Biomarker Identification of Complex Diseases/Disorders: Methodological Parallels to Parameter Estimation

Abstract

Biomarkers offer significant potential for diagnosis and treatment of complex disorders such as asthma, epilepsy, autism, Parkinson’s, and Alzheimer’s, as well as many others. In many cases, however, there is little consensus on what an appropriate biomarker would be. Consequently, biomarker identification is an important area of research for which a link between physiological measurements and the presence/absence or severity of a disorder can be established. This is nontrivial due to both the curse of dimensionality and because the number of measurements per trial often exceeds the number of trial participants. Overfitting of potential biomarkers is thus a significant problem that needs to be addressed. This paper highlights similarities between the biomarker identification problem and the parameter estimation problem, more specifically the regularization used for avoiding overfitting. Parallels between the underlying methodologies are pointed out and opportunities for advancing the systems’ concepts are discussed. Finally, a candidate biomarker for diagnosis of autism spectrum disorder is identified from a data set comprising metabolic measurements from four separate clinical trials to illustrate the procedure outlined in this work.

Application of Systems Engineering Principles and Techniques in Biological Big Data Analytics: A Review

“…In another example, a simpler but more robust version of the metabolic network models was chosen to address the limited number of points or measurements [72]. Grivas et al [73] also noted the importance of reducing model complexity by selecting models with fewer parameters, especially in clinical studies where participant recruitment is an obstacle or when data collection has already concluded. An extension to this idea is to determine or estimate some model parameters using literature data.…”

“…As a result, regularization has also been commonly used in deep learning to address overfitting where the objective function to be minimized often includes a loss function that penalizes the prediction error, and a regularization function that penalizes model complexity. These applications can be found in many references [42,65,73,[79][80][81]. A more detailed discussion on various regularization techniques employed in the literature can be found in [82].…”

“…In these cases, biomarker measurements taken at specific time points, such as "before" and "after," are commonly used to evaluate the effects of treatment. As a result, models used as biomarkers are generally time-invariant and depict a steady-state system [73].…”