Proceedings of the 1st Workshop on Healthcare AI and COVID-19, ICML 2022 Held in Baltimore, Maryland, USA on 22 July 2022 Published as Volume 184 by the Proceedings of Machine Learning Research on 21 July 2022. Volume Edited by: Peng Xu Tingting Zhu Pengkai Zhu David A. Clifton Danielle Belgrave Yuanting Zhang Series Editors: Neil D. Lawrence https://proceedings.mlr.press/v184/ Thu, 09 Feb 2023 06:19:56 +0000 Thu, 09 Feb 2023 06:19:56 +0000 Jekyll v3.9.3 Because the COVID-19 virus is highly transmissible, leading to a worldwide increment of new infections and deaths daily, the development of an automated tool to identify COVID-19 using CT images has attracted much attention. Significantly, deep metric learning can be deployed to cluster and classify the fine-grained CT images, which aims to learn a mapping from the original objects to a discriminative feature embedding space. Previous deep metric learning works have been proposed to construct various structures of loss, mine hard samples, or introduce regularization constraints, \etc. In general, traditional loss functions of deep metric learning methods are based on constraining the distance of the triplet embeddings in the feature space. Instead of focusing on the previous research directions, in this work, we pay attention to exploring confusing triplet embeddings, for the reason that confusing triplet embeddings perform a side effect on the majority of deep triplet-based metric learning methods. By considering the spatial relation of triplet embedding, and conducting theoretical analysis in the feature space, we propose an approach to recognize the confusing triplet embeddings and construct a Confusing Triplet Embedding Learning (CTEL) method by adding a hard constraint on the confusing triplet embeddings. The extensive experiments indicate that our proposed CTEL method achieves more excellent performance on two medical CT image datasets and two fine-grained standard image datasets compared with many state-of-the-art methods. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/yuan22a.html https://proceedings.mlr.press/v184/yuan22a.html Computer-assisted imagery analysis based on chest X-ray images plays a crucial role in the clinical diagnosis and screening of COVID-19. However, the radiographic features of chest X-rays are highly complex and irregular in shape. Moreover, the size and location of the lesion regions vary greatly with infection stages and patients, thus dramatically increasing the difficulty of COVID-19 identification. A lightweight adaptive self-attention network is developed to address this problem, namely ASA-CoroNet. It firstly extracts underlying features using a depthwise separable convolution-based backbone, then further identifies lesion regions through an adaptive self-attentive module, and finally utilizes a homogeneous vector capsule layer to map the obtained features into capsule vectors to instantiate detection objects accurately. Extensive experimental results demonstrate that the proposed model outperforms the state-of-the-art methods and obtains competitive results on limited datasets. More importantly, the trainable params of the proposed model are reduced by 7x compared to the state-of-the-art capsule network. In addition, we also interpret the proposed model using different class activation techniques and confirm the validity of the three components through numerous ablation studies. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/wu22a.html https://proceedings.mlr.press/v184/wu22a.html A crucial aspect of managing a public health crisis is to effectively balance prevention and mitigation strategies, while taking their socio-economic impact into account. In particular, determining the influence of different non-pharmaceutical interventions (NPIs) on the effective use of public resources is an important problem, given the uncertainties on when a vaccine will be made available. In this paper, we propose a new approach for obtaining optimal policy recommendations based on Epicast models, which can characterize the disease progression under different interventions, and a look-ahead reward optimization strategy to choose the suitable NPI at different stages of an epidemic. Given the time delay inherent in any Epicast model and the exponential nature especially of an unmanaged epidemic, we find that such a look-ahead strategy infers non-trivial policies that adhere well to the constraints specified. Using two different Epicast models, namely SEIR and EpiCast, we evaluate the proposed algorithm to determine the optimal NPI policy, under a constraint on the number of daily new cases and the primary reward being the absence of restrictions. