Toward informatics-enabled preparedness for natural hazards to minimize health impacts of climate change

Abstract

Natural hazards (NHs) associated with climate change have been increasing in frequency and intensity. These acute events impact humans both directly and through their effects on social and environmental determinants of health. Rather than relying on a fully reactive incident response disposition, it is crucial to ramp up preparedness initiatives for worsening case scenarios. In this perspective, we review the landscape of NH effects for human health and explore the potential of health informatics to address associated challenges, specifically from a preparedness angle. We outline important components in a health informatics agenda for hazard preparedness involving hazard-disease associations, social determinants of health, and hazard forecasting models, and call for novel methods to integrate them toward projecting healthcare needs in the wake of a hazard. We describe potential gaps and barriers in implementing these components and propose some high-level ideas to address them.

INTRODUCTION

Natural hazards (NHs) pose a seasonal destructive threat to populations globally.^1,2 Climate change has strong linkages to the increasing frequency and intensity of NHs.³ Since the 1960s, the annual average deaths from NHs dropped drastically,¹ heightening focus on those living with comorbidities and mental trauma, difficulties thriving, and ultimately unready for another disaster. Marginalized and vulnerable populations, seen as having unmet social needs and intersectional experience with inequities, bear a disproportionate burden recovering from NHs.⁴ Yet capture and coverage of patient-level social determinants of health (SDoH) have lagged.^5–8 With the increasing frequency of NHs disrupting communities, health informatics innovations are needed to support planning and adapting to NHs.

Climate changes in local environments can be separated into two general groups of phenomena: (1) changes that create conditions for more frequent and intense acute NHs and (2) gradual “slow-burn” changes that destabilize weather and land pattern (eg, glacial melt, sea-level rise, chronic aerosol exposures). While health effects of floods and coastal surges have been well-documented,^9–11 climate change and compounded disasters may reveal previously under-reported effects. Forecasted heatwaves contribute to wildfire risks and long-term consequences of drought, food shortages, and worker safety risks.^12–14 Poor mental health status spikes dramatically with disasters, adding mental trauma and suicide risks even among those indirectly affected.¹³ The Coronavirus Disease 2019 (COVID-19) pandemic reminded the world that prolonged occupational stress leads to burn-out and capacity shortage.¹⁵ Despite known effects of NHs (Table 1), health systems need a holistic data-driven approach to recognize pattern changes from climate change and impacts on SDoH and research information needs to empower policy and preparedness actions.

Table 1.

Known impacts of climate change on natural hazards associated with health effects

Natural hazard	Impact of climate change on natural hazards	Health effects	Vulnerable or underserved populations at risk
Heat and heat waves	More frequent dangerous heat levels with rising humidity are predicted in many regions of the world, posing a threat to human and agricultural crop health.12 Urban areas are at higher risk of experiencing heat waves due to the “urban heat island” phenomenon.20,21	Acute effects: Increased hospitalization,22–24 dehydration,23 heat stroke,22–24 heat stress,23,24 acute renal disease,22,23 respiratory disease,22,23 preterm birth,24 mental health,24 pediatric emergency trauma,22 poorer ambulance response times22 Less is known about the long-term health effects.	Vulnerable populations: >65 years old (older women are at even higher risk),22,24,25 very young children,24 preexisting conditions (eg, respiratory, cardiovascular, or metabolic disease).20,22,24,25 Racial disparities and SDoH: African American,25,26 nonnative speakers,24 lower socioeconomic status,23–26 military occupation,22 outdoor workers,23,24 manual laborers23,24
Hurricanes and floods	Hurricanes and flooding are expected to increase in intensity and number.16,27,28 Hurricanes intensity is influenced by decreasing vertical wind shear and increasing ocean temperatures, which occurs with more greenhouse gasses.27	Acute effects: Injuries, infection, drowning, increased risk of airborne and water borne exposure of chemicals and toxins through damage and flooding to factories, agricultural areas and animal facilities.29,30 Long-term effects: exacerbation of preexisting chronic illnesses (eg, diabetes and renal failure),29 exposure to mold,10 cardiovascular disease,29 respiratory disease,29 adverse pregnancy outcomes (eg, gestational hypertension, renal disease, preterm birth, stillbirths),29 mental illness and substance abuse (eg, depression and post-traumatic stress disorder).29	Vulnerable populations: elderly,29 preexisting conditions.29 Racial disparities and SDoH: lower socioeconomic status,9,10,29 limited access to health care,9,10 displacement of individuals.28 The United States is at risk of a 26.4% increase in flooding by 2050, with Black communities in the Southern states at highest risk.28
Wildfires	The population exposed to wildfire smoke has increased dramatically in the last 20 years.31 Parts of the world are projected to see a rise in area burned due to wildfire by 40%–100%.12,32	Acute effects: Increased hospitalizations,31 cardiovascular diseases (eg, cardiac ischemia, myocardial infarction),31,33–36 metabolic diseases (eg, type II diabetes),31 respiratory illnesses (eg, asthma or chronic obstructive pulmonary disease), adverse pregnancy outcomes (eg, low birthweight, preterm birth),37,38 all-cause mortality.39 Long-term health effects of wildfires include eye and respiratory tract irritation, reduced lung function, bronchitis, exacerbation of asthma and heart failure, and premature death.40	Vulnerable populations: >65 old,31,41 children,31,41 those with preexisting conditions (eg, cardiovascular or respiratory disease),31,41 pregnant women and their fetus.31,41 Racial disparities and SDoH: population groups living in low-income areas and outdoor workers.31,41

