michaelajao/hospitalisation-PINN: Hospitalisation-PINN

A Hybrid Physics-Informed Neural Network: SEIRD Model for Forecasting COVID-19 ICU Demand in England This repository contains the official implementation of the research presented in the book chapter: Ajao-Olarinoye, M., Palade, V., He, F., Wark, P.A., Mukandavire, Z., & Mousavi, S. (2025). A Hybrid Physics-Informed Neural Network: SEIRD Model for Forecasting COVID-19 Intensive Care Unit Demand in England. In Onyekpe, U., Palade, V., & Wani, M.A. (Eds.), Recent Advances in Deep Learning Applications: New Techniques and Practical Examples (1st ed., pp. 231-256). Chapman and Hall/CRC. https://doi.org/10.1201/9781003570882 Overview COVID-19 poses a complex and unprecedented challenge to public health and society, requiring scientific efforts to understand, model, diagnose, and control the disease. This research develops a novel hybrid approach combining Physics-Informed Neural Networks (PINNs) with epidemiological models to forecast COVID-19 dynamics and healthcare resource demand in England. Key Features Hybrid PINN-SEIRD Model: Integration of physics-informed neural networks with an extended SEIRD compartmental model featuring asymptomatic/symptomatic divisions, hospital admissions, and critical care compartments Parameter Estimation: Automated estimation of key epidemiological parameters including transmission rate (β), hospitalization ratio (ω), recovery rates, and mortality rates Healthcare Resource Forecasting: Prediction of ICU demand, mechanical ventilator requirements, and hospital admission rates Deep Learning Models: Implementation of multiple architectures including: Recurrent Neural Networks (RNN, LSTM, GRU) Sequence-to-Sequence models (Seq2Seq) with and without attention mechanisms Transformer-based models Autoformer and Informer architectures Simulated Annealing Hyperparameter Tuning: Metaheuristic optimization algorithm for automated hyperparameter search with temperature-based acceptance probability Attention Mechanisms: Multiple attention types (Dot Product, Scaled Dot Product, General, Additive, Concat) for Seq2Seq models Ensemble Methods: Greedy optimization and simulated annealing for model ensembling Physics-Constrained Learning: Direct integration of physical laws (epidemic differential equations) into the neural network training process Time-Varying Parameters: Support for time-dependent epidemiological parameters to capture intervention effects