DPSPDNT - Dissertações de mestrado
Permanent URI for this collection
Browse
Browsing DPSPDNT - Dissertações de mestrado by advisor "Antunes, Marília"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Classification methods applied to familial hypercholesterolemia diagnosis in pediatric agePublication . Albuquerque, João David Ferreira de Castro; Antunes, Marília; Bourbon, MafaldaIntroduction: Familial Hypercholesterolemia (FH) is an inherited disorder of lipid metabolism, characterized by increased low density lipoprotein cholesterol (LDLc) levels. The resulting severe dyslipidemia leads to the early development of atherosclerosis, representing a major risk factor for cardiovascular disease (CVD). The early diagnosis of FH is associated with a significant reduction in CVD risk, supporting the introduction of precocious and more aggressive therapeutic measures. There are different clinical criteria available for the diagnosis of FH, although only genetic testing can confirm the diagnostic. Simon Broome (SB) criteria for FH diagnosis are among the most frequently used in clinical setting, and are based on family history, presence of physical signs, and LDLc and total cholesterol (TC) levels. When compared to genetic diagnosis results however, SB criteria present a high false positive rate, which constitutes a heavy burden in terms of healthcare costs, and limits the access to the genetic study of a larger universe of potential FH cases. Aim: The main purpose of this work was to develop alternative classification methods for FH diagnosis, based on different biochemical indicators, with improved ability to screen for FH cases in comparison to SB criteria. Two different models were developed for this purpose: a logistic regression (LR), and a decision tree (DT) model. Methods: Serum concentrations of TC, LDLc, high density lipoprotein cholesterol (HDLc), triglycerides (TG), apolipoproteins AI (apoAI) and B (apoB), and lipoprotein(a) (Lp(a)) were determined, and genetic diagnosis was performed, in a sample of 252 participants in the Portuguese FH Study, at pediatric age (2-17 years). All patients met the clinical criteria for dyslipidemia, and were not under hypolipidemic medication during the evaluation period. LR and DT models were fitted to sample data. For the LR model, two different cutoff points were defined, through receiver operating characteristics (ROC) curve analysis, following Yoden index and minimum p-value (min p) methods. The DT was built based on entropy reduction, or information gain measures. A modified version of the DT method was implemented, consisting in the sequential exclusion of predictor variables as they are introduced in the model. This allows producing a classification rule that uses single cutpoints for biomarkers, simplifying its interpretation. Different operating characteristics (OC) were estimated for all models: accuracy (Acc), sensitivity (Se), specificity (Spe), positive predictive value (PPV ) and negative predictive value (NPV ). These OC were calculated by generating a confusion matrix, considering molecular study results as the true state of the disease. The best performing LR and DT models were compared with SB biochemical criteria for FH diagnosis, through bootstrap resampling techniques. Median and mean values of the OC for 200 bootstrap samples were used for predictive performance comparison. Results: The logit function for the LR final model was expressed as g(π) = -7:083 + 0:086 X LDLc -0:041 X TG - 0:037X apoAI. The best performing DT model included the variables LDLc, TG, apoAI, apoB and HDLc, by descending order of importance. Between the different classification methods, Acc, Spe and PPV were higher in the DT model, followed by the LR model with the cut point value (c) defined by the min p method (c = 0:35). The lower values in these OC are found for SB criteria (p < 0:01). Higher Se and NPV on the other hand, are achieved by SB criteria, and the LR model with the cutpoint value calculated by Youden index (c = 0:17). However, the LR model using this cutpoint achieves significantly higher Acc, Spe and NPV than SB criteria (p < 0:01). Conclusions: Both LR and DT models seem to be a valid alternative to traditional clinical criteria for FH diagnosis. It seems possible to adjust the cutoff value in the LR model for similar Se levels as the ones observed in SB criteria, with significantly less false positive retention. To be validated by additional data, this would undoubtedly indicate this method as preferable between the two, and can have a very important impact in terms of cost-effectiveness. By avoiding the repetition of predictor variables, and providing single cutoff values for each biomarker, the modified DT model assumes a structure that typically resembles medical criteria, and can therefore be easily used in clinical practice. It seems that, in spite using different methodological approaches, both LR and DT models are able to divide the sample according to the most relevant biochemical characteristics for FH diagnosis. According to both classification methods, presence of FH is directly related to LDLc levels, and inversely related to TG and ApoAI concentrations, by this order of importance. The preferred classification model, as well as model specifications, may vary as a function of the OC that are considered more important, and context in which it is applied.