SDoH: social determinants of health.

Observational systems deployed in the ocean and land and satellite remote sensing systems inform early warning systems and evacuation to safety.^16,17 Such systems generate high volumes of high-resolution data, requiring analytic tools like artificial intelligence (AI) to extract research insights. Yet, communities experiencing the intersection of multiple health inequities and social needs often bear a disproportionate burden recovering from disasters.¹⁸ Further improvements are needed to capture knowledge about SDoH and design strategies for resilience for those most vulnerable.^7,19

In this perspective article, we focus on acute NHs (Table 1) that exacerbate the deterioration of SDoHs and present what biomedical/health informatics should support to integrate research capacities into preparedness.

TOWARD AN INFORMATICS AGENDA: GAPS, BARRIERS, AND RECOMMENDATIONS

From a preparedness perspective, the central need in the wake of an NH is the ability to send early warning messages and appropriately allocate first-response resources (eg, provider support, medications, debris removal). NHs are stochastic and forecasting healthcare utilization is highly complex owing to local environments and community structures. Traditional forecasting systems are agnostic of the local burden of disease and social needs in the communities. We posit that any such comprehensive forecasting system must comprise the following components:

• C1. Localized syndromic surveillance models for NHs.

• C2. Regular updates to SdoH and associated geographical distributions of vulnerable/underserved populations.

• C3. High-resolution spatiotemporal models to predict NHs in the future.

• C4. Methods to integrate syndromic surveillance models, demographic distributions, and NH prediction models.

Before we elaborate on these components, we note that vulnerability can be defined differently depending on the context of the hazard of concern.^42–45 For the purposes of this perspective, we rely on definitions of vulnerable and underserved populations by CDC and CMS to highlight groups most broadly at-risk to NHs (Figure 1). We acknowledge that marginalized communities often experience multiple health inequities—such as discrimination and barriers to basic resources—and historical underrepresentation within biomedical research. Highlighting notable gaps in research and data infrastructure, we propose high-level ideas to address barriers as a long-term call-to-action for the informatics community.

Figure 1.

At-risk populations in a natural hazard are populations at greater risk of negative health outcomes due to disparities in social determinants of health and/or physical health compared to the majority who are facing the same natural disaster.⁴⁴ Adapted from: PUBLIC HEALTH WORKBOOK—To Define, Locate, and Reach Special, Vulnerable, and At-risk Populations in an Emergency⁴⁶ and “Serving Vulnerable and Underserved Populations.” U.S. Centers for Medicare & Medicaid Services, Department of Health and Human Services.⁴⁷

C1. Localized syndromic surveillance models for NHs

Healthcare resource allocation directly depends on the types of conditions that will see a major uptick. Therefore, it is imperative to build models that capture distributions of conditions in the wake of an NH by type. Retrospective case studies and survey-based estimates document salient trends from specific NHs by type (Table 1). As NHs reach greater intensity and duration, it is not apparent how historical syndromic patterns observed during prior events may under-extrapolate disease burden into new geographic settings for future projections. Expert intuitions may not capture all second-order effects of burden (eg, drug overdose deaths reached a record high in the United States during the COVID-19 pandemic⁴⁸ despite having no direct relationship to COVID-19).

