Deep Learning -Derived Optimal Aviation Strateg ies to Control Pandemics Syed Rizvi1 Akash Awasthi1 Maria J. Peláez2 Zhihui Wang234 Vittorio Cristini2456

2025-05-06 0 0 1.95MB 32 页 10玖币

侵权投诉

Deep Learning-Derived Optimal Aviation Strategies to Control

Pandemics

Syed Rizvi1,†, Akash Awasthi1,†, Maria J. Peláez2, Zhihui Wang2,3,4, Vittorio Cristini2,4,5,6,

Hien Van Nguyen1, Prashant Dogra2,3,*

1Department of Electrical and Computer Engineering, University of Houston, Houston, TX 77004,

USA

2Mathematics in Medicine Program, Department of Medicine, Houston Methodist Research

Institute, Houston, TX 77030, USA

3Department of Physiology and Biophysics, Weill Cornell Medical College, New York, NY

10065, USA

4Neal Cancer Center, Houston Methodist Research Institute, Houston, TX 77030, USA

5Department of Imaging Physics, The University of Texas M.D. Anderson Cancer Center,

Houston, TX 77230, USA

6Physiology, Biophysics, and Systems Biology Program, Graduate School of Medical Sciences,

Weill Cornell Medicine, New York, NY 10065, USA

†These authors contributed equally.

*Correspondence should be addressed to:

Prashant Dogra, PhD

Assistant Research Professor of Mathematics in Medicine

Department of Medicine, Houston Methodist Research Institute, Houston, TX, USA

Assistant Professor of Research in Physiology and Biophysics

Department of Physiology and Biophysics, Weill Cornell Medical College, New York, NY, USA

pdogra@houstonmethodist.org

Keywords: COVID-19, pandemic, deep learning, graph neural network, artificial

intelligence, aviation policy

ABSTRACT

The COVID-19 pandemic has affected countries across the world, demanding drastic

public health policies to mitigate the spread of infection, leading to economic crisis as a collateral

damage. In this work, we investigated the impact of human mobility (described via international

commercial flights) on COVID-19 infection dynamics at the global scale. For this, we developed

a graph neural network-based framework referred to as Dynamic Connectivity GraphSAGE

(DCSAGE), which operates over spatiotemporal graphs and is well-suited for dynamically

changing adjacency information. To obtain insights on the relative impact of different geographical

locations, due to their associated air traffic, on the evolution of the pandemic, we conducted local

sensitivity analysis on our model through node perturbation experiments. From our analyses, we

identified Western Europe, North America, and Middle East as the leading geographical locations

fueling the pandemic, attributed to the enormity of air traffic originating or transiting through these

regions. We used these observations to identify tangible air traffic reduction strategies that can

have a high impact on controlling the pandemic, with minimal interference to human mobility.

Our work provides a robust deep learning-based tool to study global pandemics and is of key

relevance to policy makers to take informed decisions regarding air traffic restrictions during

future outbreaks.

1. INTRODUCTION

In December 2019, the first COVID-19 cases were detected in the Wuhan region of China,

from where the SARS-CoV-2 virus rapidly spread worldwide, infecting people and causing the

World Health Organization to declare COVID-19 a pandemic in March 2020 (1). As of October

2022, the infection has accounted for over 614 million cases worldwide, with over 6.5 million

deaths (2). Since neither vaccines nor therapeutic drugs were available at the onset of the

pandemic, governments around the world responded by implementing stringent public health

policies to control the spread of infection. These measures included, but were not limited to, social

distancing, use of face masks, and travel restrictions. Data analytics and modeling-based studies

sought to explore the impact of public health policies on pandemic dynamics and were thus

leveraged to optimize the implementation of such policies for maximal impact.