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/thiagarajan22a.html https://proceedings.mlr.press/v184/thiagarajan22a.html As urgency over the coronavirus disease 2019 (COVID-19) increased, many datasets with chest radiography (CXR) and chest computed tomography (CT) images emerged aiming at the detection and prognosis of COVID-19. Over the last two years, thousands of studies have been published, reporting promising results. However, a deeper analysis of the datasets and the methods employed reveals issues that may hamper conclusions and practical applicability. We investigate three major datasets commonly used in these studies, detect problems related to the existence of duplicates, address the specificity of classes within those datasets, and propose a way to perform external validation via cross-dataset evaluation. Our guidelines and findings contribute towards a trust-worthy application of Machine Learning in the context of image-based diagnosis, as well as offer a more accurate assessment of models applied to the prognostication of diseases using image datasets and pave the way towards models that can be relied upon in the real world. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/silva22a.html https://proceedings.mlr.press/v184/silva22a.html Remote physiological measurement, \textit{e.g.}, heart rate and respiratory rate measurement, becomes more and more important when contact instrument-based measurement is inaccessible and non-preferable under the COVID-19 pandemic. Non-contact camera based physiological measurement enables Telehealth, remote health monitoring and smart hospital applications. Remote physiological signal measurement has challenges such as environment illumination variations, head motion, facial expression, etc. We propose a convolutional neural network to jointly estimate heart rate and respiratory rate with camera video as input in a multitask fashion, which leverages the correlation between heart rate and respiratory rate. Specifically, we propose a novel loss function which integrates the frequency correlation between heart rate and respiratory rate to improve robustness of both heart rate and respiratory rate estimation. Furthermore, we propose a post processing filter based on correlation between heart rate and respiratory rate which further improve prediction accuracy. Extensive experiments demonstrate that our proposed system significantly improves heart rate and respiratory rate measurement accuracy. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/ren22a.html https://proceedings.mlr.press/v184/ren22a.html The application of the state-of-the-art biomedical named entity recognition task faces a few challenges: first, these methods are trained on a fewer number of clinical entities (e.g., disease, symptom, proteins, genes); second, these methods require a large amount of data for pre-training and prediction, making it difficult to implement them in real-time scenarios; third, these methods do not consider the non-clinical entities such as social determinants of health (age, gender, employment, race) which are also related to patients’ health. We propose a Machine Learning (ML) pipeline that improves on previous efforts in three ways: first, it recognizes many clinical entity types (diseases, symptoms, drugs, diagnosis, etc.), second, this pipeline is easily configurable, reusable and can scale up for training and inference; third, it considers non-clinical factors related to patient’s health. At a high level, this pipeline consists of stages: pre-processing, tokenization, mapping embedding lookup and named entity recognition task. We also present a new dataset that we prepare by curating the COVID-19 case reports. The proposed approach outperforms baseline methods on four benchmark datasets with macro-and microaverage F1 scores around 90, as well as using our dataset with a macro-and micro-average F1 score of 95.25 and 93.18 respectively. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/raza22a.html https://proceedings.mlr.press/v184/raza22a.html Dozens of SARS-CoV-2 vaccines have been approved for public use, yet there remains a risk that the virus evolves to escape vaccine protection. This motivates the development of universal vaccines capable of protecting against current and potentially new strains of the virus. A key challenge is the lack of computational tools to design new viral proteins capable of vaccine escape, which could serve as good targets for the development of universal vaccines. Here, we designed VAE capable of generating SARS-CoV-2 spike proteins with variable immune visibility to the cell-mediated immune response. We compared our model with two simpler generative models; a random-mutator and an 11-gram language model. All three models can generate stable, structurally valid sequences, yet only the VAE model can generate low immunogenicity sequences that interpolate smoothly along the principal variance directions of known natural sequences. This model provides an effective computational tool for the generation of spike protein sequences useful for universal vaccine design. We provide its source code at https://github.com/hcgasser/SpikeVAE. