Healthcare Sustainability: Hospitalization Rate Forecasting with Transfer Learning and Location-Aware News Analysis

Jing Chen, Germán G. Creamer, Yue Ning, Tal Ben-ZviSustainability 2023 · DOI

TL;DR

This paper introduces TLSS, a novel AI framework that forecasts hospitalization rates during emerging pandemics like COVID-19 by combining historical disease data, news sentiment, and semantic analysis. By leveraging transfer learning from existing diseases and incorporating location-aware news insights, TLSS offers a more accurate and robust prediction system, crucial for sustainable healthcare management.

What Problem Does This Paper Address?

Forecasting hospitalization rates is vital for public health systems to manage resources and plan for long-term sustainability, especially during emerging pandemics. However, this is challenging due to limited historical data for new diseases and the failure of existing methods to incorporate crucial external factors like news, policies, and geolocation. This paper tackles the problem of accurately predicting hospitalization rates during the COVID-19 pandemic, aiming to provide actionable insights for healthcare sustainability by bridging the gap in data scarcity and the underutilization of diverse information sources.

Methods

The core of the proposed solution is the Transfer Learning Architecture with Dynamic Location-Aware Sentiment and Semantic Analysis (TLSS). This framework integrates several key components:

  • Heterogeneous Transfer Learning (HTL): TLSS learns general transmission patterns from existing infectious diseases (like influenza) using a source model (Cola-GNN) and transfers this knowledge to predict hospitalization rates for a new pandemic (COVID-19). This addresses the "data scarcity" problem by leveraging knowledge from richer, related datasets.
  • Multimodal Data Learning: This module captures temporal dependencies of hospitalization rates and encodes news sentiment and semantic features. It employs a dynamic location-aware attention mechanism to model public emotions and opinions across different locations by analyzing news articles. This allows the model to understand how local news and sentiments influence hospitalization trends.
  • Non-Linear Correlation Tests: Before building the model, the authors use White Test, Granger Causality Test, and Brownian Distance Correlation to establish a significant relationship between COVID-19 infection cases and hospitalization rates, justifying the use of infection data.
  • Sentiment and Semantic Analysis: VADER (Valence Aware Dictionary for sEntiment Reasoning) is used for sentiment analysis, and Sentence-BERT (SBERT) is used for semantic analysis of news articles to extract relevant features.

Datasets

The study utilizes several datasets for training and evaluation:

  • Hospitalization Rate Data: Collected from the CDC COVID-19 Reported Patient Impact and Hospital Capacity by State Time Series, covering daily new admissions in 50 US states from August 1, 2020, to September 30, 2022.
  • COVID-19 Cases Data: Obtained from CDC US-COVID-19-Cases, providing daily new confirmed cases in 50 US states.
  • News Data:
    • Country-level COVID-19 Original News Data from Refinitiv Real-time News.
  • State-level COVID-19 Original News Data from the Global Database of Events, Language, and Tone (GDELT).
  • Pre-trained News Sentiment and Semantic Data: These are derived from the news articles using VADER and SBERT, respectively, for both country and state levels.

The data was split into training (80%), validation (10%), and testing (10%) sets chronologically.

Key Insights and Ideas

  • Transfer Learning for Data Scarcity: The core idea of using transfer learning from existing diseases (like flu) to predict hospitalization rates for a new pandemic (COVID-19) is crucial. This effectively overcomes the challenge of limited historical data for novel diseases.
  • News as a Predictive Feature: The research highlights that news sentiment and semantic content are not just descriptive but can be predictive features for hospitalization rates, especially for longer lead times. This emphasizes the impact of public opinion and media on health trends.
  • Location Matters: The "location-aware" aspect of the model is key. It demonstrates that understanding how news and sentiments vary geographically and influence local hospitalization trends improves prediction accuracy.
  • Non-Linear Relationships are Significant: The use of non-linear correlation tests confirms that the relationship between infection cases and hospitalization rates is not purely linear, justifying the need for more sophisticated modeling approaches.

Key Results

  • Superior Performance: TLSS consistently outperformed various state-of-the-art baseline models (AR, ARMA, VAR, RNN, LSTNet, Cola-GNN) in hospitalization rate forecasting across different lead times (1, 7, and 14 days).
  • Significant Improvement with Exogenous Variables: Adding COVID-19 cases, news sentiment, and semantic features generally improved forecasting accuracy, particularly for longer lead times.
  • Ablation Study Findings:
    • The transfer learning module significantly improved predictions for lead times of 7 and 14 days.
  • News sentiment and semantic analysis modules also contributed positively, especially for longer lead times, with sentiment showing a slightly stronger influence for a one-day lead time.
  • Non-Linear Correlation: The non-linear correlation tests showed extremely significant relationships between COVID-19 cases and hospitalization rates across all 50 US states.

Limitations and Future Work

  • Data Dependency: The model's performance is inherently tied to the quality and availability of the input data, particularly news articles.
  • Generalizability: While TLSS shows promise, its application to future pandemics will depend on the availability of similar data and the ability to adapt the transfer learning approach.
  • Future Work: The authors suggest refining the AI analytics framework for greater flexibility and continuous updating. They also plan to provide more comprehensive, readable, and visualized forecasting results for broader application. Additionally, they aim to extend the exploration to other ESG factors like governance and climate risks, which might also impact hospitalization rates.

Summary updated

Abstract

Monitoring and forecasting hospitalization rates are of essential significance to public health systems in understanding and managing overall healthcare deliveries and strategizing long-term sustainability. Early-stage prediction of hospitalization rates is crucial to meet the medical needs of numerous patients during emerging epidemic diseases such as COVID-19. Nevertheless, this is a challenging task due to insufficient data and experience. In addition, relevant existing work neglects or fails to exploit the extensive contribution of external factors such as news, policies, and geolocations. In this paper, we demonstrate the significant relationship between hospitalization rates and COVID-19 infection cases. We then adapt a transfer learning architecture with dynamic location-aware sentiment and semantic analysis (TLSS) to a new application scenario: hospitalization rate prediction during COVID-19. This architecture learns and transfers general transmission patterns of existing epidemic diseases to predict hospitalization rates during COVID-19. We combine the learned knowledge with time series features and news sentiment and semantic features in a dynamic propagation process. We conduct extensive experiments to compare the proposed approach with several state-of-the-art machine learning methods with different lead times of ground truth. Our results show that TLSS exhibits outstanding predictive performance for hospitalization rates. Thus, it provides advanced artificial intelligence (AI) techniques for supporting decision-making in healthcare sustainability.
BibTeX
@inproceedings{chen2023healthcare,
  author = {Jing Chen and Germán G. Creamer and Yue Ning and Tal Ben-Zvi},
  title = {{Healthcare Sustainability: Hospitalization Rate Forecasting with Transfer Learning and Location-Aware News Analysis}},
  booktitle = {Sustainability},
  year = {2023},
  doi = {10.3390/su152215840},
}
Figure