Meta-Analysis of the Clinical Parameters of the Patients with Metabolic Syndrome and its Secondary Complications

Janusz Wojtusiak, Ryszard S. Michalski, Maggie Simanivanh, Zobair Younossi, Ancha Baranova

This is a collaborative project between

Machine Learning and Inference Laboratory, Department of Health Administration and Policy, George Mason University, Fairfax, VA

Molecular and Microbiology Department, College of Science,George Mason University, Fairfax, VA

Translational Reseach Institute, Inova Hospital, VA

Abstract: Natural induction generates hypotheses from data in the forms directly corresponding to equivalent natural language descriptions in order to make them easy to understand and interpret. For the latter reason it is particularly attractive for medical applications in which the interpretability of results of learning is essential. This study describes AQ21 algorithm implementing a form of the natural induction, and presents early results from its application to an important medical problem. Specifically, the problem concerns discovery of disease patterns in aggregated data describing patients with metabolic syndrome, associated with central obesity, insulin-resistance, hypertension, decreased level of high-density lipoprotein cholesterol (“good cholesterol”), and several other medical symptoms.. The program discovered patterns that are simple and appear to be clinically significant.

Methods: An important characteristic of the data used for this meta-analysis is its aggregated form. All the data fed into the AQ21 natural induction system were retrieved from peer-reviewed papers published in the leading medical journals. The format of the data reported in such journals excludes access to the individual measurements of the clinical parameters, as the individual's privacy is protected by HIPAA (Health Insurance Portability and Accountability Act) and similar protection policies. Therefore, we applied AQ21 natural induction program to aggregated data published in the medical journals, such as Hepatology , Obesity Research , and International Journal of Obesity . The data consist of summaries of 16 studies over the total of 12 groups of patients with present symptoms and 8 control groups of patients. A single study is described in terms of the mean of parameters measured over a group of patients. The total number of different parameters measured in these studies is 152. However, different parameters are measured in different studies. An output attribute defines one of three possible diseases, namely NAFLD (non-alcoholic fatty liver disease), SS (simple steatosis), and NASH (nonalcoholic steatohepatitis) and control groups for these diseases (used as a contrast set for learning).

Examples of rules generated :

[class=NAFLD,SS,NASH]
<= [bmi>=26.85: 8,2]
|_
[AST<=27.2]
[ADIPONECTIN>=7.25]
: p=0, n=6
: p=8,n=0,q=0.816,cx=25

[class=NAFLD,SS,NASH]
<= [HOMA>=2.27: 9,2]
|_
[fast_ins<=13.17]
[LEPTIN>=14.25]
[ADIPONECTIN>=7.25]
: p=0, n=5
: p=9,n=0,q=0.972,cx=35

[class=NAFLD,SS,NASH]
<= [ADIPONECTIN<=6.18: 8,1]
: p=8,nmin=0,nmax=1,q=0.695,cx=5

Conclusion: AQ21 is the newest implementation of AQ learning methodology that integrates several novel features which are either not present in other machine learning and data mining programs or present separately. The presented initial application of AQ21 to aggregated data representing different studies on the diseases associated with metabolic syndrome allowed to obtain clinically relevant conclusions, including one stating an importance of adiponectin, that might be added to the currently used panels of non-invasive diagnostic markers.

List of published studies used :

1) Pagano C, Soardo G, Esposito W, Fallo F, Basan L, Donnini D, Federspil G, Sechi LA, Vettor R. Plasma adiponectin is decreased in nonalcoholic fatty liver disease. Eur J Endocrinol. 2005 Jan;152(1):113-8.

2) Bugianesi E, Pagotto U, Manini R, Vanni E, Gastaldelli A, de Iasio R, Gentilcore E, Natale S, Cassader M, Rizzetto M, Pasquali R, Marchesini G. Plasma adiponectin in nonalcoholic fatty liver is related to hepatic insulin resistance and hepatic fat content, not to liver disease severity. J Clin Endocrinol Metab. 2005 Jun;90(6):3498-504.

