Neural Network Training with Highly Incomplete Datasets published in Machine Learning: Science and Technology

Working principles for training neural networks with highly incomplete dataset: vanilla (upper panel) vs GapNet (lower panel) (Image by Yu-Wei Chang.)
Neural Network Training with Highly Incomplete Datasets
Yu-Wei Chang, Laura Natali, Oveis Jamialahmadi, Stefano Romeo, Joana B. Pereira, Giovanni Volpe
Machine Learning: Science and Technology 3, 035001 (2022)
arXiV: 2107.00429
doi: 10.1088/2632-2153/ac7b69

Neural network training and validation rely on the availability of large high-quality datasets. However, in many cases only incomplete datasets are available, particularly in health care applications, where each patient typically undergoes different clinical procedures or can drop out of a study. Since the data to train the neural networks need to be complete, most studies discard the incomplete datapoints, which reduces the size of the training data, or impute the missing features, which can lead to artefacts. Alas, both approaches are inadequate when a large portion of the data is missing. Here, we introduce GapNet, an alternative deep-learning training approach that can use highly incomplete datasets. First, the dataset is split into subsets of samples containing all values for a certain cluster of features. Then, these subsets are used to train individual neural networks. Finally, this ensemble of neural networks is combined into a single neural network whose training is fine-tuned using all complete datapoints. Using two highly incomplete real-world medical datasets, we show that GapNet improves the identification of patients with underlying Alzheimer’s disease pathology and of patients at risk of hospitalization due to Covid-19. By distilling the information available in incomplete datasets without having to reduce their size or to impute missing values, GapNet will permit to extract valuable information from a wide range of datasets, benefiting diverse fields from medicine to engineering.

Laura Natali presented her half-time seminar on 1 April 2022

Opponent Bernhard Mehlig (left), Laura Natali (center), and PhD supervisor Giovanni Volpe (right). (Photo by L. Perez.)
Laura Natali completed the first half of her doctoral studies and she defended her half-time on the 1st of April 2022.

The presentation was held in hybrid format, with part of the audience in the Von Bahr room and the rest connected through zoom. The half-time consisted in a presentation about her past and planned projects and it was followed by a discussion and questions proposed by her opponent Bernhard Mehlig.

The presentation started with a description of her concluded projects about employing neural networks in an epidemic agent-based model, published in Improving epidemic testing and containment strategies using machine learning accepted in Machine Learning: Science and Technology. It continued with her second project, about handling incomplete medical datasets with neural networks, available online as a preprint Neural Network Training with Highly Incomplete Datasets on ArXiv. In the last section, she outlined the proposed continuation of her PhD, with an ongoing project for combining artificial active matter with neural networks.

Presentation by L. Natali at Spatial Data Science 2020, 11 June 2021

Comparison of different evolution regimes of disease spreading: free evolution (bottom left half) vs network strategy (top right half). (Image by Laura Natali.)
Improving epidemic testing and containment strategies using machine learning. 
Laura Natali, Saga Helgadottir, Onofrio M. Maragò, Giovanni Volpe.
Submitted to SDS2020
Date: 11 June
Time: 16:15 (CEST)

Abstract: 
Containment of epidemic outbreaks entails great societal and economic costs.  Cost-effective containment strategies rely on efficiently identifying infected individuals, making the best possible use of the available testing resources. Therefore, quickly identifying the optimal testing strategy is of critical importance. Here, we demonstrate that machine learning can be used to identify which individuals are most beneficial to test, automatically and dynamically adapting the testing strategy to the characteristics of the disease outbreak. Specifically, we simulate an outbreak using the archetypal susceptible-infectious-recovered (SIR) model and we use data about the first confirmed cases to train a neural network that learns to make predictions about the rest of the population. Using these prediction, we manage to contain the outbreak more effectively and more quickly than with standard approaches. Furthermore, we demonstrate how this method can be used also when there is a possibility of reinfection (SIRS model) to efficiently eradicate an endemic disease.

Improving epidemic testing and containment strategies using machine learning accepted in Machine Learning: Science and Technology

Comparison of different evolution regimes of disease spreading: free evolution (bottom left half) vs network strategy (top right half).
Improving epidemic testing and containment strategies using machine learning
Laura Natali, Saga Helgadottir, Onofrio M. Maragò, Giovanni Volpe
Machine Learning: Science and Technology, 2 035007 (2021)
doi: 10.1088/2632-2153/abf0f7
arXiv: 2011.11717

Containment of epidemic outbreaks entails great societal and economic costs. Cost-effective containment strategies rely on efficiently identifying infected individuals, making the best possible use of the available testing resources. Therefore, quickly identifying the optimal testing strategy is of critical importance. Here, we demonstrate that machine learning can be used to identify which individuals are most beneficial to test, automatically and dynamically adapting the testing strategy to the characteristics of the disease outbreak. Specifically, we simulate an outbreak using the archetypal susceptible-infectious-recovered (SIR) model and we use data about the first confirmed cases to train a neural network that learns to make predictions about the rest of the population. Using these prediction, we manage to contain the outbreak more effectively and more quickly than with standard approaches. Furthermore, we demonstrate how this method can be used also when there is a possibility of reinfection (SIRS model) to efficiently eradicate an endemic disease.