To this end, Kraemer et al. analyzed the effectiveness of public health interventions in

China during the early stages of the pandemic, concluding that the interventions significantly

mitigated the early spread of COVID-19 (4). Adiga et al. utilized human mobility maps to assess

the interplay between mobility data, pandemic dynamics, and public health policy in the United

States and India, analyzing potential scenarios of delayed lockdowns and school reopenings (5).

International channels of human mobility were also examined to explain early pandemic dynamics;

Adiga et al. utilized international air traffic data as a mobility indicator and assessed the

effectiveness of flight cancellation policies on the time of arrival of the pandemic to other countries

(6).

Further works used mechanistic modeling, machine learning, and deep learning-based

approaches to forecast pandemic dynamics and evaluate the impact of public health policies on

infection incidence and spread. Kai et al. forecasted the impact of mask mandates on the spread of

the pandemic (7), while Anastassopoulou et al. estimated infection spread parameters using a

compartmental epidemiological model (8). Chang et al. and Yang et al. integrated population

mobility data into epidemiological models to account for the role of population movement in

driving the evolution of the pandemic, with the goal of guiding public policy (9,10). A simple

seasonal ARIMA model was utilized by Chintalapudi et al. to forecast registered and recovered

cases at the onset of lockdown mandates in Italy (11). Ahmar et al. proposed the SutteARIMA

model for short-term COVID-19 case forecasting that combined the ARIMA and the α-Sutte

indicator (12), and found it to be more suitable for short-term case forecasting than ARIMA (13).

Chimmula et al. utilized a deep learning-based approach, applying LSTM (14) networks to forecast

Canadian COVID-19 cases and predict the possible stopping points of the pandemic (15). In a

novel approach, Dogra et al. leveraged the Elliott Wave principle of financial mathematics to

explain and forecast the trends in global COVID-19 cases based on human emotion as a driving

factor of the pandemic (3). However, a major limitation of the above works is that they do not

incorporate the complexity of spatial relationships and interactions between different geographical

locations in determining the outcomes of the pandemic.

To this end, Graph Neural Networks (GNNs) provide a deep learning-based framework

that can capture the rich relational information among elements in a network or graph and can thus

be leveraged to study the influence of global population mobility on COVID-19 dynamics.

Spatiotemporal GNNs are a special class of GNNs that simultaneously consider spatial and

temporal information when processing graph inputs, and have been widely applied to problems

such as traffic forecasting (16–22). Spatiotemporal modeling with GNNs has also been applied to

the problem of pandemic forecasting; Kapoor et al. examined county-level COVID-19 forecasting

within the United States by constructing 100 large-scale graph snapshots of US counties, with

nodes representing counties and edges representing human mobility between nodes on each day

(23). Wang et al. also constructed dynamic mobility graphs using inter-region mobility data at the

state-level, and proposed a Recurrent Message Passing (RMP) GNN for mobility-informed

infection forecasting (24). Gao et al. developed Spatiotemporal Attention Network (STAN) (25),

which involved static edges using both demographic similarity and geographical distance between

different locations, and integrated real-world evidence from medical claims into node features to

forecast pandemic dynamics. Other works (26,27) constructed models with GNN and LSTM layers

to capture both spatial and temporal dependencies in data and predict future cases on European

infection data. Sesti et al. devised a GNN-LSTM architecture that operated over a static adjacency

graph that was constructed using geographical social connectivity data between different countries

(26). Panagopoulos et al., on the other hand, applied a Message-Passing Neural Network (MPNN)

to graph snapshots at each timestep of an input window of data, concatenating the output

representations and classifying them into future case predictions (27). We observe that these works

focus on forecasting COVID-19 cases and pandemic dynamics but lack explainability experiments

that could give insight into why a GNN model made a particular prediction. Existing explainability

methods for GNNs are defined over static graphs; Pope et al. devised adaptations of common

explainability techniques for GNNs, including saliency maps, class activation maps, and excitation

backpropagation (28). GNNExplainer (29) introduced a model-agnostic method for finding

subgraph explanations of a graph by maximizing mutual information between a graph and its

subgraph explanation. These methods, however, are not defined over spatiotemporal graphs, and

therefore are limited in their applicability to mobility data that dynamically changes over a

temporal dimension.

