirrhosis between May 1994 and September 2005.
The performance of SPAN and that of any other methods were compared and the measures of performance were sensitivity, specificity, and accuracy.
Results: In the results using SPAN, the findings identified by the risk factors for liver cirrhosis were HbsAg, AntiHCV, Family history, platelet and α-FP. And we found that the sensitivity using SPAN
were much higher than those of other methods in various data sets.
Conclusions: In conclusion, as long as it works, the performance of SPAN should make sense in the context of medical diagnosis and prognosis. Also, It was known that SPAN had an advantage that its decision rules are
usually more interpretable than those of other methods.