Press release on Machine learning can help slow down future pandemics

Comparison of different evolution regimes of disease spreading: free evolution (bottom left half) vs network strategy (top right half). (Image by Laura Natali.)

The article Improving epidemic testing and containment strategies using machine learning has been featured in the News of the Faculty of Science of Gothenburg University.

Here the links to the press releases:
Swedish: Maskininlärning kan bidra till att bromsa framtida pandemier
English: Machine learning can help slow down future pandemics

The articles was also featured in:
AI ska bromsa framtidens pandemier Metal Supply (23/04/2021)
El papel de la inteligencia artificial para frenar futuras pandemias El Nacional.cat. (16/04/2021)
AI could be critical to preventing future pandemics – study Health Tech World. (16/04/2021)
Machine Learning Slows Down Future Pandemics MedIndia. (15/04/2021)
Machine Learning May Be Key to Avoiding the Next Possible Pandemic News18.com (15/04/2021)
Så kan AI bromsa nästa pandemi – svensk forskningförfinar testningen Computer Sweden (15/04/2021)
AI could prevent future pandemics Electronics360 (14/04/2021)
L’IA peut contribuer à limiter la propagation des infections lors des futures épidémies (étude) Ecofin Telecom. (14/04/2021)
Machine Learning can help slow down future pandemics:Study SocialNews.xyz (14/04/2021)
Machine learning can help slow down future pandemics —ScienceDaily Sortiwa Trending Viral News Portal (14/04/2021)
AI mot smittspridning Sveriges Radio Vetenskapsradion. (14/04/2021)

Laura Natali joins the Soft Matter Lab

Laura Natali starts her PhD at the Physics Department of the University of Gothenburg on 1st March 2020.

Laura has a Master degree in Physics, curriculum in Physics of Biosystems, from the University of Rome “La Sapienza”, where he submitted a Master thesis whose results can be found here.

In her PhD, she will focus on microswimmers and active polymers employing machine learning techniques.

Seminar on polymers under local active forces: a simplified stochastic model for motor-induced translocations by Laura Natali from Università di Roma “La Sapienza”, Nexus, 22 January 2020

Polymers under local active forces: a simplified stochastic model for motor-induced translocations.
Seminar by Laura Natali from the University of Rome “La Sapienza”, Rome, Italy.

Molecular motors are a wide class of enzymes that can transport even large macromolecules by converting chemical into mechanical energy, through the process of ATP-hydrolysis. Among those, active nanopore translocation is a common mechanism to carry proteins across biological membranes. The phospholipid bilayer is impenetrable for the folded protein, that requires to be unfolded and pulled across channels such as alpha-hemolysin [1]. Exploiting the translocation process, we can acquire information about the target proteins’ structure, such as the intermediate configurations between the folded and denatured structures [2]. Here follows the interest in modeling the motor proteins complex in a simple simulation setup.
First of all, we characterized the active polymer’s features in free space and subsequently, we confined it inside a nanopore model. The study in free space aims to investigate how activity affects both the global and the local polymer properties. The chosen model for the active force is an Active Ornstein-Uhlenbeck Particle, a model closely related to Active Brownian Particles [3]. The presence of the catalytic head affects the end-to-end distance of the polymer, which describes its degree of compactness. The active chain can be studied through the Rouse mode decomposition [4], which allowed us to analyze the dependence of the second moment of the end-to-end distance as a function of the persistence time of the activity. We considered also the distance between consecutive monomers, which provides an insight into the local effects of the active force and how it is transmitted along the chain.
In this work, we were inspired by an experiment employing the ClpXP complex [5], a protein-degradation machine responsible for unfolding and digesting malfunctioning proteins through unidirectional transport across the nanopore. In the model we propose, we enclose the polymer chain in a confining potential, simulating the effect of the nanopore. The energy generated by the molecular motor translocates the polymers, which means they cross the pore from side to side. In our setup, the effect of the molecular motor [6] is represented as a potential barrier combined with a region that makes active the monomers that are crossing it. The translocation pathways have a step-wise profile typical of their biological counterparts.

References:

  1. J. Nivala, D. B. Marks and M. Akeson, Nature Biotechnol., 2013, 31, 247
  2. M. Bacci, M. Chinappi, C. M. Casciola and F. Cecconi, J. Phys.Chem. B, 2012, 116, 4255–4262
  3. L. Caprini and U. M. B. Marconi, Soft matter, 2018, 14, 9044–9054.
  4. P. E. Rouse, J. Chem. Phys., 1953, 21, 1272.
  5. T. A. Baker and R. T. Sauer, Biochimica et Biophysica Acta (BBA) – Molecular Cell Research, 2012, 1823, 15–28.
  6. L. Caprini, U. Marini Bettolo Marconi, A. Puglisi and A. Vulpiani, J. Chemi. Phys., 2019, 150, 024902.

Place: Nexus room, Fysik Origo, Fysik
Time: 22 January, 2020, 11:00