3) Kim SG, Kim HY, Seo JA, Lee KW, Oh JH, Kim NH, Choi KM, Baik SH, Choi DS. Relationship between serum adiponectin concentration, pulse wave velocity and nonalcoholic fatty liver disease. Eur J Endocrinol. 2005 Feb;152(2):225-31.

4) Kaser S, Moschen A, Cayon A, Kaser A, Crespo J, Pons-Romero F, Ebenbichler CF, Patsch JR, Tilg H. Adiponectin and its receptors in non-alcoholic steatohepatitis. Gut. 2005 Jan;54(1):117-21.

5) Vuppalanchi R, Marri S, Kolwankar D, Considine RV, Chalasani N. Is adiponectin involved in the pathogenesis of nonalcoholic steatohepatitis? A preliminary human study. J Clin Gastroenterol. 2005 Mar;39(3):237-42.

6) Musso G, Gambino R, Durazzo M, Biroli G, Carello M, Faga E, Pacini G, De Michieli F, Rabbione L, Premoli A, Cassader M, Pagano G. Adipokines in NASH: postprandial lipid metabolism as a link between adiponectin and liver disease. Hepatology. 2005 Nov;42(5):1175-83.

7) Lewandowski K, Szosland K, O'Callaghan C, Tan B, Randeva H, Lewinski A. Adiponectin and resistin serum levels in women with polycystic ovary syndrome during oral glucose tolerance test: A significant reciprocal correlation between adiponectin and resistin independent of insulin resistance indices. Mol Genet Metab. 2005 May;85(1):61-9.

8) Qi L, Rifai N, Rimm E, Hu Frank, Liu S. Dietary glycemic index, glycemic load, cereal fiber, and plasma adiponectin concentration in diabetic men. Diabetes Care. 2005 May;28(5):1022-8.

9) Iwashima Y, Katsuya T, Ishikawa K, Kida I, Ohishi M, Horio T, Ouchi N, Ohashi K, Kihara S, Funahashi T, Rakugi H, Ogihara. Association of hypoadiponectinemia with smoking habit in men. Hypertension. 2005 Jun;45(6):1094-100.

10) Matsuda Y, Tanioka T, Yoshioka T, Nagano T, Hiroi T, Yoshikawa K, Okabe K, Nagamine I, Takasaka Y. Gender differences in association of plasma adiponectin with obesity reflect resultant insulin resistance in non-diabetic Japanese patients with schizophrenia. Psychiatry Clin Neurosci. 2005 Jun;59(3):266-73.

11) Farvid M, Ng T, Chan D, Barrett P, Watts G. Association of adiponectin and resistin with adipose tissue compartments, insulin resistance, and dyslipidaemia. Diabetes Obes Metab. 2005 Jul;7(4):406-13.

12) Furler S, Gan S, Poynten A, Chisholm D, Campbell L, and Kriketos A. Relationship of adiponectin with insulin sensitivity in humans, independent of lipid availability. Obesity ( Silver Spring ). 2006 Feb;14(2):228-34.

13) Maeda K, Ishihara K, Miyake K, Kaji Y, Kawamitsu H, Fujii M, Sugimura K, Ohara T. Inverse correlation between serum adiponectin concentration and hepatic lipid content in Japanese with type 2 diabetes. Metabolism. 2005 Jun;54(6):775-80.

14) Matsubara M. Plasma adiponectin decrease in women with nonalcoholic fatty liver. Endocr J. 2004 Dec;51(6):587-93.

15) Mendez-Sanchez N, Chavez-Tapia N, Villa A, Sanchez- Lara, Zamora-Valdes D, Ramos M, Uribe M. Adiponectin as a protective factor in hepatic steatosis. World J Gastroenterol. 2005 Mar 28;11(12):1737-41.

16) Pischon T, Girman C, Rifai N, Hotamisligil G, Rimm E. Association between dietary factors and plasma adiponectin concentrations in men. Am J Clin Nutr. 2005 Apr;81(4):780-6.