Perspectives of Deep Learning for Neonatal Sepsis Detection
Time: Fri 2023-08-25 13.00
Video link: https://kth-se.zoom.us/j/63793229166
Doctoral student: Antoine Honoré , Teknisk informationsvetenskap
Opponent: Professor Guy Carrault, Université de Rennes 1
Supervisor: Saikat Chatterjee, Teknisk informationsvetenskap; Professor Eric Herlenius, Karolinska Institutet
Newborns, whether born at term or preterm, are highly vulnerable and face life-threatening situations during their initial weeks of life every year. Even with hospitalization in a neonatal intensive care unit (NICU) and careful clinical monitoring, identifying infection-related incidents like sepsis is a challenging task. Advances in patient monitoring systems has enabled the collection of a large amount of data at the patient bedside. However, the databases are often distributed in independent sub-systems, which makes it difficult to present a complete patient history to the medical practitioners. The medical data remain difficult to interpret for medical practitioners, due to the quantity of data and their complex structure. To enable reliable predictions from large datasets arising in clinical contexts, leveraging the flexibility of deep learning models was shown to be a promising approach.
In this thesis, we evaluate existing, and we introduce prediction algorithms for neonatal sepsis detection (NSD). We study the performances of the binary classification algorithms in retrospective cohort studies, using data obtained from local hospitals in the Stockholm region. The first part of the thesis chiefly focuses on introducing both clinical and technical background. The clinical background introduces the interplay between prematurity, sepsis and cardio-respiratory interactions, which constitutes a large part of the challenges faced for NSD.The technical background introduces the theoretical tools that are used to train and evaluate binary classification algorithms. The second part of the thesis presents algorithms evaluated on NSD tasks with data obtained from NICUs in the Stockholm region.
We first study the state-of-the-art Naive Bayes (NB) algorithm and its applicability to NSD on our in-house dataset. We study NB in setups with varying population and the vital signs-based features. Using a set of 24 features with very different distributions, and even though NB makes very strong assumptions on data distribution, it can reach up to 0.82 median area-under-receiver-operating-curve (AUROC) on our NSD tasks. Further, we show that normalizing flows, which are artificial neural network based distributions, can model the data distribution more accurately, and reach a slightly better median AUROC of 0.87.NB and NFs-based classifiers make the simplifying assumption that consecutive frames of features are uncorrelated. To account for the correlation in consecutive feature samples, we evaluated the performances of a neural network-based recurrent models (RNNs). We show that these models are useful for NSD on very low birth weight (VLBW) infants, as they help improve the specificity of binary classification from 0.77 for logistic regression to 0.97 for a long-short term memory unit based RNN. We however also show that these models suffer from a degradation of their recall performances, which results in similar AUROC as NB.
Motivated by the need to develop and evaluate more accurate algorithms for NSD, we study the performances of the celebrated hidden Markov models (HMMs) and propose some extensions. We first discuss the results of HMMs coupled with normalizing flows (FlowHMM) for modeling the emission distributions. For the FlowHMM training, we used the expectation-maximization (EM) algorithm for training, and proposed an additional discriminative training step. Overall, classical HMMs with Gaussian Mixtures for the emission distributions outperformed the FlowHMM, with or without the additional discriminative learning step. Finally, we propose a factorial HMM model coupled with conditional normalizing flows and trained with variational inference. The model is capable of inferring independent latent processes from observed data. The model is evaluated on synthetic data, which allows to empirically demonstrate the abilities of the model to identify the latent factorial Markov model. Lastly, we evaluate the model on the benchmark task of recovering discrete latent processes from a public household power consumption dataset. This paves way for further evaluation of classification models on physiological dataset in more complex settings. In particular, factorial HMMs enable the joint detection of several both related and un-related diseases from physiological data. This thus motivates the collection of data from larger cohort of patients, in order to obtain a sufficient number of samples on a variety of conditions for premature infants.