A comprehensive database of various NHs and associated distributions of conditions across time post-NH, ranging from a week to several months, is needed. Recent data sets provide the spatial-temporal information about NH-affected zones.^1,49 What remains missing is health outcomes data combined with individual- and community-level SDoH. As of now, the US National Centers for Environmental Information tracks deaths and injuries from NHs and the National Syndromic Surveillance Program (NSSP) collects chief complaint data from nearly 6000 healthcare facilities and emergency departments.⁵⁰ However, NSSP only constitutes 1/7th of US outpatient clinics;⁵¹ a collaboration between local healthcare facilities and state health information exchange may provide a fuller snapshot, although the roughly 10% of Americans without interactions with healthcare systems may be under-represented. Augmenting traditional surveillance data with social media platform may mitigate this representation issue.^52,53 Social media usage and preferences vary between demographics, so it is crucial to consider the platforms’ diversity and representativeness. For example, in the United States, 46% of Hispanic Americans and only 16% of white Americans use WhatsApp⁵⁴; Facebook, Twitter, and Instagram are amenable to crisis informatics research⁵⁵ and connectivity cold-spot detection⁵⁶ across demographics. This syndromic distribution database should be retrospectively collected for previous NHs and mined in real time to build a rich up-to-date resource for new NHs.

C2. Regular updates to SDoH and associated geographical distributions of vulnerable/underserved populations

Information collected from C1 may not be sufficient to extrapolate disease patterns beyond the location and population characteristics at the time. As such, it is important to gather appropriate snapshots of population characteristics. Different metrics have been adapted to use static representations of spatial risks,^4,57 compute composite indices,⁵⁸ and identify vulnerable population areas. This supports a known emergency response use case: identifying geographic areas with high concentrations of vulnerable/underserved people for triage.^42,59 Unfortunately, census-based population estimates gathered every 10 years become less accurate as time progresses.^60,61 Thus, there needs to be a focus on regular capture of changes in measures of SdoH in a population.

Regular, explicit updates of SDoH attributes are becoming more practical, and integration of SDoH data into clinical records is progressing.^62–64 SDOH data from a population should ideally be captured at a rate sufficient to prepare for predictable NH periods, such as seasonal risks which might occur annually. Leveraging EMRs and HL7-FHIR infrastructure, public health surveillance can track SDoH variations across visits and healthcare facilities, which can then be aggregated for community-level measures. However, EMR-based coding of SDoH needs data quality and process improvement.^65,66 Tools (eg, for clinical note processing) and vocabulary are needed to screen for discrimination by gender identity, employment and occupation status, and education opportunities. With appropriate incentives in-place,⁶⁷ the data collection strategy would inform healthcare stakeholders with fresh snapshots of the patient population vulnerabilities, driving policy changes to promote disaster resilience.

C3. High-resolution spatiotemporal models to predict NHs in the future

C1 addresses the disease burden of NHs, but it lacks the capability to predict healthcare utilization events. Modeling NH risk is central to leveraging the other components to estimate future healthcare utilization. Recent disaster datasets provide spatial information and damage assessments of past NHs.^1,49 While this is useful for follow-up inquiries for affected areas, it does not help with extending predictions to new areas. A full discussion of methods that forecast NHs is out-of-the-scope for this article but a recent survey by Ward et al² highlights different approaches taken in this area. Recently, AI methods have been proposed to predict storm duration, severe wind, and severe hail in the near term.⁶⁸ Deep neural networks have seen a major resurgence in unstructured data analysis and the same appears to hold true for weather forecasting.⁶⁹ Jacques-Dumas et al⁷⁰ use a convolutional neural network with transfer learning to predict extreme long-lasting heat waves with a 15-day lead time. Using the famous U-net convolutional architecture for image segmentation, Weyn et al⁷¹ generate 6-week subseasonal forecasts in 3 min, demonstrating NH prediction with a 4-day forecast for hurricane Irma, retrospectively. This work has been improved to model more variables and at 8 times higher resolution using the vision transformer architecture as the backbone.⁷² Application of the latest AI advances for NH prediction is still nascent and there is a serious call to create new benchmarks to rigorously test methods in this area.⁷³