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/phillips22a.html https://proceedings.mlr.press/v184/phillips22a.html Since the global outbreak of the coronavirus 2019 pandemic, hundreds of works have been published, analyzing and modeling multiple aspects of the disease. Several of them venture into predictive and modeling tasks, such as mortality prediction and patient severity scoring, using machine-learning (ML) algorithms. An important limitation for most of these works is the fact that they do not consider the multiple temporal aspects of this pandemic, especially regarding disease profile and distributional changes over the months. Such temporal effects are mostly due to multiple interactions between different and novel viral strains, combined with mass vaccination campaigns targeting different groups or patterns (e.g., prioritizing older individuals and those with comorbidity first) and availability of different vaccines. These temporal effects result in impaired model effectiveness and classification errors. In this paper, using a large dataset with over 10,000 patients from 39 hospitals in Brazil admitted during a period of more than 20 months, we provide an overview of the multiple forms of temporal drift that happened during the pandemic and the magnitude of their effects on model effectiveness. Our analyses encompass changes in the severely ill patients’ profile as well as how mortality rates have changed over time. We also investigate how the importance of different predictive variables change and shift over time. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/paiva22a.html https://proceedings.mlr.press/v184/paiva22a.html IBy using a recently released chest X-ray (CXR) image database for COVID-19 positive cases along with Normal, Lung Opacity (Non-COVID lung infection), and Viral Pneumonia images, this study compares the performance of SOTA deep learning models in detecting COVID-19 CXR images. Pre-trained deep learning models are retrained under several combinations of optimizers, learning rate schedulers, and loss functions. Our study shows that these SOTA deep learning models perform well if the models and parameters are selected meticulously. Overall, EfficientNet is superior to others especially across different optimizers. Regarding the loss function, the integration of cosine embedding similarity and cross entropy is slightly better than cross entropy itself while we adopt the SGD optimizer. In terms of optimizer, SGD constantly performs well while Adam and AdamW are unstable across different models. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/liu22a.html https://proceedings.mlr.press/v184/liu22a.html Monitoring and detecting epidemics are essential for protecting humanity from extreme harm. However, it must be done in real time for accurate epidemic detection to use limited resources efficiently and save time preventing the spread. Nevertheless, previous studies have focused on predicting the number of confirmed cases after the disease has already spread or when the relevant data are provided. Moreover, it is difficult to give the reason for predictions made using existing methods. In this study, we investigated how to detect and alert infectious diseases that might develop into pandemics soon, even before the information about a specific disease is aggregated. We propose an explainable method to detect an epidemic. This method uses only global news data, which are easily accessible in real time. Hence, we convert the news data to a graph form and cluster the news themes to curate and extract relevant information. The experiments on previous epidemics, including COVID-19, show that our approach allows the explainable real-time prediction of an epidemic disease and guides decision-making for prevention. Code is available at https://github.com/sungnyun/Epidemics-Detection-GKG. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/kim22a.html https://proceedings.mlr.press/v184/kim22a.html In this paper, we introduce Temporal Multiresolution Graph Neural Networks (TMGNN), the first architecture that both learns to construct the multiscale and multiresolution graph structures and incorporates the time-series signals to capture the temporal changes of the dynamic graphs. We have applied our proposed model to the task of predicting future spreading of epidemic and pandemic based on the historical time-series data collected from the actual COVID-19 pandemic and chickenpox epidemic in several European countries, and have obtained competitive results in comparison to other previous state-of-the-art temporal architectures and graph learning algorithms. We have shown that capturing the multiscale and multiresolution structures of graphs is important to extract either local or global information that play a critical role in understanding the dynamic of a global pandemic such as COVID-19 which started from a local city and spread to the whole world. Our work brings a promising research direction in forecasting and mitigating future epidemics and pandemics. Our source code is available at https://github.com/bachnguyenTE/temporal-mgn. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/hy22a.html https://proceedings.mlr.press/v184/hy22a.html The COVID-19 pandemic hits worldwide with a significant number of deaths and poses a major threat to public health. Accurate predictions of the risk of death and medical interventions are crucial for the survival of infected patients and the distribution of limited medical resources. Although machine learning classifiers can be used to predict mortality and medical interventions, it is problematic to employ the methods because training data are limited whose attributes may be missing and classes may be imbalanced. To effectively cope with these problems, we construct HexaGAN with a hint mechanism to predict the survival of the patients and medical interventions such as intubation and supplemental oxygen. In experiments, our method outperforms combinations of existing techniques for limited data problems. Notably, our method showed about twice higher performance than benchmarks in predicting deceased patients correctly. We anticipate that our approach could help provide appropriate treatments on time, allocate limited medical resources efficiently, and ultimately reduce the mortality rate of COVID-19 patients. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/hwang22a.html https://proceedings.mlr.press/v184/hwang22a.html A particular challenge for disease progression modeling is the heterogeneity of a disease and its manifestations in the patients. Existing approaches often assume the presence of a single disease progression characteristics which is unlikely for neurodegenerative disorders such as Parkinson’s disease. In this paper, we propose a hierarchical time-series model that can discover multiple disease progression dynamics. The proposed model is an extension of an input-output hidden Markov model that takes into account the clinical assessments of patients’ health status and prescribed medications. We illustrate the benefits of our model using a synthetically generated dataset and a real-world longitudinal dataset for Parkinson’s disease. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/ceritli22a.html https://proceedings.mlr.press/v184/ceritli22a.html Calibrating complex epidemiological models to observed data is a crucial step to provide both insights into the current disease dynamics, i.e. by estimating a reproductive number, as well as to provide reliable forecasts and scenario explorations. Here we present a new approach to calibrate an agent-based model – Epicast – using a large set of simulation ensembles for different major metropolitan areas of the United States. In particular, we propose: a new neural network based surrogate model able to simultaneously emulate all different locations; and a novel posterior estimation that provides not only more accurate posterior estimates of all parameters but enables the joint fitting of global parameters across regions. Thu, 21 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/anirudh22a.html https://proceedings.mlr.press/v184/anirudh22a.html Because the COVID-19 virus is highly transmissible, leading to a worldwide increment of new infections and deaths daily, the development of an automated tool to identify COVID-19 using CT images has attracted much attention. Significantly, deep metric learning can be deployed to cluster and classify the fine-grained CT images, which aims to learn a mapping from the original objects to a discriminative feature embedding space. Previous deep metric learning works have been proposed to construct various structures of loss, mine hard samples, or introduce regularization constraints, \etc. In general, traditional loss functions of deep metric learning methods are based on constraining the distance of the triplet embeddings in the feature space. Instead of focusing on the previous research directions, in this work, we pay attention to exploring confusing triplet embeddings, for the reason that confusing triplet embeddings perform a side effect on the majority of deep triplet-based metric learning methods. By considering the spatial relation of triplet embedding, and conducting theoretical analysis in the feature space, we propose an approach to recognize the confusing triplet embeddings and construct a Confusing Triplet Embedding Learning (CTEL) method by adding a hard constraint on the confusing triplet embeddings. The extensive experiments indicate that our proposed CTEL method achieves more excellent performance on two medical CT image datasets and two fine-grained standard image datasets compared with many state-of-the-art methods. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/yuan22a.html https://proceedings.mlr.press/v184/yuan22a.html Computer-assisted imagery analysis based on chest X-ray images plays a crucial role in the clinical diagnosis and screening of COVID-19. However, the radiographic features of chest X-rays are highly complex and irregular in shape. Moreover, the size and location of the lesion regions vary greatly with infection stages and patients, thus dramatically increasing the difficulty of COVID-19 identification. A lightweight adaptive self-attention network is developed to address this problem, namely ASA-CoroNet. It firstly extracts underlying features using a depthwise separable convolution-based backbone, then further identifies lesion regions through an adaptive self-attentive module, and finally utilizes a homogeneous vector capsule layer to map the obtained features into capsule vectors to instantiate detection objects accurately. Extensive experimental results demonstrate that the proposed model outperforms the state-of-the-art methods and obtains competitive results on limited datasets. More importantly, the trainable params of the proposed model are reduced by 7x compared to the state-of-the-art capsule network. In addition, we also interpret the proposed model using different class activation techniques and confirm the validity of the three components through numerous ablation studies. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/wu22a.html https://proceedings.mlr.press/v184/wu22a.html A crucial aspect of managing a public health crisis is to effectively balance prevention and mitigation strategies, while taking their socio-economic impact into account. In particular, determining the influence of different non-pharmaceutical interventions (NPIs) on the effective use of public resources is an important problem, given the uncertainties on when a vaccine will be made available. In this paper, we propose a new approach for obtaining optimal policy recommendations based on Epicast models, which can characterize the disease progression under different interventions, and a look-ahead reward optimization strategy to choose the suitable NPI at different stages of an epidemic. Given the time delay inherent in any Epicast model and the exponential nature especially of an unmanaged epidemic, we find that such a look-ahead strategy infers non-trivial policies that adhere well to the constraints specified. Using two different Epicast models, namely SEIR and EpiCast, we evaluate the proposed algorithm to determine the optimal NPI policy, under a constraint on the number of daily new cases and the primary reward being the absence of restrictions. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/thiagarajan22a.html https://proceedings.mlr.press/v184/thiagarajan22a.html As urgency over the coronavirus disease 2019 (COVID-19) increased, many datasets with chest radiography (CXR) and chest computed tomography (CT) images emerged aiming at the detection and prognosis of COVID-19. Over the last two years, thousands of studies have been published, reporting promising results. However, a deeper analysis of the datasets and the methods employed reveals issues that may hamper conclusions and practical applicability. We investigate three major datasets commonly used in these studies, detect problems related to the existence of duplicates, address the specificity of classes within those datasets, and propose a way to perform external validation via cross-dataset evaluation. Our guidelines and findings contribute towards a trust-worthy application of Machine Learning in the context of image-based diagnosis, as well as offer a more accurate assessment of models applied to the prognostication of diseases using image datasets and pave the way towards models that can be relied upon in the real world. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/silva22a.html https://proceedings.mlr.press/v184/silva22a.html Remote physiological measurement, \textit{e.g.}, heart rate and respiratory rate measurement, becomes more and more important when contact instrument-based measurement is inaccessible and non-preferable under the COVID-19 pandemic. Non-contact camera based physiological measurement enables Telehealth, remote health monitoring and smart hospital applications. Remote physiological signal measurement has challenges such as environment illumination variations, head motion, facial expression, etc. We propose a convolutional neural network to jointly estimate heart rate and respiratory rate with camera video as input in a multitask fashion, which leverages the correlation between heart rate and respiratory rate. Specifically, we propose a novel loss function which integrates the frequency correlation between heart rate and respiratory rate to improve robustness of both heart rate and respiratory rate estimation. Furthermore, we propose a post processing filter based on correlation between heart rate and respiratory rate which further improve prediction accuracy. Extensive experiments demonstrate that our proposed system significantly improves heart rate and respiratory rate measurement accuracy. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/ren22a.html https://proceedings.mlr.press/v184/ren22a.html The application of the state-of-the-art biomedical named entity recognition task faces a few challenges: first, these methods are trained on a fewer number of clinical entities (e.g., disease, symptom, proteins, genes); second, these methods require a large amount of data for pre-training and prediction, making it difficult to implement them in real-time scenarios; third, these methods do not consider the non-clinical entities such as social determinants of health (age, gender, employment, race) which are also related to patients’ health. We propose a Machine Learning (ML) pipeline that improves on previous efforts in three ways: first, it recognizes many clinical entity types (diseases, symptoms, drugs, diagnosis, etc.), second, this pipeline is easily configurable, reusable and can scale up for training and inference; third, it considers non-clinical factors related to patient’s health. At a high level, this pipeline consists of stages: pre-processing, tokenization, mapping embedding lookup and named entity recognition task. We also present a new dataset that we prepare by curating the COVID-19 case reports. The proposed approach outperforms baseline methods on four benchmark datasets with macro-and microaverage F1 scores around 90, as well as using our dataset with a macro-and micro-average F1 score of 95.25 and 93.18 respectively. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/raza22a.html https://proceedings.mlr.press/v184/raza22a.html Dozens of SARS-CoV-2 vaccines have been approved for public use, yet there remains a risk that the virus evolves to escape vaccine protection. This motivates the development of universal vaccines capable of protecting against current and potentially new strains of the virus. A key challenge is the lack of computational tools to design new viral proteins capable of vaccine escape, which could serve as good targets for the development of universal vaccines. Here, we designed VAE capable of generating SARS-CoV-2 spike proteins with variable immune visibility to the cell-mediated immune response. We compared our model with two simpler generative models; a random-mutator and an 11-gram language model. All three models can generate stable, structurally valid sequences, yet only the VAE model can generate low immunogenicity sequences that interpolate smoothly along the principal variance directions of known natural sequences. This model provides an effective computational tool for the generation of spike protein sequences useful for universal vaccine design. We provide its source code at https://github.com/hcgasser/SpikeVAE. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/phillips22a.html https://proceedings.mlr.press/v184/phillips22a.html Since the global outbreak of the coronavirus 2019 pandemic, hundreds of works have been published, analyzing and modeling multiple aspects of the disease. Several of them venture into predictive and modeling tasks, such as mortality prediction and patient severity scoring, using machine-learning (ML) algorithms. An important limitation for most of these works is the fact that they do not consider the multiple temporal aspects of this pandemic, especially regarding disease profile and distributional changes over the months. Such temporal effects are mostly due to multiple interactions between different and novel viral strains, combined with mass vaccination campaigns targeting different groups or patterns (e.g., prioritizing older individuals and those with comorbidity first) and availability of different vaccines. These temporal effects result in impaired model effectiveness and classification errors. In this paper, using a large dataset with over 10,000 patients from 39 hospitals in Brazil admitted during a period of more than 20 months, we provide an overview of the multiple forms of temporal drift that happened during the pandemic and the magnitude of their effects on model effectiveness. Our analyses encompass changes in the severely ill patients’ profile as well as how mortality rates have changed over time. We also investigate how the importance of different predictive variables change and shift over time. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/paiva22a.html https://proceedings.mlr.press/v184/paiva22a.html IBy using a recently released chest X-ray (CXR) image database for COVID-19 positive cases along with Normal, Lung Opacity (Non-COVID lung infection), and Viral Pneumonia images, this study compares the performance of SOTA deep learning models in detecting COVID-19 CXR images. Pre-trained deep learning models are retrained under several combinations of optimizers, learning rate schedulers, and loss functions. Our study shows that these SOTA deep learning models perform well if the models and parameters are selected meticulously. Overall, EfficientNet is superior to others especially across different optimizers. Regarding the loss function, the integration of cosine embedding similarity and cross entropy is slightly better than cross entropy itself while we adopt the SGD optimizer. In terms of optimizer, SGD constantly performs well while Adam and AdamW are unstable across different models. