Skip to main content
ARS Home » Plains Area » Grand Forks, North Dakota » Grand Forks Human Nutrition Research Center » Dietary Prevention of Obesity-related Disease Research » Research » Publications at this Location » Publication #71559
Title: MULTIPLE LINEAR REGRESSION AND FINITE MIXTURE DISTRIBUTION MODELLING FOR SEQUENTIAL ANALYSIS OF HEMATOLOGICAL DATA

Author
item MCLAREN, C - MOORHEAD STATE UNIV
item KAMBOUR, E - UNIV QUEENSLAND
item MCLACHLAN, G - CASE WESTERN UNIV
item Lukaski, Henry
item LI, X - UNIV QUEENLSNAD
item BRITTENHAM, G - CASE WESTERN UNIV
item MCLAREN, GORDON - VETERANS AFFAIRS

Submitted to: Book Chapter
Publication Type: Book / Chapter
Publication Acceptance Date: 9/25/1996
Publication Date: N/A
Citation: N/A

Interpretive Summary: Assessment of an individual's iron nutritional status usually involves biochemical analyses of the patient's blood. The values of the biochemical tests then are compared with ranges of values derived from people with adequate iron status. Because a deviation within the range of normal values may represent a potentially important change in the individual's iron status, a new approach was developed to identify when variation in blood measurements is an important predictor of future risk of developing iron deficiency anemia. Statistical methods were developed to derive patient-specific reference values that account for normal physiologic variation and fluctuations in laboratory methods. Application of these methods to patients undergoing controlled induction of and iron supplementation for correction of induced anemia indicate the validity of this new model. These methods will make possible an automated examination of laboratory data with rapid and reliable identification of individuals whose laboratory measurements have changed from past values and indicate early pre-clinical onset of iron deficiency anemia. This method will be useful for use by medical professionals to provide more rigorous monitoring of iron nutritional status of patients than is currently available.

Technical Abstract: In practice, physicians mentally compare a laboratory result with previous values and use their clinical judgement to determine the significance of any change. We evaluated statistical methods to detect significant changes in patient-specific sequential measurement of standard hematologic laboratory tests. Multiple regression modeling was used with construction of (1 - alpha) 100% confidence regions for changes in mean values over time. A simultaneous significance test for these changes was derived. For comparison of patient-specific distributions of laboratory measurements, sequential analyses of double-truncated log-normal mixtures were used. The expectation-maximization (EM) algorithm was used for parameter estimation. A resampling approach was applied to test for a mixture of distributions using the likelihood ratio statistic. To avoid problems with convergence of the EM algorithm due to poor choice of starting values or multiple roots of the likelihood equation, a systematic procedure permitting selection of multiple starting values for the parameters was devised. We studied 11 healthy humans who were depleted or iron by serial phlebotomy to iron-deficiency anemia, then treated with oral iron supplements to replete iron stores and correct the anemia. Application of sequential patient-specific analyses of hemoglobin, hematocrit, mean cell volume showed that significant departures from past values could be identified even when values were within population reference ranges. Sequential alterations in red blood cell volume distributions during development of iron-deficiency anemia could be characterized and quantified. These methods provide a valuable tool for diagnostic evaluation of developing anemia and serial monitoring of response to therapy.