To address these shortcomings, we developed the Dynamic Connectivity GraphSAGE

(DCSAGE) architecture to provide an explainable deep learning-based modeling approach for

analyzing mobility-driven regional impact on pandemic dynamics. Our work differs from previous

works in that we utilize aviation data as an indicator of international human mobility in the

COVID-19 pandemic and focus on the interpretability of our modeling results. In contrast to static-

adjacency approaches, we designed our GNN architecture to directly accept dynamic adjacency

information that varies on a day-to-day basis. We used sensitivity analysis to quantify the impact

of nodes in our spatiotemporal graphs, providing insights into the influence of different

geographical regions on pandemic dynamics. From these experiments, we identified the relevance

of human mobility (via international flights) in determining the relative impact of various

geographical regions, which we quantify as the degree to which a region causes changes in the

case predictions of other regions. We use the insights gained on relative impact of different regions

to identify tangible strategies for limiting aviation to reduce the impact of highly influential nodes.

Fig. 1. DCSAGE model architecture. A one-day graph Gt is input into the model, consisting of a set of node features

Xt (daily COVID-19 cases) and adjacency information At (daily flight data). Two sequential GraphSAGE layers

process the data to learn spatial representations of the input graph, which is then fed into two stacked LSTM cells that

learn spatial relationships over time. After processing seven days of input, the LSTM cell hidden state outputs (h1t+7,

h2t+7) are concatenated with the input node features (Xt to Xt+7) and passed into the final output layer to predict the

next day cases for each node (𝑦̂).

2. METHODS

2.1. DCSAGE model development

To address the challenges of modeling dynamically changing adjacency information in

spatiotemporal graphs, we introduced a novel deep learning architecture, referred to as Dynamic

Connectivity GraphSAGE (DCSAGE), which utilizes the GraphSAGE message-passing

framework (31) to learn spatial relationships between nodes in our graph. As shown in Fig. 1 (see

end of manuscript), DCSAGE is a recurrent graph architecture composed of GraphSAGE and

LSTM layers, which exploit the spatiotemporal information present in the data. The input for

DCSAGE at timestep t is a weighted, directed graph Gt, which comprises ten nodes representing

the partitioning of the globe into ten geographical locations of interest in the context of the COVID-

19 pandemic. The ten nodes are North America, South America, Oceania (i.e., Australia and

neighboring island nations), Africa (Egypt excluded), Middle East (includes Egypt), Eastern

Europe, Western Europe, Central Asia, South Asia, and Southeast Asia. DCSAGE uses timeseries

data for daily COVID-19 infections and international flights for the ten nodes, obtained from

public databases (32–34). The infection timeseries comprises each node’s feature, while the flight

timeseries data is used to weight the edges. The key components of the DCSAGE architecture are

described below:

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DeepLearning-DerivedOptimalAviationStrategiestoControlPandemicsSyedRizvi1,†,AkashAwasthi1,†,MariaJ.Peláez2,ZhihuiWang2,3,4,VittorioCristini2,4,5,6,HienVanNguyen1,PrashantDogra2,3,*1DepartmentofElectricalandComputerEngineering,UniversityofHouston,Houston,TX77004,USA2MathematicsinMedicineProgram,Depar...

展开>> 收起<<

Deep Learning -Derived Optimal Aviation Strateg ies to Control Pandemics Syed Rizvi1 Akash Awasthi1 Maria J. Peláez2 Zhihui Wang234 Vittorio Cristini2456.pdf

共32页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep Learning -Derived Optimal Aviation Strateg ies to Control Pandemics Syed Rizvi1 Akash Awasthi1 Maria J. Peláez2 Zhihui Wang234 Vittorio Cristini2456

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: