Biomathematics Oriented Machine Learning System for Reconstructing Temporal Profiles of Biological or Clinical Markers

Abstract

Time series reconstruction algorithms are widely used to create temporal profiles from data series. However, in many clinical fields, e.g., viral kinetics, the data is noisy and sparse, making it difficult to use standard algorithms. We developed PROFILASE, which combines advanced multi-objective genetic algorithm search with machine learning architecture to harvest experts' decision-making considerations. Furthermore, PROFILASE implements additional scoring considerations, more biological in nature, thus further exploits domain expertise. We tested our system against a standard bottom-up algorithm by reconstruction of time series sparsely sampled with noise from simulated profiles. PROFILASE obtained RMS distance 2.5 fold lower (P<0.0001) than the standard algorithm, 93% correct identification rate of segment number and 88% correct profile classification rate (versus 68%). The additional considerations were found to have a significant effect on the success of reconstruction. Finally, PROFILASE was generalized to evaluate additional considerations from different fields, thus allowing better understanding of other diseases