Elaine Zaunseder

The development and continuous optimization of newborn screening (NBS) programs remains an important and challenging task due to the low prevalence of screened diseases and high sensitivity requirements for screening methods. Recently, different machine learning (ML) methods have been applied to support NBS. However, most studies only focus on single diseases or specific ML techniques making it difficult to draw conclusions on which methods are best to implement. Therefore, we performed a systematic literature review of peer-reviewed publications on ML-based NBS methods. Overall, 125 related papers, published in the past two decades, were collected for the study, and 17 met the inclusion criteria. We analyzed the opportunities and challenges of ML methods for NBS including data preprocessing, classification models and pattern recognition methods based on their underlying approaches, data requirements, interpretability on a modular level, and performance. In general, ML methods have the potential to reduce the false positive rate and identify so far unknown metabolic patterns within NBS data. Our analysis revealed, that, among the presented, logistic regression analysis and support vector machines seem to be valuable candidates for NBS. However, due to the variety of diseases and methods, a general recommendation for a single method in NBS is not possible. Instead, these methods should be further investigated and compared to other approaches in comprehensive studies as they show promising results in NBS applications.

Comprehensive whole-body models (WBMs) accounting for organ-specific dynamics have been developed to simulate adult metabolism, but such models do not exist for infants. Here, we present a resource of 360 organ-resolved, sex-specific models of newborn and infant metabolism (infant-WBMs) spanning the first 180 days of life. These infant-WBMs were parameterized to represent the distinct metabolic characteristics of newborns and infants, including nutrition, energy requirements, and thermoregulation. We demonstrate that the predicted infant growth was consistent with the recommendation by the World Health Organization. We assessed the infant-WBMs' reliability and capabilities for personalization by simulating 10,000 newborns based on their blood metabolome and birth weight. Furthermore, the infant-WBMs accurately predicted changes in known biomarkers over time and metabolic responses to treatment strategies for inherited metabolic diseases. The infant-WBM resource holds promise for personalized medicine, as the infant-WBMs could be a first step to digital metabolic twins for newborn and infant metabolism.

Isovaleric aciduria (IVA) is a rare disorder of leucine metabolism and part of newborn screening (NBS) programs worldwide. However, NBS for IVA is hampered by, first, the increased birth prevalence due to the identification of individuals with an attenuated disease variant (so-called "mild" IVA) and, second, an increasing number of false positive screening results due to the use of pivmecillinam contained in the medication. Recently, machine learning (ML) methods have been analyzed, analogous to new biomarkers or second-tier methods, in the context of NBS. In this study, we investigated the application of machine learning classification methods to improve IVA classification using an NBS data set containing 2,106,090 newborns screened in Heidelberg, Germany. Therefore, we propose to combine two methods, linear discriminant analysis, and ridge logistic regression as an additional step, a digital-tier, to traditional NBS. Our results show that this reduces the false positive rate by 69.9% from 103 to 31 while maintaining 100% sensitivity in cross-validation. The ML methods were able to classify mild and classic IVA from normal newborns solely based on the NBS data and revealed that besides isovalerylcarnitine (C5), the metabolite concentration of tryptophan (Trp) is important for improved classification. Overall, applying ML methods to improve the specificity of IVA could have a major impact on newborns, as it could reduce the newborns' and families' burden of false positives or over-treatment.

Abstract. This paper proposes a novel approach to facilitate air quality aware decision making and to support planning actors to take effective measures for improving the air quality in cities and regions. Despite many improvements over the past decades, air pollutants such as particulate matter (PM), nitrogen dioxide (NO2) and ground-level ozone (O3) pose still one of the major risks to human health and the environment. Based on both a general analysis of the air quality situation and regulations in the EU and Germany as well as an in-depth analysis of local management practices requirements for better decision making are identified. The requirements are used to outline a system architecture following a co-design approach, i.e., besides scientific and industry partners, local experts and administrative actors are actively involved in the system development. Additionally, the outlined system incorporates two novel methodological strands: (1) it employs a deep neural network (DNN) based data analytics approach and (2) makes use of a new generation of satellite data, namely Sentinel-5 Precursor (Sentinel-5P). Hence, the system allows for providing areal and high-resolution (e.g., street-level) real-time and forecast (up to 48 hours) data to inform decision makers for taking appropriate short-term measures, and secondly, to simulate air quality under different planning options and long-term actions such as modified traffic flows and various urban layouts.

This study proposes a Feed Forward Neural Net (FFNN) to forecast renewable energy generation of marine wind parks located in Denmark. The neural network uses historical weather and power generation data for training and applies the learned pattern to forecast wind energy production. Furthermore, the study shows how to improve prediction quality by leveraging specific parameters. Especially, we study the impact of the distance and direction of the weather station related to the production site in detail. In addition, we examined various parameters of the network to improve the accuracy. The proposed model distinguishes itself from other models by the fact that the optimal validation accuracy of more than 90 percent can be reached with training data sets of only a limited size, here two months of data with hourly resolution.