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/liu22a.html https://proceedings.mlr.press/v184/liu22a.html Monitoring and detecting epidemics are essential for protecting humanity from extreme harm. However, it must be done in real time for accurate epidemic detection to use limited resources efficiently and save time preventing the spread. Nevertheless, previous studies have focused on predicting the number of confirmed cases after the disease has already spread or when the relevant data are provided. Moreover, it is difficult to give the reason for predictions made using existing methods. In this study, we investigated how to detect and alert infectious diseases that might develop into pandemics soon, even before the information about a specific disease is aggregated. We propose an explainable method to detect an epidemic. This method uses only global news data, which are easily accessible in real time. Hence, we convert the news data to a graph form and cluster the news themes to curate and extract relevant information. The experiments on previous epidemics, including COVID-19, show that our approach allows the explainable real-time prediction of an epidemic disease and guides decision-making for prevention. Code is available at https://github.com/sungnyun/Epidemics-Detection-GKG. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/kim22a.html https://proceedings.mlr.press/v184/kim22a.html In this paper, we introduce Temporal Multiresolution Graph Neural Networks (TMGNN), the first architecture that both learns to construct the multiscale and multiresolution graph structures and incorporates the time-series signals to capture the temporal changes of the dynamic graphs. We have applied our proposed model to the task of predicting future spreading of epidemic and pandemic based on the historical time-series data collected from the actual COVID-19 pandemic and chickenpox epidemic in several European countries, and have obtained competitive results in comparison to other previous state-of-the-art temporal architectures and graph learning algorithms. We have shown that capturing the multiscale and multiresolution structures of graphs is important to extract either local or global information that play a critical role in understanding the dynamic of a global pandemic such as COVID-19 which started from a local city and spread to the whole world. Our work brings a promising research direction in forecasting and mitigating future epidemics and pandemics. Our source code is available at https://github.com/bachnguyenTE/temporal-mgn. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/hy22a.html https://proceedings.mlr.press/v184/hy22a.html The COVID-19 pandemic hits worldwide with a significant number of deaths and poses a major threat to public health. Accurate predictions of the risk of death and medical interventions are crucial for the survival of infected patients and the distribution of limited medical resources. Although machine learning classifiers can be used to predict mortality and medical interventions, it is problematic to employ the methods because training data are limited whose attributes may be missing and classes may be imbalanced. To effectively cope with these problems, we construct HexaGAN with a hint mechanism to predict the survival of the patients and medical interventions such as intubation and supplemental oxygen. In experiments, our method outperforms combinations of existing techniques for limited data problems. Notably, our method showed about twice higher performance than benchmarks in predicting deceased patients correctly. We anticipate that our approach could help provide appropriate treatments on time, allocate limited medical resources efficiently, and ultimately reduce the mortality rate of COVID-19 patients. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/hwang22a.html https://proceedings.mlr.press/v184/hwang22a.html A particular challenge for disease progression modeling is the heterogeneity of a disease and its manifestations in the patients. Existing approaches often assume the presence of a single disease progression characteristics which is unlikely for neurodegenerative disorders such as Parkinson’s disease. In this paper, we propose a hierarchical time-series model that can discover multiple disease progression dynamics. The proposed model is an extension of an input-output hidden Markov model that takes into account the clinical assessments of patients’ health status and prescribed medications. We illustrate the benefits of our model using a synthetically generated dataset and a real-world longitudinal dataset for Parkinson’s disease. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/ceritli22a.html https://proceedings.mlr.press/v184/ceritli22a.html Calibrating complex epidemiological models to observed data is a crucial step to provide both insights into the current disease dynamics, i.e. by estimating a reproductive number, as well as to provide reliable forecasts and scenario explorations. Here we present a new approach to calibrate an agent-based model – Epicast – using a large set of simulation ensembles for different major metropolitan areas of the United States. In particular, we propose: a new neural network based surrogate model able to simultaneously emulate all different locations; and a novel posterior estimation that provides not only more accurate posterior estimates of all parameters but enables the joint fitting of global parameters across regions. Wed, 20 Jul 2022 00:00:00 +0000 https://proceedings.mlr.press/v184/anirudh22a.html https://proceedings.mlr.press/v184/anirudh22a.html