62. digital high resolution melt for first pass screening for sepsis
Name: Mridu Bhashini Sinha
Grad Year: 2017
Early and accurate diagnosis of bacterial infections remains a challenge in clinical neonatology. Infectious disease is the 3rd leading cause of neonatal death worldwide, affecting 22% of newborns each year. Identification of pathogenic bacteria in clinical samples still relies primarily on growth-based detection and phenotypic identification. These methods have low sensitivity and specificity and can take 48-120 hours to finalize results, thus abetting untimely diagnosis and empiric antibiotic use. Recently, digital High Resolution Melt dHRM has demonstrated detection of bacteria associated with clinically relevant polymicrobial blood infections. The dHRM platform generates a curve of change in fluorescence versus temperature by sequentially melting the sample at controlled temperatures and measuring the sample?s fluorescence. Previous work has used support vector machines to classify bacteria from a pre-defined library of melt curve signatures. However, classification with respect to a pre-defined library may not be enough for clinical decision support. Statistical information is more relevant to clinical decision makers over a hard yes-no decision. There is also a need to identify new organisms that are not a part of our library. To address these challenges, we aim to develop a soft-decision supervised machine learning framework that provides conditional probabilities of each class given the data we acquired. This approach also enables anomaly detection for new microbes and improperly prepared samples. The use of this Bayesian framework will allow for incorporation of informed priors based on the clinical information in the prediction model. The melt curves for samples in general behave continuously as a function of temperature. Specifically, we model our measured melt curve as a weighted sum of Legendre polynomials of temperature plus additive independent Gaussian noise. In our training dataset, for each class, we estimate the unknown weights and noise variance using standard model fitting procedures. We perform model selection to identify the optimal number of weights using the Akaike information criterion. We use a Bayesian framework for the test dataset, so that we can exploit prior information from the clinical phenotype about which class the sample may belong to. The likelihood and prior probability associated with each class is then used to calculate the posterior. Future work concerns information theoretic techniques may be used to estimate upper bounds on the number of classes of bacteria that can be discerned; these techniques will also provide insight into the size of train set for which melt curves can be classified with acceptable confidence.
Industry Application Area(s)
Life Sciences/Medical Devices & Instruments