Informed Machine Learning for Epidemics - From Data Analysis to Time-Series Forecasting

Aug 5, 2026

Upload by Ezra

This research investigates how machine learning can be used in epidemiology to develop strategies addressing epidemic challenges. Epidemics are treated both as complex physical phenomena and as stressors on healthcare systems, motivating two complementary directions: data analysis to assess patient conditions and time-series forecasting to predict epidemiological trends. A tree-based pipeline evaluates infection severity from patient records, while forecasting methods handle data scarcity and noisy, incomplete signals through domain knowledge guidance, extensive deep learning reviews, and techniques such as augmentation and decorrelation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages117

Reading Time4 h 55 min

AccessFree to download

Introduction

Contribution

Background

Machine Learning

Knowldge Injection

Epidemics

Tree-based Data Analysis for Infection Disease

Feature Importance in Tree-based models

Feature Selection with Boruta

Data Augmentation

Decorrelation

Informed Deep Learning for Epidemic Forecasting

Universal Ordinary Differential Equation

Renewal Equations

Transfer Learning

Model-Agnostic Meta Learning

Conclusion

Alma Mater Studiorum-Università di Bologna DOTTORATO DI RICERCA IN COMPUTER SCIENCE AND ENGINEERING Ciclo 36 Settore Concorsuale: 09/H1-SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI Settore Scientifico Disciplinare: ING-INF/05-SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI INFORMED MACHINE LEARNING FOR EPIDEMICS: FROM DATA ANALYSIS TO TIME-SERIES FORECASTING Presentata da: Federico Baldo Coordinatore Dottorato Ilaria Bartolini Supervisore Michela Milano Esame finale anno 2024 Abstract This research dives deep into the application of Machine Learning (ML) in epidemiology, aiming to craft innovative strategies to address challenges presented by epidemic phenomena. Recognizing the dual nature of epidemics—as complex physical occurrences and as catalysts of pressure on healthcare systems—we focus on two main areas: data analysis for patient condition assessment and time-series forecasting for epidemiological trend prediction. In the realm of data analysis, we introduced a tree-based ML pipeline apt at evaluating the severity of infections using patient records. On the forecasting front, we addressed the challenges posed by the scarcity of data and the intricate nature of epidemics, developing methodologies for predicting epidemiological trends. In this setting, we exploit domain knowledge to guide the learning process, ensuring accurate and reliable predictions even when confronted with incomplete and noisy data. Our contributions include an extensive review of deep learning models incorporating constraint-based domain knowledge to enhance performance. Additionally, we provide a thorough overview of deep learning applications in epidemic forecasting, highlighting a wide set of approaches and their e↵ectiveness. Concerning data analysis, we propose a decision tree-based approach for assessing Sars-CoV-2 patient risk, alongside the deployment of a data augmentation technique and a decorrelation method to enhance the robustness and reliability of the analysis. In the realm of time-series forecasting, we conducted a comprehensive exploration of Physics Informed Neural Networksand their applications in modeling epidemic phenomena. Furthermore, we introduced a novel methodology that integrates deep neural networks with renewal processes for time-series forecasting. Both these methods are further enhanced with the addition of transfer learning, which allows the approximation of an accurate predictor with scarce data. The approaches presented exhibit enhanced performance compared to other deep learning methods, marking a promising avenue for future research. Indeed, this research showcases some of the potential applications of ML in epidemiology, motivating future extensions and the signiﬁcance of interdisciplinary approaches in addressing the complex challenges presented by epidemics. Contents 1 Introduction 1 1.1 Contribution ............................ 2 2 Background 4 2.1 Machine Learning ......................... 4 2.1.1 Decision Tree ....................... 5 2.1.2 Artiﬁcial Neural Networks ................ 10 2.2 Knowldge Injection ........................ 12 2.2.1 Constraint-based Knowledge Injection in Deep Neural Networks ......................... 13 2.2.2 Knowledge Injection Methods: A Taxonomy ...... 14 2.2.3 Domain Knowledge Injection Approaches ........ 15 2.2.4 Physics-Informed Machine Learning ........... 30 2.3 Epidemics ............................. 33 2.3.1 Epidemiological Models ................. 34 2.3.2 Deep Learning for Epidemics .............. 36 2.3.3 Epidemiological Data ................... 43 3 Tree-based Data Analysis for Infection Disease 45 3.1 Feature Importance in Tree-based models ............ 46 3.2 Feature Selection with Boruta .................. 46 3.3 Data Augmentation ........................ 47 3.4 Decorrelation ........................... 48 4 Informed Deep Learning for Epidemic Forecasting 56 4.1 Universal Ordinary Di↵erentia

FAQs

What two main goals does the research focus on in epidemiology?

It targets (1) data analysis for evaluating patient condition severity and (2) time-series forecasting to predict epidemiological trends.

How does the study improve infection risk assessment from patient data?

It proposes a tree-based machine learning pipeline and adds robustness through feature-related strategies such as selection, data augmentation, and decorrelation.

What methods are used for epidemic trend forecasting under limited and noisy data?

The work explores Physics Informed Neural Networks and introduces a renewal-process-based approach, further enhanced with transfer learning to approximate accurate predictors from scarce data.

Aug 5, 2026

Upload by Ezra