C4. Methods to integrate syndromic surveillance models, demographic distributions, and NH prediction models

If we know the risk of a particular NH occurring in an area (C3), based on the prior syndromic distribution for that NH (C1) and the current vulnerable demographic snapshot of that area (C2), healthcare utilization in the wake of that NH can be projected, analogous to the COVID-19 SEIR models.⁷⁴ This involves developing novel methods to integrate different models and distributions (C1–C3) to map to utilization across disaster management phases. We posit that extra resource needs can be directly tied to uptick in diseases and hence we set out to model the distribution of diseases in a particular location at a given time of the year. Future distributions of diseases estimated from the combination of C1, C2, and C3 will enable local facilities to associate upticks in diseases with potential extra resources needed. Estimating model uncertainty arising out of the three component estimates will be challenging. In complex systems involving multiple interacting variables each carrying noise, uncertainty snowballs and lead to unreliable projections (eg, overshooting utilization leads to resource wastage). These aspects of uncertainty and complexity are the two main pillars identified from a consensus study of digital technologies and environmental sciences.⁷⁵ In terms of specific methods, computer vision, causal inference, uncertainty quantification, transfer learning, and time series analysis have been put forward by AI scientists⁷⁶ to handle climate change in general; these methods are pertinent in the context of joint modeling needed to integrate C1–C3.

CONCLUDING REMARKS

In this perspective, we reviewed climate change-induced NHs and associated disease burden. Subsequently, we identified gaps and barriers in appropriate resource allocation in the wake of such NHs and presented essential components central to an informatics strategy in mitigating adverse human health impacts. We conclude with some important considerations surrounding cost burden, interoperability, and privacy.

We note that federal and state health agencies may have to incur a major cost burden to materialize a resource planning system that aligns with our agenda. There is a value conflict between health system profit margins, adequate staffing, and organizational investment into disaster risk reduction for climate change response. However, the significant financial burden of handling climate and sensitive health outcomes during future climate crises may dwarf the upfront costs.⁷⁷ Since climate crises disproportionately affect marginalized and vulnerable populations, there is merit in gauging the economic tradeoff of addressing the health harms in relation to the costs of inaction disproportionately felt by marginalized groups.⁷⁸

This perspective has an implicit focus on the USA. However, the ongoing COVID-19 pandemic has demonstrated that international data sharing and collaboration are essential for rapid advances. As pointed out by a recent G7 health ministers' communique,² striving toward standards for data sharing is essential while complying with international patient privacy regulations. Federated machine learning⁷⁹ approaches ought to be considered to minimize concerns of data sharing and privacy breaches. Within the United States, at best the area deprivation index describes vulnerability as fine as the census block-group scale.⁸⁰ Inadvertent private health information disclosures can occur if SDoH data are captured at the block-group scale. Although AI methods are bound to play a crucial operational role, unscrupulous use of AI algorithmic decision-making may exacerbate disparities.^81,82 Deliberate considerations of the use of particular AI methods⁸³ should be a top priority to avoid detrimental outcomes in mitigating the effects of climate change on human health.

FUNDING

This work was supported by US National Center for Advancing Translational Sciences Grants UL1TR001998 and U24TR002306 (including the supplement U24TR002306-04S3). The Office of Data Science Strategy, US NIH, provided funding for AYL as a Data and Technology Advancement (DATA) National Service Scholar. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

AUTHOR CONTRIBUTIONS

JP initiated the work and wrote the first draft of the paper with NOR contributing substantially with literature review, tables, and figures. RK conceived the perspective, including the high-level agenda proposed, and coordinated the writing efforts. All remaining authors contributed equally to the content, drafting, and revision and are hence listed in the alphabetical order.

Data Availability

No new data were generated or analyzed in support of this research.